Improving Fault Detection in Acoustic Signals through Explainable AI

Sammanfattning

Fault detection using acoustic signals has emerged as a crucial approach in industrial predictive maintenance due to its non-intrusive nature and cost-effectiveness. Recent advances in machine learning (ML) and deep learning (DL) models have demonstrated promising results in detecting anomalies from acoustic data. Despite their effectiveness, these models function as black boxes, offering little insight into which features contribute to anomaly detection. We propose a SHAP-weighted reconstruction error method that operates in two stages. First, SHAP analysis identifies frequency-time regions that characterize anomalous machine sounds by computing feature contributions to reconstruction errors. Second, these SHAP-derived importance weights are used to amplify errors in anomaly-prone frequency bands while suppressing irrelevant regions. The method is evaluated on multiple machine types, including Slider, Fan, Pump, valve, ToyCar and Toy conveyor using the DCASE 2020 Task 2 dataset. Experimental results demonstrate that the proposed SHAP-weighted method consistently improves anomaly detection across all evaluated machine types. For the Valve machine, which exhibited lower baseline performance, the method achieved significant improvements in AUC and recall. The study will be conducted on real-world industrial datasets, evaluating the effectiveness of ML and DL models in terms of feature importance, anomaly attribution, and classification accuracy. The main contributions of this work are: (1) using SHAP to identify frequency-time regions where anomalies occur; (2) a SHAP-weighted reconstruction error method that improves detection performance by amplifying errors in important frequency bands; (3) Evaluation across multiple machine types and IDs.Overall, our results demonstrate that XAI is not just for explanation—it can also be used to make anomaly detection better.