1 AUTOMATIC CLASSIFYING METHOD FOR WEDGE 2 TIGHTNESS BY SUPPORT VECTOR MACHINE AND 3 ARTIFICIAL NEURAL NETWORK 4 5 Running head: Automatic wedge tightness by Support Vector Machine and 6 Artificial Neural Network Technique 7 8 Thanachai Poombansao1*, Waree Kongprawechnon1, Somsak 9 Kittipiyakul1, Chonlada Theerawon2, Muntita Charoenlarp1 and It 10 Chunsangnate1 11 12 1 School of Information, Computer, and Communication Technology (ICT), 13 Sirindhorn International Institute of Technology, Thammasat University, Khlong Luang, 14 Pathum Thani, 12121, Thailand. Tel. 0-2986-9009; Fax. 0-2986-9112; 15 E-mail: t.poombansao@gmail.com 16 2 17 2 National Electronics and Computer Technology Center (NECTEC), Khlong Luang, Pathum 18 Thani, 12121, Thailand. Tel. 0-2564-6900; Fax. 0-2564-6901-3; 19 *Corresponding author 20 21 Abstract 22 This study proposed an automatic classifying system for wedge tightness of a 23 generator’s wedge. It consists of 4 processes called data collection, 24 preprocessing, feature extraction, and classification. The aim of this study is to 25 classify and verify the tightness of a generator’s wedge for the Electricity 26 Generating Authority of Thailand (EGAT) by using Machine Learning 27 algorithms, called Support Vector Machine (SVM) and Artificial Neural 28 Network (ANN). The linear and radial basis function (RBF) are selected for 29 SVM classifier. The evaluation is completed by using a 10-fold cross validation 30 technique to give high accuracy and a low number of False Negatives (FN). 31 From the simulation results, the signals extracted in the frequency domain has 32 less FN than the time domain, and ANN gives the best performance among 33 classifiers. 34 Keywords: Wedge tightness signal, Support Vector Machine, Artificial Neural 35 Network, classification 2 3 36 37 Introduction 38 Nowadays, the use of electricity is increasing rapidly around the world due to the 39 exponential growth of new technology and population. Electricity is involved in all 40 daily activities such as industries, hospitals, and houses. Therefore, more usage of 41 power plants is needed to support the demand of users, which cause the deterioration 42 of generators. The information of a generator such as the tightness of a generator’s 43 wedge is very important to evaluate the safety of the generator. Normally, human 44 judgment has been used to determine the wedge performance inside the generator by 45 inspection and evaluation. A human can check the wedge tightness by tapping with a 46 hammer, feeling for vibration, listening to the sound by trained ears, and recording 47 the degree of looseness on a chart. A tight wedge has very little vibration or hard 48 ringing sound. On the other hand, a loose wedge has noticeably vibration and a very 49 hollow sound (Geoff Klempner and Isidor Kerszenbaum, 2004). This method is very 50 subjective because it depends on individual judgment from operators by hearing the 51 sound from a slot wedge (Summer project pal, 2011). However, it consumes both 52 time and cost because the rotor in the generator has to be removed and the generator 53 needs to shut down during operation. Hence, a robot has been used instead of a 54 human to determine the wedge performance, which is cost efficient because the rotor 55 does not need to be removed during operation. George introduced the testing of stator 56 wedge tightness in an electrical generator that contains a vibration sensor attached to 3 4 57 the stator core lamination and a mounting system for mounting the base assembly to 58 impact assembly (George, F.D. et al, 1996). At first, robots used for generators in 59 Thailand were imported robots. However, the robots are very expensive and too big 60 for some generators in Thailand. Therefore, the Electricity Generating Authority of 61 Thailand (EGAT) plans to develop its own robot that is smaller with lower cost. The 62 robot also is required to automatically classify the wedge performance of a generator. 63 The task of a robot can be divided into two main functions including wedge 64 tapping control and automatic analyze wedge tightness system. Wedge tapping 65 control is used to adjust the tapping pattern of the robot to be a single, periodic, or 66 repeating pattern. The wedge tightness signals are analyzed afterward. There are 4 67 steps to analyze wedge tightness signals including data collection, preprocessing, 68 feature extraction, and classification. 