M. Sc. Zhixin Huang

Autonomous Energy Systems (AESy)

E-Mail: zhixin.huang[at]uni-kassel[dot]de

Website: Zhixin Huang

Publikationen

[ 2026 ]
[ 2025 ]
[ 2024 ]
[ 2023 ]
[ 2022 ]
[ 2021 ]
[ 2017 ]

2026[ to top ]

TRACE: Time series representation learning with contrastive embeddings for anomaly detection in photovoltaic systems.Nivarthi, Chandana Priya; Huang, Zhixin; Gruhl, Christian; Sick, Bernhard. In Energy and AI, 23, bl 100670. Elsevier, 2026.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
- [ URL ]
Reliable anomaly detection in photovoltaic (PV) inverters is critical for ensuring operational efficiency and reducing maintenance costs in renewable energy systems. We introduce TRACE (Time series Representation learning with Autoencoder-based Contrastive Embeddings), a self-supervised contrastive learning framework for multivariate time series anomaly detection in PV systems. TRACE employs a two-stage architecture: autoencoder-based representation learning with interchangeable backbones followed by contrastive training through a Siamese network. The framework generates semantically coherent augmentations by perturbing autoencoder reconstructions and applies three negative mining strategies to create challenging contrastive pairs. Comprehensive experiments on a real-world PV inverter dataset and two industrial benchmarks demonstrate TRACE’s superiority. Autoencoder-based augmentations deliver a 21.3% relative improvement in mean F1 (0.616 vs. 0.508) over traditional perturbation methods, with TransformerAE emerging as the optimal backbone architecture. While negative sampling strategies show dataset-specific advantages, their impact remains secondary to encoder capacity. TRACE with TransformerAE and reconstruction-error negatives consistently outperforms fourteen state-of-the-art time series anomaly detection methods, achieving highest F1 scores on all the three datasets while maintaining exceptional precision up to 0.99. Visualization analysis confirms TRACE’s capacity for early fault detection up to three days before failure and interpretable embedding separation. The framework addresses the fundamental challenge of label scarcity in industrial monitoring through self-supervised learning, providing a practical and transparent solution for predictive maintenance in PV systems and broader industrial applications.
@article{nivarthi2026trace, abstract = {Reliable anomaly detection in photovoltaic (PV) inverters is critical for ensuring operational efficiency and reducing maintenance costs in renewable energy systems. We introduce TRACE (Time series Representation learning with Autoencoder-based Contrastive Embeddings), a self-supervised contrastive learning framework for multivariate time series anomaly detection in PV systems. TRACE employs a two-stage architecture: autoencoder-based representation learning with interchangeable backbones followed by contrastive training through a Siamese network. The framework generates semantically coherent augmentations by perturbing autoencoder reconstructions and applies three negative mining strategies to create challenging contrastive pairs. Comprehensive experiments on a real-world PV inverter dataset and two industrial benchmarks demonstrate TRACE’s superiority. Autoencoder-based augmentations deliver a 21.3% relative improvement in mean F1 (0.616 vs. 0.508) over traditional perturbation methods, with TransformerAE emerging as the optimal backbone architecture. While negative sampling strategies show dataset-specific advantages, their impact remains secondary to encoder capacity. TRACE with TransformerAE and reconstruction-error negatives consistently outperforms fourteen state-of-the-art time series anomaly detection methods, achieving highest F1 scores on all the three datasets while maintaining exceptional precision up to 0.99. Visualization analysis confirms TRACE’s capacity for early fault detection up to three days before failure and interpretable embedding separation. The framework addresses the fundamental challenge of label scarcity in industrial monitoring through self-supervised learning, providing a practical and transparent solution for predictive maintenance in PV systems and broader industrial applications.}, author = {Nivarthi, Chandana Priya and Huang, Zhixin and Gruhl, Christian and Sick, Bernhard}, journal = {Energy and AI}, keywords = {itegpub}, pages = 100670, publisher = {Elsevier}, title = {TRACE: Time series representation learning with contrastive embeddings for anomaly detection in photovoltaic systems}, volume = 23, year = 2026 }
%0 Journal Article %1 nivarthi2026trace %A Nivarthi, Chandana Priya %A Huang, Zhixin %A Gruhl, Christian %A Sick, Bernhard %D 2026 %I Elsevier %J Energy and AI %P 100670 %R 10.1016/j.egyai.2025.100670 %T TRACE: Time series representation learning with contrastive embeddings for anomaly detection in photovoltaic systems %U https://www.sciencedirect.com/science/article/pii/S2666546825002022 %V 23 %X Reliable anomaly detection in photovoltaic (PV) inverters is critical for ensuring operational efficiency and reducing maintenance costs in renewable energy systems. We introduce TRACE (Time series Representation learning with Autoencoder-based Contrastive Embeddings), a self-supervised contrastive learning framework for multivariate time series anomaly detection in PV systems. TRACE employs a two-stage architecture: autoencoder-based representation learning with interchangeable backbones followed by contrastive training through a Siamese network. The framework generates semantically coherent augmentations by perturbing autoencoder reconstructions and applies three negative mining strategies to create challenging contrastive pairs. Comprehensive experiments on a real-world PV inverter dataset and two industrial benchmarks demonstrate TRACE’s superiority. Autoencoder-based augmentations deliver a 21.3% relative improvement in mean F1 (0.616 vs. 0.508) over traditional perturbation methods, with TransformerAE emerging as the optimal backbone architecture. While negative sampling strategies show dataset-specific advantages, their impact remains secondary to encoder capacity. TRACE with TransformerAE and reconstruction-error negatives consistently outperforms fourteen state-of-the-art time series anomaly detection methods, achieving highest F1 scores on all the three datasets while maintaining exceptional precision up to 0.99. Visualization analysis confirms TRACE’s capacity for early fault detection up to three days before failure and interpretable embedding separation. The framework addresses the fundamental challenge of label scarcity in industrial monitoring through self-supervised learning, providing a practical and transparent solution for predictive maintenance in PV systems and broader industrial applications.

2025[ to top ]

