Feature-based performance of SVM and KNN classifiers for diagnosis of rolling element bearing faults

1. Introduction

Fault detection and diagnosis of rotating machinery play an essential role in maintenance planning, human safety, and cost reduction in modern industrial systems. As the rolling element bearings are an integral component of rotating machinery, their failure is a leading cause of machinery malfunction [1]. Bearing failures account for 30 to 40 percent of total machinery failure. These defects, if detected early enough, can help prevent accidents [2]. Various methods for diagnosing bearing faults have been employed, such as acoustic emission [3], vibration [4], motor current [5], thermography [6] and, so on. The microphone sensor of a mobile phone is used to record acoustic data and investigate bearing health issues [3]. However, the approach is limited by the weak frequency response of the integrated microphone in low-frequency bands, which is especially problematic for low voltage motors. An adaptive noise canceling method is proposed for diagnosing bearing faults in induction motors [7]. Vibration-based diagnostics is the most extensively utilized technique for early failure identification in induction motors among several diagnostic methods [8-9]. An artificial neural network (ANN) was used to estimate the bearing condition [8]. It has been discovered that successful bearing diagnosis can be achieved by applying suitable measurement and processing of motor vibration signals. Bearing defect detection based on Hidden Markov Modeling (HMM) using vibration signal is proposed in reference [10]. An amplitude demodulated signal is used for feature extraction and training HMMs to estimate normal and faulty bearings. Envelope analysis, often known as High-Frequency Resonance Technique (HFRT), is the most widely used frequency-domain approach for bearing defect diagnostics [11]. Due to mechanical components, however, the technology suffers from a low signal-to-noise ratio and the existence of a high number of frequencies. Furthermore, the procedure necessitates the determination of bearing defect frequencies in advance.

Many research findings support the machine learning approach in machinery fault diagnosis as the ML methods are more competitive than signal-based methods [12-15]. Machine learning characteristics collected from data are more objective than signal-based methods. Furthermore, the accuracy criterion of fault diagnosis is more helpful in selecting fault diagnosis methods. In the current work, vibration signals from a bearing are gathered under both healthy and faulty conditions. Various time and frequency-domain features are extracted from the data and are used to distinguish different bearing conditions using the SVM and KNN classifiers.

1.2. K-Nearest Neighbor (KNN) algorithm

The K-Nearest Neighbor (KNN) algorithm is one of the most basic machine-learning algorithms [13]. It is a method of calculating the distance between two points [14]. Due to its simplicity and ease of implementation, this is a widely used classifier. It is a non-parametric classification and regression method. As seen in Fig. 2, this algorithm assumes that related entities are close to one another.

Fig. 1Different hyperplanes separating the two classes of data points

Fig. 2A demonstration of KNN algorithm

KNN determines the distance between two points using multiple techniques, such as Euclidian and Manhattan [15], based on the idea of similarity based on proximity or distance. However, Euclidian is the most generally used among all of these ways, and it can be represented by Eq. (1):

1

$d(m,b)=\sqrt{\sum _{i=1}^n(m_i-b_i{)}^2}$

where $m$, $b$ are two points in an $n$ dimensional Euclidian space.

KNN must be run several times with different values of $K$ to determine the chosen number of $K$ (the number of neighbors) for a given dataset. The value of $K$ should be selected to decrease the number of errors when making predictions from each run.

2. Bearing data description

Fig. 3 shows the CWRU test platform, including a 2-horsepower motor, dynamometer, torque sensor, and electronic control unit [16]. Experiments are performed with bearings having EDM-created single point defects with diameters of 7 mils, 14 mils, 21 mils, 28 mils, and 40 mils (1 mil = 0.001 inches). SKF bearings are used for the 7, 14, and 21 mils diameter faults, whereas NTN bearings are used for the 28 mils and 40 mils diameter faults.

Vibration data was acquired using accelerometers installed at the 12 o'clock position of the motor casing and processed in MATLAB (.mat) format. Signals were captured at a sampling frequency of 12 kHz for drive and fan end bearings and 48 kHz for drive end bearing faults.

Fig. 3Test stand of CWRU bearing data center [16]

3. Procedure

The step-wise methodology employed in the present work is presented in this section. First of all, the bearing data corresponding to fault categories of interest is ensembled. Numerous time-domain and frequency-domain features are then extracted from this ensembled data and also ranked. Finally, these features are selected in different combinations to train ML models to obtain their fault classification accuracy. The methodology’s workflow is depicted in Fig. 4.

Fig. 4The methodology used for bearing fault classification

4. Results and discussion

After the features are extracted and the models are trained, the accuracy results of fault classification are obtained and presented in terms of scatter plot and confusion matrix with the selection of different features. The predictors that may distinguish the classes can be determined by plotting several predictors on the scatter plot. A scatter plot depicts the data before training the classifier, and the model prediction results are displayed once the classifier has been trained. RMS and skewness are considered as predictors in the present case. A confusion matrix can be used to detect the areas where the classifier has failed. The True Positive Rate (TPR) is defined as the percentage of correctly classified observations per true class. The False Negative Rate (FNR) is the proportion of observations that are wrongly categorized per true class. In the last two columns on the right, the plot gives summaries for each true class.

To train ML models, features both from the time-domain and frequency-domain are selected in the first trial. A combination of three non-linear time-domain features, namely, approximate entropy, correlation dimension, and Lyapunov exponent, are considered in the second feature set. In the last group, three frequency-domain features are used for model training. Figs. 5-7 show the scatter plots and confusion matrices for a few results of fault classification of SVM and KNN classifiers with three different feature sets. The corresponding accuracies obtained are reported in Table 2.

Fig. 5Scatter plot and confusion matrix for SVM classifier with the feature set 1

Fig. 6Scatter plot and confusion matrix for KNN classifier with the feature set 2

Fig. 7Scatter plot and confusion matrix for KNN classifier with the feature set 3

5. Conclusions

The current research, which employs machine learning techniques, demonstrates that the specific selection of fault features plays a significant role in training machine learning models for bearing defect classification. The set of frequency-domain features has the best performance in both SVM and KNN classifiers among the three feature sets. The fault classification accuracy is the lowest when just non-linear time-domain features are used for ML model training. However, the accuracy increases significantly with a combination of features from time-domain and frequency-domain. With only frequency-domain features, the accuracy further improves. This trend can be observed in both the SVM and the KNN classifiers. Also, in all three cases, KNN outperforms SVM. As a result, frequency-domain is more supportive in identifying rolling element bearing defects in terms of fault features, and the KNN model beats the SVM model.