69 Data collection is the raw data received by the robot to illustrate the wedge 70 inside the generator. Preprocessing refers to improving the raw data quality including 71 filtering, and noise cancellation. Feature extraction reduces the pattern vector to a 72 lower dimension and keeps the useful information from the original vector, called the 73 feature vector. Classification takes the feature vector as an input to define the 74 tightness of the wedge by placing the feature vector into the most appropriate class 75 (Tohka J., 2013). Many classification methods have been applied to the system such 76 as nearest neighbor classifier, wavelet transform, probabilistic classification, and 77 multiclass classification. 4 5 78 SVM classification is an alternative way for pattern recognition by using the 79 separation on the hyperplane (decision boundary) to distinguish data diversity. ANN 80 classification is the machine learning algorithm to teach to recognize and remember 81 different traits. ANN also has low computation loads, and flexibility for pattern 82 recognition. A computational neuron can produce either a linear or non-linear input 83 which allows the network to learning efficiently. ANN is configured for a specific 84 application such as pattern recognition or data classification through a learning 85 process. There are many research works with SVM and ANN for classification such 86 as Cardiac Auscultation Analysis, a screening system to diagnose disease from heart 87 sounds in digital format (Phatiwuttipat, P. et al, 2011). 88 classification methods also are used for analyzing construction materials with 89 ultrasonic waves to find the number of defects, and the result is very accurate 90 (Saechai, S. et al, 2012). The objective of this work is to develop an automatic 91 classification system for determining the tightness of a generator’s wedge. The SVM and ANN 92 This study is beneficial for identifying the tightness of a generator’s wedge 93 in a power plant. A brief description of both SVM and ANN classifiers, the 94 methodology, and the comparison between both classification results is discussed in 95 the following sections. 96 97 Material and methods 5 6 98 System Configuration 99 In the experiment, Figure 1 illustrates the developed robot by the National 100 Electronic and Computer Technology Center (NECTEC) with size 180-300 101 millimeters wide, 300 millimeters long and 20 millimeters thick. The developed 102 robot is cheaper than the cost of an imported robot (about 250,000 U.S. dollars). 103 The overall robot system is shown in Figure 2. The control box acts like a 104 brain and nervous system which is responsibility for controlling and linking all 105 functions into the entire system, including movement of a camera, movement of 106 robot, and wedge tapping. The display monitor is used for monitoring each activity in 107 the system including video camera, wedge’s result, position of the robot and the 108 graphical user interface (GUI). The joystick is used for controlling the direction of the 109 camera, the movement of robot, and tapping pulse. In addition, the robot connects to 110 the control box via cable. This study focuses on wedge tightness signal analysis. 111 The raw data of a wedge tightness signal obtained from the robot goes 112 through the process called “automatic analyze wedge tightness system” in order to 113 classify the data. The block diagram in Figure 3 illustrates the signal processing part 114 of the wedge tightness signal. 115 116 Data collection 6 7 117 Data collection comes from the robot tapping on the generator and receiving a 118 reacted signal back. “Signal conditioning” is used in order to focus and make the 119 signals even clearer. A Data Acquisition System (DAQ) is used to convert analog 120 data samples from the robot into digital numeric values to plot a graph (Figure 4). 121 Figure 4 a) and Figure 5 a) show the tight wedge data signal and the loose wedge 122 signal, respectively. The real raw data is obtained from the board calibration, which 123 has 20,000 sampling points. The robot taps every 100 ms and each sampling point has 124 a time interval of 50 µs. 125 In this experiment, there are 240 inputs for each wedge signal (tight and loose). The 126 inputs have been reduced into 2,000 sampling points to extract only significant parts. 127 The 2,000 sampling points of tight wedge signal and loose wedge signal are shown in 128 Figure 4 and Figure 5, respectively. 