Towards Cognitive Situational Awareness in Maritime Traffic Using Federated Evidential Learning.Gao, Shang; Huang, Zhixin; Al-Falouji, Ghassan; Sick, Bernhard; Tomforde, Sven. In IEEE Conference on Cognitive and Computational Aspects of Situation Management (CogSIMA), bll 9–16. IEEE, 2025.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
The increasing complexity of maritime navigation and the shift towards (semi)autonomous systems necessitate enhanced situational awareness (SA) to ensure maritime safety. This transition introduces new requirements for situation mod- elling and SA, particularly in busy waterways. To address these challenges, we present the Federated Evidential Learning for Anomaly Detection of Ship Trajectories (FEAST) framework, which integrates Federated Learning and Evidential Learning to provide a privacy-preserving, collaborative, and uncertainty- aware approach to out-of-distribution (OOD) anomaly detection in maritime traffic. FEAST uses data from the Automatic Identification System from the Kiel region, Germany, which exhibits unique characteristics of dynamic and heterogeneous maritime activity due to its connection with the traffic-dense Kiel Canal. Our extensive evaluations demonstrate that FEAST improves OOD anomaly detection by leveraging epistemic and aleatoric uncertainty estimates, outperforming baseline methods such as Denoise AutoEncoders and Variational AutoEncoders. Consequently, FEAST forms a solution to reliable and inter- pretable maritime traffic anomaly detection, supporting enhanced SA in maritime operations.
@inproceedings{gao2025towards, abstract = {The increasing complexity of maritime navigation and the shift towards (semi)autonomous systems necessitate enhanced situational awareness (SA) to ensure maritime safety. This transition introduces new requirements for situation mod- elling and SA, particularly in busy waterways. To address these challenges, we present the Federated Evidential Learning for Anomaly Detection of Ship Trajectories (FEAST) framework, which integrates Federated Learning and Evidential Learning to provide a privacy-preserving, collaborative, and uncertainty- aware approach to out-of-distribution (OOD) anomaly detection in maritime traffic. FEAST uses data from the Automatic Identification System from the Kiel region, Germany, which exhibits unique characteristics of dynamic and heterogeneous maritime activity due to its connection with the traffic-dense Kiel Canal. Our extensive evaluations demonstrate that FEAST improves OOD anomaly detection by leveraging epistemic and aleatoric uncertainty estimates, outperforming baseline methods such as Denoise AutoEncoders and Variational AutoEncoders. Consequently, FEAST forms a solution to reliable and inter- pretable maritime traffic anomaly detection, supporting enhanced SA in maritime operations.}, author = {Gao, Shang and Huang, Zhixin and Al-Falouji, Ghassan and Sick, Bernhard and Tomforde, Sven}, booktitle = {IEEE Conference on Cognitive and Computational Aspects of Situation Management (CogSIMA)}, keywords = {itegpub}, pages = {9–16}, publisher = {IEEE}, title = {Towards Cognitive Situational Awareness in Maritime Traffic Using Federated Evidential Learning}, year = 2025 }
%0 Conference Paper %1 gao2025towards %A Gao, Shang %A Huang, Zhixin %A Al-Falouji, Ghassan %A Sick, Bernhard %A Tomforde, Sven %B IEEE Conference on Cognitive and Computational Aspects of Situation Management (CogSIMA) %D 2025 %I IEEE %P 9--16 %R 10.1109/CogSIMA64436.2025.11079521 %T Towards Cognitive Situational Awareness in Maritime Traffic Using Federated Evidential Learning %X The increasing complexity of maritime navigation and the shift towards (semi)autonomous systems necessitate enhanced situational awareness (SA) to ensure maritime safety. This transition introduces new requirements for situation mod- elling and SA, particularly in busy waterways. To address these challenges, we present the Federated Evidential Learning for Anomaly Detection of Ship Trajectories (FEAST) framework, which integrates Federated Learning and Evidential Learning to provide a privacy-preserving, collaborative, and uncertainty- aware approach to out-of-distribution (OOD) anomaly detection in maritime traffic. FEAST uses data from the Automatic Identification System from the Kiel region, Germany, which exhibits unique characteristics of dynamic and heterogeneous maritime activity due to its connection with the traffic-dense Kiel Canal. Our extensive evaluations demonstrate that FEAST improves OOD anomaly detection by leveraging epistemic and aleatoric uncertainty estimates, outperforming baseline methods such as Denoise AutoEncoders and Variational AutoEncoders. Consequently, FEAST forms a solution to reliable and inter- pretable maritime traffic anomaly detection, supporting enhanced SA in maritime operations.
Efficient Bayesian Updates for Deep Active Learning via Laplace Approximations.Huseljic, Denis; Herde, Marek; Rauch, Lukas; Hahn, Paul; Huang, Zhixin; Kottke, Daniel; Vogt, Stephan; Sick, Bernhard. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECMLPKDD), bll 20–35. Springer, 2025.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
Deep active learning (AL) selects batches of instances for annotation to avoid retraining deep neural networks (DNNs) after each new label. Employing a naive top-b selection can result in a batch of redundant (similar) instances. To address this, various AL strategies employ clustering techniques that ensure diversity within a batch. We approach this issue by substituting the costly retraining with an efficient Bayesian update. Our proposed update represents a second-order optimization step using the Gaussian posterior from a last-layer Laplace approximation. Thereby, we achieve low computational complexity by computing the inverse Hessian in closed form. We demonstrate that in typical AL settings, our update closely approximates retraining while being considerably faster. Leveraging our update, we introduce a new framework for batch selection through sequential construction, updating the DNN after each label acquisition. Furthermore, we incorporate our update into a look-ahead selection strategy as a feasible upper baseline approximating optimal batch selection. Our results highlight the potential of efficient updates to advance deep AL research.
@inproceedings{huseljic2025efficient, abstract = {Deep active learning (AL) selects batches of instances for annotation to avoid retraining deep neural networks (DNNs) after each new label. Employing a naive top-b selection can result in a batch of redundant (similar) instances. To address this, various AL strategies employ clustering techniques that ensure diversity within a batch. We approach this issue by substituting the costly retraining with an efficient Bayesian update. Our proposed update represents a second-order optimization step using the Gaussian posterior from a last-layer Laplace approximation. Thereby, we achieve low computational complexity by computing the inverse Hessian in closed form. We demonstrate that in typical AL settings, our update closely approximates retraining while being considerably faster. Leveraging our update, we introduce a new framework for batch selection through sequential construction, updating the DNN after each label acquisition. Furthermore, we incorporate our update into a look-ahead selection strategy as a feasible upper baseline approximating optimal batch selection. Our results highlight the potential of efficient updates to advance deep AL research.}, author = {Huseljic, Denis and Herde, Marek and Rauch, Lukas and Hahn, Paul and Huang, Zhixin and Kottke, Daniel and Vogt, Stephan and Sick, Bernhard}, booktitle = {European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECMLPKDD)}, keywords = {itegpub}, pages = {20–35}, publisher = {Springer}, title = {Efficient Bayesian Updates for Deep Active Learning via Laplace Approximations}, year = 2025 }
%0 Conference Paper %1 huseljic2025efficient %A Huseljic, Denis %A Herde, Marek %A Rauch, Lukas %A Hahn, Paul %A Huang, Zhixin %A Kottke, Daniel %A Vogt, Stephan %A Sick, Bernhard %B European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECMLPKDD) %D 2025 %I Springer %P 20--35 %R 10.1007/978-3-032-05981-9_2 %T Efficient Bayesian Updates for Deep Active Learning via Laplace Approximations %X Deep active learning (AL) selects batches of instances for annotation to avoid retraining deep neural networks (DNNs) after each new label. Employing a naive top-b selection can result in a batch of redundant (similar) instances. To address this, various AL strategies employ clustering techniques that ensure diversity within a batch. We approach this issue by substituting the costly retraining with an efficient Bayesian update. Our proposed update represents a second-order optimization step using the Gaussian posterior from a last-layer Laplace approximation. Thereby, we achieve low computational complexity by computing the inverse Hessian in closed form. We demonstrate that in typical AL settings, our update closely approximates retraining while being considerably faster. Leveraging our update, we introduce a new framework for batch selection through sequential construction, updating the DNN after each label acquisition. Furthermore, we incorporate our update into a look-ahead selection strategy as a feasible upper baseline approximating optimal batch selection. Our results highlight the potential of efficient updates to advance deep AL research.

2024[ to top ]

