MITDCNN: A multi-modal input Transformer-based deep convolutional neural network for misfire signal detection in high-noise diesel engines

Abstract

The accurate detection of faults in diesel engines is crucial for extending their operational lifespan, ensuring safety, and yielding significant economic and societal benefits. However, existing detection methodologies encounter difficulties in the presence of substantial environmental noise and lack generalizability across various operating conditions. Traditional fault diagnosis models also predominantly rely on single data detection schemes, such as one-dimensional data input for anomaly analysis, which may not account for the inherent correlations and distribution discrepancies within the same data source presented in different formats. To tackle these challenges, we proposed a novel multi-modal input Transformer-based convolutional neural network (MITDCNN) for misfire detection in diesel engines under significant environmental noise and diverse working conditions. We collected vibration signals from engine cylinder heads at different speeds and extracted both one-dimensional amplitude vector features and two-dimensional image properties. These features were input into the multi-modal feature extraction network and subsequently processed and identified by the cross-channel feature fusion detection network. Our fusion module incorporated a dual-dimension spatial and channel attention mechanism to suppress vibration noise interference effectively, so that the network is able to counteract the impact of environmental noise and varying operating conditions on the final detection outcomes. We validated the efficacy of the proposed approach through a curated dataset and benchmark comparisons with the established algorithms. Our results demonstrate that MITDCNN's accuracy and noise robustness are notably superior to other available methods. Across four datasets generated under distinct working conditions, our approach achieves a minimum accuracy of 99.008%, even with a signal-to-noise ratio of −4dB, greatly surpassing other methods.