Data Mining for Fault Diagnosis and Machine Learning. for Rotating Machinery

Transcription

1 Key Engineering Materials Vols (2005) pp online at (2005) Trans Tech Publications, Switzerland Online available since 2005/Sep/15 Data Mining for Fault Diagnosis and Machine Learning for Rotating Machinery Gang Zhao 1, DongXiang Jiang 2, Kai Li 3, JinHui Diao 4 1, 2, 3, 4 Department of Thermal Engineering, Tsinghua University Beijing , P.R. China 1 zhaog02@mails.tsinghua.edu.cn, 2 jiangdx@tsinghua.edu.cn Keywords: faults diagnosis, machinery learning, data mining, rotating machinery Abstract. Data mining is used not only for database analyses, but also for machine learning. The data mining technique described in this paper was used for steam turbine fault diagnostics based on continuous data measurements. The classification rules are based on standardized vibration frequency data for steam turbines and field experts analyses of turbine vibration problems. The expert knowledge enables the steam turbine fault diagnosis system to be more powerful and accurate. The system can identify twenty types of standard steam turbine faults. The system was developed using 2000 simulated data sets. The data mining methods were then used to identify 20 explicit rules for the turbine faults. The method was also used with actual power plant data to successfully diagnose real faults. The results indicate that data mining can be effectively applied to diagnosis of rotating machinery by giving useful rules to interpret the data. Introduction Beginning in around 1985, the goal of rotating machinery fault diagnostics was primarily to store the vibration spectra and to provide graphical tools so that the analyst could quickly access the data and determine what might be wrong with the machine. But as the data collection devices (originally spectrum analyzers) became smaller, faster, and more portable, the amount of data to be analyzed rapidly grew. The data acquisition system could soon store hundreds of spectra. As the data acquisition systems and measurement techniques improved, the analyst was faced with mountains of data. Similar problem was developing in other area with large data warehouses with the rapid developments in digital data acquisition and storage technology. Although valuable information may be hiding in the data, the overwhelming data volume makes it difficult, if not impossible, for human beings to extract the information without powerful tools. The overwhelming amount of data resulted in the new technique of data mining, which seeks to extract knowledge from huge volumes of data through numerical analysis of the data. Data mining is not only database analysis method, but also an important machine learning tool. For machine learning, data mining is defined by Witten and Frank as the extraction of implicit, previously unknown, and potentially useful information from data [1]. This paper describes the application of data mining techniques to steam turbine fault diagnostics. Many methods have been used for data mining, with the decision tree often shown to be the most valuable form of data mining. The decision tree classifier (DTC) has been used successfully in many diverse areas such as radar signal classification, character recognition, remote sensing, medical diagnosis, expert systems, and speech recognition, to name only a few [2]. Perhaps, the All rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of the publisher: Trans Tech Publications Ltd, Switzerland, (ID: /04/08,11:10:38)

2 176 Damage Assessment of Structures VI most important feature of DTC's is their capability to break down a complex decision-making process into a collection of simpler decisions, thus providing a solution which is often easier to interpret. Fault diagnosis is based on pattern identification and classification. The first step in steam turbine fault diagnostics is pattern identification from the measured data. The next step is to interpret what the patterns indicate about the machine, but proper interpretation requires some knowledge about the machine. Decision trees provide a good approach to supervised classification and prediction in artificial intelligence and statistical pattern recognition. A tree is "grown" from data using a recursive partitioning algorithm to create a tree which (hopefully) accurately predicts classes on new data. Neural networks provide another valuable method for fault classification for rotating machinery. Crupi et al. [3] describe the use of neural networks to evaluate vibration signatures in rotating machinery and recognize the occurrence of faults. The procedure can be used to diagnose fault not considered in the training set. However, the neural networks knowledge is hidden in the network, so the rules can not be easily extracted and interpreted. Decision trees can be more effectively applied to steam turbine fault diagnosis because the fault diagnosis requires not only pattern classification, but also rule extraction and knowledge interpretation. Steam turbine fault diagnostics Steam turbine faults are generally classified into twenty types listed in Table 1 based on field experts experience and theoretical analyses. Table 1. Steam turbine fault classification. Fault No. Description Fault No. Description F0 normal F10 pedestal looseness F1 imbalance F11 foundation looseness F2 components missing F12 worn coupling F3 bent shaft F13 electricity magnet excited F4 shaft-seal rubbing F14 sub-harmonic vibration F5 axial rubbing F15 oil whirl F6 axial misalignment F16 oil whip F7 eccentricity faults F17 steam excited vibration F8 rotor crack F18 valve vibration F9 shrunk-on-disc failure F19 power disturbance In any faults diagnosis, feature extraction is an important step for detecting steam turbine faults. Features can be extracted from the frequency domain of a typical steam turbine vibration analysis. However, analysis of the steam turbine data requires a detailed understanding of the steam turbine design, operation, and maintenance. Vibration spectrum analysis is a practical and powerful tool for steam turbine fault diagnosis because it is based on a great deal of engineering experience. Although there have recently been many new methods applied to fault diagnosis, most approaches are based on or related to the vibration spectrum data. However, the fault can not be easily related to the spectrum data because the steam turbine system is very complex and influenced by numerous process parameters. The best method is to use the feature-fault relationship matrices in well-established machining reference databases, expert intelligence for the reasoning and decision-making and experimental results of signal characteristics for various working conditions. Table 3 show a fuzzy feature-faults relationship matrix for a steam turbine developed using fuzzy

