Research Of Satellite Parameter Classification Based On Machine Learning

In the orbital operation of the satellite,it is necessary to judge the on-orbit operation mode of the payload.Taking the Dark Matter Particle Explorer Satellite(DAMPE)as an example,the ground staff can determine whether the payload is in the specified working state,by monitoring the telemetry data of the payload,and then determine whether the detector performs the task according to the scheduled plan.The method used now is an expert system based on thresholds,which requires more labor costs.Moreover,the operating mode of the payload is affected by many factors,which tend to change over time,and the threshold is expensive to be re-set.According to machine learning theory,this paper abstracts the operation mode judgment of payload into the multi-class classification problem in machine learning.The classifier model is trained by telemetry data,to classify the satellite telemetry parameters as an auxiliary tool for the operation mode judgment.The research mainly focuses on the following two aspects in the process of establishing the classification model:1.The class imbalance problem.In the data samples from satellites,the samples corresponding to some operating modes may be relatively rare,and the distribution of samples in each class is unbalanced,which may make it difficult for the classifier to accurately classify rare classes.2.Data stream classification and concept drift problem.Satellite payload related parameters may change over time,such changes need to be detected and dealt with,and classification model updated on this basis.Aimed at the machine learning classification method of satellite payload parameters and the problems that need to be faced in practical applications,the real telemetry data of dark matter particle explorer satellite is extracted,preprocessed and its characteristics are simply analyzed.Using satellite real telemetry data,several classification algorithm,including k neighbor algorithm,naive Bayes algorithm,and decision tree classification algorithm,are used to classify satellite telemetry parameters.Among these methods,The classification accuracy of the decision tree classification algorithm is quite high,but the classification effect of the algorithm on rare classes is very poor.For the problem of class imbalance,this article discusses several methods to solve the classification of imbalanced data.Including re-sampling method,cost sensitive learning method,classifier integration method,training set partitioning method and feature selection method.The actual effects of these methods in satellite parameter classification are verified and compared by experiments.The experimental results show that both the re-sampling method and the cost-sensitive weighting method can significantly improve the classifier's classification performance for rare classes in the experimental data samples.Relatively speaking,the weighting method is often used to improve the effect.obvious.The XGBoost algorithm,especially after sample weighting,has reached a fairly high level of f1-macro evaluation,much higher than several other algorithms.Aiming at the problem of data stream classification and concept drift,a simple data stream classification model based on adaptive window and XGBoost algorithm is designed.This method is easy to implement and extend.Experiments verify the effectiveness of this classification model.The method based on machine learning method for classifying satellite parameters presented in this paper has great practical application value,easy to expand and transplant,it reduce cost,reduce the workload of ground monitoring staff.The potential information in the data can also be mined in the classification process to provide relevant reference for relevant experts.