ConvECA-Net: A lightweight convolutional neural network for fault diagnosis of tapered roller bearings

Abstract

The paper focuses on detecting faults in tapered roller bearings to prevent unplanned shutdowns, accidents, and financial losses. Tapered roller bearings are an indispensable component used in mechanical applications, where handling of combined loads is required in rotating machinery. This study proposes a novel framework ConvECA-Net for advanced fault diagnosis of tapered roller bearings. This innovative architecture combines Efficient Channel Attention (ECA), adaptive kernel size strategy, and Leaky ReLU activation. This fault classification network is designed to classify five fault conditions in the tapered roller bearings. This proposed architecture is compared with deep learning algorithms such as ANN and ResNet50 and traditional machine learning algorithms such as SVM and RF. The proposed ConvECA-Net achieves a better classification accuracy of 95.07% and requires only 563 K trainable parameters and a model size of 2.2 MB. A systematic component-wise ablation study is conducted to validate that each component of this architectural design, namely ECA attention, adaptive kernel sizing, and Leaky-ReLU activation, individually and collectively contribute to the overall performance of this diagnostic model, as it achieves superior performance compared to its baseline variant by 5.45%. Cross-condition robustness is further established by evaluating this diagnostic model on three different rotational speeds and various load levels. Evaluation of noise robustness on this diagnostic model for various levels of Gaussian white noise and pink noise, further establishes its robustness of the proposed model. Further, Statistical validation of this diagnostic model is conducted by running this experiment ten times, using paired t-tests, Shapiro-Wilk normality tests, and stratified 5-fold cross-validation (94.82% ± 0.38%). Finally, analysis of computational efficiency of this diagnostic model reveals that it achieves 192 MFLOPs and an inference latency of 0.82 ms/sample, making it suitable for industrial condition monitoring systems of rotating machinery to prevent the plant shutdown.