129 130 Preprocessing 131 Preprocessing is used for the preparation of signal. Preprocessing refers to 132 operations performed on the raw data to improve its quality such as filtering, and 133 noise cancellation. The wedge tightness signal response is reduced by using a 134 threshold to capture the significant part. The threshold is set to 2,000 sampling points 135 to remove the unwanted part of the signal. 136 7 8 137 Feature extraction 138 Feature extraction extracts the significant features from the raw data. Fast 139 Fourier Transform (FFT) is used for feature extraction in the frequency domain. FFT 140 is a very useful tool for signal analysis to obtain the features for classification (Kim et 141 al., Unpublished). FFT computes the same as the Discrete Fourier Transform (DFT) 142 by letting 143 as follows: be complex numbers. The DFT can be defined by the formula 144 145 where 146 By contrast, FFT takes approximately 147 which takes approximately 148 experiment to obtain the input features of classifiers as follows. 149 150 151 152 operations which is faster than DFT, operations. There are 2 methods used in this i) Time domain classifier: 2,000 data points in the time domain are used in this experiment, as shown in Figure 6. ii) Frequency domain classifier: 2,000 data points are transformed to the frequency domain by using Fast Fourier Transform (FFT), as shown in Figure 7. 153 154 Classification 8 9 155 The input sample is separated into 2 classes including 120 tight wedge data 156 signals and 120 loose wedge data signals. The classification methods used in this 157 experiment are Support Vector Machine (SVM) and Artificial Neural Network 158 (ANN). 159 160 Support Vector Machine (SVM) 161 Support Vector Machine (SVM) is one of the important models for pattern 162 recognition and classification. The main concept of SVM is to transform an input 163 signal into feature space and divide the clear gap between classes with an optimal 164 hyperplane (decision boundary) and maximum margin (support vector) (Avci, E., 165 2009), as shown in Figure 8 (Chatunapalak, I. et al., 2012). Since most of the data are 166 non-linear and non-separable, a kernel function is used for mapping a non-linear 167 input space into a linear space (higher dimensional feature space) to perform the 168 separation (Ahmad, A.R. et al., 2009). Figure 9 illustrates how to transform a non- 169 linear data point to a higher dimensional feature space by using SVM with a kernel 170 function. 171 Many kernel functions are used for SVM classifiers such as linear, polynomial, radial 172 basis unction (RBF), Fourier, etc. The kernel parameters in a kernel function must be 173 set properly to maximize the accuracy of SVM classification. This study selected 174 linear and RBF kernels for SVM classification. 9 10 175 SVM classification trained and evaluated input samples by a data splitting 176 method. The training and testing samples generally split into 90% and 10% of the 177 input samples for each class, respectively. The input, training, and testing samples are 178 provided in Table 1. The binary classifications of SVM are performed for 179 classification. Each data input is prepared in the form of row vector of size 180 where 181 columns of features (2,000 features), as shown in the formula as follows: represents the number of input samples. Each input sample contains 182 183 The output of a wedge tightness signal depends on the data class. The 184 numbers 1 and 0 are set as a target for class 1 and class 2 (tight and loose wedge), 185 respectively. The output vector for binary classification respectively is presented as 186 follows: 187 188 The proper kernel function is selected to train the non-linear SVM classifier. The 189 linear and RBF kernels are used to train classifiers and test results because the RBF 190 kernel has good classification according to many studies. The RBF kernel gives high 191 classification performance in this study. Ten-fold cross validation is applied for 192 classification, as shown in Figure 10. The datasheets are randomly divided into 10 193 subsets. Each subset is constructed and trained by using 9 out of 10 (90%) folds and 10 11 194 tested 1 out of 10 (10%) to obtain a cross validation estimate of its error rate. The 195 process is repeated in the same steps for 10 subsets and averaged for classification 196 accuracy from all data (Delen et al., 2005). The proposed SVM classification used 197 MATLAB for simulations. 