Joint Threshold Learning Convolutional Networks for Intelligent Fault Diagnosis Under Nonstationary Conditions

Rotating machines are essential components in manufacturing, power generation, transportation, and aerospace industries. Nevertheless, most existing diagnosis methodologies are developed based on the assumption that machinery operates under stable conditions, which may limit their ability to uncover sufficient discriminative features when confronted with high noise levels. To tackle this issue, this article proposes a joint threshold learning convolutional network (JT-LCN) for rotating machinery diagnosis under nonstationary conditions. The major contributions of this research work can be summarized and highlighted as follows: 1) proposing a novel plug-and-play joint-thresholding module (JTM) that uses both soft and hard threshold mechanisms for intelligent signal denoising; 2) constructing an end-to-end network architecture, termed JT-LCN, which achieves a lightweight design through the use of depthwise separable convolution (DSC) and effectively purifies and identifies discriminative fault-related information by progressively using the JTM; and 3) introducing a dynamic self-knowledge distillation approach to enhance the generalization ability of the proposed network while minimizing computational costs and run-time memory requirements. Comprehensive experimental results conclusively demonstrate that the developed JT-LCN outperforms competition state-of-the-art approaches.