Towards Interpretable Autoencoders for Anomaly Detection: An Application of Gradient-Based XAI Methods

Abstract Wind turbines are essential in the transition to renewable energy, helping to reduce greenhouse gas emissions and combat climate change. However, their efficient operation and maintenance pose significant challenges, particularly in anomaly detection due to the lack of interpretability in deep learning models. Existing models often function as “black boxes,” making it difficult to understand the reasoning behind their predictions. This paper proposes the application of gradient-based Explainable AI (XAI) methods to improve the interpretability of autoencoders used for anomaly detection in wind turbines. Using 11.5 years of SCADA data from a wind farm in Italy, a Multi-Layer Perceptron (MLP) autoencoder is developed to model normal operational behavior and detect anomalies. The novelty of this approach lies in the use of SmoothGrad to smooth gradients and analyze the Jacobian matrix, revealing the relationships between sensor inputs. The results demonstrate that these XAI methods enhance the transparency and trustworthiness of the anomaly detection system, supporting more effective maintenance strategies for wind turbines. Specifically, the improved interpretability allows for targeted inspections, where maintenance teams can prioritize and focus on the most critical areas identified by the model, thereby reducing downtime and maintenance costs.