Research On Fault Diagnosis Of AUV’s Sensor And Thruster Based On Deep Learing

There are many types of AUV faults.The AUV fault diagnosis method widely used today is mainly based on the signal method,that is,the fault feature extraction is performed on the signal collected by the AUV sensor,and the fault is based on the difference between the sensor signal and the AUV theoretical state value.classification.This signal-based fault diagnosis method has high requirements for the design expert of the fault observer.It requires not only a solid signal processing technology but also a deep understanding of the working characteristics of the target carrier,but for the work.In the deep water AUV,the working environment is very complicated and varied.If only the expert’s offline design fault observer is used to detect the AUV running state,it is easy to ignore some new faults and minor faults.In this paper,the focus of research is on the fusion of deep learning technology and AUV sensor and thruster fault diagnosis.A new idea of data-driven AUV fault diagnosis is proposed to increase the autonomy and robustness of AUV.In this paper,the modeling method of the motion characteristics and motion model of general AUV is firstly described.At the same time,an AUV dynamics modeling method based on improved Elman neural network is proposed,and the advantages of this method in the field of AUV are analyzed.The complexity of AUV overall system fault diagnosis is analyzed and it is concluded that the AUV fault diagnosis idea should be considered from the perspective of decomposition.At the same time,the typical fault types of AUV thrusters and sensors are pointed out.The propeller type of propeller is selected as the propeller of this paper.Winding,wear and deviation and open circuit faults of the sensor TCM-5.A semi-physical simulation model is established based on the AUV of the tunnel,and the propeller and sensor fault samples of the AUV are obtained by semi-physical simulation.Secondly,in order to solve the coupling effect of the fault features to a certain extent,all the fault samples in this paper do not directly use the residual of the measured state value of the sensor and the theoretical state value of the AUV as the input signal,but the FFT transform of the residual signal.The first 2000 sets of Fourier coefficients are used as input signals.Considering the over-fitting and dimension explosion problems that shallow networks are prone to solve in the case of strong nonlinear problems,a fault diagnosis method for AUV thrusters based on D-SAE deep network is proposed in the fault diagnosis of AUV thrusters.The core idea of the D-SAE algorithm is the self-encoding idea.The artificial input noise sample is used to fit the original sample to increase the robustness of the network.It is very suitable for AUV systems with more external interference.During the training process.Add the dropout algorithm to improve the network structure’s ability to resist over-fitting problems.Through comparison of multiple sets of comparison experiments,it is verified that the D-SAE network-based AUV thruster fault diagnosis method has higher calculation rate and diagnostic accuracy.Finally,considering that the local features in the AUV sensor fault signal have important information characteristics,if the number of neurons in the hidden layer of the network is increased to improve the network’s ability to extract fault features,it is easy to cause the network dimension disaster.And over-fitting.In order to solve the above problems,this paper proposes a CDFL algorithm structure with local class ability and relatively short training time for fault diagnosis of AUV sensors.The basic structure of the CDFL network proposed in this paper is a CNN network with a convolutional layer and a pooled layer,which is characterized in that the filter weights in the convolutional layer are trained by an external BP network and are used throughout the CDFL network.Subsequent backpropagation does not change its value to ensure its local recognition ability.The external BP network must have the same input layer size as the filter in the convolutional layer,and the number of neurons in the hidden layer of the BP network must be the same as the number of filters.Through comparison of multiple sets of comparison experiments,it is verified that the AUV sensor fault diagnosis method based on CDFL network proposed in this paper has higher calculation rate and diagnostic accuracy.