Initial Position Estimation of Surface Permanent Magnet Synchronous Motor Base on Domain-Adversarial Neural Networks
The low-speed sensorless control of the permanent magnet synchronous motor(PMSM)primarily depends on the motor's saliency characteristics.However,the surface-mounted permanent magnet synchronous motor(SPMSM)exhibits a low inductive saliency ratio,leading to a poor signal-to-noise ratio(SNR)in position signals.It challenges traditional high-frequency injection position estimation models in achieving high-precision estimation.Deep machine learning can extract features from subtle signals and accurately determine position even in low SNR environments.However,traditional supervised machine learning methods require abundant labeled training data.In practical applications,instantaneous load variations make it challenging to collect comprehensive samples and accurately label position tags,thus impacting the generalization ability of position estimation models trained on steady-state datasets.This paper proposes a data domain-adversarial neural network(DANN).This method enhances the traditional convolutional neural network(CNN)by introducing a domain classifier and a gradient reversal layer(GRL),transitioning from a single-stream to a dual-stream structure.The two networks learn invariant domain features,adapt data distribution,and mitigate the degraded position estimation by anti-transfer learning.Balanced three-phase high-frequency excitation voltages are applied to the three-phase windings of the motor in the stationary reference frame.The negative-sequence first-order component and the positive-sequence second-harmonic component of the high-frequency response current are combined to form a new current vector,facilitating the extraction of position and polarity information.These reconstructed current vectors are graphically depicted and employed as inputs for the CNN.Implicit position feature information is extracted from two-dimensional images,and a correspondence between the images and the rotor positions is established through training.Consequently,the CNN model accurately recognizes two-dimensional images for position estimation.This CNN network is used as the feature extractor in the DANN network,incorporating a domain discriminator to distinguish input features from the source and target domains.Furthermore,a gradient reversal layer(GRL)is introduced between the feature extractor and the domain classifier.In backpropagation,the gradient parameters of the loss function in the domain classifier are reversed,which facilitates adversarial learning and aligns the probability distribution of new test data with the training set.Experimental results indicate that traditional observers have significant errors in position estimation,with numerous deviations exceeding 90°.The CNN-based position estimation model maintains an average positional deviation of 13.2°,whereas the DANN-based model further decreases this deviation to 8.6°.When predicting new,unfamiliar data,the DANN model exhibits notably superior position estimation performance compared to the CNN model.The dual-stream DANN model can precisely estimate the initial position of motors with low saliency ratios,and position estimation results on new data are improved.Consequently,the poor data generalization ability inherent in the single-stream CNN position estimation model is mitigated.The following conclusions are drawn.By leveraging the correlation between images and positions,image recognition-based position estimation methods enable accurate position determination through high-frequency current image recognition,solving the issue of the initial position estimation with low saliency ratios.Introducing the gradient reversal layer for domain adversarial learning facilitates data distribution-adaptive transfer learning.The position can be accurately estimated even with notable disparities in data probability distributions between test data and the training set.The dual-stream DANN model demonstrates superior domain generalization capability and enhanced robustness compared to the single-stream CNN network model.This approach improves the predictive performance of deep learning in scenarios of limited training samples and high labeling costs.A robust theoretical and practical framework for industrial machine-learning applications has been established.
国家科技期刊平台
NSTL
万方数据