3 Key Engineering Materials Vols mathematics. The table relates the typical twenty steam turbine faults with ten vibration spectrum features. The alphabetic symbols used to describe the spectrum and process features are listed in Table 2. The notation n X in the second column of table 2 denotes a frequency component (or range) in the spectrum at n times the turbine s rotational speed. Table 2. Symbols for vibration frequency and process feature description. Frequency Description Process Description feature feature f1 0.01~0.39X P1 Amplitude jump during operation f2 0.4~0.49X P2 vibrations at various power load f3 0.50X P3 axial vibration f4 0.51~0.99X P4 shaft average centerline f5 1 X P5 critical speed spectrum f6 2 X P6 stable at various running speeds f7 3~5 X P7 vibration level increase during running up f8 odd of X P8 level jump during run up f9 high X P9 3x at 1/3 critical speed f10 Power line P10 half-speed whirl Table 3. Spectrum feature-fault relationship chart. Faults f1 f2 f3 f4 f5 f6 f7 f8 f9 f10 F F F F F F F F F F F F F F F F F F F F The relationships listed in Table 3 show that some faults such as an imbalance, F1, and a bent shaft, F3, can not be distinguished since they have similar spectrum features. Therefore, a second relationship matrix given in Table 4 is used to relate the process features to the steam turbine faults. Table 4 was derived directly from the author and other field expert experience, so it can be used to efficiently diagnose faults. The two relationship charts in Tables 3 and 4 provide the basis for steam

4 178 Damage Assessment of Structures VI turbine fault diagnosis. Table 4. Process feature- fault relation chart. Fault P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 F0 N N L L N N Y N N N F1 N N L L N N Y N N N F2 Y N L L N N N N N N F3 N Y L L N N N N N N F4 N N L L N N N N N N F5 N N H L N N N N N N F6 N N H L N N N N N N F7 N N L H N N N N N N F8 N N L L N N N N Y N F9 N N L L N N N Y N N F10 N N M L N N N P N N F11 N N L L N N N Y N N F12 N N L L N N N N N N F13 N N L L N N N Y N N F14 N N L L N N N N N N F15 N N L L N N N N N Y F16 N N L L Y N N N N N F17 N P L L P N N N N N F18 N N L L N Y N N N N F19 N N L L N N N N N N Y=Yes, N=No, L=Low, M=Middle, H=High, P=Possible Decision Tree Classifier Methods Decision trees are based on the following terminology. (1) A decision tree is a flow chart or diagram representing a classification system or a predictive model. The tree is structured as a sequence of simple questions with the answers to those questions tracing a path down the tree. (2) The end product is a collection of hierarchical rules that segment the data into groups, where a decision (classification or prediction) is made for each group. (3) The hierarchy is called a tree, and each segment is called a node. (4) The original segment containing the entire data set is referred to as the root node of the tree. (5) A node with all of its successors forms a branch of the node that created it. (6) The final nodes are called leaves. A decision made at each leaf is applied to all observations in the leaf. The depth of a node in a tree is the path length from the root to the node. The height of a node in a tree is the largest path length from the node to a leaf. The height of a tree is the height of its root. The Ctree software [4] was used for the fault diagnostics because the software is easy to use and the data preparation is quite straight forward. Ctree is based on C4.5 algorithm. The Node Splitting Criterion calculated the entropy to select the split between branches. While growing the tree, a predictor is chosen at any point to split a node so that the information gain is maximized after the split. The C4.5 algorithm actually uses the gain ratio (= Gain / Split Information) to select the split.

