Explainable Anomaly Detection: Counterfactual driven What-If Analysis

TL;DR

This work tackles explainability in anomaly detection for predictive maintenance by using counterfactual explanations to enable what-if analysis. It integrates a Temporal Convolutional Network (TCN) anomaly detector with CoMTE-based counterfactual generation and demonstrates the approach on the PRONOSTIA bearing dataset. A proof-of-concept shows that counterfactuals can be presented as actionable scenarios, including online interactions, while achieving high cross-validation accuracy and robust anomaly detection. The results suggest practical value for diagnosing root causes and exploring mitigations in real-time maintenance planning, paving the way for more complex, real-world deployments.

Abstract

There exists three main areas of study inside of the field of predictive maintenance: anomaly detection, fault diagnosis, and remaining useful life prediction. Notably, anomaly detection alerts the stakeholder that an anomaly is occurring. This raises two fundamental questions: what is causing the fault and how can we fix it? Inside of the field of explainable artificial intelligence, counterfactual explanations can give that information in the form of what changes to make to put the data point into the opposing class, in this case "healthy". The suggestions are not always actionable which may raise the interest in asking "what if we do this instead?" In this work, we provide a proof of concept for utilizing counterfactual explanations as what-if analysis. We perform this on the PRONOSTIA dataset with a temporal convolutional network as the anomaly detector. Our method presents the counterfactuals in the form of a what-if analysis for this base problem to inspire future work for more complex systems and scenarios.

Figures (6)

• Figure 1: Story board of counterfactuals as a what-if analysis. The user received a notification of an anomaly (a). The user may then receive an unactionable explanation (b). The user could then perform what-if analysis via counterfactuals to test a more actionable explanation (c).
• Figure 2: Example of what-if analysis for the value of an item over time
• Figure 3: Data flow model beginning with class balancing and ending with k-d tree creation. Class balancing is used to increase the size of the training data to an appropriate ratio of healthy and anomalous data. The model is trained using the balanced data and tested with the original testing data. K-d trees are the created to represent the accurately predicted data.
• Figure 4: Bearing1_1 dataset split into its two features
• Figure 5: K-d tree partitioning two-dimensional data