DEVELOPMENT OF NEURAL NETWORK FOR STOCHASTIC NOISE MODEL PARAMETERS ESTIMATING IN ACCELEROMETER SIGNAL

Abstract

MEMS accelerometers are widely used in many applications but they have significant random and systematic errors. It limits their accuracy under real-world conditions of work. Modern studies in the field of calibration and accelerometers parameter estimation are focused at increasing autonomy, adaptability and resilience to external influences. These studies are usually limited by training conditions or power of hardware resources. Thus it requires the development of estimation methods capable to operate under high uncertainty, with limited information, with processes that nature is fundamentally stochastic. For these tasks are highly relevant machine learning algorithms which capable model non-linear dependencies, to account time and spectral properties of noise processes.

In this work neural network for estimating parameters of stochastic processes in accelerometer signal is developed. The issues of designing and modeling operation of such neural network are considered. The neural network utilizes an architecture that simultaneously processes time and spectral information. The developed neural network model features two parallel branches: time and spectral. After separate information processing, the results from each branch are combined into shared latent space and fed into the final regressor. It that estimates the noise signal model parameters. Gauss-Markov processes are selected as the model for the stochastic processes. In this paper the aspects of training the developed neural network are described and recommendations for training data generation are provided. To increase training stability and efficiency, two standard training process optimization mechanisms are used — EarlyStopping and ReduceLROnPlateau. This allows to improve convergence quality, to reduce risk of overfitting and ensures the acquisition of the model with high generalization capabilities. The developed neural network model demonstrates effective ability to estimate the of stochastic processes parameters that form the random noise in MEMS-accelerometers signal.