With the assumption of sufficient labeled data, deep learning based machinery fault diagnosis methods show effectiveness. However, in real-industrial scenarios, it is costly to label the data, and unlabeled data is underutilized. Therefore, this paper proposes a semi-supervised fault diagnosis method called Bidirectional Wasserstein Generative Adversarial Network with Gradient Penalty (BiWGAN-GP). First, by unsupervised pre-training, the proposed method takes full advantage of a large amount of unlabeled data and can extract features from vibration signals effectively. Then, using only a few labeled data to conduct supervised fine-tuning, the model can perform an accurate fault diagnosis. Additionally, Wasserstein distance is used to improve the stability of the model’s training procedure. Validation is performed on the bearing and gearbox fault datasets with limited labeled data. The results show that the proposed method can achieve 99.42% and 91.97% of diagnosis accuracy on the bearing and gear dataset, respectively, when the size of the training set is only 10% of the testing set.

The compressed sensing (CS) methods have two issues when processing vibration signals of rotating machinery: (1) the data-driven method and signal priors (e.g., global/local and structure priors) are not combined well, resulting in the reconstruction not converging to high-precision results; (2) the current metrics only focus on numerical accuracy and do not consider PHM scene fitness. We propose a deep compressed sensing framework (RMVS-DCS): (1) Global-Local Feature Reconstruction (GLFR) blocks alternately reconstruct features at global and local scales. Pre-train the output layer by dictionary learning, introducing the signal structure prior to guiding the model to converge to a better one. (2) A dual evaluation system, combining the numerical indicators and prognostics and health management (PHM) task accuracy, helps the model selection in PHM scenarios. On the rotor unbalance data set, its reconstruction and prediction error are reduced by 82.60% and 89.90% (compression ratio is 0.5), demonstrating the advanced reconstruction precision and PHM compatibility of the framework.

OOD detection mechanism is important when en-countering new (out-of-distribution, OOD) faults. Current OOD detection methods are often coupled with diagnostic models, lim-iting their actual deployment. Moreover, they usually assume that OOD faults exhibit significant distributional differences from the in-distribution (ID) data, and the ability to magnify differences is insufficient, affecting the detection ability. There-fore, we propose a reconstruction-based method using deep com-pressed sensing (DCS) and contrastive learning (CL), which is plug-and-play and can magnify the differences without OOD fault samples during training. The frequency-domain signal first (1) passes through the measurement module initialized by CL to align ID semantics, then (2) is reconstructed differently by the reconstruction module centered on global-local feature recon-struction blocks with local feature frequency-domain attention (LFA-GLFR blocks). (3) With the optimization objective of joint reconstruction accuracy, sparsity, and ID error consistency, the parameter training is guided to move in the direction of increas-ing differences to achieve better discrimination. We conduct ex-periments on the Shandong University (SDU) bearing dataset and the University of Connecticut (UoC) gearbox dataset. On the former, its average AUROC, AUPRC, F1-score, and F0.5-score are 0.1303, 0.1645, 0.0850, and 0.1156 higher than those of the compared state-of-the-art (SOTA) methods, demonstrating its advancement.