【目的】针对真空泵用驱动电机在实际工作过程中散热困难，且在出现冲击载荷情况下电机温度容易突破绝缘限制的问题，本文采用神经网络法对真空泵用驱动电机进行发热预测。【方法】以一台4.5 kW的真空泵用驱动电机为例，首先，对冲击载荷下的电机进行温度仿真分析，对瞬时数据进行分类，利用历史数据对神经网络进行训练，建立真空泵用驱动电机运行数据与温度的映射关系。其次，搭建电机试验平台，对电机进行冲击试验，用试验数据修正神经网络训练模型。通过试验比较预测温度值与实际测量温度，验证修正后模型的准确性。【结果】试验结果表明，经过误差修正后，神经网络模型的温度预测精度得到了提高。误差修正后的双向长短期记忆网络在预测定子绕组端部温度方面的表现优于卷积神经网络和长短期记忆网络，决定系数达到0.979。【结论】本研究为不依赖温度传感器实现对真空泵用驱动电机受冲击载荷时的发热预测提供了一种方法和技术路线，为驱动电机的有效控制提供算法判断依据。

[Objective] To address the difficulty in heat dissipation of drive motors used in vacuum pumps during actual operation and the tendency of its temperature to exceed insulation limits under impact loads, a neural network method is used to predict the heating of vacuum pump drive motors. [Methods] Taking a 4.5 kW vacuum pump drive motor as an example, the temperature simulation analysis of the motor under impact loads was first conducted. The instantaneous data were classified, and a neural network was trained using historical data to establish the mapping relationship between the operating data and temperature of the vacuum pump drive motor. Subsequently, a motor test platform was constructed to conduct impact tests on the motor. The test data were used to correct the neural network training model. By comparing predicted temperatures with actual measured temperatures through experiments, the accuracy of the corrected model was validated. [Results] The experimental results showed that the accuracy of the neural network model for temperature prediction was improved after error correction. The bidirectional long short-term memory network with error correction outperformed the convolutional neural network and the long short-term memory network in predicting the temperature at the stator winding end, with a coefficient of determination reaching 0.979. [Conclusion] This study provides a method and a technical approach for heating prediction of vacuum pump drive motors under impact loads without relying on temperature sensors, offering an algorithmic basis for the effective control of drive motors.

辽宁省“揭榜挂帅”科技计划项目(2023JH1/11100010)