Abstract
Induction motors are widely used in the industry and failures in motors are very common. One of the faults is the broken rotor bar which is caused by several factors including mechanical stress, dynamic stress, thermal stress, environmental and residual stress. If go undetected, broken rotor bar fault can cause catastrophic failures in motors, and this can even cause harm to other components connected to the motor. These kind of failures affects the production which results in economic losses hence it is important to monitor the induction motors during production to avoid these kinds of losses. Broken rotor bar fault detection is critical in three phase induction motors for reliability of the machine in the industry. Also, early detection will save the motor and makes it possible for repairs. Although traditional diagnostic methods are effective, they lack adaptability as well as precision that is necessary for early detection of fault.
This research focuses on the performance evaluation of machine learning methods in detection of broken rotor bar fault using Motor Current Signature Analysis (MCSA). The Decision Tree Algorithm (DTA), Artificial Neural Network (ANN), Deep Learning (DL), and a hybrid model, are developed then applied in detection of broken rotor bar fault and studied to evaluate their performance. The intention of the research was to identify and recommend which machine learning method is best suitable for fault detection in induction motors so that these catastrophic failures in induction motors can at least be minimized.
In this study, the training data which is the motor current signature data for healthy and faulty conditions was collected through experimental measurement, it was then used for training of the machine learning models. For feature extraction of the frequency spectrum of the current signal, Fast Fourier Transform (FFT) was used.
IV
Based on the findings of this study, the hybrid model managed to get better accuracy on average than the other three models, the DTA had better accuracy in detection of broken rotor bar fault than ANN and DL however, they all were close. Although they all were close, the DTA showed better computational speed than the hybrid model, ANN and DL. 5000 samples of current signature were used for feature extraction and after feature extraction and pruning of data, it reduced to approximately 1000 samples. The results showed that the hybrid model followed by the DTA are more effective for detection of broken rotor bars when limited data is used for training when compared to ANN and DL.
To sum up, machine learning presents a promising approach to broken rotor bar fault detection and the hybrid model has proven to be the most effective method. The individual operation requirements such as model interpretability and amount of data available for training should be taken into account when selecting a technique to use.