198 199 Artificial Neural Network (ANN) 200 An Artificial Neural Network (ANN), usually called Neural Network (NN), is 201 a good technique for determining pattern recognition. The main concept of ANN is 202 inspired by animal nervous systems (brain), which is composed of the large number 203 of highly interconnected processing elements (neurons) working to solve specific 204 problems through a learning process. The learning process of ANN includes adaptive 205 learning, self-organization, real-time operation, and fault tolerance. ANN uses the 206 back-propagate model to solve non-linear signals. 207 The back-propagated model is a feed-forward multi layered ANN that allows 208 a signal to travel one way from input to output with no feedback (loops). Figure 11 209 illustrates the back-propagated model of ANN. According to Figure 11, the ANN 210 network consists of 3 layers including input layer, hidden layer, and output layer. The 211 input layer represents the raw data fed into the network as an input, which forwards to 212 the hidden layer. The hidden layer consists of neurons to solving complex problem by 11 12 213 calculation, and forwards the result into the output layer. The output layer takes the 214 results from the hidden layer, performs a calculation, and gives the final result. 215 The input samples and classes are the same as SVM. MATLAB is used to 216 train, validate, and test for Neural Network Pattern Recognition Protocol (nprtool) by 217 selected input and target. Training, validation, and testing is set to 85%, 5%, and 218 10%, respectively, and the number of hidden neurons is set to 10. 219 220 Results and discussion 221 Each raw data of wedge tightness signal performs preprocessing and feature 222 extraction to obtain the input features for the classifier. The number of total input 223 samples, training samples, and testing samples are mentioned in a previous section. 224 The trained classifier is evaluated by test samples. The classification accuracies are 225 analyzed by using two different types of input features from the data that were 226 obtained in the time domain and frequency domain. The efficiency of each single 227 classification is obtained by a 2×2 confusion matrix. The confusion matrix is a table 228 that shows the performance of the algorithm. An example of a confusion matrix is 229 shown in Figure 12. The first and second rows represent the data with target tight 230 wedge and loose wedge, respectively. The first and second columns represent the 231 predictor of a tight wedge and loose wedge, respectively. The true positives (TP) are 232 the number of tight wedges which are correctly classified. False positives (FP) are the 12 13 233 number of loose wedges which are misclassified from tight wedges. On the other 234 hand, true negatives (TN) are the number of loose wedges which are correctly 235 classified and false negatives (FN) shows the number of tight wedges which are 236 misclassified from loose wedges. Positive predictive value (PPV) and negative 237 predictive value (NPV) are the rate that the predictor can give a correct classification 238 for the tight wedge and loose wedge, respectively. Sensitivity and specificity are the 239 correctness prediction for tight wedge and loose wedge, respectively. Accuracy 240 (ACC) value computes the average accuracy in the classification tables. All Values in 241 the tables are the average classification accuracies of 10 confusion matrices. 242 243 Time Domain Analysis 244 The result of classification algorithms for wedge tightness signal in the time 245 domain is shown in Table 1 including SVM (linear and RBF) with kernel functions 246 and ANN classifier. 247 From Table 1, the value of specification in the time domain between 2 classes 248 (tight and loose wedge) is classified. The sensitivity of ANN (98.24%) has better 249 performance than VSM in both linear (96.08%) and RBF (96.1%) functions. The 250 specification of ANN (90.52%) also has better efficiency than the others (linear 251 88.59% and RBF 88.67%). In comparison, the SVM with RBF kernel function has an 252 accuracy of about 91.38% which is worse than that of SVM with linear function 13 14 253 (92.34%), but has twice faster average training time than that of the linear function. 