基于BP-LSTM的旋转机械振动信号异常检测模型  

Anomaly Detection Model of Rotating Machinery Vibration Signal Based on BP-LSTM

作　　者：高玉才 付忠广[1] 谢玉存 杨云溪 Gao Yucai;Fu Zhongguang;Xie Yucun;Yang Yunxi(Key Laboratory of Power Station Energy Transfer Conversion and System of Ministry of Education,North China Electric Power University,Beijing 102206,China)

机构地区：[1]华北电力大学电站能量传递转化与系统教育部重点实验室,北京102206

出　　处：《煤矿机械》2021年第8期210-213,共4页Coal Mine Machinery

基　　金：国家自然科学基金资助项目(50776029)。

摘　　要：旋转机械运行环境恶劣,振动信号易受外界干扰,因此实现振动状态的异常检测较为困难。神经网络技术能够从大量的振动数据中自动提取故障特征,相对于人工提取故障特征,工作量大为减少。结合长短时记忆(LSTM)网络对时间序列数据具有的超强感知与处理能力,提出一种用BP神经网络压缩输入数据维度从而提取故障特征、用LSTM网络进行异常检测的新方法。首先将实验台获取的振动数据划分为相同序列长度的数据样本并划分为不同的数据集,然后将数据样本进行预处理并搭建BP-LSTM网络,将建立的模型用于旋转机械振动信号异常检测。模拟仿真和实验结果表明:BP-LSTM网络模型对旋转机械运行状态的检测具有较高的精度和稳定性,该方法优于基于时域特征参数进行异常检测的支持向量机(SVM)、K近邻和LSTM等传统学习方法。Rotating machinery operates in harsh environment, and the vibration signal is susceptible to external interference, so it is difficult to detect abnormal vibration conditions. Neural network technology can automatically extract fault features from a large amount of vibration data. Compared with manually extracting fault features, the workload is greatly reduced. Combining the long and short-term memory(LSTM) network with super-strong perception and processing capabilities for time series data, a new kind of method was proposed which uses BP network to compress the input data and extract fault features, and then uses the LSTM network to perform anomaly detection. Firstly, the vibration data obtained by the experimental platform was divided into data samples of the same sequence length and divided into different data sets. The data samples were then preprocessed, the BP-LSTM network was built and used the established model to detect abnormal vibration signals of rotating machinery.Simulation and experimental results show that the BP-LSTM network model has high accuracy and stability in detecting the operating state of rotating machinery. This method is better than support vector machines(SVM), K-nearest neighbor and LSTM, which are traditional learning methods and based on time-domain feature parameters for anomaly detection.

关 键 词：LSTM网络 神经网络 旋转机械 异常检测 检测精度 

分 类 号：TH17[机械工程—机械制造及自动化]