Explainable Artificial Intelligence for Industrial Anomaly Diagnosis in Multi-sensor Data

• DIRECTORS: Xabier De Carlos and Ehki Zugasti
• UNIVERSITY: Mondragon Unibertsitatea

ABSTRACT

This thesis explores the potential of Explainable Artificial Intelligence (XAI) in the context of time-series anomaly detection and diagnosis. By enhancing the transparency of traditionally opaque models and offering immediate and intelligible explanations, we have set the stage for more informed decision-making processes across diverse sectors.

Prior to our practical exploration, we reviewed the existing literature on XAI, anomaly detection, and diagnosis. Then, our experiments start with a study of the Counterfactual Explanation Method (CEM) in time-series tasks. This investigation revealed both the advantages and limitations of CEM. Acknowledging its shortcomings, particularly its time-consuming nature. Then, we present the Real-Time Guided Counterfactual Explanations (RTGCEx) method. This innovative method is a model-agnostic approach that provides user-driven counterfactual explanations in real-time across different domains and data types.

Afterward, to avoid losing essential information that anomaly detectors might contain and that model-agnostic methods might miss, we address the challenge of creating intrinsically interpretable models. To achieve this, we first introduce the Diagnostic Fourier-based Spatio-temporal Transformer (DFSTrans). This tool combines the capabilities of 1D Convolutional Neural Networks with a Transformer-inspired structure. This model effectively learns spatial and temporal dependencies in multivariate sensor data, proving to be a potent tool for diagnosing anomalies. Recognizing the challenges associated with obtaining labeled data, we developed an unsupervised variant, termed uDFSTrans. This model incorporates a dual strategy: a multi-masking technique and a context-oriented attention mechanism, facilitating the detection and elucidation of anomalies without the necessity for labeled data.