Research of the stability of a machine learning model for processing signal of the Earth orientation device

The article deals with the solution of the problem of recognizing the artifacts of the Sun and the Moon in the signal of the photodetector of the Earth orientation device. The previously developed machine learning model, trained on clean signals (real and generated), must be tested for resistance to real signals that can be noisy and, if necessary, correct the model. A study of the stability of the machine learning model to noisy signals was carried out, a decrease in the quality of the classifier of normal and artifact signals of the photodetector of the Earth orientation device was revealed. To correct the model, the method of augmenting the initial data using artificially noisy signals was used, the types of noise characteristic of electronics (thermal, shot, and flicker noise) were selected, and the levels of the signal-to-noise ratio for devices of this class were selected. Methods for generating noise signals are selected and algorithms for correct mixing of clean signals with noise are developed. The shortcomings of the existing linear model were revealed, which appeared only when using different levels of the signal-to-noise ratio. Linear model errors turned out to be significant for the application area. A new classifier model based on decision trees was chosen. For the new model, a stability test was also carried out, and the need to expand the range of the signal-to-noise ratio for the training sample was shown. The applied method of classifying artifacts and increasing the stability of the model can be used in photodetector devices if it is necessary to provide built-in signal processing in the computing core of the device.