Genetic–Neural Adaptation of the Kalman Filter Process Noise Covariance for Dynamic Tracking in Electro-Pneumatic Systems

This paper proposes a hybrid machine-learning approach based on the genetic-neural online adaptation of the process noise covariance matrix in an adaptive Kalman filter, designed to improve the accuracy of dynamic tracking in electro-pneumatic systems. The method combines offline parameter optimisation using a genetic algorithm with online adaptation via a neural network. A two-level architecture was developed to generate the process noise covariance matrix, based on adaptive coefficients through which the statistical characteristics of the mathematical model are aligned with the current dynamic state of the electro-pneumatic system. This approach ensures that the algorithm adapts to changes in speed, acceleration, external load and motion profile, leading to a more accurate state estimation, improved feedback quality and enhanced accuracy of dynamic tracking across diverse operating modes. A key feature of the proposed method is the use of an independent reference measurement channel. Unlike current adaptive structures, in which the Kalman filter is tuned solely based on measurements from internal sensors, the proposed method determines the adaptive coefficients by optimisation based on minimising the actual tracking error, measured by an independent reference metrological system implemented using a laser interferometer.