Research On Deep Learning Based Method For Aircraft Telemetry Time Series Data Anomaly Detection And Forecasting

With the rapid development of information technology,time series data is being generated in various fields at a breathtaking rate.Especially,in the field of aerospace measurement and control,there are large amount of telemetry data in the form of time series.These data information will directly reflect the working state of each aircraft component,and is an important evidence for detecting the operation state and fault handling.In this paper,aiming at the rotational speed and temperature time series data of a certain aircraft device,using the theory and technology of deep learning,we study the anomaly detection and forecasting method for data-driven-based time series data.The main works are as follows:Considering that the measured data in the practical project is presented in the form of typical unbalanced data distribution,we investigate the issue of anomaly detection for time series data with unbalanced inter-class distribution.We propose an anomaly detection method for time series data based on deep convolution neural network.In the proposed method,firstly,the unbalanced time series data is processed by sampling method,so that the data set with insufficient data samples can have enough representative samples.By this way,the shortcomings that some existing detection techniques have lower detection accuracy due to neglecting the skewness of the data distribution can be overcame.Secondly,in order to improve the detection performance of the convolution neural network.,the original time series data is preprocessed by using the scale transformation and time slice technologies.Then we investigate the influence of different hidden layers for the detection performance of convolution neural network model,and construct a convolution neural network model with four hidden layers.Finally,simulation experiments were carried out by using the rotational speed and temperature time series data of a certain aircraft device.The results show the effectiveness of the proposed method is satisfactory.Comparing with the traditional detection methods that are based on the combination of feature extraction and classifier,the proposed method has satisfactory performance in detection accuracy.We propose a time series data forecasting method based on variant long short-term memory(LSTM)recurrent neural network.In the propos.ed method,considering the requirement of timeliness for fault forecasting,we firstly merge the inputting gate and the forgetting gate into an update gate to construct a variant LSTM structure,which simplifies the learning parameters and improves the training efficiency.Then we propose a time series data forecasting model based on the proposed variant LSTM recurrent neural network.Considering the requirements of the long sequence data forecasting,we introduce a parameter update mechanism in the forecasting stage.By this way,the network model can dynamically adjust the model parameters,thus,more suitable for the forecasting of long-term sequence data.Finally,the simulation experiments are carried out by using the rotational speed and temperature time series data of a certain aircraft device.Through comparing with several typical forecasting models,the results show that the proposed method has satisfactory forecasting effect.Our research work can provide a certain reference value for using historical data detect and predict faults in the engineering of aerospace measurement and control,and provides basis for improving the accuracy of fault diagnosis in the actual measurement proj ect.