DATA-DRIVEN FAULT DETECTION AND CLASSIFICATION DESIGN WITH DEEP LEARNING ON THREE INTERACTING TANK SYSTEM

• Main Author: Rayhana, Dita
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/77413
• Institution: Institut Teknologi Bandung

In modern control design, precise and reliable control is crucial to ensure the optimal performance, efficiency, and safety of a system. Sensors play a pivotal role within the system by enabling control algorithms to make accurate decisions. Fault detection systems are one of the vital components that should be incorporated into modern control designs. Not only do they serve as detectors, but error classification systems are also required to facilitate the diagnosis and mitigation processes. The success of Convolutional Neural Networks (CNNs) in pattern recognition and object detection has led to the adoption of the Gramian Angular Field method for representing time series data as images for CNN input. This research focuses on the design of a fault detection and classification system using Deep Learning methods, specifically CNN, Tiled CNN, and Transfer Learning with the EfficientV2M model. The faults to be detected are incipient faults, such as bias and drift. The system is expected to classify the data into three categories: normal data, biased data, and drifted data. The results of this study indicate that detectors and classifiers using two- dimensional input perform significantly better than those using time series input. The CNN method achieved a test data accuracy of 97.2%, while the Tiled CNN method achieved an accuracy of 97.08%. The system's performance using CNN and Tiled CNN methods is far superior to that of the system using Artificial Neural Networks (ANN) with time series input, which achieved only 17% accuracy. Meanwhile, the fault detection and classification system using the transfer learning model EfficientV2M yielded the most satisfactory results among the methods, with a test data accuracy of 99.05%, an average precision of 99%, a recall rate of 99%, and an F-1 score of 99%.