Threshold-driven frequency feature integration for neural network-based oscillation detection and quantification in process industries

Abstract

Oscillatory behavior in control loops is a prevalent challenge in process industries, often resulting in detrimental effects such as decreased product quality, lower throughput, and higher energy consumption. These oscillations are typically caused by factors such as valve stiction, suboptimal controller tuning, and external disturbances. This paper introduces a neural network-based method for detecting oscillations, applying data pre-processing and domain-informed feature engineering techniques to improve accuracy while minimizing computational demands. The input features to the neural network are prominent features from the Fast Fourier Transform (FFT) and FFT of Autocorrelation Function (ACF) of dynamic process data, identified based on peaks in the frequency domain data. A sensitivity analysis is performed to evaluate the impact of the number of input features on the model’s accuracy, precision, and recall. The analysis shows that the proposed method achieves a reduction of up to 80% in the number of input features compared to existing techniques in the literature, thus reducing computational time without sacrificing performance, making it suitable for online applications. The proposed algorithm achieves a 96.63% accuracy and a recall of 0.96 for detecting oscillatory behavior. In addition, algorithms are proposed in this work to quantify the oscillation period and the amplitude of the oscillation. The oscillation period is calculated based on the frequency and amplitude obtained from the FFT of ACF of dynamic process data, giving an overall accuracy of 93.15% for regular and irregular oscillations. The performance of the method for predicting the amplitude of oscillation is presented for industrial data to validate its effectiveness in real-world scenarios.