A Dual-Task Learning Approach for Bearing Anomaly Detection and State Evaluation of Safe Region

Abstract

Predictive maintenance has emerged as an effective tool for curbing maintenance costs, yet prevailing research predominantly concentrates on the abnormal phases. Within the ostensibly stable healthy phase, the reliance on anomaly detection to preempt equipment malfunctions faces the challenge of sudden anomaly discernment. To address this challenge, this paper proposes a dual-task learning approach for bearing anomaly detection and state evaluation of safe regions. The proposed method transforms the execution of the two tasks into an optimization issue of the hypersphere center. By leveraging the monotonicity and distinguishability pertinent to the tasks as the foundation for optimization, it reconstructs the SVDD model to ensure equilibrium in the model's performance across the two tasks. Subsequent experiments verify the proposed method's effectiveness, which is interpreted from the perspectives of parameter adjustment and enveloping trade-offs. In the meantime, experimental results also show two deficiencies in anomaly detection accuracy and state evaluation metrics. Their theoretical analysis inspires us to focus on feature extraction and data collection to achieve improvements. The proposed method lays the foundation for realizing predictive maintenance in a healthy stage by improving condition awareness in safe regions.