Malfunction diagnosis based on residence time distribution of radiotracer signals in industrial processes using machine learning techniques

Diagnosing problems in industrial processes has always been a complex challenge, frequently obstructed by the complex structure of these systems. The present study presents a robust methodology integrating nuclear radiotracer data with machine learning approaches to improve diagnosis. Radiotracers are used to measure residence time distribution (RTD) as a crucial diagnostic technology. Experiments utilize a Flow Rig System (FRS) to simulate industrial conditions, where a Tc-99 m radiotracer (1 mCi) is injected in Dirac form and monitored with sodium iodide scintillation detectors integrated with an ALTAIX data acquisition system (DAS). Machine learning algorithms are subsequently employed to categorize four RTD signals: normal RTD, small exchange RTD, recirculation RTD, and parallel flow RTD. Identifying these signal kinds is essential for precise system diagnostics. We utilize deep learning via Convolutional Neural Networks (CNNs) for feature extraction and an Artificial Neural Network (ANN) for classification. Additionally, the Binary Tree Growth Algorithm (BTGA) is employed to refine feature selection, improving model efficacy and decreasing processing demands. The deep learning model attains complete identification accuracy while implementing the HP classifier, which enhances processing time and precision. We simulate RTD signals for two scenarios − Perfect Mixers in Series (PMS) and Perfect Mixers with Exchange (PMSEX). We corroborate our results by comparing them with RTD simulation tools, demonstrating significant correlation and concordance. Our Results highlight the efficacy of combining advanced machine learning approaches with new real-time data modelling to enhance diagnostics efficiency and reliability in industrial operations. This method offers a revolutionary technique to improve process optimization and defect identification.