ARTIFICIAL NEURAL NETWORK BASED CLASSIFICATION FOR MONOBLOCK CENTRIFUGAL PUMP USING WAVELET ANALYSIS

Transcription

1 International Journal of Mechanical Engineering and Technology (IJMET), ISSN (Print) ISSN (Online) Volume 1 Number 1, July - Aug (2010), pp IAEME, IJMET I A E M E ARTIFICIAL NEURAL NETWORK BASED CLASSIFICATION FOR MONOBLOCK CENTRIFUGAL ABSTRACT PUMP USING WAVELET ANALYSIS V. Muralidharan Department of Mechatronics Engineering School of Mechanical Engineering Tamil Nadu, SRM University v_murali2@yahoo.co.in V. Sugumaran Department of Mechatronics Engineering School of Mechanical Engineering Tamil Nadu, SRM University P. Shanmugam Department of Mechatronics Engineering School of Mechanical Engineering Tamil Nadu, SRM University K. Sivanathan Department of Mechatronics Engineering School of Mechanical Engineering Tamil Nadu, SRM University Fault diagnosis of monoblock centrifugal pump as pattern recognition problem has three major steps. Feature extraction, Feature selection and classification. Numbers of advanced algorithms are being used for feature extraction and classification. However, all the features extracted from raw signal need not have useful information for the study. Presences of redundant features are also possible. There are many algorithms which can be used to filter such redundant features. In this paper, Discrete Wavelet Transform (DWT) is used for feature extraction and best features are selected using decision tree algorithm and classification is done using ANN and the results are presented. 28

2 Keywords: Monoblock centrifugal pump, ANN, Discrete Wavelet Transform (DWT), classification, feature extraction. 1. INTRODUCTION In a monoblock centrifugal pump, defective bearing, defect on the impeller and cavitations occurs which leads to a very serious problems. Cavitation can cause more undesirable effects, such as deterioration of the hydraulic performance (drop in head capacity and efficiency) [1]. Vibration signals are widely used in condition monitoring of centrifugal pumps [2]. Fault detection is achieved by comparing the signals of monoblock centrifugal pump running under normal and faulty conditions [4]. The faults considered in this study are bearing fault (BF), impeller fault (IF), bearing and impeller fault (BFIF) together and cavitation (CAV). By the application of Seismic or piezoelectric transducers, the vibration levels are measured for each condition. Data acquisition system is used to capture the vibration signals [3]-[5]. From the vibration signal relevant features can be extracted and classified using a classifier by various machine learning approaches. Czeslaw T. Kowalski, Teresa Orlowska-Kowalska discussed diagnosis problems of the induction motors in the case of rotor, stator and rolling bearing faults. Two kinds of neural networks (NN) were proposed for diagnostic purposes: multilayer Perceptron networks and self organizing Kohonen networks. Neural networks were trained and tested using measurement data of stator current and mechanical vibration spectra [6]. Renpu Li presented a hybrid system to extract efficiently classification rules from decision table. Rough sets were used to extract the association rules and classification was performed using neural networks [7]. Sri Kolla presented an ANN based technique to identify faults in a three-phase induction motor. The main types of faults considered are overload, single phasing, unbalanced supply voltage, locked rotor, ground fault, over voltage and under-voltage. A feed forward layered neural network structure was used. The network was trained using the back propagation algorithm. The trained network was tested with simulated fault current and voltage data [8]. 29

3 2. EXPERIMENTAL STUDIES The main idea of the study is to find whether the monoblock centrifugal pump is in good condition or in faulty condition by systematic procedure following certain steps. If the pump is found to be in faulty condition then the next step is to segregate the faults into bearing fault, impeller defect, bearing and impeller defect together and cavitation Construction The monoblock centrifugal pump for condition monitoring is taken for this study. The motor (2hp) is used to drive the pump. Piezoelectric type accelerometer is used to measure the vibration signals. The accelerometer is mounted on the pump inlet using adhesive and connected to the signal conditioning unit where the signal goes through the charge amplifier and an analog to digital converter (ADC) and the signal is stored in the memory. Then the signal is processed from the memory and it is used to extract different features Procedure The pump is allowed to rotate at a speed of 2880 rpm at normal working condition and the vibration signals are measured. The sampling frequency of 24 khz and sample length of 1024 were considered for all conditions of the pump. The sample length was chosen arbitrarily to an extent; however, the following points were considered. After calculating the wavelet transforms it would be more meaningful when the number of samples is more. On the other hand, as the number of samples increases, the computation time increases. To strike a balance, sample length of around 1,000 was chosen. The specification of the monoblock centrifugal pump is given as below. Table 1 Monoblock centrifugal pump specification In the present study the following faults were simulated. (i)cavitation (ii) Impeller fault (iii) Bearing fault (iv) Bearing and Impeller fault together. 30

