Research On Anomaly Detection And Prediction Algorithm For Multidimensional Time Series Data Based On Deep Learning

In recent years,the level of industrial automation advances rapidly under the support of information science and technology,a large amount of time series data can be collected by multi-sensor equipment.These data often contain a lot of information about working status and operating mode.By mining the hidden information in these time series data for anomaly detection and prediction,it can not only master the health degree of industrial equipment in real time,ensure the normal operation of equipment,but also provide effective guidance for the industrial production process,improve industrial capacity.Based on deep learning method and combined with actual industrial datasets and synthetic datasets,this thesis conducts research on anomaly detection and prediction of multidimensional time series data as follows:In view of the characteristics of industrial time series data,such as large amount of data and high dimension of data.This thesis proposes that the feature selection module based on extreme gradient boosting algorithm is added into the anomaly detection algorithm and prediction algorithm model.Score and sort the variables in the original datasets.By setting different thresholds for experiments,the optimal high correlation subsets of the datasets are obtained.Considering the characteristics of strong coupling and correlation of industrial time series data,this thesis proposes a multi-dimensional attention convolution gated recurrent encoder anomaly detection algorithm based on reconstruction by using encoder-decoder structure.The feature matrix is constructed by using the highly correlated time series data after feature selection by extreme gradient boosting algorithm.A fully convolutional encoder with attention mechanism is used to extract the correlation features between different time series.A Conv GRU module based on attention mechanism is used to extract temporal features between time series.Finally,the fully convolutional decoder is used to jointly decode the feature matrix to obtain the reconstructed feature matrix,set a proper threshold,and the residual feature matrix is used to detect anomalies.The experimental results of three fault datasets show that the anomaly detection effect of this algorithm is better than the other nine comparison algorithms.In view of the problem that the prediction effect of multi-dimensional time series data prediction algorithm model tends to decrease with the increase of time series dimension and length.This thesis proposes a gated recurrent double-branch prediction algorithm based on attention mechanism.The algorithm model has a double branch structure,high correlation time series data after feature selection by extreme gradient boosting algorithm is used as the input of the encoder-decoder branch.The main task of this branch is to effectively compress time series data and generate optimized hidden vector representation.Adding an attention mechanism into the GRU encoder enables the model to give different weights to non-target time series adaptively.The attention mechanism is added before the decoder,different weights are given to the hidden layer of the encoder adaptively,and the characteristics of long-term dependence between time series are extracted.The prediction branch uses fully connected neural network to effectively model the optimized hidden vector.Finally,the prospective prediction of multidimensional time series data is achieved.Experimental results of three prediction datasets show that the prediction accuracy of the proposed algorithm is higher than that of the other eight comparison algorithms.Finally,a data-driven industrial application system is constructed.Experimental results of three fault datasets show that the system is effective and robust.