Implementation of a Biologically Plausible Neural Network-Based Disturbance Observer for Real-Time Control Systems

In most automated applications, a control loop system which consists of a feedback loop and a controller is designed to make the system responses follow given commands. The system is operated in real-time by sensing errors between the commands and actual responses, and calculating control inputs to minimize these errors. Parameters for the controller (e.g. PID controller) are optimized based on a mathematical model which approximates the real system. Therefore, control performance can be degraded if there exist unexpected disturbances or discrepancies between the model and the real system. A Disturbance Observer (DOB) can be employed to address these degradations. It compares actual system responses with model-predicted responses, calculates compensation inputs, and adds these to the existing control inputs. Among various DOBs, Artificial Neural Network-based DOBs (ANN-DOB) [1] have demonstrated superior performances in many control problems. However, conventional ANN-DOBs have suffered from significant computational burdens, because they require repetitive learning and inference operations per each control cycle time (typically 50 μs to 1 ms).<br/><br/>In this study, a Biologically plausible Neural Network-based DOB (BNN-DOB) is proposed to overcome the above computational burden issues. The biological plausibility means mimicking functionalities of neurons in biological brains, where learning and inference operations occur through spike signal flows and neuron interactions without any computational processes. In the proposed BNN-DOB, the conventional ANN structure is replaced with a Spiking Neural Network (SNN) structure [2], and a Reward-modulated Spike Timing-Dependent Plasticity (R-STDP) rule [3] is adopted to train weights of neurons in the SNN. R-STDP is a modified STDP that additionally adjusts weight updates using reward signals, changing the original STDP from an unsupervised learning rule to a supervised rule. In the BNN-DOB, model – actual system response errors are used as reward signals.<br/><br/>The performance of the proposed BNN-DOB is verified using simulations under a control scenario where unexpected system nonlinearities, load condition variations, and external disturbances are added to a real system compared to an expected model. The results demonstrate that the proposed BNN-DOB can maintain the system’s command tracking performances even with added system nonlinearities and load condition changes, and reduce disturbance-induced system oscillation to 0.1 times the system oscillation when no DOB is used.<br/><br/>[1] Liu Jinkun, "Radial Basis Function (RBF) neural network control for mechanical systems: design, analysis and Matlab simulation.", <i>Springer Science & Business Media</i>, 2013.<br/>[2] Maass Wolfgang, "Networks of spiking neurons: the third generation of neural network models.", <i>Neural networks</i>, 1997.<br/>[3] Frémaux Nicolas, and Wulfram Gerstner, "Neuromodulated spike-timing-dependent plasticity, and theory of three-factor learning rules.", <i>Frontiers in neural circuits</i>, 2016.