Evaluating Applicability of Explainable AI (XAI) and Anomaly Detection Methods in Industrial Test Logs

Sammanfattning

This thesis evaluates the applicability of anomaly detection and Explainable Artificial Intelligence (XAI) methods for analysing industrial software test logs. Modern testing environmentsgenerate large volumes of log data, making manual analysis inefficient and error-prone. Machine Learning (ML) techniques such as anomaly detection can identify irregular patterns, but theirblack-box nature limits trust and interpretability. To address this, the study investigates whether combining anomaly detection with XAI can enhance transparency and usability in industrial testlogs. Additionally, Large Language Models (LLMs) were integrated as an alternative technique to generate natural language explanations.The research was conducted as a case study at Westermo Network Technologies AB (Westermo), using both proprietary and open-source log datasets. Isolation Forest (iForest) was applied for unsupervised anomaly detection, while SHapley Additive exPlanations (SHAP) and Diverse Counterfactual Explanations (DiCE) were explored for interpretability. Evaluation followed the Goal-Question-Metric (GQM) framework, measuring detection accuracy, response time, explanation fidelity, cost, and perceived usefulness. The study involved a focus group consisting of practitioners from Westermo that provided valuable qualitative feedback through discussions and a survey. Proof-of-concept implementation was done to collect quantitative data. Quantitative results show that anomaly detection processes large datasets efficiently but suffers from false positives in industrial logs and SHAP provided consistent feature attributions. Qualitative results show that LLMs were preferred by participants for interpretability, despite challenges in reproducibility and cost. The findings suggest that hybrid approaches combining anomaly detection, XAI, and LLMs hold promise for improving transparency and trust in industrial testing workflows. Future work could focus on experimenting with different types of industrial logs, training strategies for anomaly detection, other types of anomaly detection and XAI methods, and agentic frameworks.