Abstract: Vibration signals can effectively reflect the operating status of the motor, and are therefore considered an important basis for diagnosing induction motor faults. However, the original vibration signal has the problem of single features and long time series, and existing research usually extracts features based on expert experience, which is costly. In recent years, the accumulation of fault data has promoted the application of deep learning methods in fault diagnosis. A feature engineering method (MAC-LSTM) based on multi-attention mechanism and one-dimensional convolutional neural network is proposed for fault diagnosis of induction motors, which does not require any prior knowledge. Firstly, the multi-attention mechanism is used to expand the dimensionality of features, making the representation of original features more abundant. Secondly, the convolutional neural network extracts features from the time dimension and reduces the dimensionality, effectively solving the problem of the original signal timing being too long. Finally, LSTM captures the temporal dependence of the signal for fault diagnosis of induction motors. The experimental results show that MAC-LSTM has achieved excellent performance in fault diagnosis of induction motors based on vibration signals and has high generali-zation ability.

Key words: induction motor, fault diagnosis, attention mechanism, convolutional neural network, long and short term memory neural network

摘要： 振动信号能够有效地反映出电机运行的状态，因此被视为诊断感应电机故障的重要依据。然而，原始振动信号存在特征单一、时序过长的问题，已有研究通常基于专家经验提取特征，成本较高。近年来，故障数据的积累推动了基于深度学习方法在故障诊断中的运用。针对上述问题提出了一种基于多头注意力机制和一维卷积神经网络的特征工程方法（multi-attention with 1D convolutional neural network，MAC-LSTM）用于感应电机的故障诊断，该方法无须任何先验知识。多头注意力机制被用来拓展特征的维度，使得原始特征的表示更加丰富；卷积神经网络从时间维度上提取特征并降维，有效解决原始信号时序过长的问题；LSTM捕获信号的时序依赖性，用于感应电机的故障诊断。实验结果表明，MAC-LSTM在基于振动信号的感应电机故障诊断中取得了优异的性能，并且具有很高的泛化能力。

关键词: 感应电机, 故障诊断, 注意力机制, 卷积神经网络, 长短期记忆神经网络