4 The faults were introduced one at a time and the pump performance characteristic and vibration signals were taken. 3. FEATURE EXTRACTION Discrete Wavelet Transform (DWT) has been widely used and provides the physical characteristics of time-frequency domain data. DWT of different versions of different wavelet families have been considered. The following wavelet families and their sub families have been tried for the present study. Daubechies wavelet, Coiflet, biorthogonal wavelet, reversed bi- orthogonal wavelet, biorthogonal wavelet, symlets and meyer wavelet Feature definition Feature extraction constitutes computation of specific measures, which characterize the signal. The discrete wavelet transform (DWT) provides an effective method for generating features. The collection of all such features forms the feature vector. A feature vector is given by Eq. (1) A component in the feature vector is related to the individual resolutions by the following equation Eq. (2) Where, is the ith feature element in a DWT feature vector. ni is the number of samples in an w2i,j samples in an w2i,j individual sub-band, is the jth detail coefficient (high frequency component) of the ith sub-band. The wavelets considered for the present investigation are Haar (db1), Daubechies, Symlets, Coiflets, Biorthogonal, Reverse Biorthogonal and Meyer (dmey). 4. CLASSIFICATION Artificial Neural Networks (ANN) is modeled on biological neurons and nervous systems. They have the ability to learn, and has the processing elements known as 31

5 neurons which perform their operations in parallel. ANN s are characterized by their topology, weight vector and activation functions. They have three layers namely an input layer, which receives signals from the external world, a hidden layer, which does the processing of the signals and an output layer, which gives the result back to the external world. 4.1 Multi-Layer Perceptron (MLP) This is an important class of neural networks, namely the feed forward networks. Typically, the network consists of a set of input parameters that constitute the input layer. MLPs have been applied to solve some difficult and diverse problems by training them in a supervised manner with back-propagation algorithm. Each neuron in the hidden and output layer consists of an activation function, which is generally a non-linear function. The weights of the network to be trained are initialized to small random values. The weights are updated through an iterative learning process known as Error Back Propagation (BP) algorithm. Error Back Propagation process consists of two passes through the different layers of the network; a forward pass in which input patterns are presented to the input layer of the network and its effect propagates through the network layer by layer[6]-[8]. Finally, a set of outputs is produced as the actual response of the network. During the forward pass the synaptic weights if the networks are all fixed. The error value is then calculated, which is the mean square error (MSE) given by Eq. (3) Where, Where, m is the number of neurons in the output layer, is the kth component of the desired or target output vector and is the kth component of the output vector. The training process is carried out until the total error reaches an acceptable level 32

6 (threshold). If Etot < Emin the training process is stopped and the final weights are stored, which is used in the testing phase for determining the performance of the developed network. 5. RESULTS AND DISCUSSION As mentioned in section 3, all the wavelets and their sub families were taken for extracting the features from the raw signal. Thereafter, classification was performed using ANN. It was found that out of all wavelets rbio1.5 was giving comparitvely good results (100%) and hence it was considered as the best feature for the given conditions of the problem. Keeping number of neurons in hidden layer constant, the behaviour of the Root Mean Square (RMS) error and number of epcochs were calculated. By choosing best values of number of neurons in hidden layer, RMS error and number of epochs were calculated for constant learning rate and momentum. Finally, considering all the trials the best values of RMS error and corresponding epochs were found. The characteristics of different parameters were studied and plotted. a. No. of neurons in hidden layer Vs RMS error 33

7 b. No. of neurons in hidden layer Vs No. of epochs c. Momentum Vs RMS error 34

8 d. Learning rate Vs RMS error e. Learning rate Vs No. of Epochs 35

9 f. Momentum Vs No. of epochs Figure 1 Characteristics of different parameter for the developed network 5.1 ANN architecture Network type The neural network model definitions and model architecture is as follows: Transfer function No. of nodes in input layer : 3 No. of hidden layers : 1 No. of neurons in hidden layer: 5 No. of neurons in Output layer: 1 Training rule Training tolerance : 0.1 Learning rule Momentum learning step size : 0.1 Momentum learning rate : 0.9 No. of epochs : 1746 : Feed Forward Back Propagation : Sigmoid transfer function in hidden and output layer : Back propagation RMS Error : Training termination : Momentum learning method : Minimum mean square error 36

10 6. CONCLUSION From the above results and discussion, one can confidently say that feature extraction using wavelets (DWT) as well as ANN algorithm for classification were found to be good candidates for practical applications of fault diagnosis of monoblock centrifugal pump. 7. ACKNOWLEDGEMENT The authors express their sincere gratitude to Dr. K. P. Soman, Dr. K.I. Ramachandran and Mr. N. R. Sakthi vel Amrita School of Engineering, Coimbatore, India for giving us an insight into wavelets and data mining techniques. 8. REFERENCES [1] L. Alfayez, D.Mba, G.Dyson (2005), The application of acoustic emission for detecting incipient cavitation and the best efficiency point of a 60 kw mono block centrifugal pump, NDT&E International, 38, [2]Jonathan P. Peck, John Burrows (1994), On-line condition monitoring of rotating equipment using neural networks, ISA Transactions, 33, [3] Czesław cempel (1988), Vibroacoustical diagnostics of machinery - An outline, Mechanical systems and Signal Processing,2, [4] H.Q.Wang, P.Chen and Mie (2007), Fault diagnosis of centrifugal pump using symptom parameters in frequency domain, The CGIR Ejournal, 9, [5] Huaqing Wang and Peng chen (2007), Sequential condition diagnosis for centrifugal pump system using fuzzy neural network, Neural Information Processing-Letters and Reviews, 2, [6]Czeslaw T.Kowalski, Teresa Orlowska-Kowalska (2003), Neural networks application for induction motor faults diagnosis, Mathematics and Computers in Simulation, 63, [7] Renpu Li and Zheng-ou Wang (2004), Mining classification rules using rough sets and neural networks, European Journal of Operational Research, 157, [8]Sri Kolla and Logan Varatharasa (2000), Identifying three-phase induction motor faults using artificial neural networks, ISA Transactions,39,