Answers are grounded in this paper and cite their sources.
Accurately identifying early weak faults in rotating machinery is essential for safe operation and intelligent condition monitoring. However, practical industrial diagnosis is frequently constrained by limited fault samples, strong background noise, and variations in speed and load. Conventional deep-learning models commonly employ fixed convolution kernels or computationally intensive architectures, limiting their adaptability to time-varying vibration patterns and increasing the risk of overfitting under small-sample conditions. To address these challenges, this paper proposes a lightweight DDSC-T-HRC-LSTM-HAA diagnostic framework integrating dynamic depth-wise separable temporal convolution (DDSC-T), hierarchical residual-connected long short-term memory (HRC-LSTM), hierarchical adaptive attention (HAA), and an auxiliary convolutional neural network (CNN) branch. DDSC-T uses input-dependent weights to aggregate parallel depth-wise separable temporal-convolution branches with different receptive fields, thereby adaptively extracting local transient and resonance-related features. HRC-LSTM subsequently models hierarchical dependencies among repetitive fault responses through stacked recurrent layers and residual projections. HAA sequentially recalibrates the learned representations along the channel and temporal dimensions, while the auxiliary CNN branch preserves complementary shallow morphological information through lightweight learnable fusion. Comprehensive evaluations on the public Politecnico di Torino rolling-bearing dataset (T1) and the laboratory QPZZ-II gearbox dataset (T2) yield overall accuracies of 99.7 % and 98.3 %, respectively. For the challenging 150 μm bearing-defect categories, the recall remains above 99.2 %. In direct transfer from the 100 Hz-0 N source condition to unseen speed–load conditions, the proposed model achieves condition-averaged Accuracy and Macro-F1 values of 95.2 % and 94.7 %, respectively, and retains 91.5 % Accuracy under the most challenging combined speed–load shift. Under additive white Gaussian noise, the accuracy remains above 98% at signal-to-noise ratios no lower than 5 dB and reaches 61.08 % at 0 dB. With only 0.31 M trainable parameters and 0.48 GFLOPs, the model achieves a single-segment inference latency of 9.8 ms on the NVIDIA Jetson Xavier NX, demonstrating a favorable balance among diagnostic performance, cross-condition robustness, and embedded deployment efficiency.
No takes yet. Share an insight, caveat, or question.
Fan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: