PID Tuning Using Neural Network Classification of Self-Excited Oscillations

This paper proposes a method for tuning PID controllers based on the shape of self-excited oscillations in a system. The approach utilizes a Modified Relay Feedback Test (MRFT) to excite oscillations and a Neural Network (NN) classifier to identify their shapes. The novelty of this work lies in the application of shape-based tuning, categorized into triangular, sinusoidal, wavy, and curved triangular waveforms, each with distinct Tuning Rules (TR)s for integrating and non-integrating systems. A feedforward neural network is developed to classify the MRFT-induced oscillation shapes, enabling the application of appropriate TRs. This classifier shows remarkable accuracy with noise-free signals, and through retraining with noisy data, maintains high performance, outperforming traditional linear discriminant analysis. The study concludes that the NN-based classification significantly enhances the precision of PID tuning by accurately identifying the oscillation shape, thereby ensuring the application of the most effective TR.