Vibration based fault diagnosis of monoblock centrifugal pump using decision tree

Abstract

Monoblock centrifugal pumps are widely used in a variety of applications. In many applications the role of monoblock centrifugal pump is critical and condition monitoring is essential. Vibration based continuous monitoring and analysis using machine learning approaches are gaining momentum. Particularly artificial neural networks, fuzzy logic were employed for continuous monitoring and fault diagnosis. This paper presents the use of C4.5 decision tree algorithm for fault diagnosis through statistical feature extracted from vibration signals of good and faulty conditions.

Introduction

Centrifugal pump plays an important role in industries and it requires continuous monitoring to increase the availability of the pump. The pumps are the key elements in food industry, waste water treatment plants, agriculture, oil and gas industry, paper and pulp industry, etc. In a monoblock centrifugal pump, bearing, seal and impeller are the critical components that directly affect the desired pump characteristics. In a monoblock centrifugal pump, defective bearing, defective seal, defect on the impeller and cavitation cause more number of serious problems such as abnormal noise, leakage, high vibration, etc. Cavitation can cause more undesirable effects, such as deterioration of the hydraulic performance (drop in head-capacity and efficiency), damage of the pump by pitting, erosion and structural vibration (Alfayez, Mba, & Dyson, 2005). Vibration signals are widely used in condition monitoring of centrifugal pumps (Peck & Burrows, 1994). Fault detection is achieved by comparing the signals of monoblock centrifugal pump running under normal and faulty conditions. The faults considered in this study are bearing fault (BF), seal fault (SF), impeller fault (IF), bearing and impeller fault (BFIF) together and cavitation (CAV). In conventional condition monitoring, the commonly used method is vibration analysis in frequency domain through Fast Fourier Transform (FFT). Seismic or piezo-electric transducers are used to measure the vibration. The level of vibration can be compared with historical baseline value to assess the severity. Interpreting the vibration signal is a complex process that requires specialized training and experience. Commonly used technique is to examine the individual frequencies present in the signal. These frequencies correspond to certain mechanical component or certain malfunction. By examining these frequencies and their harmonics, the analyst can identify the location, type of problem and the root cause as well (Cempel, 1988). Nowadays, the application of machine learning for fault diagnosis is more common as an alternative to conventional methods. It is largely due to increased availability of computational resources and development in algorithms. Also for complex systems involving many components, it is difficult to compute characteristic fault frequencies. Even if characteristic frequencies are available the vibration signals are highly non-stationary in nature and FFT based methods are suited only for stationery processes. In the machine learning approach the data acquisition system is used to capture the vibration signals (Kong and Chen, 2004, McFadden and Smith, 1984, Rajakarunakaran et al., 2008). From the vibration signal relevant features can be extracted and classified using a classifier (Koo and Kim, 2000, Sanz et al., 2007, Wang and Hu, 2006).

Alfayez et al. (2005) presented where acoustic emission has been applied for detecting incipient cavitation and determining the best efficiency point (BEP) of a centrifugal pump are based on net positive suction head (NPSH) and performance tests. Peck and Burrows (1994) discussed rule based expert system using historical vibration data monitored from compressors, pumps and electric motors and a heuristic artificial neural network system were designed and evaluated to extract and identify useful patterns and trends in the vibration signals. Cempel (1988) presented vibroacoustical (VA) diagnostic in technical diagnostics and at different stages of the machinery lifetime. Vibroacoustical (VA) diagnostic consists of two parts. The first part was the creation or choice of a proper VA symptom to a given fault and the establishment of the most effective type of “condition-symptom” relation for a given case. The second part was condition recognition and forecasting where the confidence level depends on the applied model and the inference method.

Wang and Chen (2007) proposed the synthetic detection index with fuzzy neural network to evaluate the sensitivity of non dimensional symptom parameters for detecting faults in centrifugal pump. Rajakarunakaran et al. (2008) developed a model for the fault detection of centrifugal pumping system using two different artificial neural network approaches, namely feed forward network with back propagation algorithm and binary adaptive resonance network (ART1). The performance of the developed back propagation and ART1 model were tested for a total of seven categories of faults in the centrifugal pumping system. Classification accuracy of 99.3% was achieved. Wang and Hu (2006) used fuzzy logic principle as classifier with the features extracted from the vibration signals of the pump.

Kong and Chen (2004) proposed a new combined diagnostic system for triplex pump based on wavelet transform, fuzzy logic, neural network. The developed diagnostic system consists of four parts. The first part was wavelet transform in which multi resolution analysis was employed. The second part was for asymptotic spectrum estimation of the characteristic variable. The third part was employed for characteristic variable fuzzified in simulating fuzzy inference using incomplete information. The fourth part was the neural network trained with fuzzified characteristic variable for triplex pump failure diagnosis. Yuan and Chu (2006) discussed the fault diagnosis based on support vector machine. It is binary tree classifier composed of several two class classifiers. The effectiveness of the method is verified by the application to the fault diagnosis for turbo rotor pump. Zhang, Asakura, Xu, and Xu (2003) introduced a fault diagnosis system using fuzzy neural network based on the series of standard fault pattern pairings between fault symptoms and fault. Fuzzy neural networks were trained to memorize these standard pattern pairs and it adopts bidirectional association to produce 97.3% classification accuracy (see Fig. 1).