5 Key Engineering Materials Vols The Stopping Criteria stops the node splitting and identifies the node as a leaf node if any one of the following criterion is met: (1) The number of records in the node is less than some pre-specified limit. (2) The node purity is more than some pre-specified limit p which means that the proportion of records in the node with class equal to the majority class is p or more. (3) The node depth is more than some pre-specified limit. (4) The predictor values for all the records are identical. Tree Pruning is based on the pessimistic error rate at the node. Each node has a 50% error rate confidence interval with its upper limit taken as the pessimistic error rate. If the pessimistic error rate of a node is less than that of the subtree rooted at that node, the node is pruned. Rule Generation is based on final tree geometry. The path from root to each leaf node gives a rule for that leaf node. Thus, a tree with k leaf nodes has a set of k rules. Then individual rules are pruned by dropping clauses one by one from that rule. The decision to drop a clause is based on the outcome of a statistical independence test. The test evaluates whether keeping a clause is independent of the final decision of the rule. If it is independent, then the clause is NOT contributing towards the final decision and it is dropped to simplify the rule. The independence tests are based on the Chi-square test and Fisher's exact test. Table 5. Classification results for various purities. Maximum purity Training set Test set misclassification rate misclassification rate 100% 0% 1.8% 95% 0% 1.73% 90% 0% 1.6% 80% 0% 1.33% 75% 0% 0.50% 70% 0% 1.19% 65% 0% 1.9% Application of data mining to steam turbine fault diagnosis A numerical simulation was developed based on the two relationship matrices in Tables 3 and 4 to test the decision tree classifier. The simulation firstly generated one hundred data points including spectrum features and process features for each type of steam turbine fault for a total of 2000 data points. Then, the data set was randomly divided into training and test sets. Next, the Ctree software was used to analyze the data set and to grow the decision tree. The pruning technique was used to generate a stable tree. The maximum purity of the tree was adjusted to get better results. For example, for a maximum purity rate of 100%, the misclassification rate for the training set was 0% and for the test set was 1.8%. However, when maximum purity was reduced to 75%, the misclassification rate for test set was reduced to 0.50%. Table 5 lists the results for various purities.

6 180 Damage Assessment of Structures VI Table 6. Classification Tree Information for a purity of 75%. Tree Information Item Value Tree Information Item Value Number of Training observations 993 Total Number of Nodes 40 Number of Test observations 1007 Number of Leaf Nodes 22 Number of Predictors 20 Number of Levels 12 Class Variable Faults Training Data Misclassification rate 0.00% Number of Classes 20 Test data Misclassification rate 0.50% Table 7. Test set results. Rule ID Fault Class Length Support Confidence Capture 1 F % 100.0% 100.0% 2 F % 100.0% 100.0% 3 F % 100.0% 100.0% 4 F % 100.0% 100.0% 5 F % 100.0% 100.0% 6 F % 100.0% 100.0% 7 F % 100.0% 100.0% 8 F % 100.0% 100.0% 9 F % 100.0% 100.0% 10 F % 100.0% 100.0% 11 F % 100.0% 100.0% 12 F % 93.8% 100.0% 13 F % 100.0% 100.0% 14 F % 98.2% 100.0% 15 F % 100.0% 100.0% 16 F % 94.3% 100.0% 17 F % 100.0% 100.0% 18 F % 100.0% 100.0% 19 F % 100.0% 100.0% 20 F % 100.0% 100.0% Simulation Results Tables 6, 7, and 8 list the classification results for the simulated steam turbine faults data. Table 6 describes the resulting decision tree for a maximum purity of 75%. The misclassification rate is sufficiently low for common engineering applications. The decision tree was then used to develop the if-then rules used by engineers to analyze and interpret the fault diagnosis results. The method can automatically extract the knowledge from the data as part of a fault diagnosis expert system. Table 7 summarizes the rule results for the test set, including the support, confidence and capture rates. The support rate measures how widely applicable the rule is in the training set. The confidence rate measures the accuracy of the rule. The capture indicates how many records of a fault were correctly captured by the rule. The twenty rules after pruning correspond to the twenty types of faults. Most of the confidence rates were 100%, with only 3 confidence rates less than 100% due to misclassification of the test data. Table 8 lists the specific rules for each fault type. The rules agree well with spectrum analysis theory. In addition, many process features from the field

