Research On Key Technologies For Abnormal Detection Of Satellite Attitude Sensors By Data Driven Methods

With the increasing maturity of satellite technology and the increasing demand for satellites,the number of satellites launched into space is increasing.As more and more incidents of satellite attitude sensor system failures accumulate,anomaly detection can detect possible fault symptoms as early as possible,and provide timely warnings before component performance deteriorates or even malfunctions occur.Therefore,the anomaly detection technology for satellite attitude sensors has received increasing attention and research;On the other hand,the integration and complexity of satellites are gradually increasing,and the dimensions and volume of telemetry data are rapidly expanding.With the enhancement of satellite computing power,telemetry data prediction technology based on data-driven analysis and prediction for massive data has become a key research direction.Simulation models established on the ground often have significant errors between simulation data and actual data due to limitations in experimental conditions and understanding of the space environment,making it difficult to achieve high-precision autonomous anomaly detection.If this part of the error can be reduced,using a combination of physical and predictive models for anomaly detection will have interpretability,high accuracy,and scalability at the same time.Therefore,based on the existing data driven technology,this paper proposes the key parameter prediction model and anomaly detection method that are respectively applicable to the sun sensor and magnetometer.For the sun sensor current with less influencing factors,the system identification method is used to establish a nonlinear black box equivalent system between the simulation current and the actual current.Specifically,for the dynamic linear and static nonlinear relationship between the simulated current and the actual current,the three-layer structure of the Hammerstein Wiener model is used,the self constructed biorthogonal wavelet is used in the model input nonlinear module to extract the signal characteristics and nonlinear,the lipschitz entropy is used to select the dynamic linear module parameters suitable for the signal,and the piecewise linear function is used in the output nonlinear module to nonlinear map the intermediate variables to the output,Identify linear and nonlinear characteristics separately.By using the momentum gradient descent method to train model parameters,the convergence speed was improved and oscillation was reduced.This article establishes a black box model between the simulated current and the actual current,and proposes a corresponding anomaly detection threshold generation method.Finally,the feasibility of this approach is demonstrated on the actual current data of a certain model.For magnetic data with complex influencing factors,the focus is on studying the residual between the actual magnetic data and the output of the basic magnetic model.For the two tasks of missing value supplementation and continuous sequence prediction,time series residual prediction models are designed separately.In response to the common problem of missing data,the actual magnetic data is first subjected to field removal preprocessing.A missing value supplement method is proposed,which first corrects the transformation matrix and then uses the TCN-SE model to predict the residual,solving the problem of difficulty in high-precision filling.Then,a periodic expansion convolution model based on periodic information is proposed to predict the continuous residual sequence,and Bayesian optimization is used for hyperparameter optimization,which solves the problem that common time series models cannot accurately establish global dependence.At the same time,the difference between extrapolated data and telemetry data is used to define fault tolerance for real-time anomaly detection,and the historical dataset is sequentially updated to achieve online model updates.Finally,a certain type of telemetry magnetic data is selected for example verification,and the proposed model has achieved good compensation effect on missing data.The prediction error sequence on the continuous residual sequence test set is approximately white noise,and the simulated diffusion drift anomaly is detected,which proves the effectiveness of the method.