Resource-Efficient Network for Intelligent Fault Diagnosis Using Tree-Structured Parzen Estimator

The fault diagnosis algorithm can timely detect the faults of devices by monitoring their signals, avoiding greater losses. Deep learning has been widely applied in the field of fault diagnosis. However, manually adjusting the network structure and hyperparameters requires expert knowledge and is laborious and time-consuming. In addition, the model structures of most existing deep learning methods are complex and redundant, which is not conducive to operation on hardware devices with limited computational resources. This paper proposes a resource-efficient fault diagnosis method based on Tree-structured Parzen Estimator (TPE). Additionally, a backbone network based on global average pooling (GAP) is proposed. For specific fault diagnosis tasks, the TPE automatically search for hyperparameters, resulting in an optimal model that not only requires low computational resources but also maintains high accuracy levels. Experimental results show that the proposed method can maintain high diagnostic accuracy with lower multiply-accumulate operations (MACs).