254 From the result, ANN classifier provides the best efficiency, compared to both SVM 255 (linear and RBF) classifiers in the time domain. However, a false negative is very 256 important for wedge tightness because the misclassification of a loose wedge can 257 cause uncertainly in a generator. Since false negatives still exist in the time domain 258 (1.25% for ANN, 1.97% for SVM with linear function, and 1.98% for SVM with 259 RBF function), frequency domain classification is applied to decrease the number of 260 false negatives. 261 262 Frequency Domain Analysis 263 The classification result in frequency domain is shown in Table 2 including 264 SVM (linear and RBF) with kernel functions and ANN classifiers. From Table 2, the 265 number of false negatives for SVM with RBF function is 0.08%, which is the worst 266 classification, compared to the other methods. The numbers of false negatives in 267 ANN and SVM with linear function are reduced to 0%, which is a perfect 268 classification for wedge tightness. The training time decreases when using SVM with 269 linear function (11.182s), SVM with RBF function (7.095s), and ANN (2s) methods. 270 271 Conclusion 14 15 272 The automatic classification method for wedge tightness by support vector machine 273 and artificial neural network provides good recognition and classification results in 274 the frequency domain. The classification result shows that the ANN classification 275 method has the best performance, which has the highest accuracy and shortest 276 training time. The result shows that the frequency domain signal has more accuracy 277 than the time domain signal. By comparison between 2 SVM methods, the linear 278 function provides a better solution in terms of accuracy, but consumes more time than 279 the RBF function does. 280 For further study, binary-classification (tight or loose wedge) from this 281 experiment may not be sufficient to identify a suspected loose wedge generator in a 282 real environment. The multi-classification technique can be developed in the future to 283 predict how many degrees of looseness in a suspected loose wedge for a generator. 284 For instance, multi-class support vector machine has been used by an automatic 285 screening method to detect glaucoma for early treatment and preserve eye sight 286 (Vejjanugraha, P. et al, 2013). 287 wedge tightness needs Graphic User Interface (GUI) for ease of use and 288 understanding. Moreover, the automatic classifying method for 289 290 Acknowledgements 15 16 291 This research is financially supported by School of Information, Computer, and 292 Communication Technology (ICT) and the National Research University Project, 293 Thailand office of Higher Education Commission (NRU). The author would like to 294 thank the National Electronics and Computer Technology Center (NECTEC) and 295 Electricity Generating Authority of Thailand (EGAT) for advice and data in this 296 research. 297 298 299 300 301 302 303 304 305 306 307 16 17 308 309 310 References 311 Geoff Klempner and Isidor Kerszenbaum (2004). Stator winding mechanical tests. In: 312 Operation and maintenance of large turbo-generators. John Wiley & Sons, 313 Inc, Hoboken, New Jersey, p. 485-487. 314 Summer project pal (2011). A new method for Stator Slot Wedge Testing of Large 315 …… Generators. Available from: http://topicideas.net/how-to-a-new-method-for- 316 …… stator-slot-wedge-testing-of-large-generators. Access date: Jan, 2001. 317 George, F.D., Mark, W.F., and Harry, L.S., inventors; Westinghouse Electric 318 ...…….Corporation, assignee. February 27, 1996. System and method for impact 319 …… testing wedge tightness in an electrical generator. U.S. patent no. 5,493,894. 320 Tohka, J. (2013). Introduction to Pattern Recognition. Department of Signal 321 ……….Processing, Tempere University of Technology. 322 Phatiwuttipat, P., Kongprawechnon, W., and Tungpimolrut, K. (2012). Cardiac 323 …........Auscultation Analysis System with Neural Network and SVM technique. 324 ………International Institute of Technology, Thammasat University, Pathum 325 ………Thani, Thailand. 17 18 326 Saechai, S., Kongprawechnon, W., and Sahamitmongkol, R. (2012). Test System for 327 ………Defect Detection in Construction Material with Ultrasonic Waves by Support 328 ………Vector Machine and Neural Network, SCIS-ISIS 2012, Kobe, Japan. 329 Kim, J.S., Kim, I.B., and Yoo, K.S. An experimental comparison of ultrasonic pulse 330 ………velocity and frequency domain signal of concrete, unpublished. 