An Intelligent Multi-Sensor Information Fusion Fault Diagnosis Method of Three-Phase Motors Based on Game Mapping Learning

Mechanical and electrical faults of three-phase motors can be comprehensively recognized by using multi-sensor data. Existing intelligent fault diagnosis methods, however, are short of an explicit learning mechanism to effectively mine key fault information and fuse multi-sensor features, thereby limiting their diagnosis performance. To overcome these problems, this paper proposes an intelligent multi-sensor information fusion fault diagnosis method based on game mapping learning for three-phase motors. By automatically extracting fault features from different sensor data and adaptively fusing them, the proposed method can accurately recognize various faults of three-phase motors. First, multiple parallel self-learning feature mapping networks are used to automatically extract fault features from different input data from multi-sensor sources. Then, a sensor source discriminator is constructed to form a game-playing relationship between it and self-learning feature mapping networks, aiming to refine fault features and make them aggregate by fault categories. Meanwhile, for ensuring the spatial separability of different types of fault features, a sample difference metric loss function is introduced to the optimization objective. Finally, a fault pattern recognizer is employed to fuse multi-sensor features and classify motor faults. Three-phase motor fault simulation experiments are designed and carried out in this paper. The multi-sensor data, including vibration, current and sound signals, are obtained to verify the proposed method. First, the selection of some hyper-parameters is discussed, and some implementation details of the network are determined. Then, the ablation experiments are performed and the experimental results are as follows. (1) The additions of game-playing learning strategy and sample difference metric loss function improve the diagnostic accuracy of the network. (2) The combined effect of game-playing learning strategy and sample difference metric loss function makes the average accuracy of the proposed method over 99%. Next, the comparison between the proposed method and the state-of-the-art methods shows that, the proposed method has the highest accuracy and the best stability in mechanical and electrical fault diagnosis of three-phase motors. Meanwhile, the above results also explain that the lack of an explicit learning mechanism in fusing multi-sensor features will lead to poor diagnosis accuracy. Finally, the aggregation process of the fault features from different sensor data is visualized. The results show that, the game-playing learning strategy and the sample difference metric loss function make the fault features aggregate by fault categories, thus realizing the effective fusion of multi-sensor source information. The following conclusions can be drawn from the experimental analyses: (1) Game-playing learning can guide the network to automatically extract fault features from multi-sensor data and make them aggregate within the class, thus avoiding the influence of irrelevant measurement noise or redundant sensor information, and improving the accuracy and robustness of three-phase motor fault diagnosis results. (2) The sample difference metric loss function makes different types of multi-sensor fault features distributed in clusters and clear boundaries, which is conducive to accurately distinguish various mechanical or electrical faults of three-phase motors. (3) Compared with the existing feature-level fusion diagnosis methods, the proposed method has an explicit learning mechanism in the process of feature extraction and fusion, and accordingly its diagnosis performance is better. © 2023 Chinese Machine Press. All rights reserved.