Performing real-time fault diagnosis with virtual sensor and digital twin methods

In this thesis study, innovative approaches were developed to improve the fault diagnosis processes of asynchronous motors, integrating modern technologies such as virtual sensors, digital twins, surrogate learning, and augmented reality in a comprehensive structure. The study stands out particularly for modeling physical vibration signals from physical current signals, thereby creating a virtual vibration sensor. A combination of theoretical and experimental studies was conducted, offering a comprehensive solution for proactive and predictive maintenance processes. In the theoretical studies, machine learning, standalone deep learning, Ensemble Learning with Virtual Sensor Model, and surrogate learning approaches were utilized to develop fault diagnosis models on pre-existing datasets. Virtual vibration sensors were created by modeling raw physical current data (Ia, Ib, Ic) into raw vibration signals. For all other classification approaches, classical data processing techniques such as Fourier Transform, Windowing, and Band Power Analysis were applied to extract meaningful features from the data. In this phase, virtual vibration sensors achieved the best performance with 97.86% accuracy, while the RBF surrogate model demonstrated significant success with 92.78% accuracy and 93.33% precision. The Ensemble Learning with Virtual Sensor Model achieved an accuracy of 91.11%, effectively combining the strengths of different models. Standalone deep learning models, on the other hand, demonstrated solid performance in sequential data analysis with 88.89% accuracy.In the experimental studies conducted in a laboratory setting, real-time data acquisition systems collected three-phase current and three-axis vibration signals from asynchronous motors. These data were used to test the performance of all models trained during the theoretical studies. In the experimental studies, virtual vibration sensors demonstrated strong performance with 96.35% accuracy, while the Ensemble Learning with Virtual Sensor Model achieved 92.06% accuracy and 92.08% precision, yielding successful results. The RBF surrogate model achieved the best performance with 97.78% accuracy in experimental studies. Additionally, digital twins were developed, and augmented reality (AR) visualization was provided. The digital twin was continuously updated with sensor data, enabling users to visually monitor motor faults. The AR system accelerated maintenance processes by allowing real-time visualization of fault regions, making it possible to identify issues quickly and accurately. This system contributed to reducing unplanned downtime by enabling fast and precise fault diagnosis. This thesis study offers a fast, cost-effective, proactive, and user-friendly system for asynchronous motor fault diagnosis by modeling physical vibration signals from physical current signals and creating a virtual vibration sensor. The developed methods provide significant solutions in terms of accuracy, precision, and cost efficiency in industrial maintenance processes and contribute substantially to the digital transformation of industrial maintenance with digital twin and augmented reality technologies.