A Multidepth Step-Training Convolutional Neural Network for Power Machinery Fault Diagnosis Under Variable Loads

Abstract

Due to the operation conditions of variable loads, it is challenging to achieve high-accuracy fault diagnosis of power machinery. The attention mechanism is widely used in this issue because of its ability to capture domain-invariant features of vibration signals. However, when the problem is specific to thermal engine diagnosis, the attention collapse can be caused by the interaction between load patterns and fault patterns. Consequently, the deep features converge to decrease the network generalization. To address this issue, this research employs the ensemble learning of crowd intelligence strategy, which is opposite to the attention mechanism of elite strategy. A multidepth step-training convolutional neural network (MDNN) is proposed. The multidepth architecture enhances feature diversity, and the step-training feature ensemble incorporates features into decision-making, thus overcoming feature convergence. The MDNN is tested using two datasets: a light-duty rotor-bearing test rig (electromechanical system) and a heavy-duty diesel engine test rig (thermodynamic machinery). According to the results, for the load-varying diesel engine, the attention mechanism exacerbates feature convergence, whereas MDNN effectively mitigates it. Meanwhile, with the mixture of four engine loads, the diagnosis accuracy of the attention mechanism-based network falls sharply to 54.27% from 59.20%, while the MDNN rises to 95.46%. The results offer a promising method for load-varying fault diagnosis of thermodynamic machinery and give a comprehensive understanding of the importance of avoiding feature convergence in the prognostic diagnosis of diesel engines.