Research On Permanent Magnet Synchronous Traction Machine Demagnetization Diagnosis Based On Convolutional Neural Network

With the accelerated pace of the construction of China’s urbanization and the further promotion of civil engineering for the installation of additional auxiliary facilities in old urban residential communities,the number and density of elevators as special equipment are constantly increasing.With the advantages of high structural integration,good speed regulation and smooth operation,permanent magnet synchronous traction machine(PMSTM)is gradually becoming the mainstream application of elevator power source.The permanent magnet in the PMSTM will inevitably demagnetize under the influence of multiple factors coupling such as alternating demagnetizing magnetic field,high temperature and vibration.In view of the complicated inspection process of PMSTM and the limitation of expert experience in demagnetization data processing,this paper applies convolutional neural network to the demagnetization diagnosis of PMSTM based on end-to-end demagnetization diagnosis,and the work done is as follows:In view of the complicated magnetic performance inspection procedure of permanent magnets in PMSTM and the hidden hazards of disturbing the original signal timing performance in the two-dimensional convolution scheme,an end-to-end one-dimensional convolution scheme is adopted in this paper to process the demagnetization feature signals.A representative demagnetization feature database is constructed by collecting multiple demagnetization sensitive feature data of PMSTM under typical operating conditions,a one-dimensional convolutional neural network model is built,the demagnetization degree intervals are divided to determine the types of network outputs,and the processed demagnetization feature data are input into the network for training to realize the discrimination of demagnetization degree of PMSTM.In view of the problem that a single sensor cannot reliably respond to the status of permanent magnets leading to poor diagnosis,this paper adopted a one-dimensional multiscale separable convolutional network model,which considered the fusion of multi-sensor information to process the demagnetization signals from multi-channel inputs.In view of the high complexity of the convolutional network caused by multi-channel fault signal input,a lightweight diagnostic network model was built using a depthwise separable convolution network framework to reduce the parameters.In addition,this paper used a multi-scale feature extraction module to increase the richness of the extracted features and introduced a channel attention mechanism to consider the contribution weight of different demagnetization-sensitive features to the demagnetization of the PMSTM.The model performance was verified with the help of visualization to improve the demagnetization diagnosis rate of PMSTM.In view of the scarcity of reference data for permanent magnet synchronous traction machine operation in the complex operating environment of elevators,which cannot adequately guarantee the generalization capability of demagnetization diagnosis network,an improved generative adversarial network was used for sample data set expansion to compensate for the missing demagnetization data that cannot be measured due to practical constraints.In this paper,an adversarial training method was used to make the generator model learn the demagnetization feature data distribution,the Wasserstein distance was used to improve the network performance,and conditional information was introduced to expand the demagnetization feature data in bulk on demand.In addition,this paper evaluated the quality of the generated demagnetization samples with the help of visualization and an independent classifier to verify the feasibility of the data,and the expanded data was used for diagnostic model training to improve the demagnetization diagnosis of PMSTM.