7 Key Engineering Materials Vols experts experience are integrated into the rules to improve the classification process. Table 8. Rules derived from the classification tree. Rule IF Then 1 f F14 2 f3 <.1051, P9 = Y F8 3 f3 <.1051, f , P9 = N F6 4 f , f3 <.1051, f6 <.47008, P7 = Y, P9 = N F1 5 f1 <.84369, f3 <.1051, f , f6 <.47008, P6 = N, P9 = N F10 6 f1 < ,f3 <.1051, f6 <.47008, P6 = Y, P9 = N F18 7 f1 <.84369, f , f3 <.1051, f6 <.47008, P6 = Y, P9 = N F0 8 f , f3 <.1051, f6 <.47008, P1 = N, P7 = N,P9 = N F3 9 f , f3 <.1051,f6 <.47008,P1 = Y,P7 = N,P9 = N F2 10 f1<.18795,f2< ,f3<.1051,f5<.47041,f6<.47008, P6=N, P9=N F19 11 f1<.8437,f1.188,f2.186,f3<.105,f5<.47041,f6<.47008,p6=n,p9=n F12 12 f2 <.37561, f , f5 < F17 13 f1<.188,.009 f2<.20,f3<.105, f5<.470,f6<.470,p6=n,p9=n F11 14 f , f5 <.47041, f F f1<.395,f2 <.0947, f3<.105,f5<.47041, f6<.47008, P6=N,P9=N F13 16 f1 <.84369, f F f1<.84,f2<.19,f2.095,f3<.11,f5<.470,f6<.47,p3=h,p6=n,p9=n F f1<.84,f2 <.186,f2.095,f3<.11,f5<.47,f6<.47,P3=L,P6=N,P9=N F4 19 f1<.18795,f ,f3<.1051,f5< ,f6<.470,p5=n,p6=n,p9=n F15 20 f1<.18795,f2.3756,f3<.105,f5<.09472,f6<.470,p5=y,p6=n,p9=n F16 Application to power plant data The rules automatically extracted using the simulated data mining accurately summarize the rules for faults in rotating machinery such as steam turbines. The method was tested with real data by implementing the program as part of a power plant steam turbine diagnostic system. The original data is from remote online monitors and diagnostic systems installed in several power plants of the Shandong Electricity Power Company, China [5]. The data was measured by vibration transducers from the Bently Nevada Company. The data was then transformed to the frequency domain using Fast Fourier Transforms. Cases 1-4 are for various bearings at two different 300 MW units. The data extracted from the spectra was analyzed using the fuzzy logic algorithm used to generate Table 3. After the process data was added into the data sets, the data was analyzed using the rules extracted from the training set. The diagnostic results are listed in Table 9 which lists the faults identified from the spectrum data by the fuzzy logical algorithm. Table 10 lists the process data and diagnosis results. Case 1 is due to from steam excited vibration. Case 2 is due to axial misalignment. Case 3 is due to electrical-magnet excitation. Case 4 indicates a normal operating condition. The diagnostic results were then used to identify measures to suppress the vibrations. The results demonstrate that these methods can be effectively applied steam turbine fault diagnosis in actual power plants.

8 182 Damage Assessment of Structures VI Table 9. Spectrum feature data. Case No. f1 f2 f3 f4 f5 f6 f7 f8 f9 f10 Case Case Case Case Table 10. Process data and classification results. Case No. P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Predicted faults Case 1 N Y L L N N N N N N F17 Case 2 N N H L N N N N N N F6 Case 3 N N L L N N N N N N F13 Case 4 N N L L N N Y N N N F0 Conclusions Data mining method was used to classify simulated data and real data into known classes for steam turbine fault diagnostics. The use of the simulated data enabled the system to directly capture the field experts knowledge into the resulting classification rules. The classification rules were automatically extracted from the data sets for use by engineers to diagnose and interpret steam turbine faults. The simulation results and the results using actual data from operating power plants shows that the data mining methods can be effectively applied to steam turbine fault diagnostics. The automatic extraction of the classification rules shows that these machine learning methods can be applied to large turbo-machinery databases and can include engineering knowledge and field experience. The results can then be used for fault diagnosis of large rotating machines, such as steam turbines. Reference [1] I.H. Witten and E. Frank: Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations (Academic Press. USA, 2000). [2] S.R. Safavian and D. Landgrebe: IEEE Transactions on Systems, Man, and Cybernetics Vol. 21 (1991), p. 660 [3] V. Crupi, E. Guglielmino and G. Milazzo G: Journal of Vibration and Control Vol. 10 (2004), p [4] A. Saha: Ctree in Excel. http: // [5] D. Jiang, H. Sun and X. Zhan: 5th International Conference Acoustical and Vibratory Surveillance Methods and Diagnostic Techniques, France (2004)