Abstract: Due to vibration signals’ non-stationary characteristic of complex machinery such as reciprocating compressor and aeroengine, the single feature threshold alarm method frequently utilized at present has shortage of inaccurate alarm. For this reason, a variational auto-encoder-based method for fault early warning is proposed according to the response signals’ characteristics of mechanical response point. The method is based on high-dimensional feature parameters of the mechanical vibration signals. It is used to self-learn the statistical distribution model of high-dimensional features by variational auto-encoder, and the normal working condition model is used as a benchmark model. The fault early warning is realized by comparing the difference between real-time working condition model and benchmark model with the adaptive early warning threshold. The advantages of proposed method are verified by comparing with the single feature threshold alarm method and the state subspace-based early warning method. The results demonstrate that the method can improve the accuracy of mechanical fault early warning and greatly advance the alarm time point of the fault. It has high timeliness and strong adaptability.

Key words: variational auto-encoder, machinery, fault early warning, statistical distribution

摘要： 由于往复压缩机、航空发动机等复杂机械振动信号呈现非平稳性，目前应用较多的单特征门限报警方法存在报警准确率低的问题。针对该问题，依据机械响应点的响应信号特点，提出一种基于变分自编码器的故障预警方法。该方法基于机械振动信号的高维特征参数，利用变分自编码器自学习出高维特征的统计分布模型，将正常工况模型作为基准模型，通过计算实时工况模型与基准模型间的差异度，并将其与自适应预警阈值相比较，实现故障预警。通过与单特征门限报警方法、基于状态子空间的预警方法进行比较，验证了该方法的优越性。实验结果表明，该方法能够提高机械故障预警的准确率并大幅提前故障的报警时间点，具有较高的时效性和较强的适应性。

关键词: 变分自编码器, 机械, 故障预警, 统计分布