Research On LSTM Based Anomaly Detection And Fault Early Warning For In-orbit Satellites

In-orbit satellite has high civil and military value and its cost is high,therefore,it is very important to extend its service life.The in-orbit satellite is so far away in deep space that it is difficult to fix it.At present,the only way to reduce the occurrence of faults and extend the service life of in-orbit satellites is monitoring the telemetry data and alternating to use the redundant software and hardware according to control instructions.In order to monitor the status of in-orbit satellites systematically,thousands of kinds of telemetry data are collected for real-time monitoring of in-orbit satellites,so that anomalies and fault omens can be detected in time and major faults can be prevented through control instructions.Faults or failures mean performance degradation or failure,and anomaly detection is one of the important means of fault early warning.In-orbit satellite anomaly detection and fault early warning has the following characteristics,including huge amounts of telemetry data,a great variety of telemetry channels,complex intra-channel and inter-channel relationships.Besides that,low misdetection rate,low false alarm rate and interpretability are required so that the needs of real-time monitoring of in-orbit satellites can be satisfied.Existing anomaly detection and fault early warning approaches have their own limitations,such as poor generalizability,dependency on labeled sample data,and disadvantages in detecting anomalies within predefined limits and anomalies of unknown types.Therefore,new anomaly detection and fault early warning methods are urgently needed to satisfy the increasing demands for reliability and interpretability of in-orbit satellites monitoring.Based on the application background of anomaly detection and fault early warning of in-orbit satellites,this paper focuses on the challenges and difficulties of anomaly detection and fault early warning of in-orbit satellites,comprehensively uses machine learning,knowledge engineering and other related knowledge,systematically and deeply studies the anomaly detection and fault early warning methods based on Long Short-Term Memory networks(LSTM).Specifically,we studied how to combine LSTM with oneclass support vector machines for anomaly detection,a multi-task learning based encoderdecoder model,a deep-learning based point-wise classification fault detection method and how to combine LSTM with ontology for fault early warning.The main research results and innovations of this paper are as follows.(1)Since anomaly detection methods merely based on LSTM prediction errors rely too much on the model prediction performance,a novel anomaly detection method DALEO(Detecting Anomalies using LSTM and Ensembled One-class support vector machines)is proposed.This method integrates two unsupervised anomaly detection methods,LSTM and One-Class Support Vector Machines(OC-SVM),in a new way:firstly,multiple One-Class Support Vector Machines are used to obtain the ensemble outputs of high precision and high recall according to different voting thresholds,and then these ensemble results are integrated into the two stages of LSTM-based anomaly detection method according to their characteristics.The experimental results demonstrate the advantage of DALEO in improveing anomaly detection performance and its effectivenss in dealing with telemetry data of irregular intervals.(2)Traditional anomaly detection methods need additional anomaly diagnosis procedures to evaluate the detected anomalies and can not find non-abnormal fault omens.To solve the above problems,a Multi-Task Learning based Encoder-Decoder(MTLED)model is proposed.‘temporal feature matrix’ is introduced into MTLED model,so that features can be extracted for each time point.The task-shared encoder obtains the temporal feature matrix through point-wise feature extraction and then the temporal feature matrix and multiple decoders are utilized to realize anomaly detection,anomaly diagnosis,event detection respectively.Owing to the features extracted for each sampling point,point-wise anomaly detection can be realized in an end-to-end way with the decoder for anomaly detection.Experiments show the feasibility of MTLED and its advantages in improving the performance of each task when they are solved together,as well as its generalizability in anomaly detection.(3)It is difficult for the existing fault detection methods to obtain high performance and low detection delay simultaneously,therefore,a Deep-learning based Point-wise Classification method for Fault Detection(DPCFD)is proposed in this paper.DPCFD mainly consists of two deep learning models: Sequence State Generator(SSG)and Deeplearning based Point-wise Classification Model(DPCM).SSG model is used to generate channel real-time state sequence corresponding to any single telemetry sequence.‘Channel real-time states’ refer to the real-time qualitative descriptions of channel data.The real-time state sequence of each channel and the original telemetry sequence are used as the input of DPCM model,and these inputs are grouped according to the correlation between channels.Through the grouped feature extraction and concatenation,as well as a variety of feature enhancement means,DPCM model obtains the fault detection results through point-wise classification.Experimental results show the effectiveness of SSG model and the advantages of DPCM in acheveing both high detection performance and low detection delay.(4)In order to solve the problem that the fault warning results obtained from deep learning method are lack of interpretability and do not make full use of the prior knowledge,a fault warning method CDOFW(Combining Deep learning and Ontology for Fault Warning)is proposed.By constructing the ontology Onto AC and Onto ACF,the knowledge about satellite fault is modeled systematically.In order to reduce the computing resources of overall monitoring,a monitoring framework based on fault modes is designed.CDOFW integrate the prior knowledge such as fault threshold into the deep learning model for fault early warning.Besides that,ontologies are used for fault early warning and fault determination in CDOFW.The reasoning process of ontology will output ‘fault phenomenon’ which can be seen as the explanation of fault early warnings.Simulation results show that more reliable and interpretable fault early warning results can be obtained with CDOFW,while the overall computing resources is reduced.