Electrical motors are the most widely used electrical machines in the industry. Induction motors (IM) are the most popular electrical motors in industrial applications. IMs are preferred due to their low cost, high reliability, and simplicity. However, IMs are not free from fault despite their advantages. The faults of the IM can be broadly classified into two groups mechanical and electrical. Mechanical faults of IMs include air gap eccentricity, misalignment, bearing faults, and gearbox-related faults. The electrical faults are related to the stator and rotor. Stator-related faults consist of open-circuit, short-circuit and insulation failure, whereas rotor-related faults consist of broken rotor bars and end rings. The distribution of the faults of IMs is given in Fig. 1 [1].

Unexpected faults of IMs can lead to interruptions in production lines, significant financial losses and reduced revenues. The accurate detection and diagnosis of faults of IMs at an early stage prevent undesired results and long downtimes [2].

The fault diagnosis methods of induction motor can be classified as knowledge-based, model-based, and signature-extraction-based. The knowledge-based methods are based on machine learning models. The mathematical models of induction motors are used in the model-based methods, whereas the signature-extraction-based method extracts relevant signatures from monitoring signals. The most widely utilized monitoring signals for the fault detection of induction motors include stator current, voltage, air-gap torque, vibration, angular speed, instantaneous power, and magnetic flux signals.

The monitoring signals can be processed in the time domain, frequency [3, 4], or time-frequency [5–8] domains to detect the faults of induction motors. Scalar indices such as root-mean-square (rms), crest factor, kurtosis, spectral kurtosis, skewness, peak value, peak-to-peak value, shape factor, impulse factor, and clearance factor are employed in the time-domain analysis to assess the health of IMs [9]. The frequency domain methods provide successful results in the analysis of stationary signals. Still, they are ineffective in analysing non-stationary signals where the spectrum and period of the signals change. Time-frequency analysis methods are preferred in the analysis of non-stationary signals. The commonly used signal processing methods are the Fast Fourier Transform (FFT) [10], Wavelet Transform (WT) [11], Empirical Mode Decomposition (EMD) [12, 13], Discrete Wavelet Transform (DWT) [14, 15], and Wavelet Packet Transform (WPT) [16] and their variants.

In recent years, machine learning (ML) based fault detection methods have overcome the limitations above. An ML process operates in three stages: 1) It takes the collected data, 2) Finds patterns in the collected data, and 3) Predicts the new patterns [17]. Various ML algorithm-based methods provide early detection and prevent faults [17, 20]. An experimental comparison of various ML-based techniques for detecting rotor faults of induction motors is presented in [18]. The features are extracted from the current signals. To verify the ML-based classifies, stator current signals of IM are tested for different operation conditions. A comprehensive review of condition monitoring of induction motors based on ML methods is presented in [19]. The widely used ML-based methods are principal component analysis (PCA) [21], artificial neural networks (ANN) [22–24], support vector machines (SVM) [25], decision trees (DT) [26], k-nearest neighbors (k-NN), singular value decomposition (SVD) [27], and random forest (RF) methods [28].

Due to the short air-gap distance of IMs, the changes in the air gap are more important than in other machines. Therefore, IMs' diagnosis and detection of the eccentricity faults (which affect the flux distribution in the air gap and may lead to mechanical stress) are important. Various diagnostic methods have been proposed to detect eccentricity faults. This study compares the performance of four ML methods, including k-NN, DT, SVM, and RF, in detecting the eccentricity faults of induction motors based on three-axis vibration signals. One of the most crucial issues for fault detection methods of IMs is determining the proper characteristic features. The purpose of the study is to compare the performances of the ML methods and determine the most effective features with the highest accuracy in the detection of eccentricity fault. The paper is organized as follows: Section 2 briefly reviews the classification methods and statistical features. Sections 3 and 4 describe the implementation of the eccentricity fault and data collection system. The results and comparison of the ML methods and determination of the most effective statistical feature are discussed in Section 5. The results are concluded in Section 6.