A Parallel Dual-Branch Deep Learning Framework Integrating VMD Feature Engineering and Multiscale Temporal Feature Fusion for Small-Sample Fault Diagnosis

Rolling bearings are crucial in mechanical systems, where failures pose safety risks and economic losses. However, their rarity makes timely fault detection and data collection challenging. To address this, a parallel dual-branch fault diagnosis model was proposed. One branch applied variational mode decomposition (VMD) with the hybrid particle swarm optimization (HPSO) algorithm to automatically and optimally decompose the raw signals. Then, continuous wavelet transform (CWT) was used on the decomposed modes to create stacked feature maps, which was followed by the normalization with a ${1} \times {1}$ convolution to accomplish the feature engineering. Next, the improved convolutional neural networks (CNNs) and the bidirectional long short-term memory (BiLSTM) models were used in sequence to extract features from both the frequency and temporal domains. In the other branch, the original signal was downsampled at three different step sizes to capture multiscale temporal representations, and the multilayer perceptrons (MLPs) were employed to learn temporal features from different time scales, which were fused into one feature output thereafter. Eventually, the time-frequency and the temporal features from both branches were concatenated and passed to the output for fault diagnosis. The proposed feature engineering method and the parallel dual-branch fault diagnosis model were evaluated on small datasets containing 20 samples from Case Western Reserve University (CWRU), achieving an impressive 100% accuracy. Moreover, the model was compared with recent mainstream models on the Paderborn (PU) dataset. The results demonstrated that it excelled at learning key features from small samples, highlighting its potential for small-sample rolling bearing fault diagnosis.