An intelligent fault diagnosis method for a diesel engine valve based on improved wavelet packet-Mel frequency and convolutional neural network

Considering the diesel engine vibration signals have the characteristics of the non-stability and non-linearity due to its compact-complex structure, strong noise and especially unstable operating conditions, we proposes an novel method based on improved wavelet packet-Mel frequency and convolutional neural network (CNN) to extract features and diagnose faults of diesel engine valve. Firstly, the wavelet packet transform is applied with the purpose of decomposing vibration signal and reconstructing each wavelet packet coefficient. Secondly, an improved Mel frequency cepstrum method is used to extract features from the reconstructed vibration signals. MFC algorithm is a well-known feature extraction technique widely used for speech recognition. Then, feature matrixes are constituted to obtain more definite and comprehensive time-frequency distributed representation, of which the row represents the average Mel frequency cepstrum coefficients and the column represents the frequency bands of wavelet packet decomposition in ascending order. Finally, a deep hierarchical CNN structure constructed by convolution layers, max-pooling layers and fully-connected layers is trained using a standard backpropagation, of which the input of first layer with 256 neurons is the above 2D feature matrixes and the output of final layer with 3 neurons is the number of vibration signal states. The experimental results of the fault diagnosis for the diesel engine valves show that the proposed method has the good diagnosis performance for diesel engine valve clearance faults.