Counterfactual Explanation for Auto-Encoder Based Time-Series Anomaly Detection

TL;DR

The paper tackles interpretability in auto-encoder–based time-series anomaly detection by introducing a three-module framework: an AE detector, a feature selector, and a gradient-based counterfactual explainer. It jointly identifies the most influential features responsible for anomalies and provides counterfactuals that explain why a sample is flagged, evaluated via validity, sparsity, and distance. Experiments on the SKAB benchmark and real industrial data show that the proposed approach achieves actionable, sparse explanations with competitive anomaly detection performance, improving diagnostic interpretability. The work advances practical anomaly analysis by enabling targeted, signal-level explanations that facilitate root-cause analysis and maintenance planning, with potential for broader industrial adoption.

Abstract

The complexity of modern electro-mechanical systems require the development of sophisticated diagnostic methods like anomaly detection capable of detecting deviations. Conventional anomaly detection approaches like signal processing and statistical modelling often struggle to effectively handle the intricacies of complex systems, particularly when dealing with multi-variate signals. In contrast, neural network-based anomaly detection methods, especially Auto-Encoders, have emerged as a compelling alternative, demonstrating remarkable performance. However, Auto-Encoders exhibit inherent opaqueness in their decision-making processes, hindering their practical implementation at scale. Addressing this opacity is essential for enhancing the interpretability and trustworthiness of anomaly detection models. In this work, we address this challenge by employing a feature selector to select features and counterfactual explanations to give a context to the model output. We tested this approach on the SKAB benchmark dataset and an industrial time-series dataset. The gradient based counterfactual explanation approach was evaluated via validity, sparsity and distance measures. Our experimental findings illustrate that our proposed counterfactual approach can offer meaningful and valuable insights into the model decision-making process, by explaining fewer signals compared to conventional approaches. These insights enhance the trustworthiness and interpretability of anomaly detection models.

Figures (6)

• Figure 1: Our proposed methods has 3 modules, 1) Anomaly detector, 2) Feature selector, and 3) Counterfactual explainer. The samples that are classified anomalous by the anomaly detector (module 1) are explained though the feature selector (module 2) and the counterfactual explainer (module 3). The explainer (module 3) uses the selected features from the feature selector and the input sample.
• Figure 2: Industrial data: UMAP embedding learnt on no-fault and anomalous data from the test set. Later the generated counterfactual is projected into the same embedding.
• Figure 3: Plot of counterfactual explanation generated by our approach for industrial dataset. This plotted sample was of correlation loss anomaly. Signal 7 and signal 8 in blue show the input and signal 7 in orange shows the explanation.
• Figure 4: Plot showing the counterfactual explanations provided by our approach and the anomalous samples. Only the high impact features that were explained are plotted.
• Figure 5: Plot showing the explanations provided by reconstruction, counterfactual(CF) based (i.e., without feature selector) and our approach (i.e., with feature selector). Additionally the input sample is plotted.