LaplaceCNN-based CAN bus fault diagnosis under class imbalance dataset

The controller-area network (CAN) bus has a wide range of applications. It plays a key role in information transfer. In fact, a failure of the CAN bus causes information transfer to fail. Therefore, fault diagnosis of the CAN bus is necessary. In addition, CAN bus failures generally result from sample imbalance problems. Therefore, this study proposes a CAN bus fault diagnosis (CANFD) method that integrates a Laplace network (LaplaceCNN) and uses a class-imbalance dataset. The method first utilises the adaptive synthetic (ADASYN) oversampling algorithm to process the class-imbalance CANFD dataset. LaplaceCNN is trained on the processed balanced dataset to diagnose CAN bus faults. To collect the frequency-domain characteristics and feed them into the deep network, LaplaceCNN uses Laplace wavelet convolution as the initial layer of the network. Multiscale wide convolutional and dense residual blocks are proposed to enhance the feature extraction capability of LaplaceCNN and stabilise the model performance. Introducing focal loss further improves the performance of LaplaceCNN on a class-imbalanced dataset. Experiments demonstrate that the proposed method performs better in terms of fault diagnosis.