An improved feature extraction method using texture analysis with LBP for bearing fault diagnosis

Highlights

• In this study, a novel approach based on Texture analysis with LBP is proposed to feature extraction.

• One advantage is that this method uses all data points for feature extraction.

• It is fast and can be used in real-time application.

• High accuracies are achieved for bearing fault classification.

• Original data is used.

Abstract

Bearings are one of the most widespread components used for energy transformation in machines. Mechanical wear and faulty bearings reduce the efficiency of rotating machines and thus increase energy consumption. The feature extraction process is an essential part of fault diagnosis in bearings. In order to diagnose the fault caused by the bearing correctly, it is necessary to determine an effective feature extraction method that best describes the fault.

In this study, a new approach based on texture analysis is proposed for diagnosing bearing vibration signals. Bearing vibration signals were first converted to gray scale images. It can be understood from the images that the signals of different bearing failures form different textures. Then, using these images, LBP (Local Binary Pattern) and texture features were obtained. Using these features, different machine learning models and bearing vibration signals are classified. Three different data sets were created to test the proposed approach. For the first data set, the signals composed of very close velocities were classified. 95.9% success rate was observed for the first data set. The second data set consists of faulty signals at different parts of the bearing (inner ring, outer ring and ball) measured in the same RPM. The type of fault has been determined, and a 100% success rate was obtained for this data set. The final data set is composed of the fault size dimensions (mm) of different ratios. With the proposed approach, a 100% success rate was obtained in the classification of these signals. As a result, it was observed that the obtained feature had promising results for three different data types and was more successful than the traditional methods.

Introduction

In the industrial automation systems of recent years, machine motion is generally provided by rotational force. Bearings are a mechanical component commonly used in motor systems that perform this rotational motion and are used to reduce friction. Early detection and diagnosis of rotating machines, deteriorating condition, low efficiency and avoidance of unexpected failures are becoming increasingly important in these systems. The main reasons for the failure of rotating machines are generally due to bearing faults. For example, metal bearing failures in induction motors constitute 40% of the faults in the system [1]. Therefore, several techniques have been developed for the health monitoring of bearings to prevent such failures early. Apart from these techniques, vibration-based fault analysis has proved to be more advantageous in revealing bearing failure. Furthermore, it is impossible to prevent wear due to the constant friction of the mechanical components. For this reason, a condition monitoring based on bearing diagnostics should be applied to rotating machines in automation systems [2]. When the current literature is examined, the methods based on vibration analysis and current analysis can be seen to be the most applied fault monitoring methods. The data obtained in these studies are analyzed by methods such as time–space [3], [4], [5], frequency space [6], [7], time–frequency space [8], [9], [10] and then supported by methods such as artificial intelligence techniques [11], [12], [13], [14], [15]. Nowadays, in order to define bearing failures better, more studies on improving the method of feature extraction are started. Van and Kang developed a new feature extraction technique based on non-local means de-noising of non-local means and empirical mode decomposition in order to obtain more accurate error information in the feature extraction step. Later, during the feature selection phase, the hybrid distance evaluation technique (DET) was combined with utilizing the advantages of particle swarm optimization models. As a result, a comparison was made between three popular classifier types: K-nearest neighbors, probabilistic neural network, and support vector machine classifiers are employed as the classifier. They obtained the classification with the success of 98.58% with DET-PSO-KNN model [16]. Zhang et al. proposed a new intelligent diagnostic method that can automatically learn the bearing fault characteristics. To perform automatic feature extraction, they developed a subset based deep auto-encoder (SBTDA.) In order to obtain the appropriate configuration, they optimized several key parameters with the particle swarm optimization (PSO) algorithm. They used three publicly available data sets with the proposed method and stated that they achieved superior success compared to other studies with 99.65%, 99.66% and 99.60% mean test accuracy, respectively [17]. Han et al. have proposed a new method for fault diagnosis using the Complementary Ensemble Empirical Mode Decomposition (CEEMD) method as feature extraction and Teager energy operator for signal enhancement. They achieved better characteristic profile by using the Teager energy operator and the CEEMD combination. The authors stated that the application of the model with simulation data and experimental data was successful in distinguishing weak error characteristics from noise [18]. Li et al. proposed a new signal processing scheme for the fault detection of the train axle bearings based on the multi-scale morphological filter (MMF) feature selection. They stated that more than 30 vibration signals were calculated for the axle bearings with different conditions and the features which can reflect the fault characteristics more effectively and representative were selected by using the maximum relevance and minimum redundancy principle. In the experimental results, they showed that the method they proposed had superior performance in extracting fault characteristics of faulty train axle bearings. Also, they compared MMF with MMF based on kurtosis criteria and MMF based on spectral kurtosis criteria. Based on the proposed feature selection, the MMF method observed that these two methods left behind in the detection of train axle bearing failures [19]. Zarei et al. propose an intelligent method based on artificial neural networks (ANN) to detect bearing failures of induction motors. In this method, the vibration signal has first passed the removing non-bearing fault component (RNFC) filter to eliminate the fault components that are not caused by the bearing, and then they performed fault classification using a second neural network using pattern recognition techniques. They suggested that the proposed method in the three-phase asynchronous motor results in a confirmation of its ability despite the low quality (noisy) of the vibration signal measured in fault detection [20]. Ahmed et al. stated that they submitted an original article based on a compressive sampling (CS) using the multiple measurement vector (MMV) and feature order for bearing fault classification. The MMV-based CS used a large amount of bearings to reduce the vibration signal by generating sampled signals from the raw vibration signals. They tested the model for the classification of bearing failures by three classification algorithms (LRC, ANN and SVM). They stated that they achieved a high degree of performance [21].

The feature extraction step is the most crucial part of fault diagnosis in bearings. In order to diagnose the fault caused by the bearing correctly, it is necessary to determine an effective feature extraction method that best describes the fault. In recent trends, deep learning methods have been preferred intensively for feature extraction besides traditional methods [22], [23], [24]. This study aims to propose a new feature extraction approach in the classification of bearing vibration signals. Therefore, in this study, an approach based on texture analysis for bearing vibration signals is presented. Firstly, bearing vibration signals were transformed to gray scale images. Then, LBP (Local Binary Pattern) is used to obtain texture features. In the final stage, different machine learning models and actual bearing vibration signals are classified to test the success of the system validation. The most important advantages of the proposed method are; the first is that it uses all the values on the vibration signal, and the second one is that it is simple and real-time applications. According to the results, it was observed that the proposed method provides stable features for classification of vibration signals.

In this study, in the first part, literature information about the study subject and general information about bearing vibration signals classification are presented. In the second part, information about the data set used in the study is given. In the third part, proposed feature extraction approach is explained. In the fourth section, the results obtained are given in detail. In the fifth section, the results are discussed.