Application Of Self-Attention LSTM In Time Series Analysis

In the age of big data,medical,financial,climate,aerospace,transportation and other fields will generate a lot of time series data,especially in the field of transportation,the research on the time series data of axle failures can avoid the huge losses caused by the failures,which is of great significance to the development of social economy.The time series of axle failures will be affected by noise during the collection process,and the data obtained has the characteristics of non-linearity,instability,and imbalance in categories.Ordinary deep learning methods cannot record data information useful for classification in the past.In order to better classify time series data,this paper proposes a hybrid classification model of axle failure based on time series data CEEMDAN-MPRes Net-self-attention LSTM(CEEMDAN-MPRes Net-SALSTM).First of all,time series data signals are collected by acoustic emission sensor.Then,in order to reduce the influence of noise on the data,the time series data is decomposed by the ensemble empirical mode decomposition algorithm with adaptive noise,and several IMF signal components with different frequencies and a residual component are obtained;Next,in order to better retain the data features and reduce the impact of the change of the input vector sequence on the classification results,it is proposed to add a self-attention mechanism to the LSTM model;in order to reduce the time cost of training the model,the MPRes Net model is used to first feature the data set rough extraction,reduce the dimensionality of the time series,and then send the extracted feature vector to the self-attention LSTM model to further extract features.Finally,classification is performed through the softmax layer.After the above theoretical and experimental research,the axle failure time series data collected by the acoustic emission sensor is used as the experimental object,and the hybrid classification model of CEEMDAN-MPRes Net-self-attention LSTM is used for ablation experiments with other models.It is concluded that the model has obvious classification accuracy,the accuracy rate reaches 99%,and the loss rate is also lower than other models.