改进LSTM神经网络在电机故障诊断中的应用

Application of improved LSTM neural network in motor fault diagnosis

摘要:

三相异步电机因其结构简单、维护方便、可靠性高等特点被广泛应用到工业生产中，所以保证三相异步电机在生产环境中的安全与稳定运行具有十分重要的意义。传统的三相异步电机故障诊断均采用特征电流法，但在实际应用中由于特征谐波难以分离，从而导致无法判断；采用先进的长短期记忆(LSTM, Long Short-term Memory)神经网络以及最新提出的RAdam优化器，在电机正常运转时对其运行特性进行实时采集，通过双峰谱线插值法以及滑窗法提取谐波之后，对电机输出结果进行时序预测和比对；最后以工程中实际电机数据为例，通过测量其故障运行实际数据，验证了该算法的可行性；经实验测试可得，相比于传统神经网络，该算法具有更好的故障检测能力。

Abstract:

Conventional asynchronous motors are widely used in industrial production due to their simple structure, convenient maintenance, and high reliability. Therefore, it is of great significance to ensure the safe and stable operation of the frequency converter in the production environment. Motor fault diagnosis uses the characteristic current method, but in practical applications, the characteristic harmonics are separated, which makes it impossible to judge; the advanced long short-term memory (LSTM, long short-term memory) neural network and the newly proposed RAdam optimizer are used. When the motor is running normally, its operating characteristics are collected in real time. After the harmonics are extracted by the double-peak spectral interpolation method and the sliding window method, the output results of the motor are time series predicted and compared; finally, the actual motor data in the project is taken as an example. The feasibility of the algorithm is verified by measuring the actual data of its fault operation; it can be obtained through experimental tests, and it is used in traditional neural networks, and the algorithm has better fault detection capabilities;