XES3MaP: Explainable Risks Identified from Ensembled Stacked Self-Supervised Models to Augment Predictive Maintenance

Understanding the reasons behind a model's predictions is as important as achieving accurate results. The limited adoption of AI methods in fields like Energy and Industry is mainly attributed to a lack of trust, which is a crucial factor in user acceptance. Explainable AI is a recent approach to address this issue and enable the rapid deployment of AI in complex domains. This paper presents a framework for explainable anomaly detection and risk prognostics that utilizes an Ensembled Stacked Self-Supervised Model and Shapley additive approach to generate local and global explanations. The quality of the explanations is evaluated with the assistance of human experts to enhance model performance and streamline an automated online approach. The explanation produced is evaluated utilizing local accuracy metric. The proposed framework is tested on real-world industrial cooling system data, on which out of the ten anomalies identified, eight were successfully linked to maintenance actions, while two were attributed to random sensor measurement disturbances. This prognostic outcome establishes a new benchmark in the field.

I. Introduction

Predictive maintenance (PdM) serves as the backbone of industrial operations. It utilizes sensor data to safeguard asset health by identifying, analyzing, and tracking degradation, ultimately predicting failure progression. The three key activities leveraged in PdM are: i.

Anomaly Detection: Identifying deviations from normal system behavior through data analysis.

ii.

Failure Prognosis: Estimating the remaining useful life (RUL) of an asset before failure occurs.

iii.

Diagnostics: Classifying and pinpointing the root cause of equipment failure.