Valve fault diagnosis of internal combustion engine based on an improved stacked autoencoder

The valve train fault is a common mechanical fault of internal combustion engines (ICEs) due to the valve clearance usually oversized because of the wear of valve mechanism, material deformations, and long continuous running hours. Feature extraction dependent on the expertise and experience too much in traditional fault diagnosis. In this study, a stacked autoencoder (SAE) is proposed for adaptive and hierarchical feature extraction in cylinder vibration signals. The capability of feature mining in SAE is enhanced after unsupervised layer-by-layer pre-training and supervised fine-tuning. Further, the dropout trick and the batch normalization trick are introduced to prevent over-fitting and accelerate model convergence. The harmonic search (HS) algorithm is proposed to obtain the optimal hyper-parameter values in the SAE model, and achieve adaptive adjustment of the model structure. The diesel engine vibration data consisting of seven valve health states is employed to verify the effectiveness of the proposed method, the results demonstrate that the proposed method outperforms original SAE and many conventional fault diagnosis algorithms in terms of the classification accuracy.