Satellite anomaly detection method based on Bayesian neural network

Abstract

The invention discloses a satellite anomaly detection method based on a Bayesian neural network, and the method comprises the steps: different from an anomaly detection method employing a conventionalneural network, introducing the Bayesian idea into the neural network, and enabling the weight of the network not to be a single value, but to accord with certain probability distribution. The Bayesian thought gives uncertainty to the neural network, and gives better mathematical explanation to the neural network which is a black box model. The method comprises the following steps of firstly, creating a traditional long-short-term neural network according to satellite data; secondly, introducing a Bayesian thought, establishing a Bayesian long-short-term neural network, performing approximateinference by using a dropout method, and learning a network weight by minimizing KL divergence between approximate distribution and posteriori distribution of the network weight; and then, outputtinga network result in a Monte Carlo sampling approximate weight distribution mode; calculating the uncertainty of an anomaly detection classification result by adopting two measurement modes of prediction entropy and mutual information; finally, further judging manually the classified samples with high uncertainty or through an auto-encoder, so that the accuracy of anomaly detection can be better improved.

Description

technical field [0001] The invention relates to a novel satellite abnormal data detection method based on a Bayesian neural network, which is a method for detecting satellite abnormal data and belongs to the cross field of engineering application and information science. Background technique [0002] A satellite is a complex system composed of many interrelated and mutually restrictive components. It is a fusion of multidisciplinary technologies such as telemetry perception, wireless communication, and navigation control. Satellites have opened the door for human beings to explore the vast universe. Since the satellite is in the harsh outer space environment such as solar radiation for a long time, unpredictable abnormalities or failures may occur during its in-orbit operation. Taking measures in advance to detect these unpredictable abnormalities or failures in time can ensure the long-term stable operation of the satellite. Therefore, anomaly detection plays an important r...

AI Technical Summary

Problems solved by technology

[0005] Satellite anomaly detection relies heavily on model accuracy, and an overconfident decision may underestimate the danger posed by anomalous telemetry data

Moreover, the data set in this field is unbalanced (satellite telemetry data has far more normal data than abnormal data), which will lead to overfitting of traditional models on the sampled categories

Images