Condition monitoring and fault diagnosis of diesel engines are of great significance for safety production and maintenance cost control. The digital twin method based on data-driven and physical model fusion has attracted more and more attention. However, the existing methods lack deeper integration and optimization facing complex physical systems. Most of the algorithms based on deep learning transform the data into the substitution of the physical model. The lack of interpretability of the deep learning diagnosis model limits its practical application. The attention mechanism is gradually developed to access interpretability. In this study, a digital twin auxiliary approach based on adaptive sparse attention network for diesel engine fault diagnosis is proposed with considering its signal characteristics of strong angle domain correlation and transient non-stationary, in which a new soft threshold filter is designed to draw more attention to multi decentralized local fault information dynamically in real time. Based on this attention mechanism, the distribution of fault information in the original signal can be better visualized to help explain the fault mechanism. The valve failure experiment on a diesel engine test rig is conducted, of which the results show that the proposed adaptive sparse attention mechanism model has better training efficiency and clearer interpretability on the premise of maintaining performance.

The diesel engine has been a significant component of large-scale mechanical systems for the intelligent manufacturing industry. Because of its complex structure and poor working environment, it has trouble effectively acquiring the representative fault features. Further, fault diagnosis of the diesel engine faces great challenges. This paper presents a new fault diagnosis method for the detection of diesel engine faults under multiple operation conditions instead of conventional methods confined to a single condition. First, an adaptive correlation threshold process is designed as a preprocessing unit to enhance data quality by weakening non-impact region characteristics. Next, a feature extraction method for sound signals based on the Mel frequency cepstrum (MFC) is improved and introduced into the machinery fault diagnosis. Then, the combination of the improved feature and vibrational mode decomposition (VMD) is proposed to incorporate VMD into an effective adaptive decomposition of non-stationary signals to combine it with an excellent feature representation of the vibration signal. Finally, the vector quantization algorithm is adopted to reduce the feature dimensions and generate codebook model bases, which trains the K-Nearest Neighbor classifiers. Five comparative methods were carried out, and the experimental results show that the proposed method offers a good effect of the common valve clearance fault of diesel engines under different conditions.