A Domain-Adversarial Wide-Kernel Convolutional Neural Network for Noisy Domain Adaptive Diesel Engine Misfire Diagnosis

In practical application scenarios, it is significant to take corresponding measures and diagnose diesel engine misfire faults. However, since the diesel engine noise is changeable and random, the existing methods obtain limited performance for diesel engine misfire fault diagnosis between different noise domains. In order to handle the above problems, this article proposes a domain-adversarial wide-kernel convolutional neural network (DAWDCNN) for noisy domain adaptive diesel engine misfire diagnosis. First, we utilize five 1-D convolution-pooling layers as feature extractor, of which the first layer contains a wide convolutional kernel. Then, extracted deep features from source domain data are fed into the fault classifier. Similarly, the extracted deep features from source and target domains are fed into the domain discriminator. Moreover, domain-adversarial training is employed as a training method of DAWDCNN. Also, we introduce a gradient reversal layer (GRL) to the backpropagation of the network in order to speed up the model training. Compared with the conventional phased domain-adversarial training method, GRL effectively improved the training speed by up to nearly 50%. It is also shown that DAWDCNN converges the fastest and has the most stable accuracy after convergence by comparing the accuracy-epoch curves of the test phases. Finally, for 11 noisy domain adaptability tasks from both one- and hybrid-cylinder diesel engine misfire datasets, DAWDCNN obtained better generalization performance of noisy domain adaptability than other existing methods. For example, on the four datasets, general average accuracies for other existing methods, such as random forest (RF), long short term memory (LSTM), deep neural networks (DNNs), convolutional neural network (CNN)_1d, CNN_1d-2d, DNN_maximum mean discrepancy (MMD), CNN_1d_MMD, and WDCNN_MMD, were 45.785%, 47.339%, 96.152%, 95.423%, 96.366%, 63.678%, 92.151%, and 97.519%, respectively, while DAWDCNN obtained 99.900% general average accuracy.