KR20120027733A - Rotating machinery fault diagnostic method and system using support vector machines - Google Patents

Abstract

PURPOSE: A method and an apparatus for diagnosing the defects of rotation machinery using a support vector machine is provided to enable a person unskilled to diagnose defects on rotation machinery using a specific vector and SVM algorithm. CONSTITUTION: A method for diagnosing the defects of rotation machinery using a support vector machine is as follows. Signals are measured using multiple acceleration sensors and displacement sensors installed in components of rotation machinery(20). The signals are pre-processed to extract specific values from the measured signals. Specific vectors are computed from the extracted specific values(30). It is diagnosed whether the rotation machinery has defects or not using the specific vectors and SVM algorithm(80).

Description

Rotating Machinery Fault Diagnostic Method and System Using Support Vector Machines

According to the present invention, signals detected by a plurality of acceleration sensors and displacement sensors installed in components of a rotating machine to be diagnosed are input, and the input signals are preprocessed through frequency conversion or wavelet conversion, and the like. A method and apparatus for diagnosing a defect of a rotating machine capable of constructing a feature vector from the values and classifying the normal or multiple abnormal defects of the rotating machine from a pre-trained SVM classification model using the feature vector and the support vector machine algorithm. will be.

Rotating machines such as motors, pumps, blowers, turbines, etc. experience various abnormal conditions depending on the design characteristics, installation conditions, operating conditions, as well as the number of years of operation, especially due to eccentricity, misalignment and poor bearing performance due to mass unbalance. Vibration is triggered, which can lead to reduced performance and shortened life of rotating machinery, which can, in severe cases, even shut down power plant systems in which the rotating machinery is installed.

Therefore, in order to maintain optimum performance for major rotating machines, tests, inspections, and maintenance are conducted at regular intervals, but predictive maintenance and condition diagnosis are essential.

In general, the diagnosis of a rotating machine diagnoses the state of the device through signal processing and analysis such as vibration and sound emission. However, fault diagnosis requires a certain level of expertise and diagnostic experience and is not an area that non-experts can easily diagnose. In addition, since it may be difficult to distinguish between normal and defective states only by measuring signals, signal analysis through frequency conversion or wavelet conversion is generally required.

In addition, there are many techniques for diagnosing abnormalities by using pattern data such as pattern recognition using characteristic data such as signal average values for the diagnosis of rotating machines. Even in this case, a method using the same feature data is used.

The pattern recognition technique used in the conventional method uses a fuzzy, neural network, or K-means clustering algorithm, but in this case, the SVM (Support Vector Machine) method is known to be much better in terms of binary classification and generalization.

In addition, it is very important to use the feature data because the diagnosis result by the pattern recognition method is influenced not only by the performance of the pattern recognition method itself but also by the input feature data.

The problem to be solved by the present invention is one or more pre-trained SVM classification model using SVM algorithm that shows excellent performance in terms of binary classification and generalization and the signal detected by the sensors installed in the components of the rotating machine to be diagnosed. And classify one of the normal or abnormal defects of the rotating machine by using the optimal feature data of the SVM classification model.

Another problem to be solved by the present invention consists of a plurality of SVM classifiers that classify one of the defect types belonging to each defect type group by using a feature vector and an SVM algorithm consisting of feature values obtained from the measured signals. Even if you are not an expert, it is easy to diagnose, which is to reduce the maintenance cost.

The problem solving means of the present invention comprises the steps of measuring the respective signals from a plurality of acceleration sensors and displacement sensors installed in the components of the rotating machine to be diagnosed, and preprocessing the measured signals through frequency conversion or wavelet conversion, and the like; And generating a feature vector from the feature values of the preprocessed signals, and diagnosing a normal or multiple abnormal defects of the machine from the pre-trained SVM classification model using the feature vector and the SVM algorithm. To provide a method for diagnosing defects.

Still another object of the present invention is a plurality of acceleration sensors and displacement sensors for acquiring a signal from the components of the rotating machine to be diagnosed, frequency conversion or wavelet conversion of the signals measured by the plurality of acceleration sensors and displacement sensors Means for preprocessing, means for obtaining a feature vector comprising the feature values of the preprocessed signals, and normal or multiple abnormalities of the rotating machine from the SVM classification model previously learned using the obtained feature vector and SVM algorithm. The present invention provides a defect diagnosis device for a rotating machine comprising means for diagnosing a defect.

Another problem solving means of the present invention consists of one SVM classifier classifying one of the normal or abnormal defect types through a plurality of pre-trained SVM classification model, or when classified as the abnormal state, two or three defect types A defect diagnosis method of a rotating machine comprising an SVM classifier classified into one of a plurality of defect type groups, and a plurality of SVM classifiers classified into one of a defect type belonging to each defect type group when classified into each defect type group, and To provide a device.

The present invention utilizes one or more SVM classification models and SVM classification models that have been learned in advance using signals detected by sensors installed in the components of a rotating machine to be diagnosed, and SVM algorithms that exhibit excellent performance in terms of binary classification and generalization. By using the optimal feature data of the classification of one of the normal or abnormal defects of the rotating machine has an advantageous effect that can easily determine the normal and abnormal defects.

Another effect of the present invention consists of a plurality of SVM classifiers that classify one of the defect types belonging to each defect type group by using a feature vector and an SVM algorithm consisting of feature values obtained from the measured signals. If not, it can be easily diagnosed, thereby reducing maintenance costs.

