Control System Response Improvement via Denoising Using Deep Neural Networks

Abstract

Noise is an inseparable part of control systems. Every sensor reading used for determining the state of a control system is corrupted with noise., therefor increasing the signal to noise ratio of sensor readings can significantly improve the performance of systems. The proposed filter in this paper captures the underlying probability distribution of the noise-free input signal in the training stage and therefor is capable of refining the corrupted input signal regardless of the distribution of the added noise. In order to acquire better data-driven based results in the suggested approach, it has been decided to use different neural network structures and stack these layers to form a hybrid multilayer filter. The properties of the stacked neural network sublayers guarantee a robust and general solution for the intended purpose. The key elements of the proposed filter are two auto-encoders., a dense neural network and a convolutional layer. Effects of every sublayer on the corrupted input signal., along with fine-tuning of these sublayers are discussed in detail. Afterwards in order to assess the generality and the robustness of the method., the proposed filter is exposed to a non-Gaussian noise. Finally., the proposed filter is tested on a linear and a nonlinear system. The comparison between systems” output to the reconstructed signal with that of the original noise-free signal., suggests the substantial improvement in the performance of the given systems. This improvement is in terms of systems” output resemblance to the reference noise-free output signal when the reconstructed signal is applied to the systems.