Dianxin kexue (Jul 2020)
An LSTM auto-encoder based anomaly detection for industrial system
Abstract
In the context of the industrial internet,automatic and effective anomaly detection methods are of great significance to the safe and stable production of industrial systems.Traditional anomaly detection methods have the disadvantages of requiring a large number of labeled samples,and not adapting to high-dimensional time series data.Aiming at these limitations,an industrial system anomaly detection method based on LSTM (long short-term memory)auto-encoder was proposed.Firstly,to address the limitation of relying on labeled samples,an encoder used to learn the features and patterns of a large number of normal samples in an unsupervised manner.Then,anomaly detection was performed via reconstructing samples and calculating the reconstruction error.Secondly,to adapt to high-dimensional time series data,a BiLSTM (bidirectional LSTM) was used as an encoder,and then the potential characteristics of multi-dimensional time series data were mined.Experiments based on a real paper industry data set which demonstrate this method has improved the existing unsupervised anomaly detection methods in various indicators,and the overall accuracy of the detection has reached 93.4%.