Spatial-Temporal Attention Graph Neural Network with Uncertainty Estimation for Remaining Useful Life Prediction.Huang, Zhixin; Nivarthi, Chandana Priya; Gruhl, Christian; Sick, Bernhard. In International Joint Conference on Neural Networks (IJCNN), bll 1–9. IEEE, 2024.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
In the increasingly complex industrial system health management domain, accurate prediction of remaining useful life plays an essential role. This paper analyzes the methods to improve the predictive performance of remaining useful life from three aspects: optimizing model structures, augmenting uncertainty estimation in predictions, and transitioning normalization methods. Based on our analysis, we propose a novel model, the Uncertainty Spatial-Temporal Attention Graph Neural Network (USTAGNN), which consists of three primary components: sensor graph construction, a spatio-temporal feature extractor, and a probabilistic prediction module. The feature extractor leverages graph neural networks and temporal convolutional networks as a foundation to extract spatial and temporal features, further enhanced by attention mechanisms, spectral normalization, and residual connections to bolster its distance awareness. Following extensive experimental comparisons, we utilized the parameter-driven dynamic adjacency matrix for sensor graph construction and the deep kernel Gaussian process for precise uncertainty estimation. USTAGNN tries to resolve issues not thoroughly addressed in existing research, such as comparative analyses of sensor graph construction methods, accurate uncertainty estimation, and the model’s generalization under different preprocessing conditions. The proposed model demonstrated state-of-the-art performance on various subsets of the C-MAPSS dataset, achieving up to a 35.9% improvement in prediction score.
@inproceedings{huang2024spatial, abstract = {In the increasingly complex industrial system health management domain, accurate prediction of remaining useful life plays an essential role. This paper analyzes the methods to improve the predictive performance of remaining useful life from three aspects: optimizing model structures, augmenting uncertainty estimation in predictions, and transitioning normalization methods. Based on our analysis, we propose a novel model, the Uncertainty Spatial-Temporal Attention Graph Neural Network (USTAGNN), which consists of three primary components: sensor graph construction, a spatio-temporal feature extractor, and a probabilistic prediction module. The feature extractor leverages graph neural networks and temporal convolutional networks as a foundation to extract spatial and temporal features, further enhanced by attention mechanisms, spectral normalization, and residual connections to bolster its distance awareness. Following extensive experimental comparisons, we utilized the parameter-driven dynamic adjacency matrix for sensor graph construction and the deep kernel Gaussian process for precise uncertainty estimation. USTAGNN tries to resolve issues not thoroughly addressed in existing research, such as comparative analyses of sensor graph construction methods, accurate uncertainty estimation, and the model’s generalization under different preprocessing conditions. The proposed model demonstrated state-of-the-art performance on various subsets of the C-MAPSS dataset, achieving up to a 35.9% improvement in prediction score.}, author = {Huang, Zhixin and Nivarthi, Chandana Priya and Gruhl, Christian and Sick, Bernhard}, booktitle = {International Joint Conference on Neural Networks (IJCNN)}, keywords = {itegpub}, pages = {1–9}, publisher = {IEEE}, title = {Spatial-Temporal Attention Graph Neural Network with Uncertainty Estimation for Remaining Useful Life Prediction}, year = 2024 }
%0 Conference Paper %1 huang2024spatial %A Huang, Zhixin %A Nivarthi, Chandana Priya %A Gruhl, Christian %A Sick, Bernhard %B International Joint Conference on Neural Networks (IJCNN) %D 2024 %I IEEE %P 1--9 %R 10.1109/IJCNN60899.2024.10650622 %T Spatial-Temporal Attention Graph Neural Network with Uncertainty Estimation for Remaining Useful Life Prediction %X In the increasingly complex industrial system health management domain, accurate prediction of remaining useful life plays an essential role. This paper analyzes the methods to improve the predictive performance of remaining useful life from three aspects: optimizing model structures, augmenting uncertainty estimation in predictions, and transitioning normalization methods. Based on our analysis, we propose a novel model, the Uncertainty Spatial-Temporal Attention Graph Neural Network (USTAGNN), which consists of three primary components: sensor graph construction, a spatio-temporal feature extractor, and a probabilistic prediction module. The feature extractor leverages graph neural networks and temporal convolutional networks as a foundation to extract spatial and temporal features, further enhanced by attention mechanisms, spectral normalization, and residual connections to bolster its distance awareness. Following extensive experimental comparisons, we utilized the parameter-driven dynamic adjacency matrix for sensor graph construction and the deep kernel Gaussian process for precise uncertainty estimation. USTAGNN tries to resolve issues not thoroughly addressed in existing research, such as comparative analyses of sensor graph construction methods, accurate uncertainty estimation, and the model’s generalization under different preprocessing conditions. The proposed model demonstrated state-of-the-art performance on various subsets of the C-MAPSS dataset, achieving up to a 35.9% improvement in prediction score.
LiST: An All-Linear-Layer Spatial-Temporal Feature Extractor with Uncertainty Estimation for RUL Prediction.Huang, Zhixin; Gruhl, Christian; Sick, Bernhard. In IEEE Conference on Industrial Electronics and Applications (ICIEA), bll 1–7. IEEE, 2024.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
In the context of Remaining Useful Life (RUL) prediction for industrial systems, the pursuit of prediction accuracy must be balanced against the hardware costs of model operation and the reliability of prediction results. To resolve these challenges, we introduce LiST, an all-linear-layer spatial-temporal feature extractor integrated with uncertainty estimation, specifically designed for processing sensor multivariate time series (MTS) data. Unlike traditional linear models that flatten MTS and thus neglect their spatial-temporal dependencies, LiST's linear layers act on both the sensor and time dimensions of MTS that can extract spatial and temporal features like GNN and RNN models. Through performance comparisons on four RUL prediction datasets, LiST uses only 66.1% of the parameters, achieves comparable accuracy to state-of-the-art GNN and RNN models, obtains the best results on two datasets with up to a 21.6% improvement in Score, and enhances training efficiency by 3.2 times. Additionally, LiST can predict RUL with uncertainty estimation and precisely disentangle epistemic and aleatoric uncertainties, thus enhancing the model's practicality and reliability in real-world industrial applications.
@inproceedings{huang2024list, abstract = {In the context of Remaining Useful Life (RUL) prediction for industrial systems, the pursuit of prediction accuracy must be balanced against the hardware costs of model operation and the reliability of prediction results. To resolve these challenges, we introduce LiST, an all-linear-layer spatial-temporal feature extractor integrated with uncertainty estimation, specifically designed for processing sensor multivariate time series (MTS) data. Unlike traditional linear models that flatten MTS and thus neglect their spatial-temporal dependencies, LiST's linear layers act on both the sensor and time dimensions of MTS that can extract spatial and temporal features like GNN and RNN models. Through performance comparisons on four RUL prediction datasets, LiST uses only 66.1% of the parameters, achieves comparable accuracy to state-of-the-art GNN and RNN models, obtains the best results on two datasets with up to a 21.6% improvement in Score, and enhances training efficiency by 3.2 times. Additionally, LiST can predict RUL with uncertainty estimation and precisely disentangle epistemic and aleatoric uncertainties, thus enhancing the model's practicality and reliability in real-world industrial applications.}, author = {Huang, Zhixin and Gruhl, Christian and Sick, Bernhard}, booktitle = {IEEE Conference on Industrial Electronics and Applications (ICIEA)}, keywords = {itegpub}, pages = {1–7}, publisher = {IEEE}, title = {LiST: An All-Linear-Layer Spatial-Temporal Feature Extractor with Uncertainty Estimation for RUL Prediction}, year = 2024 }
