Intelligent fault diagnosis of roller bearings using a transfer learning-based dual-input spectrogram–scalogram convolutional neural network model

Abstract

This research proposes DiSCNet (Dual-input Spectrogram–Scalogram Convolutional Network), a dual-input convolutional neural network (CNN) designed for robust and highly accurate bearing fault diagnosis using time–frequency representations of vibration signals. Unlike traditional single-input models, DiSCNet employs a dual-branch architecture that processes both spectrograms, derived via the short-time Fourier transform, and scalograms, computed using the continuous wavelet transform, in parallel. Each modality is passed through a pre-trained ResNet-50 backbone to extract rich 2048-dimensional deep features, which are concatenated and fed into a custom classification head consisting of fully connected layers with dropout regularization to enhance generalization. The model is trained end-to-end on matched spectrogram–scalogram image pairs generated from vibration signals acquired experimentally from a bearing test rig. Training utilizes the Adam optimizer and cross-entropy loss, resulting in a low validation loss of 0.0183. Extensive evaluation against seven state-of-the-art CNN models – AlexNet, GoogLeNet, MobileNetV3, EfficientNet-B0, DenseNet-121, ResNet-50, and ConvNeXt-Tiny – demonstrates DiSCNet's superior performance. It achieves a validation accuracy of 99.38%, with outstanding results across additional metrics: F1-score (0.9932), precision (0.9946), recall (0.9938), Matthews correlation coefficient (0.9937), and Cohen's kappa (0.9936). The results highlight the advantages of integrating spectral and wavelet-domain features into a unified deep learning framework. By effectively leveraging complementary information from both domains, DiSCNet enhances feature discriminability and class separability, making it a highly promising approach for practical industrial condition-monitoring and fault-diagnosis applications.