Deep Contrastive Transfer Learning for Rotating Machinery Fault Diagnosis

Abstract

Diagnostic models based on deep transfer learning hold the potential to apply diagnostic knowledge across relevant machinery. However, the existing methods suffer from several drawbacks: 1) most methods use wavelet convolutional kernels to extract features, but the extracted features are redundant and may diminish the diagnostic accuracy; 2) traditional domain alignment methods cannot effectively utilize both low- and high-order measure to match global or local feature information; and 3) contrastive learning in transfer learning requires the assignment of labels to the target domain, but target domain pseudo-labels may not be reliable. To address these issues, a novel domain adaptation method named time–frequency adaptive contrastive neural network (TFACNN) is proposed. This method uses wavelet convolutional kernels to extract features from the time–frequency domain. It then uses sparse constraint and kurtosis constraint to extract features corresponding to the fault frequencies. It utilizes a generalized distribution matching measure to align features, minimizing distribution differences between the two domains. Additionally, an adaptive threshold is designed to ensure the quality of target domain pseudo-labels, thereby improving the performance of contrastive learning. The proposed method is validated on two different datasets, which include rotating machinery fault diagnosis dataset and turning process monitoring dataset, demonstrating its superiority over other methods for rotating machinery fault diagnosis.