%0 Conference Paper %1 huang2024list %A Huang, Zhixin %A Gruhl, Christian %A Sick, Bernhard %B IEEE Conference on Industrial Electronics and Applications (ICIEA) %D 2024 %I IEEE %P 1--7 %R 10.1109/ICIEA61579.2024.10665236 %T LiST: An All-Linear-Layer Spatial-Temporal Feature Extractor with Uncertainty Estimation for RUL Prediction %X In the context of Remaining Useful Life (RUL) prediction for industrial systems, the pursuit of prediction accuracy must be balanced against the hardware costs of model operation and the reliability of prediction results. To resolve these challenges, we introduce LiST, an all-linear-layer spatial-temporal feature extractor integrated with uncertainty estimation, specifically designed for processing sensor multivariate time series (MTS) data. Unlike traditional linear models that flatten MTS and thus neglect their spatial-temporal dependencies, LiST's linear layers act on both the sensor and time dimensions of MTS that can extract spatial and temporal features like GNN and RNN models. Through performance comparisons on four RUL prediction datasets, LiST uses only 66.1% of the parameters, achieves comparable accuracy to state-of-the-art GNN and RNN models, obtains the best results on two datasets with up to a 21.6% improvement in Score, and enhances training efficiency by 3.2 times. Additionally, LiST can predict RUL with uncertainty estimation and precisely disentangle epistemic and aleatoric uncertainties, thus enhancing the model's practicality and reliability in real-world industrial applications.
Time-Series Representation Learning via Heterogeneous Spatial-Temporal Contrasting for Remaining Useful Life Prediction.Huang, Zhixin; He, Yujiang; Nivarthi, Chandana Priya; Gruhl, Christian; Sick, Bernhard. In International Conference on Pattern Recognition (ICPR), bll 1–21. Springer, 2024.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
Classical contrastive learning paradigms rely on manual augmentations like cropping, masking, dropping, or adding noise randomly to create divergent sample views from original data. However, the choice of which method to manipulate samples is often subjective and may destroy the latent pattern of the sample. In response, this paper introduces a novel contrastive learning paradigm without choosing sample view augmentation methods, termed Heterogeneous Spatial-Temporal Representation Contrasting (HSTRC). Instead of sample view augmentation, we employ dual branches with a heterogeneous spatial-temporal flipped structure to extract two distinct hidden feature views from the same source data, which avoids disturbing the original time series. Leveraging a combination of cross-branch spatial-temporal contrastive and projected feature contrastive loss functions, HSTRC can effectively extract robust representations from unlabeled time series data. Remarkably, by only fine-tuning the fully connected layers on top of extracted representations by HSTRC, we achieve the best performance across several Remaining Useful Life prediction datasets, showing up to 19.2% improvements over the state-of-the-art supervised learning methods and classical contrastive learning paradigms. Besides, further intensive experiments demonstrate HSTRC's effectiveness in active learning, out-of-distribution testing, and transfer learning scenarios.
@inproceedings{huang2024time, abstract = {Classical contrastive learning paradigms rely on manual augmentations like cropping, masking, dropping, or adding noise randomly to create divergent sample views from original data. However, the choice of which method to manipulate samples is often subjective and may destroy the latent pattern of the sample. In response, this paper introduces a novel contrastive learning paradigm without choosing sample view augmentation methods, termed Heterogeneous Spatial-Temporal Representation Contrasting (HSTRC). Instead of sample view augmentation, we employ dual branches with a heterogeneous spatial-temporal flipped structure to extract two distinct hidden feature views from the same source data, which avoids disturbing the original time series. Leveraging a combination of cross-branch spatial-temporal contrastive and projected feature contrastive loss functions, HSTRC can effectively extract robust representations from unlabeled time series data. Remarkably, by only fine-tuning the fully connected layers on top of extracted representations by HSTRC, we achieve the best performance across several Remaining Useful Life prediction datasets, showing up to 19.2% improvements over the state-of-the-art supervised learning methods and classical contrastive learning paradigms. Besides, further intensive experiments demonstrate HSTRC's effectiveness in active learning, out-of-distribution testing, and transfer learning scenarios.}, author = {Huang, Zhixin and He, Yujiang and Nivarthi, Chandana Priya and Gruhl, Christian and Sick, Bernhard}, booktitle = {International Conference on Pattern Recognition (ICPR)}, keywords = {itegpub}, pages = {1–21}, publisher = {Springer}, title = {Time-Series Representation Learning via Heterogeneous Spatial-Temporal Contrasting for Remaining Useful Life Prediction}, year = 2024 }
%0 Conference Paper %1 huang2024time %A Huang, Zhixin %A He, Yujiang %A Nivarthi, Chandana Priya %A Gruhl, Christian %A Sick, Bernhard %B International Conference on Pattern Recognition (ICPR) %D 2024 %I Springer %P 1--21 %R 10.1007/978-3-031-78398-2_1 %T Time-Series Representation Learning via Heterogeneous Spatial-Temporal Contrasting for Remaining Useful Life Prediction %X Classical contrastive learning paradigms rely on manual augmentations like cropping, masking, dropping, or adding noise randomly to create divergent sample views from original data. However, the choice of which method to manipulate samples is often subjective and may destroy the latent pattern of the sample. In response, this paper introduces a novel contrastive learning paradigm without choosing sample view augmentation methods, termed Heterogeneous Spatial-Temporal Representation Contrasting (HSTRC). Instead of sample view augmentation, we employ dual branches with a heterogeneous spatial-temporal flipped structure to extract two distinct hidden feature views from the same source data, which avoids disturbing the original time series. Leveraging a combination of cross-branch spatial-temporal contrastive and projected feature contrastive loss functions, HSTRC can effectively extract robust representations from unlabeled time series data. Remarkably, by only fine-tuning the fully connected layers on top of extracted representations by HSTRC, we achieve the best performance across several Remaining Useful Life prediction datasets, showing up to 19.2% improvements over the state-of-the-art supervised learning methods and classical contrastive learning paradigms. Besides, further intensive experiments demonstrate HSTRC's effectiveness in active learning, out-of-distribution testing, and transfer learning scenarios.
PrOuD: Probabilistic Outlier Detection Solution for Time Series Analysis on Real-world Photovoltaic Inverters.He, Yujiang; Huang, Zhixin; Vogt, Stephan; Sick, Bernhard. In Energies (MDPI), 17(1), bl 64. MDPI, 2024.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
- [ URL ]
Anomaly detection methods applied to time series are mostly viewed as a black box, which solely provides a deterministic answer to the detected target. Without a convincing explanation, domain experts can hardly trust the detection results and conduct a further time series diagnosis in real-world applications. To overcome the challenge, we mathematically analyzed the sources of anomalies and novelties in multivariate time series as well as their relationships from the perspective of Gaussian-distributed non-stationary noise. Furthermore, we proposed mathematical methods to generate artificial time series and synthetic anomalies, with the goal of solving the problem that it is difficult to train and evaluate models in real-world applications due to the lack of sufficient data. In addition, we designed extbfProbabilistic extbfOutlier extbfDetection (PrOuD), which is a general solution to provide interpretable detection results to assist domain experts in time series analysis. PrOuD can convert a predictive uncertainty of a trained model about a time series value into an estimated uncertainty of the detected outlier through Monte Carlo Estimation. The experimental results obtained on both artificial time series and real-world photovoltaic inverter data demonstrated that the proposed solution could detect emerging anomalies accurately and quickly. The implemented PrOuD demo case shows the potential to makes the detection results of the existing detection methods more convincing so that domain experts can more efficiently complete their tasks, such as time series diagnosis and anomalous pattern clustering.