Sanz et al. (2007) presented a technique for monitoring the condition of rotating machinery from vibration analysis that combines the capability of wavelet transform to treat the transient signals with the ability of auto associative neural networks to extract features of datasets in an unsupervised mode. Trained and configured networks with wavelet transform coefficients of non faulty signals are used as a method to detect the novelties or anomalies of faulty signals. Koo and Kim (2000) introduced wigner distribution for analyzing vibration signals and developed an expert system for vibration monitoring and diagnostics for rotating machines using back propagation neural network. A classification accuracy of 81.25% was achieved in finding abnormalities of pump. Yuan and Chu (2007) presented a method that jointly optimises the feature selection and support vector machine parameters. A hybrid vector that describes both the fault features and the support vector machine parameters was taken as the constraint condition. This new method can select the best fault features in a shorter time and improve the performance of support vector machine classifier.

Artificial neural network gives 97.3% classification accuracy. It is a very good result. However, training of an artificial neural network classifier is complex and time consuming. The robustness and effectiveness of fuzzy classifier depends on the rules suggested by the experts or algorithms. Support vector machines also perform very well in classification. Feature selection has to be carried out through some other algorithms. It increases the computational time.

To overcome the above difficulties, researchers are having constant lookout for a classifier which will give very high classification accuracy with simple operation and does feature selection and classification simultaneously. C4.5 decision tree algorithm seems to be an algorithm which satisfies these conditions and it is being used in many applications. To quote a few, Sun, Chen, and Li (2007) used principal component analysis (feature selection), C4.5 decision tree and back propagation neural network (classification) to fault diagnosis of rotating machinery such as turbines and compressors. From the result, C4.5 and PCA-based diagnosis method has higher accuracy and needs less training time than back propagation neural network BPNN. Sugumaran, Muralidharan, and Ramachandran (2007) illustrated the use of a decision tree that identified the best features from a given set of samples for classification. They used Proximal Support Vector Machine (PSVM), which has the capability to efficiently classify the faults using statistical features. The classification results of PSVM and SVM were compared and found that the classification efficiency of SVM was a little less than that of PSVM.

Polat and Günes (2009) proposed a novel hybrid classification system based on C4.5 decision tree classifier and one-against-all approach to classify the multi-class problems including dermatology, image segmentation and lymphography. Firstly C4.5 decision tree has been used and achieved 84.48%, 88.79%, and 80.11% classification accuracies for dermatology, image segmentation, and lymphography datasets, respectively. The proposed method based on C4.5 decision tree classifier and one-against-all approach obtained 96.71%, 95.18%, and 87.95% for above datasets, respectively. Sugumaran and Ramachandran (2007) discussed condition monitoring of roller bearing using decision tree. Statistical features like minimum value, standard error and kurtosis, etc., were extracted from vibration signals there from a rule set is formed for fuzzy classifier. This paper presented the use of decision tree to generate the rules automatically from the feature set.

Data mining has been successfully applied to medical field such as dermatology, image segmentation and lymphography (Polat & Günes, 2009). Some data mining algorithms have also been applied to fault diagnosis of machines and induction motors (Tran, Yang, Oh, & Tan, 2009). For example, wavelet transform technique was used in fault diagnosis of rotating machinery (Chen and Mo, 2004, Kong and Chen, 2004, Sanz et al., 2007). SVM, Fuzzy logic and neural network are widely used as classifiers. Walsh transform and SVM were used in the fault diagnosis of shaft (Xiang, Zhou, An, Peng, & Yang, 2008). Genetic programming was used in condition monitoring to detect the fault in rotating machinery (Zhang, Jack, & Nandi, 2005). Decision tree algorithm was used as a classifier to detect the fault in rotating machinery, induction motor, bearing, shaft and gears (Sugumaran et al., 2007, Sugumaran and Ramachandran, 2007, Sun et al., 2007). This is the only algorithm which can do both feature extraction and classification simultaneously. Therefore, the decision tree C4.5 algorithm is used in this paper for fault diagnosis of monoblock centrifugal pump.

The rest of the paper is organised as follows. In Section 2, experimental setup and experimental procedure is described. Section 3 presents feature extraction from the time domain signal. In Section 4, feature selection using decision tree is discussed. Section 5 describes the training of the classifier and the classification accuracy is tested and subsequently Section 6 presents results of the experiment. Conclusions are presented in the final section.