Research On Fault Diagnosis Method Of Brushless DC Motor

Brushless DC Motor(BLDCM)is an important power device,which is with excellent characteristics and widely used in various fields.Turn-to-turn short circuit and rotor demagnetization are common faults of motors.The occurrence of faults influences the operating efficiency of the motor,and in severe cases damages the motor and causes accidents.Therefore,in order to reduce losses and prevent accidents,in this thesis,fault diagnosis research on the short-circuit faults of the windings of the BLDCM and the partial rotor demagnetization faults is conducted.Firstly,starting from the structural principle of the BLDCM,the relationship between the electromagnetic torque of the motor and the current is analyzed.Via the mathematical analysis of the inter-turn short circuit and the rotor demagnetization,the correlation between the DC bus current and the phase current and the fault is explored.The experimental angle is verified.Secondly,from the forward propagation process and back propagation process of the convolutional neural network,the mathematical principle of the convolutional neural network propagation process is explained,the evaluation index of the diagnosis model is clarified,and the mathematics of the forward and backward propagation of the residual structure network model is expounded.Then,model training and fault diagnosis strategies are built based on residual network.Thirdly,a signal acquisition platform is constructed for BLDCM.Data under different load conditions of the motor is obtained by adjusting the load torque.A BLDCM fault diagnosis data set is established based on the data of different load conditions under various conditions.Finally,in order to expand the feature learning ability of the network at different time sizes,a multi-scale fusion convolution is constructed.The network applies a multi-scale learning strategy to automatically extract features at different time scales,and introduces an inter-layer residual jump connection mechanism in each parallel convolution path.The residual network introduces forward information through the identity mapping between layers,reduces the disappearance of the gradient,and increaes the representational power of the model.Furthermore,in order to improve the training process of the model and prevent the model from overfitting,Batch Nomalization,Adam optimization algorithm and Dropout regularization strategy are used to optimize the model through these methods.Comparative experiments on the model are performed using accuracy and other indicators to verify its diagnostic performance.