@article{he2023proud, abstract = {Anomaly detection methods applied to time series are mostly viewed as a black box, which solely provides a deterministic answer to the detected target. Without a convincing explanation, domain experts can hardly trust the detection results and conduct a further time series diagnosis in real-world applications. To overcome the challenge, we mathematically analyzed the sources of anomalies and novelties in multivariate time series as well as their relationships from the perspective of Gaussian-distributed non-stationary noise. Furthermore, we proposed mathematical methods to generate artificial time series and synthetic anomalies, with the goal of solving the problem that it is difficult to train and evaluate models in real-world applications due to the lack of sufficient data. In addition, we designed extbf{Pr}obabilistic extbf{Ou}tlier extbf{D}etection (PrOuD), which is a general solution to provide interpretable detection results to assist domain experts in time series analysis. PrOuD can convert a predictive uncertainty of a trained model about a time series value into an estimated uncertainty of the detected outlier through Monte Carlo Estimation. The experimental results obtained on both artificial time series and real-world photovoltaic inverter data demonstrated that the proposed solution could detect emerging anomalies accurately and quickly. The implemented PrOuD demo case shows the potential to makes the detection results of the existing detection methods more convincing so that domain experts can more efficiently complete their tasks, such as time series diagnosis and anomalous pattern clustering.}, author = {He, Yujiang and Huang, Zhixin and Vogt, Stephan and Sick, Bernhard}, journal = {Energies (MDPI)}, keywords = {itegpub}, number = 1, pages = 64, publisher = {MDPI}, title = {PrOuD: Probabilistic Outlier Detection Solution for Time Series Analysis on Real-world Photovoltaic Inverters}, volume = 17, year = 2024 }
%0 Journal Article %1 he2023proud %A He, Yujiang %A Huang, Zhixin %A Vogt, Stephan %A Sick, Bernhard %D 2024 %I MDPI %J Energies (MDPI) %N 1 %P 64 %R 10.3390/en17010064 %T PrOuD: Probabilistic Outlier Detection Solution for Time Series Analysis on Real-world Photovoltaic Inverters %U https://www.mdpi.com/1996-1073/17/1/64 %V 17 %X Anomaly detection methods applied to time series are mostly viewed as a black box, which solely provides a deterministic answer to the detected target. Without a convincing explanation, domain experts can hardly trust the detection results and conduct a further time series diagnosis in real-world applications. To overcome the challenge, we mathematically analyzed the sources of anomalies and novelties in multivariate time series as well as their relationships from the perspective of Gaussian-distributed non-stationary noise. Furthermore, we proposed mathematical methods to generate artificial time series and synthetic anomalies, with the goal of solving the problem that it is difficult to train and evaluate models in real-world applications due to the lack of sufficient data. In addition, we designed extbfProbabilistic extbfOutlier extbfDetection (PrOuD), which is a general solution to provide interpretable detection results to assist domain experts in time series analysis. PrOuD can convert a predictive uncertainty of a trained model about a time series value into an estimated uncertainty of the detected outlier through Monte Carlo Estimation. The experimental results obtained on both artificial time series and real-world photovoltaic inverter data demonstrated that the proposed solution could detect emerging anomalies accurately and quickly. The implemented PrOuD demo case shows the potential to makes the detection results of the existing detection methods more convincing so that domain experts can more efficiently complete their tasks, such as time series diagnosis and anomalous pattern clustering.
Multi-Task Representation Learning with Temporal Attention for Zero-Shot Time Series Anomaly Detection.Nivarthi, Chandana Priya; Huang, Zhixin; Gruhl, Christian; Sick, Bernhard. In International Joint Conference on Neural Networks (IJCNN), bll 1–10. IEEE, 2024.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
Ensuring the reliability of critical industrial systems across various sectors is crucial. It is essential to detect deviations from regular behaviour to mitigate disruptions and preserve infrastructure integrity. However, accurately labelling anomaly datasets is challenging due to their rarity and manual annotation subjectivity. The conventional approach of training separate models for each dataset entity further complicates model development. This paper presents a novel Multi-task Learning framework combining LSTM Autoencoder with temporal attention mechanism (MTL-LATAM) for effective time series anomaly detection. Multitask learning models improve adaptability and generalizability, leading to reduced runtime and compute power while supporting zero-shot evaluation. These models offer flexibility in detecting emerging anomalies. Additionally, we introduce a dynamic thresholding mechanism to incorporate temporal context for anomaly detection and provide visualizations of attention weights to enhance interpretability. The study compares MTL- LATAM, with other multi-task models, evaluates multi-task versus single-task models and assesses the performance of the proposed frame- work in zero-shot learning scenarios. The findings indicate MTL- LATAM’s effectiveness across real-world and open-source datasets, achieving 95% and 97% task synergy. The results underscore the superior performance of multi-task models in zero-shot tasks compared to individual models trained exclusively on their respective datasets.
@inproceedings{nivarthi2024multi, abstract = {Ensuring the reliability of critical industrial systems across various sectors is crucial. It is essential to detect deviations from regular behaviour to mitigate disruptions and preserve infrastructure integrity. However, accurately labelling anomaly datasets is challenging due to their rarity and manual annotation subjectivity. The conventional approach of training separate models for each dataset entity further complicates model development. This paper presents a novel Multi-task Learning framework combining LSTM Autoencoder with temporal attention mechanism (MTL-LATAM) for effective time series anomaly detection. Multitask learning models improve adaptability and generalizability, leading to reduced runtime and compute power while supporting zero-shot evaluation. These models offer flexibility in detecting emerging anomalies. Additionally, we introduce a dynamic thresholding mechanism to incorporate temporal context for anomaly detection and provide visualizations of attention weights to enhance interpretability. The study compares MTL- LATAM, with other multi-task models, evaluates multi-task versus single-task models and assesses the performance of the proposed frame- work in zero-shot learning scenarios. The findings indicate MTL- LATAM’s effectiveness across real-world and open-source datasets, achieving 95% and 97% task synergy. The results underscore the superior performance of multi-task models in zero-shot tasks compared to individual models trained exclusively on their respective datasets.}, author = {Nivarthi, Chandana Priya and Huang, Zhixin and Gruhl, Christian and Sick, Bernhard}, booktitle = {International Joint Conference on Neural Networks (IJCNN)}, keywords = {itegpub}, pages = {1–10}, publisher = {IEEE}, title = {Multi-Task Representation Learning with Temporal Attention for Zero-Shot Time Series Anomaly Detection}, year = 2024 }
%0 Conference Paper %1 nivarthi2024multi %A Nivarthi, Chandana Priya %A Huang, Zhixin %A Gruhl, Christian %A Sick, Bernhard %B International Joint Conference on Neural Networks (IJCNN) %D 2024 %I IEEE %P 1--10 %R 10.1109/IJCNN60899.2024.10651344 %T Multi-Task Representation Learning with Temporal Attention for Zero-Shot Time Series Anomaly Detection %X Ensuring the reliability of critical industrial systems across various sectors is crucial. It is essential to detect deviations from regular behaviour to mitigate disruptions and preserve infrastructure integrity. However, accurately labelling anomaly datasets is challenging due to their rarity and manual annotation subjectivity. The conventional approach of training separate models for each dataset entity further complicates model development. This paper presents a novel Multi-task Learning framework combining LSTM Autoencoder with temporal attention mechanism (MTL-LATAM) for effective time series anomaly detection. Multitask learning models improve adaptability and generalizability, leading to reduced runtime and compute power while supporting zero-shot evaluation. These models offer flexibility in detecting emerging anomalies. Additionally, we introduce a dynamic thresholding mechanism to incorporate temporal context for anomaly detection and provide visualizations of attention weights to enhance interpretability. The study compares MTL- LATAM, with other multi-task models, evaluates multi-task versus single-task models and assesses the performance of the proposed frame- work in zero-shot learning scenarios. The findings indicate MTL- LATAM’s effectiveness across real-world and open-source datasets, achieving 95% and 97% task synergy. The results underscore the superior performance of multi-task models in zero-shot tasks compared to individual models trained exclusively on their respective datasets.

