Intelligent Systems Research Group

The field of anomaly detection comprises methods for identifying observations that either deviate from most observations or otherwise do not conform to an expected state of normality. Especially in real-time applications, effective anomaly detection procedures are vital to detect critical system states while limiting false alarms. To ensure the practicality of any anomaly detection system, uncertainty quantification becomes increasingly important, especially in the context of IoT with a multitude of different data streams from sensors.

Key Challanges in Anomaly Detection:

• Non-Stationary Data: Systems evolve over time, requiring adaptive models that maintain accuracy and uncertainty guarantees.

• Noisy and High-Dimensional Data: Distinguishing true anomalies from noise in complex datasets remains difficult.

• Scalability: Real-time processing demands efficient algorithms that balance computational cost and accuracy.

• Interpretability: Stakeholders need explainable models, yet many advanced ML methods (e.g., deep learning) lack transparency.

• Reliability: Ensuring that uncertainty estimates remain well-calibrated and trustworthy over time is critical, especially in dynamic environments where distribution shifts, concept drift, and data scarcity can degrade reliability

Our Approach:

Our research builds on top of the conformal inference framework for creating anomaly detection systems with reliable and principled uncertainty quantification. Conformal inference is model-agnostic and non-parametric and a promising approach to control model-uncertainties. Uncertainty-quantified anomaly detection in temporal settings is a necessity for many industry applications, but challenges like non-stationarity, noise, scalability, interpretability, reliability must be addressed to unlock its full potential.
Fruthermore, our research focuses on Online Machine Learning and Platform Design and Development.

In contrast to anomaly detection, which focuses on identifying deviations from normal behavior, fault diagnosis aims to determine the specific type and root cause of faults or failures in a system. It typically relies on labeled data from known fault conditions to classify and analyze faults accurately.

Cross-domain fault diagnosis addresses scenarios where labeled data from the target domain (e.g., a specific machine or operating condition) is scarce or unavailable. Transfer learning is a powerful approach to tackle this problem by leveraging knowledge from a source domain, such as a similar machine or operating condition, to enhance diagnosis in the target domain. This is achieved by adapting the source domain’s features, models, or learned representations to the target domain, minimizing the need for extensive labeled data.

Key Challenges in Cross-Domain Fault Diagnosis:

• Domain Discrepancies: Differences in operating conditions, sensors, or environments can lead to significant variations in feature distributions between the source and target domains.
• Lack of Labeled Data: Target domains often lack sufficient labeled data for effective adaptation, making it challenging to learn robust representations.
• Model Generalization: Ensuring models generalize well across diverse domains without overfitting to specific domain.
• Class Imbalance: Fault data often exhibits severe class imbalances, with certain fault types being underrepresented.
• Interpretability: Transferred models may exhibit reduced interpretability due to the complexity of domain adaptation mechanisms, limiting trust and applicability in critical systems.

Our Approach:

Our research project was focused on fault diagnosis in wind turbines using SCADA and vibration data collected from multiple turbines. A variety of anomaly detection models were developed to process this data, each generating an anomaly score that reflects the severity of detected anomalies—higher scores indicating more severe deviations. These scores were aggregated into a new dataset, termed the Anomaly Space, where each wind turbine component is assigned corresponding anomaly scores. To enable fault diagnosis, labels were provided by diagnosticians, who documented failures over several years, noting when and where they occurred and when they were resolved. A supervised learning classifier was then trained on this labeled anomaly space to deliver accurate fault diagnoses.

The design of the Anomaly Space, which focuses on deviations from normal behavior for each wind turbine individually, inherently reduces cross-domain discrepancies. By standardizing anomaly detection for individual turbines, the feature distribution between turbines becomes more consistent, facilitating cross-domain fault diagnosis. This approach tackles one of the key challenges in transfer learning by narrowing the gap between source and target domains.

Weber K., Preisach C.
Supervised Transfer Learning Framework for Fault Diagnosis in Wind Turbines
Proceedings of the Upper-Rhine Artificial Intelligence Symposium 2024, Offenburg, Germany. [pdf]

Hennhöfer O., Preisach C.
Leave-One-Out-, Bootstrap- and Cross-Conformal Anomaly Detection
Proceedings, 15th IEEE International Conference on Knowledge Graph ICKG'24, Abu Dhabi, UCE, 2024. [pdf]

Hennhöfer O.
Python-Library: unquad
Conformal Anomaly Detection. [link]