Fuzzy Rule-Based Anomaly Explanation in Micro-electromechanical Systems

This paper illustrates the application of a fuzzy rule-based system to enhance explainability, thereby reducing the time required to identify anomalous data behavior. The proposed algorithm, integrated with an autoencoder model, is employed in the field of anomaly detection, which consists of searching for small amounts of anomalies within extensive datasets. The data analyzed is sourced from the production of a MEMS (Micro-Electromechanical System) based inertial sensor employed in the automotive industry. Therefore, the paper addresses real-world challenges encountered in industry, with the suggested method’s results validated by domain experts. The article places a strong emphasis on explainability through the utilization of a fuzzy rule-based system, which greatly facilitates and shortens the time of decision and intervention in the industrial environment. The fuzzy system is trained by the bacterial memetic algorithm, which combines the bacterial evolutionary algorithm with the Levenberg-Marquardt local search technique, thus providing an efficient optimization for the model. Leveraging the white-box behavior of the fuzzy system, the trained model is then utilized to generate comprehensive linguistic interpretations, which can be readily understood by MEMS experts.