2023[ to top ]

Active Learning with Fast Model Updates and Class-Balanced Selection for Imbalanced Datasets.Huang, Zhixin; He, Yujiang; Herde, Marek; Huseljic, Denis; Sick, Bernhard. In Workshop on Interactive Adapative Learning (IAL), ECML PKDD, bll 28–45. 2023.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
- [ URL ]
Active learning strategies aim to minimize the number of queried samples for model training. However, two challenges in pool-based deep active learning on imbalanced datasets are observed in experiments: (1) the declining performance of active learning strategies due to imbalanced class distribution; (2)~the lack of sample diversity in acquisition batches due to the absence of timely model updates. This paper proposes the AL-FaMoUS, a general solution combining fast model updates and class-balanced minibatch selection to the active learning process. Furthermore, a simplification of the AL-FaMoUS, which selects one single sample in each acquisition minibatch, is experimentally evaluated on four image and three time-series imbalanced datasets. The results demonstrate that the implemented AL-FaMoUS outperforms the other adopted AL strategies, including uncertainty sampling and BALD solely combined with either the fast model update or the class balance selection strategy, in terms of accuracy and Macro F1 score.
@inproceedings{huang2023active, abstract = {Active learning strategies aim to minimize the number of queried samples for model training. However, two challenges in pool-based deep active learning on imbalanced datasets are observed in experiments: (1) the declining performance of active learning strategies due to imbalanced class distribution; (2) the lack of sample diversity in acquisition batches due to the absence of timely model updates. This paper proposes the AL-FaMoUS, a general solution combining fast model updates and class-balanced minibatch selection to the active learning process. Furthermore, a simplification of the AL-FaMoUS, which selects one single sample in each acquisition minibatch, is experimentally evaluated on four image and three time-series imbalanced datasets. The results demonstrate that the implemented AL-FaMoUS outperforms the other adopted AL strategies, including uncertainty sampling and BALD solely combined with either the fast model update or the class balance selection strategy, in terms of accuracy and Macro F1 score.}, author = {Huang, Zhixin and He, Yujiang and Herde, Marek and Huseljic, Denis and Sick, Bernhard}, booktitle = {Workshop on Interactive Adapative Learning (IAL), ECML PKDD}, keywords = {itegpub}, pages = {28–45}, title = {Active Learning with Fast Model Updates and Class-Balanced Selection for Imbalanced Datasets}, year = 2023 }
%0 Conference Paper %1 huang2023active %A Huang, Zhixin %A He, Yujiang %A Herde, Marek %A Huseljic, Denis %A Sick, Bernhard %B Workshop on Interactive Adapative Learning (IAL), ECML PKDD %D 2023 %P 28--45 %T Active Learning with Fast Model Updates and Class-Balanced Selection for Imbalanced Datasets %U https://ceur-ws.org/Vol-3470/paper5.pdf %X Active learning strategies aim to minimize the number of queried samples for model training. However, two challenges in pool-based deep active learning on imbalanced datasets are observed in experiments: (1) the declining performance of active learning strategies due to imbalanced class distribution; (2)~the lack of sample diversity in acquisition batches due to the absence of timely model updates. This paper proposes the AL-FaMoUS, a general solution combining fast model updates and class-balanced minibatch selection to the active learning process. Furthermore, a simplification of the AL-FaMoUS, which selects one single sample in each acquisition minibatch, is experimentally evaluated on four image and three time-series imbalanced datasets. The results demonstrate that the implemented AL-FaMoUS outperforms the other adopted AL strategies, including uncertainty sampling and BALD solely combined with either the fast model update or the class balance selection strategy, in terms of accuracy and Macro F1 score.
Spatio-Temporal Attention Graph Neural Network for Remaining Useful Life Prediction.Huang, Zhixin; He, Yujiang; Sick, Bernhard. In Computational Science and Computational Intelligence (CSCI), bll 99–105. IEEE, 2023.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
RUL prediction plays a crucial role in the health management of industrial systems. Given the increasing complexity of systems, data-driven predictive models have attracted significant research interest. Upon reviewing the existing literature, it appears that many studies either do not fully integrate both spatial and temporal features or employ only a single attention mechanism. Furthermore, there seems to be inconsistency in the choice of data normalization methods, particularly concerning operating conditions, which might influence predictive performance. To bridge these observations, this study presents the Spatio-Temporal Attention Graph Neural Network. Our model combines graph neural networks and temporal convolutional neural networks for spatial and temporal feature extraction, respectively. The cascade of these extractors, combined with multihead attention mechanisms for both spatio-temporal dimensions, aims to improve predictive precision and refine model explainability. Comprehensive experiments were conducted on the CMAPSS dataset to evaluate the impact of unified versus clustering normalization. The findings suggest that our model performs state-of-the-art results using only the unified normalization. Additionally, when dealing with datasets with multiple operating conditions, cluster normalization enhances the performance of our proposed model by up to 27%.
@inproceedings{huang2023spatio, abstract = {RUL prediction plays a crucial role in the health management of industrial systems. Given the increasing complexity of systems, data-driven predictive models have attracted significant research interest. Upon reviewing the existing literature, it appears that many studies either do not fully integrate both spatial and temporal features or employ only a single attention mechanism. Furthermore, there seems to be inconsistency in the choice of data normalization methods, particularly concerning operating conditions, which might influence predictive performance. To bridge these observations, this study presents the Spatio-Temporal Attention Graph Neural Network. Our model combines graph neural networks and temporal convolutional neural networks for spatial and temporal feature extraction, respectively. The cascade of these extractors, combined with multihead attention mechanisms for both spatio-temporal dimensions, aims to improve predictive precision and refine model explainability. Comprehensive experiments were conducted on the CMAPSS dataset to evaluate the impact of unified versus clustering normalization. The findings suggest that our model performs state-of-the-art results using only the unified normalization. Additionally, when dealing with datasets with multiple operating conditions, cluster normalization enhances the performance of our proposed model by up to 27%.}, author = {Huang, Zhixin and He, Yujiang and Sick, Bernhard}, booktitle = {Computational Science and Computational Intelligence (CSCI)}, keywords = {Graph_Neural_Network}, pages = {99–105}, publisher = {IEEE}, title = {Spatio-Temporal Attention Graph Neural Network for Remaining Useful Life Prediction}, year = 2023 }
%0 Conference Paper %1 huang2023spatio %A Huang, Zhixin %A He, Yujiang %A Sick, Bernhard %B Computational Science and Computational Intelligence (CSCI) %D 2023 %I IEEE %P 99--105 %R 10.1109/CSCI62032.2023.00022 %T Spatio-Temporal Attention Graph Neural Network for Remaining Useful Life Prediction %X RUL prediction plays a crucial role in the health management of industrial systems. Given the increasing complexity of systems, data-driven predictive models have attracted significant research interest. Upon reviewing the existing literature, it appears that many studies either do not fully integrate both spatial and temporal features or employ only a single attention mechanism. Furthermore, there seems to be inconsistency in the choice of data normalization methods, particularly concerning operating conditions, which might influence predictive performance. To bridge these observations, this study presents the Spatio-Temporal Attention Graph Neural Network. Our model combines graph neural networks and temporal convolutional neural networks for spatial and temporal feature extraction, respectively. The cascade of these extractors, combined with multihead attention mechanisms for both spatio-temporal dimensions, aims to improve predictive precision and refine model explainability. Comprehensive experiments were conducted on the CMAPSS dataset to evaluate the impact of unified versus clustering normalization. The findings suggest that our model performs state-of-the-art results using only the unified normalization. Additionally, when dealing with datasets with multiple operating conditions, cluster normalization enhances the performance of our proposed model by up to 27%.

