Cross-load robustness evaluation of deep and handcrafted learning frameworks for bearing fault diagnosis

Abstract

This paper introduces a comprehensive leakage-free methodology for bearing fault diagnosis by three of the most representative feature learning approaches: Raw Signal Convolutional Neural Network (Raw CNN), Fast Fourier Transform (FFT) based hybrid MLP-CNN, and handcrafted statistical descriptor based Multilayer Perceptron (MLP). The vibration data collected from the Case Western Reserve University (CWRU) bearing set for four operating conditions (0, 1, 2, and 3 HP) was used. To eliminate information leakage, all models were trained under the base-line condition 0 and tested under the unseen conditions 1, 2, and 3 HP, which were cross loads to assess the robustness of the models objectively. The Raw CNN achieved accuracies of 59.18%, 59.84%, and 61.92%, while the FFT-based MLP-CNN improved performance to 79.92%, 77.01%, and 65.22%. The handcrafted feature based MLP gave the best accuracy of 99.47%, 89.43%, and 78.92% for each of the three datasets. In order to support the evaluation, computational complexity, external validation with the Paderborn University bearing dataset, comparison with alternative machine learning classifiers, SHAP, and Permutation Feature Importance (PFI) analyses were performed. The external validation yielded 98.76% accuracy for classification, further confirming the generalization capability of the proposed framework. The results show that the handcrafted statistical descriptors, along with a simple MLP, is an interpretable, computationally efficient, and robust method for cross-load bearing fault diagnosis.