Abstract: A stack autoencoder (SAE) is proposed to extract deep features hierarchically from complex raw signals, in view of the lack of professional knowledge will weaken the efficiency of handcrafted feature extraction in mechanical fault diagnosis. The SAE can mine deep features via layer-by-layer pre-training, fine-tuning, etc., moreover, the dropout regularization layer and the batch normalization layer are introduced before each hidden layer in the network to prevent over-fitting and accelerate convergence. Aiming at the value of hyperparameters in SAE network, firstly, the appropriate range of values for each hyperparameter is obtained via a series of experiments, then, the harmony search (HS) algorithm is proposed within the range to optimize the hyperparameters to achieve adaptive adjustment of the network structure and improve feature extraction. The experimental results show that the proposed HS-SAE scheme outperforms original SAE and many traditional fault diagnosis algorithms in terms of the fault classification accuracy when testing with the diesel engine vibration data consisting of seven valve health states.

Key words: autoencoder, feature extraction, parameter optimization, fault diagnosis