Segmentation, correction, and classification of abnormal sensor data in mechanical engineering based on multi-task learning

Abstract

Rolling bearings and hydraulic internal pumps are the two most commonly used fault diagnosis devices in mechanical engineering. However, harsh industrial environments not only harm their health but also the sensors used for monitoring. Abnormal sensor data problems are common in practice and significantly affect data-based fault detection methods. Therefore, this study jointly investigates the anomaly detection of sensor data and the fault detection of engineering components. The related issues are divided into three tasks: classification, correction, and segmentation of abnormal sensor data. A multi-task learning framework based on the teacher-student structure is then proposed to fulfill these tasks in one shot. The designed feature corrector corrects abnormal representations, while the correction attention guides the classifier to focus on the normal parts. A semantic segmentation model is integrated to achieve novel and comprehensive anomaly detection. The proposed multi-task framework is validated using rolling bearing and hydraulic pump datasets. The experimental results show that the jointly trained models outperform those that are trained independently.