Accurate diagnosis of rolling bearings can prevent unexpected failures in rotating machinery. Most of the previous studies on bearing fault diagnosis have primarily emphasised achieving high classification accuracy under constant operating conditions, particularly when using sensors near the fault source. By contrast, industrial realities involve varying operating loads and constrained sensor placement, resulting in discrepancies between actual faulty signals and those captured by sensors. This study addresses these challenges through a novel three-part framework: (i) a lightweight hybrid convolutional neural network with gated recurrent unit (HCNN-GRU) architecture to assess the impact of varying sensor placements and loads, (ii) a novel hybrid exponential linear unit-rectified linear unit (ELU-ReLU) activation that resolves dying-ReLU effects and stabilises features and (iii) a novel progressive five-step transfer-learning strategy enabling robust adaptation across load domains. An experimental setup has also been developed to acquire vibration data under various sensor placements and load variations, thereby providing a realistic test environment for evaluating the proposed model. In the proposed network, smaller filters, reduced depth, hybrid activation and a GRU layer collectively enhance efficiency, reducing the number of parameters to 740.4K compared to 25.6M in ResNet-50. The proposed model achieved nearly 99% accuracy on hanoi university of science and technology (HUST) dataset and 100% accuracy on the case western reserve university (CWRU) dataset. This study presents a comprehensive comparative analysis of five deep learning architectures: the Proposed HCNN-GRU model, ResNet-50, signal enhancement in frequency domain and time domain domain-two-dimensional convolutional neural network (SEFT-2DCNN), MobileNet and convolutional neural network-long short-term memory (CNN-LSTM), for bearing fault diagnosis under varying load conditions. On the test-rig dataset, the proposed model achieved 99.5%–100% accuracy across all load conditions, while reducing training time by approximately 64%, 33% and 5% compared to the CNN-LSTM, ResNet-50 and MobileNet models, respectively. Although the training time of the SEFT-2DCNN model was 51% less, its accuracy was lower than that of the proposed HCNN-GRU model.
Srivastava et al. (Thu,) studied this question.