331 Avci, E. (2009). A new intelligent diagnosis system for the heart valve disease by 332 ………Using genetic-SVM classifier. Expert Systems with Applications, 36: 10618- 333 ………10626. 334 Ahmad, A.R., Khalid, M., Gaudin, C.V., and Poisson, E. (2004). Comparison of 335 ………support vector machine and neural network in character level discriminant 336 ………training for online word recognition. UNITEN Students Conference on 337 ………Research and Development, Malaysia. 338 Chatunapalak, I., Phatiwuttipat, P., Kongprawechnon, W., and Tungpimolrut, K. 339 ………(2012). Childhood musical murmur classification with support vector machine 340 ………technique. In Proceeding of the 9th International Conference on Electrical 341 ………Engineering/Electronics, Computer, Telecommunications and Information 342 ………Technology (ECTI-CON 2012), Hua Hin, Thailand. 343 Delen, D., Walker, G., and Kadam, A. (2005). Predicting breast cancer survivability: 344 ………a comparison of three data mining methods. Artificial Intelligence in 345 ………Medicine, 34: 10618-10626. 18 19 346 Vejjanugraha, P., Kongprawechnon, W., Kondo, T., Tungpimolrut, K., Kaneko, H., 347 ………and Sintuwong, S. (2013). An Automatic Screening Method For Detecting 348 ………Glaucoma 349 ………University, Khlong Luang, Pathum Thani, Thailand. Using Multi-class Support 350 351 352 353 354 355 356 357 358 359 360 361 362 19 Vector Machine, Thammasat 20 363 Table 1. Classification results in the time domain Methods Kernel TP FP FN TN PPV NPV Sensitivity Specification ACC Training Function (%) (%) (%) (%) (%) (%) (%) (%) (%) time (S) ANN - 46.25 5 1.25 47.51 90.55 97 98.24 90.52 93.75 2.000 SVM Linear 48.04 5.71 1.97 44.31 89.4 95.79 96.08 88.59 92.34 7.460 RBF 48.03 6.66 1.98 43.34 87.84 95.69 96.1 86.67 91.38 3.857 364 365 366 367 368 369 370 371 20 21 372 Methods 373 Kernel TP FP FN TN PPV NPV Sensitivity Specification ACC Training Function (%) (%) (%) (%) (%) (%) (%) (%) (%) time (S) ANN - 56.67 0 0 43.33 100 100 100 100 100 2.000 SVM Linear 50 0 0 50 100 100 100 100 100 11.182 RBF 49.92 0.46 0.08 49.55 99.12 99.84 99.84 99.09 99.47 7.095 Table 2. Classification results in the frequency domain 374 21 22 375 376 377 378 Figure 1. Imported robot for testing. 379 380 381 382 383 384 22 23 385 386 Control box Control Movement of camera PLC + Joystick Automatic Analysis of wedge tightness system Display Monitor + Manual Camera Control Movement of robot 387 388 Figure 2. Overall robot system 389 390 391 392 393 394 23 24 395 396 DATA COLLECTION PREPROCESSING FEATURE EXTRACTION CLASSIFICATION TIME ANALYSIS Wedge tightness signal response Noise Removal Time Domain FREQUENCY ANALYSIS FFT Transform Frequency Domain 397 398 Figure 3. Block diagram of signal processing part 399 400 401 402 403 404 405 406 407 24 SVM or ANN CLASSIFICATION RESULTS 25 408 409 (a) (b) 410 411 412 Figure 4. a) The original sampling points of tight wedge signal b) The 2,000 sampling points of tight wedge signal 413 414 415 25 26 416 417 (a) (b) 418 419 420 Figure 5. a) The original sampling points of loose wedge signal b) The 2,000 sampling points of loose wedge signal 421 422 423 26 27 424 425 (a) 426 (b) 427 428 429 430 Figure 6. a) Tight wedge signal in the time domain b) Loose wedge signal in the time domain 431 432 433 434 27 28 435 436 437 (a) 438 (b) 439 440 441 442 Figure 7. a) Tight wedge signal in the frequency domain b) Loose wedge signal in the frequency domain 443 444 445 446 28 29 447 448 449 450 Figure 8. Basic concept of support vector machine 451 452 453 454 455 456 457 458 29 30 459 460 461 462 Figure 9. Transformation non-linear data points to higher dimension linear space 463 464 465 466 467 468 469 470 30 31 471 472 Have not been tested sets Test sets Tested sets 10% 474 10% 10% 10% 10% 10% 10% 10% 10% 10% 10% 10% 10% 10% 473 10% 10% 10% Round 1 10% 10% 10% Round 2 Figure 10. Ten-fold cross-validation procedure 475 476 477 478 479 480 481 31 10% … 10% 10% 10% 10% 10% 10% 10% 10% 10% Round 10 32 482 483 484 485 486 Figure 11. Back-Propagation model of artificial neural network 487 488 489 490 491 32 33 492 493 494 Condition positive Condition negative Test outcome positive True Positive (TP) False Positive (FP) Positive Predictive Value (PPV) TP TP+FP Test outcome negative False Negative (FN) True Negative (TN) Negative Predictive Value (NPV) TN FN+TN Sensitivity TP TP+FN Specification TN FP+TN Accuracy TP+TN TP+FP+FN+TN 495 496 Figure 12. Example of 2×2 confusion matrix 33