Taxonomy of Induction-Motor Mechanical-Fault Based on Time-Domain Vibration Signals by Multiclass SVM Classifiers

Intell Ind Syst (2016) 2:269–281 DOI 10.1007/s40903-016-0053-x ORIGINAL PAPER Taxonomy of Induction-Motor Mechanical-Fault Based on Time-Domain Vibration Signals by Multiclass SVM Classifiers Purushottam Gangsar1 · Rajiv Tiwari1 Received: 10 July 2016 / Revised: 18 August 2016 / Accepted: 19 August 2016 / Published online: 3 October 2016 © Springer Science+Business Media Singapore 2016 Abstract In the present work, faults in induction motors (IM) have been diagnosed by multiclass support vector machine (SVM) algorithms based on time domain vibration signals. The main focus is to classify mechanical faults of induction motors, i.e. the bearing fault, unbalanced rotor, bowed rotor and rotor misalignment at different rotational speeds and diverse loading conditions. In this work, an induction motor test setup was used to generate vibration signals of seeded mechanical faults. For the effective fault diagnosis, one-versus-one multiclass SVM approach with the Gaussianradial basis function (RBF) kernel has been used. For the fault classification, firstly optimum statistical features from higher statistical moments have been selected. Also the selection of SVM kernel parameters, numbers of feature datasets and optimum ratio of training-to-testing data have been performed. The SVM classifier is trained and tested at the same rotational speeds as the measured data as well as innovatively tested at intermediate rotational speeds for which measured data was not available. It is observed that classification accuracy gradually increases with the increase of the rotational speed and with the increase of the load on the IM. Keywords Induction motor (IM) · Mechanical faults · Multi-fault classification · Support vector machine (SVM) · Radial basis kernel (RBF) · One-versus-one (OVO) approach B 1 Rajiv Tiwari Department of Mechanical Engineering, Indian Institute of Technology Guwahati, Guwahati, Assam 781 039, India Introduction Induction motors or squirrel-cage motors are nowadays commonly used prime movers in a wide field of light and heavy duty rotating machine applications in all types of industry. This is due to their simple construction, handiness, reliability, low maintenance requirements, low cost, adaptability to wide variety of operation conditions and availability of power converters using adequate control approach. However, induction motors are prone to many types of faults in industry that causes a complete motor failure, which in turn may cause huge production losses in terms of the production cost, production time, wastage of raw material and also chances of human injuries. Induction motors usually falls out of service not just due to the age or operating hours but due to the unusual mechanical load, power supply deviation, improper or inadequate lubrications, more heat and ineffectual sealing. Hence, the early detection of faults in induction motors is very important in order to prevent the complete failure of motor and unexpected huge production losses in industry [1]. In recent years, many condition based monitoring (CBM) techniques have been employed for early detection of faults in IM, i.e. machine current signature analysis (MCSA), vibration, air gap torque, acoustic noise measurement, instantaneous angular speed and temperature measurement [2]. In order to find the most effective CBM techniques for the detection of various faults in IM, Timusk et al. [3] compared the vibration CBM with the MCSA for the detection of broken rotor bar (BRB) and bearing faults for different load conditions and speeds, and finally concluded that the vibration based CBM was the most effective technique for the bearing fault detection while the MCSA was the best for the BRB detection. Kral et al. [4] suggested that the vibration based CBM technique was reliable technique for the detection of bearing and other mechanical faults in the IM. Many 123 270 researchers have been suggested to use vibration signal for condition monitoring and fault diagnosis of mechanical as well as electrical faults in IM, due to its easy measurability, high accuracy and effectiveness in signal analysis that represent the actual machine condition, among others. Condition monitoring of machines can be performed through signal analysis in time, frequency and time–frequency domain [5]. Various artificial intelligence techniques have been successfully used for the fault taxonomy in IM, such as artificial neural network (ANN), fuzzy logic, principal component analysis (PCA) and linear discriminant analysis (LDA), classification and regression tree (CART), and immune genetic system [6,7]. A relatively new AI method, SVM is also used as an artificial intelligence technique for the fault detection of IM [5,8]. In order to compare these AI methods for fault diagnosis, Samanta [9] performed the gear fault detection using ANN and SVM with genetic algorithm, and concluded that the SVM can perform well in comparison with ANN even with smaller number of samples and also training time is less in case of the SVM. Silva and Pederiva [10] performed the induction motor fault detection using artificial intelligence techniques, like SVM, fuzzy logic and ANN, and concluded that the SVM has a good generalization, among others. Nowadays, SVM has gained popularity for fault diagnosis over other methods, due to its generalization capabilities and computational cost. The SVM has been extended to multiclass-classification from binary classification and also various kernels have been developed to handle nonlinear problem through the SVM. In order to compare multiclass SVM techniques, Hsu and Lin [11] and Hsu et al. [12] presented a comparison among different methods of multiclass SVM through different kernels, and concluded that oneversus-one method is most effective for the classification, among one-versus-one (OVO), one-versus-all (OVA) and direct-acyclic graph SVM (DAGS). RBF kernels produced better result among other kernel such as, linear, polynomial and sigmoid. For effective fault classification through the SVM, data preparation, i.e. feature extraction and selection, is a critical step. Feature extraction methods such as principal component analysis (PCA) and independent component analysis (ICA) and feature selection techniques such as the genetic algorithm (GA) and decision tree have been introduced in the literature. In order to evaluate the work related to fault diagnosis using SVM, Tiwari and Bordoloi [13] performed multi-fault classification of gears based on the SVM. Genetic algorithm, grid-search method and artificial bee colony algorithm were used for optimizing SVM parameters. Li et al. [14] performed fault diagnosis of rolling element bearings by the SVM. Improved ant colony optimization (IACO) algorithm was used for optimization of the SVM parameter. Baccarini et al. [8] presented a practical industrial application of the SVM for mechanical faults diagnostic of IMs based 123 Intell Ind Syst (2016) 2:269–28 (...truncated)