A non-invasive deep learning-based mobile platform application for motor fault detection

Electric motor faults result in performance loss and disruptions in industrial processes, reducing operational efficiency and increasing maintenance costs. These faults typically require maintenance or repair efforts that can be expensive and time-consuming. Therefore, it is crucial for the manufacturing sector to minimize faults and fault risks and to ensure that motors operate reliably and efficiently. In this thesis, a deep learning based fault detection system is proposed for motor fault classification without the need for any sensor connections into or onto the motor assembly. It is especially focused on the electrical and mechanical faults of the asynchronous motor, one of the most used motors in the industry. The mobile application platform, which can capture vibration data using the three-axis accelerometer sensor in smartphones, was developed using Flutter. The deep models were trained using vibration data captured in three axes (x, y, z) by placing a smartphone with the application installed onto the electric motor for which fault detection is desired. Subsequently, test data were used to detect the fault status and class of the electric motor. Different scenarios and conditions were created to test the adequacy of accelerometer sensors in smartphones for detecting electric motor faults. Fault classification with vibration data obtained under various motor operating speeds in both loaded and unloaded conditions achieved accuracy rates exceeding 90%. The most significant original contribution of the study is the implementation of an application that enables electric motor fault classification without the need for any internal or external motor connections and with the use of minimal hardware resources. The system, which performs AI-supported electric motor fault detection with a mobile application, represents a fast and cost-effective approach by reducing the need for expert knowledge to a minimum, thanks to the use of deep learning architectures and the absence of additional sensor connections.