An Efficient Lightweight Network With Improved Heterogeneous Convolution for Bearing Fault Diagnosis

Deep convolutional neural networks (DCNNs) are prevalent in intelligent bearing fault diagnosis, yet their computational complexity hinders industrial deployment. This paper presents a novel lightweight deep convolutional neural network (DCNN) designed for computationally efficient bearing fault diagnosis. The proposed model innovates by integrating channel-pruned heterogeneous convolutions for compact feature extraction, an attention mechanism for adaptive feature weighting, and dropout regularization to enhance generalization. This combination achieves an exceptional balance between performance and efficiency. On two public bearing datasets, our method achieves high diagnostic accuracies (97.79%, 96.78%) with an extremely small model size of only 0.02M parameters and 1.64M FLOPs. Comparative analyses demonstrate that it outperforms existing lightweight and classical models by achieving the highest accuracy, precision, and recall with the minimal computational footprint, making it exceptionally suitable for real-world industrial implementations.