Research On Deep Neural Network Based Method For Time Series Data Anomaly Detection And Forecasting

Time series is an important form of structured data and has important applications in many fields.In the field of aeronautics and astronautics,a large number of telemetry data are generated during the operation of spacecraft.These data are presented in the form of time series,which can directly reflect the working state and operation mode of each component.By using time series data mining and analysis technology to detect and predict the telemetry time series data,it is helpful to master the health status of spacecraft in real time and carry out health management,which is an effective means to ensure the normal operation of spacecraft and improve the efficiency.In this paper,based on the characteristics of deep neural network,taking the time series data of a certain equipment in a practical project as the research object,the anomaly detection and multi-step-ahead prediction method of time series data are deeply studied.The main works are as follows:In practical engineering,most of time series data are in the form of inter-class imbalance,and the minority class samples are far less than the majority class samples,which makes the learning model unable to learn enough features from the minority class.To solve this problem,we propose a Cost-Sensitive Hybrid Network(CSHN)model to solve the anomaly detection problem of time series data with severe skew between classes.The proposed model consists of Convolutional Neural Network(CNN)and Gated Recurrent Unit(GRU)network,and combines with a cost-sensitive loss function.This model integrates the characteristics both CNN and GRU.The former is provided with strong local feature learning ability and the latter has better sequence characteristic learning ability.The cost-sensitive loss function is used to substitute general cross-entropy loss function,to solve the problem that the detection accuracy of the minority class will be inaccuracy caused by the skew data distribution.In the proposed cost-sensitive loss function,we use different penalty factors to penalty the misclassification of the network model.The simulation experiments are conducted on the speed and temperature datasets of a component of the spacecraft,as well as the UCR datasets,the experimental results show that this method has good performance.In particular,for imbalanced datasets,the detection accuracy of the proposed method is improved obviously.The issue of multi-step-ahead time series prediction is a daunting challenge of predictive modeling,especially the prediction of long-term time series data.The prediction performance of the prediction model becomes worse and worse with the increase of the prediction length.To solve this problem,we propose a multi-output iterative prediction model based on stacked LSTM neural network(MO-LSTMs).The model consists of multiple LSTM hidden layers and dropout stack,each hidden layer not only contains different neural units and the memory state of the cells in each layer are reset,so this model not only improves the single LSTM network nonlinear representation ability,and solves the problem that the single LSTM network structure is difficult to maintain the time-sequence characteristics between samples in the training process.Meanwhile,using the dropout algorithm also improves the generalization ability and robustness of the proposed model.Using the MO-LSTMs model,we establish a multi-step-ahead time series data prediction method.The method utilizes the strategy of multi-output iterative prediction to reduce the errors accumulation and errors propagation for long-term time series prediction.It also reduces the computational complexity of the iterative strategy.The simulation experiments are conducted on the speed and temperature datasets of a component of the spacecraft,the results show that this method has good performance for the prediction of long-term time series data.