2022[ to top ]

Design of Explainability Module with Experts in the Loop for Visualization and Dynamic Adjustment of Continual Learning.He, Yujiang; Huang, Zhixin; Sick, Bernhard. In Workshop on Interactive Machine Learning Workshop (IMLW), AAAI, bll 1–6. 2022.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
- [ URL ]
Continual learning can enable neural networks to evolve by learning new tasks sequentially in task-changing scenarios. However, two general challenges should be overcome in further research before we apply this technique to real-world applications. Firstly, newly collected novelties from the data stream in applications could contain anomalies that are meaningless for continual learning. Instead of viewing them as a new task for updating, we have to filter out such anomalies to reduce the disturbance of extremely high-entropy data for the progression of convergence. Secondly, fewer efforts have been put into research regarding the explainability of continual learning, which leads to a lack of transparency and credibility of the updated neural networks. Elaborated explanations about the process and result of continual learning can help experts in judgment and making decisions. Therefore, we propose the conceptual design of an explainability module with experts in the loop based on techniques, such as dimension reduction, visualization, and evaluation strategies. This work aims to overcome the mentioned challenges by sufficiently explaining and visualizing the identified anomalies and the updated neural network. With the help of this module, experts can be more confident in decision-making regarding anomaly filtering, dynamic adjustment of hyperparameters, data backup, etc.
@inproceedings{he2022design, abstract = {Continual learning can enable neural networks to evolve by learning new tasks sequentially in task-changing scenarios. However, two general challenges should be overcome in further research before we apply this technique to real-world applications. Firstly, newly collected novelties from the data stream in applications could contain anomalies that are meaningless for continual learning. Instead of viewing them as a new task for updating, we have to filter out such anomalies to reduce the disturbance of extremely high-entropy data for the progression of convergence. Secondly, fewer efforts have been put into research regarding the explainability of continual learning, which leads to a lack of transparency and credibility of the updated neural networks. Elaborated explanations about the process and result of continual learning can help experts in judgment and making decisions. Therefore, we propose the conceptual design of an explainability module with experts in the loop based on techniques, such as dimension reduction, visualization, and evaluation strategies. This work aims to overcome the mentioned challenges by sufficiently explaining and visualizing the identified anomalies and the updated neural network. With the help of this module, experts can be more confident in decision-making regarding anomaly filtering, dynamic adjustment of hyperparameters, data backup, etc.}, author = {He, Yujiang and Huang, Zhixin and Sick, Bernhard}, booktitle = {Workshop on Interactive Machine Learning Workshop (IMLW), AAAI}, keywords = {itegpub}, pages = {1–6}, title = {Design of Explainability Module with Experts in the Loop for Visualization and Dynamic Adjustment of Continual Learning}, year = 2022 }
%0 Conference Paper %1 he2022design %A He, Yujiang %A Huang, Zhixin %A Sick, Bernhard %B Workshop on Interactive Machine Learning Workshop (IMLW), AAAI %D 2022 %P 1--6 %T Design of Explainability Module with Experts in the Loop for Visualization and Dynamic Adjustment of Continual Learning %U https://arxiv.org/abs/2202.06781 %X Continual learning can enable neural networks to evolve by learning new tasks sequentially in task-changing scenarios. However, two general challenges should be overcome in further research before we apply this technique to real-world applications. Firstly, newly collected novelties from the data stream in applications could contain anomalies that are meaningless for continual learning. Instead of viewing them as a new task for updating, we have to filter out such anomalies to reduce the disturbance of extremely high-entropy data for the progression of convergence. Secondly, fewer efforts have been put into research regarding the explainability of continual learning, which leads to a lack of transparency and credibility of the updated neural networks. Elaborated explanations about the process and result of continual learning can help experts in judgment and making decisions. Therefore, we propose the conceptual design of an explainability module with experts in the loop based on techniques, such as dimension reduction, visualization, and evaluation strategies. This work aims to overcome the mentioned challenges by sufficiently explaining and visualizing the identified anomalies and the updated neural network. With the help of this module, experts can be more confident in decision-making regarding anomaly filtering, dynamic adjustment of hyperparameters, data backup, etc.
Fast Bayesian Updates for Deep Learning with a Use Case in Active Learning.Herde, Marek; Huang, Zhixin; Huseljic, Denis; Kottke, Daniel; Vogt, Stephan; Sick, Bernhard. In arXiv e-prints, bl arXiv:2210.06112. 2022.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
- [ URL ]
Retraining deep neural networks when new data arrives is typically computationally expensive. Moreover, certain applications do not allow such costly retraining due to time or computational constraints. Fast Bayesian updates are a possible solution to this issue. Therefore, we propose a Bayesian update based on Monte-Carlo samples and a last-layer Laplace approximation for different Bayesian neural network types, i.e., Dropout, Ensemble, and Spectral Normalized Neural Gaussian Process (SNGP). In a large-scale evaluation study, we show that our updates combined with SNGP represent a fast and competitive alternative to costly retraining. As a use case, we combine the Bayesian updates for SNGP with different sequential query strategies to exemplarily demonstrate their improved selection performance in active learning.
@article{herde2022fast, abstract = {Retraining deep neural networks when new data arrives is typically computationally expensive. Moreover, certain applications do not allow such costly retraining due to time or computational constraints. Fast Bayesian updates are a possible solution to this issue. Therefore, we propose a Bayesian update based on Monte-Carlo samples and a last-layer Laplace approximation for different Bayesian neural network types, i.e., Dropout, Ensemble, and Spectral Normalized Neural Gaussian Process (SNGP). In a large-scale evaluation study, we show that our updates combined with SNGP represent a fast and competitive alternative to costly retraining. As a use case, we combine the Bayesian updates for SNGP with different sequential query strategies to exemplarily demonstrate their improved selection performance in active learning.}, author = {Herde, Marek and Huang, Zhixin and Huseljic, Denis and Kottke, Daniel and Vogt, Stephan and Sick, Bernhard}, journal = {arXiv e-prints}, keywords = {itegpub}, pages = {arXiv:2210.06112}, title = {Fast Bayesian Updates for Deep Learning with a Use Case in Active Learning}, year = 2022 }
%0 Journal Article %1 herde2022fast %A Herde, Marek %A Huang, Zhixin %A Huseljic, Denis %A Kottke, Daniel %A Vogt, Stephan %A Sick, Bernhard %D 2022 %J arXiv e-prints %P arXiv:2210.06112 %T Fast Bayesian Updates for Deep Learning with a Use Case in Active Learning %U https://arxiv.org/abs/2210.06112 %X Retraining deep neural networks when new data arrives is typically computationally expensive. Moreover, certain applications do not allow such costly retraining due to time or computational constraints. Fast Bayesian updates are a possible solution to this issue. Therefore, we propose a Bayesian update based on Monte-Carlo samples and a last-layer Laplace approximation for different Bayesian neural network types, i.e., Dropout, Ensemble, and Spectral Normalized Neural Gaussian Process (SNGP). In a large-scale evaluation study, we show that our updates combined with SNGP represent a fast and competitive alternative to costly retraining. As a use case, we combine the Bayesian updates for SNGP with different sequential query strategies to exemplarily demonstrate their improved selection performance in active learning.
Active Learning in Multivariate Time Series Anomaly Detection.Huang, Zhixin. In Organic Computing -- Doctoral Dissertation Colloquium 2021, S. Tomforde, C. Krupitzer (reds.), bll 113–124. kassel university press, 2022.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
Multivariate time series anomaly detection is an important research field of industry, science, and finances. In practical applications, unlabeled data is abundant, and labeled data is rare. Experts’ experience plays a key role in data labeling, while experts often have limited time, are expensive, and are concerned primarily with finding the potential anomalies. Active learning, which has been widely used to select the most informative data for labeling through experts, can improve training with less labeling effort. However, most active learning algorithms have not been evaluated on datasets with imbalanced class frequencies, such as anomaly detection. Although some semi-supervised and active learning methods were proposed to handle univariate time series data, not much work has been done on multivariate time series classification. Therefore, this article identifies critical challenges and proposes potential solutions for active learning in multivariate time series anomaly detection. We combine all possible solutions into a framework that conforms to the scope of the observation/control system in organic computing.
@incollection{huang2022active, abstract = {Multivariate time series anomaly detection is an important research field of industry, science, and finances. In practical applications, unlabeled data is abundant, and labeled data is rare. Experts’ experience plays a key role in data labeling, while experts often have limited time, are expensive, and are concerned primarily with finding the potential anomalies. Active learning, which has been widely used to select the most informative data for labeling through experts, can improve training with less labeling effort. However, most active learning algorithms have not been evaluated on datasets with imbalanced class frequencies, such as anomaly detection. Although some semi-supervised and active learning methods were proposed to handle univariate time series data, not much work has been done on multivariate time series classification. Therefore, this article identifies critical challenges and proposes potential solutions for active learning in multivariate time series anomaly detection. We combine all possible solutions into a framework that conforms to the scope of the observation/control system in organic computing.}, author = {Huang, Zhixin}, booktitle = {Organic Computing – Doctoral Dissertation Colloquium 2021}, editor = {Tomforde, Sven and Krupitzer, Christian}, keywords = {itegpub}, pages = {113–124}, publisher = {kassel university press}, title = {Active Learning in Multivariate Time Series Anomaly Detection}, year = 2022 }
%0 Book Section %1 huang2022active %A Huang, Zhixin %B Organic Computing -- Doctoral Dissertation Colloquium 2021 %D 2022 %E Tomforde, Sven %E Krupitzer, Christian %I kassel university press %P 113--124 %T Active Learning in Multivariate Time Series Anomaly Detection %X Multivariate time series anomaly detection is an important research field of industry, science, and finances. In practical applications, unlabeled data is abundant, and labeled data is rare. Experts’ experience plays a key role in data labeling, while experts often have limited time, are expensive, and are concerned primarily with finding the potential anomalies. Active learning, which has been widely used to select the most informative data for labeling through experts, can improve training with less labeling effort. However, most active learning algorithms have not been evaluated on datasets with imbalanced class frequencies, such as anomaly detection. Although some semi-supervised and active learning methods were proposed to handle univariate time series data, not much work has been done on multivariate time series classification. Therefore, this article identifies critical challenges and proposes potential solutions for active learning in multivariate time series anomaly detection. We combine all possible solutions into a framework that conforms to the scope of the observation/control system in organic computing.
LwTool: A data processing toolkit for building a real-time pressure mapping smart textile software system.Guo, Tao; Huang, Zhixin; Cheng, Jingyuan. In Pervasive and Mobile Computing, 80, bl 101540. Elsevier, 2022.
- [ Abstract ]
- [ BibTeX ]
- [ EndNote ]
- [ URL ]
Pressure mapping smart textile is a new type of sensing modality that transforms the pressure distribution over surfaces into digital ”image” and ”video”, that has rich application scenarios in Human Activity Recognition (HAR), because all human activities are linked with force change over certain surfaces. To speed up its application exploration, we propose a toolkit named LwTool for the data processing, including: (a) a feature library, including 1830 ready-to-use temporal and spatial features, (b) a hierarchical feature selection framework that automatically picks out the best features for a new application from the feature library. As real-time processing capability is important for instant user feedback, we emphasize not only on good recognition result but also on reducing time cost when selecting features. Our library and algorithms are validated on Smart-Toy and Smart-Bedsheet applications, an 89.7% accuracy for Smart-Toy and an 83.8% accuracy for Smart-Bedsheet can be achieved (10-fold cross-validation) using our feature library. Adopting the feature selection algorithm, the processing speed is increased by more than 3 times while maintaining high accuracy for both two applications. We believe our method could be a general and powerful toolkit in building real-time recognition software systems for pressure mapping smart textile.
@article{guo2022lwtool, abstract = {Pressure mapping smart textile is a new type of sensing modality that transforms the pressure distribution over surfaces into digital ”image” and ”video”, that has rich application scenarios in Human Activity Recognition (HAR), because all human activities are linked with force change over certain surfaces. To speed up its application exploration, we propose a toolkit named LwTool for the data processing, including: (a) a feature library, including 1830 ready-to-use temporal and spatial features, (b) a hierarchical feature selection framework that automatically picks out the best features for a new application from the feature library. As real-time processing capability is important for instant user feedback, we emphasize not only on good recognition result but also on reducing time cost when selecting features. Our library and algorithms are validated on Smart-Toy and Smart-Bedsheet applications, an 89.7% accuracy for Smart-Toy and an 83.8% accuracy for Smart-Bedsheet can be achieved (10-fold cross-validation) using our feature library. Adopting the feature selection algorithm, the processing speed is increased by more than 3 times while maintaining high accuracy for both two applications. We believe our method could be a general and powerful toolkit in building real-time recognition software systems for pressure mapping smart textile.}, author = {Guo, Tao and Huang, Zhixin and Cheng, Jingyuan}, journal = {Pervasive and Mobile Computing}, keywords = {SmartTextile}, pages = 101540, publisher = {Elsevier}, title = {LwTool: A data processing toolkit for building a real-time pressure mapping smart textile software system}, volume = 80, year = 2022 }
%0 Journal Article %1 guo2022lwtool %A Guo, Tao %A Huang, Zhixin %A Cheng, Jingyuan %D 2022 %I Elsevier %J Pervasive and Mobile Computing %P 101540 %R 10.1016/j.sysarc.2021.102387 %T LwTool: A data processing toolkit for building a real-time pressure mapping smart textile software system %U https://www.sciencedirect.com/science/article/abs/pii/S1574119222000013 %V 80 %X Pressure mapping smart textile is a new type of sensing modality that transforms the pressure distribution over surfaces into digital ”image” and ”video”, that has rich application scenarios in Human Activity Recognition (HAR), because all human activities are linked with force change over certain surfaces. To speed up its application exploration, we propose a toolkit named LwTool for the data processing, including: (a) a feature library, including 1830 ready-to-use temporal and spatial features, (b) a hierarchical feature selection framework that automatically picks out the best features for a new application from the feature library. As real-time processing capability is important for instant user feedback, we emphasize not only on good recognition result but also on reducing time cost when selecting features. Our library and algorithms are validated on Smart-Toy and Smart-Bedsheet applications, an 89.7% accuracy for Smart-Toy and an 83.8% accuracy for Smart-Bedsheet can be achieved (10-fold cross-validation) using our feature library. Adopting the feature selection algorithm, the processing speed is increased by more than 3 times while maintaining high accuracy for both two applications. We believe our method could be a general and powerful toolkit in building real-time recognition software systems for pressure mapping smart textile.

