Condition monitoring and automatic fault diagnosis of induction motor using non stationary signals

Abstract

newline Unlike traditional stationary techniques, non-stationary signal analysis can effectively capture the dynamic behavior of induction motors under varying operating conditions such as transient states and load fluctuations. This makes it particularly suitable for real-time monitoring, as it can detect faults even during motor startup, acceleration, or other non-steady-state operations, which are often missed by stationary methods. Non-stationary techniques like time-frequency analysis are capable of identifying early-stage faults by analyzing transient signals and specific frequency variations. By incorporating advanced machine learning algorithms for automatic fault diagnosis, the system becomes capable of analyzing complex data patterns on its own, making the diagnostic process more efficient and significantly less dependent on human intervention. This approach enhances motor reliability, minimizes downtime, reduces maintenance costs, and enables predictive maintenance strategies. It also contributes to extending the life of the motor and optimizing its overall operational performance. newline newlineThis thesis utilizes both vibration and acoustic signal data for bearing fault detection, and current signal data recorded in a controlled laboratory environment for diagnosing rotor faults. Advanced machine learning techniques have been employed to automate the fault diagnosis process, improving reliability, reducing maintenance expenses, and enabling predictive maintenance applications. newline newlineThe detection and classification of bearing faults in induction motors have been addressed using advanced signal processing and Machine Learning (ML) techniques through two distinct approaches. In the first approach, a two-level Wavelet Packet Transform (WPT) is applied for signal filtering to extract meaningful vibration features. From both original and wavelet-decomposed signals, eleven statistical features are computed. Feature selection is performed using ANOVA F-test and Mutual Information methods. These selected features are t