Abstract: To solve the difficulties that faced by traditional signal processing methods in processing underwater acoustic signals containing complex noise components, which in turn affected the fault diagnosis accuracy of valve leakage in subsea Christmas trees, a non-contacting fault diagnosis method based on improved deep neural network (DNN) was proposed. First of all,based on the underwater valve leakage theories, and under the COMSOL simulation environment, combining with the axisymmetric free jet model and the quadrupole acoustic source far-field sound pressure solving algorithm, the valve leakage jet acoustic field model was established.Secondly, based on the basic theories of noise reduction, the basic model of DNN for noise reduction was established, and the time-frequency domain signal wasput into the constructed DNN model for noise reduction processing to minimize the noise components in the signals.Finally, the underwater acoustic signals after noise reduction processing were used and put into the convolutional neural network to achieve the purpose of fault diagnosis. In order to verify the effectiveness of the fault diagnosis method based on improved deep neural network (DNN), the underwater acoustic experiments were designed and performed. Experimental and research results show that the noise composition of the underwater acoustic signals after DNN denoising processing are significantly reduced, the normalized mean squared error value is reduced from 0.4992 to 0.0110, the decrease is 97.80%; and the fault diagnosis accuracy of the signals after DNN denoising processing reaches 98.89%, which proves that the method can effectively diagnose the leakage fault of the valve.
Key words: deep neural network(DNN); subsea tree valve; noise reduction model; valve leakage jet acoustic field; valve leakage simulation acoustic experiment; fault diagnosis accuracy