Time Series Data Reconstruction And Anomaly Detection Based On Attention Mechanism

With the development of the Internet of Things(Io T),the time series data recorded from various sensors is becoming more and more complete,and the sampling frequency is getting higher and higher.The industry needs an accurate and fast method to perform data mining on time series data for downstream tasks.Time series data reconstruction refers to learning the representation of a set of data by ignoring the signal "noise" by training the model,and trying to generate a representation from the simplified encoding as close to its original input as possible.And anomaly detection is an application of time series reconstruction,moreover one of hot topic in aerospace.This paper mainly applies the attention mechanism which has been widely used in natural language processing tasks for time series data processing,including reconstruction,anomaly detection,relationship extraction and optimization of related models.The main work of this article is as follows:1)Verify the effect of the attention mechanism in time series data tasks,and embed the attention mechanism into LSTM autoencoder.First,one-step prediction is determined,and then the compare the reconstruction effect of the sub-dataset under different conditions of the high storage data set.Compare reconstruction effects for three other different datasets for various scenes.At the same time,the comparison of reconstruction index on absolute error and DTW(dynamic time wrapping)error is given.Finally,with the help of dynamic threshold segmentation,anomaly detection is performed on the NASA spacecraft dataset.2)Aiming at the low speed of LSTM model and its inability to extract long-distance relationships,Transformer which has become mainstream model in natural language processing is introduced.Reconstruction experiments show that Transformer is dominant on some datasets,and computation time is greatly reduced on all datasets.The anomaly detection experiments show that the Transformer can almost achieve the same anomaly detection accuracy,and the classification experiments show that the Transformer can extract the relationship between artificial instructions in the entire sequence,especially in the abnormal sequence,which is helpful for the human factor analysis of the spacecraft failure.3)For the volume and speed bottlenecks of the Transformer model in the actual deployment scenario of anomaly detection,optimize models from the perspective of algorithms and engineering.The model was successfully compressed using the distillation network,and the accuracy was higher than that of directly trained networks.Use mixed-precision training to speed up the training process and low-precision inference to speed up the inference process.Experiments show that for time series data,the compression and acceleration of the Transformer model really work.