2021[ to top ]

Transport efficiency of biofunctionalized magnetic particles tailored by surfactant concentration.Reginka, Meike; Hoang, Hai; Efendi, Özge; Merkel, Maximilan; Huhnstock, Rico; Holzinger, Dennis; Dingel, Kristina; Sick, Bernhard; Bertinetti, Daniela; Herberg, Friedrich; Ehresmann, Arno. In Langmuir, 37(28), bll 8498–8507. ACS, 2021.
- [ BibTeX ]
- [ EndNote ]
- [ URL ]
@article{reginka2021transport, author = {{Reginka}, Meike and {Hoang}, Hai and {Efendi}, {\"O}zge and {Merkel}, Maximilan and {Huhnstock}, Rico and {Holzinger}, Dennis and {Dingel}, Kristina and {Sick}, Bernhard and {Bertinetti}, Daniela and {Herberg}, Friedrich and {Ehresmann}, Arno}, journal = {Langmuir}, keywords = {itegpub}, number = 28, pages = {8498–8507}, publisher = {ACS}, title = {Transport efficiency of biofunctionalized magnetic particles tailored by surfactant concentration}, volume = 37, year = 2021 }
%0 Journal Article %1 reginka2021transport %A Reginka, Meike %A Hoang, Hai %A Efendi, Özge %A Merkel, Maximilan %A Huhnstock, Rico %A Holzinger, Dennis %A Dingel, Kristina %A Sick, Bernhard %A Bertinetti, Daniela %A Herberg, Friedrich %A Ehresmann, Arno %D 2021 %I ACS %J Langmuir %N 28 %P 8498--8507 %R 10.1021/acs.langmuir.1c00900 %T Transport efficiency of biofunctionalized magnetic particles tailored by surfactant concentration %U https://pubs.acs.org/doi/abs/10.1021/acs.langmuir.1c00900 %V 37
The Problem with Real-World Novelty Detection -- Issues in Multivariate Probabilistic Models.Gruhl, Christian; Hannan, Abdul; Huang, Zhixin; Nivarthi, Chandana; Vogt, Stephan. In Workshop on Self-Improving System Integration (SISSY), ACSOS, bll 204–209. IEEE, 2021.
- [ BibTeX ]
- [ EndNote ]
@inproceedings{gruhl2021problem, author = {Gruhl, Christian and Hannan, Abdul and Huang, Zhixin and Nivarthi, Chandana and Vogt, Stephan}, booktitle = {Workshop on Self-Improving System Integration (SISSY), ACSOS}, keywords = {itegpub}, pages = {204–209}, publisher = {IEEE}, title = {The Problem with Real-World Novelty Detection – Issues in Multivariate Probabilistic Models}, year = 2021 }
%0 Conference Paper %1 gruhl2021problem %A Gruhl, Christian %A Hannan, Abdul %A Huang, Zhixin %A Nivarthi, Chandana %A Vogt, Stephan %B Workshop on Self-Improving System Integration (SISSY), ACSOS %D 2021 %I IEEE %P 204--209 %R 10.1109/acsos-c52956.2021.00055 %T The Problem with Real-World Novelty Detection -- Issues in Multivariate Probabilistic Models
Toward Application of Continuous Power Forecasts in a Regional Flexibility Market.He, Yujiang; Huang, Zhixin; Sick, Bernhard. In International Joint Conference on Neural Networks (IJCNN), bll 1–8. IEEE, 2021.
- [ BibTeX ]
- [ EndNote ]
@inproceedings{he2021application, author = {He, Yujiang and Huang, Zhixin and Sick, Bernhard}, booktitle = {International Joint Conference on Neural Networks (IJCNN)}, keywords = {itegpub}, pages = {1–8}, publisher = {IEEE}, title = {Toward Application of Continuous Power Forecasts in a Regional Flexibility Market}, year = 2021 }
%0 Conference Paper %1 he2021application %A He, Yujiang %A Huang, Zhixin %A Sick, Bernhard %B International Joint Conference on Neural Networks (IJCNN) %D 2021 %I IEEE %P 1--8 %R 10.1109/ijcnn52387.2021.9533626 %T Toward Application of Continuous Power Forecasts in a Regional Flexibility Market
Uncertainty and Utility Sampling with Pre-Clustering.Huang, Zhixin; He, Yujiang; Vogt, Stephan; Sick, Bernhard. In Workshop on Interactive Adaptive Learning (IAL), ECML PKDD. 2021.
- [ BibTeX ]
- [ EndNote ]
@inproceedings{huang2021uncertainty, author = {Huang, Zhixin and He, Yujiang and Vogt, Stephan and Sick, Bernhard}, booktitle = {Workshop on Interactive Adaptive Learning (IAL), ECML PKDD}, keywords = {itegpub}, note = {(accepted)}, title = {Uncertainty and Utility Sampling with Pre-Clustering}, year = 2021 }
%0 Conference Paper %1 huang2021uncertainty %A Huang, Zhixin %A He, Yujiang %A Vogt, Stephan %A Sick, Bernhard %B Workshop on Interactive Adaptive Learning (IAL), ECML PKDD %D 2021 %T Uncertainty and Utility Sampling with Pre-Clustering

2017[ to top ]

TPM Framework: a Comprehensive Kit for Exploring Applications with Textile Pressure Mapping Matrix.Zhou, Bo; Cheng, Jingyuan; Mawandia, Ankur; He, Yujiang; Huang, Zhixin; Sundholm, Mathias; Yildrim, Muhammet; Cruz, Heber; Lukowicz, Paul. In International Conference on Mobile Ubiquitous Computing, Systems, Services and Technologies (UBICOMM). IARIA, 2017.
- [ BibTeX ]
- [ EndNote ]
@inproceedings{zhou2017tpm, author = {Zhou, Bo and Cheng, Jingyuan and Mawandia, Ankur and He, Yujiang and Huang, Zhixin and Sundholm, Mathias and Yildrim, Muhammet and Cruz, Heber and Lukowicz, Paul}, booktitle = {International Conference on Mobile Ubiquitous Computing, Systems, Services and Technologies (UBICOMM)}, keywords = {not_ies}, publisher = {IARIA}, title = {TPM Framework: a Comprehensive Kit for Exploring Applications with Textile Pressure Mapping Matrix}, year = 2017 }
%0 Conference Paper %1 zhou2017tpm %A Zhou, Bo %A Cheng, Jingyuan %A Mawandia, Ankur %A He, Yujiang %A Huang, Zhixin %A Sundholm, Mathias %A Yildrim, Muhammet %A Cruz, Heber %A Lukowicz, Paul %B International Conference on Mobile Ubiquitous Computing, Systems, Services and Technologies (UBICOMM) %D 2017 %I IARIA %T TPM Framework: a Comprehensive Kit for Exploring Applications with Textile Pressure Mapping Matrix