1 is a diagram illustrating a support vector machine (SVM) algorithm.
2 is an exemplary view showing an example of a defect diagnosis method of a rotary machine to which the diagnostic method of the present invention is applied.
3 is an exemplary diagram of an SVM classification model using one SVM classifier.
4 is an exemplary diagram of an SVM classification model using a plurality of SVM classifiers.
5 is a graph showing measurement data detected from a sensor installed on one side of a rotating machine having normal and abnormal defects.
6 is an exemplary diagram of characteristic values of measurement data detected from a sensor installed on one side of a rotating machine simulating steady state and ten defect types.
7 is a diagram illustrating an example of a defect diagnosis apparatus screen according to the present invention.
Explanation of the Main References
10, 11: data with class information
12: Support Vector
13, 14: hyperplane example to classify two classes
15: input data
20: input signal (measured value)
30: feature extraction and feature vector construction step
40: determining optimal feature 50: SVM learning step
60: input signal for diagnosis (measured value) 70: optimal feature extraction step
80: fault diagnosis step 100: SVM classification model
111, 121 ~ 126: SVM classifier

Hereinafter, the present invention will be described in detail. Defect diagnosis method of a rotating machine according to the present invention comprises the steps of measuring the respective signals from a plurality of acceleration sensors and displacement sensors installed in the components of the rotating machine to be diagnosed, and through the frequency conversion or wavelet conversion Pretreatment.

The method may further include generating a feature vector from feature values of the preprocessed signals, and diagnosing a normal or multiple abnormal defects of a rotating machine from a pre-trained SVM classification model using the feature vector and the SVM algorithm. .

The defect diagnosis apparatus of a rotary machine according to the present invention is a frequency conversion or wavelet for the signals measured by a plurality of acceleration sensors and displacement sensors, and a plurality of acceleration sensors and displacement sensors for obtaining signals from the components of the rotary machine to be diagnosed Means for preprocessing through conversion or the like.

In addition, means for obtaining a feature vector consisting of the feature values of the pre-processed signals, and means for diagnosing the normal or multiple abnormal defects of the rotating machine from the previously learned SVM classification model using the obtained feature vector and SVM algorithm It consists of. It looks at a specific embodiment of the present invention.

<Examples>

&Lt; Example 1 >

A first embodiment according to the present invention will be described with reference to the drawings. Embodiment 1 relates to a method for diagnosing a defect of a rotating machine according to the present invention. 1 is a diagram illustrating a support vector machine (SVM) algorithm. Figure 2 shows one example of a defect diagnosis method of a rotary machine to which the diagnostic method of the present invention is applied.

Defect diagnosis method of a rotating machine according to the present invention comprises the steps of measuring the respective signals from a plurality of acceleration sensors and displacement sensors installed in the components of the rotating machine to be diagnosed, and through the frequency conversion or wavelet conversion Pretreatment.

The method may further include generating a feature vector from feature values of the preprocessed signals, and diagnosing a normal or multiple abnormal defects of a rotating machine from a pre-trained SVM classification model using the feature vector and the SVM algorithm. .

Each technical configuration according to the present invention will be described in detail.

FIG. 1 illustrates the concept of classifying into two classes using SVMs. The hyperplane capable of distinguishing the two classes 10 and 11 has many more than the ultraplanes 13 and 14 shown in FIG. 1 but has a maximum margin. 14 is unique.

Basically, SVM deals with the binary classification problem and is divided into two classes (10, 11) with a hyperplane 14 defined by a support vector (12). Here, the support vector 12 is the data closest to the boundary among the data belonging to each class.

When the support vector 12 is determined by using the learning data 10 and 11 for which the class is known, the data other than the support vector 12 is no longer used and the data 15 that does not know the class is supported by the support vector 12. It is classified using the hyperplane 14 composed of information.

Nonlinear input space data that cannot be easily separated linearly is linearized by converting the data into a high-dimensional feature space using a kernel function. Commonly used kernel functions include one of RBF kernel, polynomial kernel, linear kernel, and sigmoid function.

2 illustrates an example of diagnosing a defect using an SVM.

In FIG. 2, a step 30 of calculating and extracting features of the defect information from the input signal 20 having defect information and generating a feature vector composed of the class of the corresponding defect information and the features of the extracted defect information is included.

The SVM classification model 100 classifying into one of the 11 defect type classes including the steady state and the configured feature vector are input to determine the optimal feature for the SVM classification model 100 (40).

After the step 50 of learning using the optimal feature vector and the SVM algorithm composed of the optimal feature and the defect type information, the SVM classifier using the information corresponding to the optimal hyperplane 14 shown in FIG. 1 may be implemented.

The SVM classifier implemented in accordance with the present invention utilizes the signal detected by the sensor installed in the components of the rotating machine by using the optimal characteristic information 40 for the steady state and various combination types of defect states stored in the database or memory of the SVM classifier. 60, if the optimal feature vector is extracted and the optimal feature vector is obtained (70), it is possible to diagnose 80 whether the rotating machine is normal or defective from the SVM classifier using the optimal feature vector as input data.

The SVM classification model 100 will be described in more detail with reference to FIGS. 3 and 4 as follows. FIG. 3 shows an example of an SVM classification model classifying one SVM classifier 111 into one of ten defect types (C2 to C11), that is, eleven classes, in a steady state (C1) and an abnormal state.

In Fig. 3, the ten defect types in an abnormal state are rotor mass unbalance (C2), bearing inner ring defect (C3), bearing outer ring defect (C4), bearing ball defect (C5), rotor contact wear type 1 (C6). , Rotor contact wear type 2 (C7), rotor axle bending (C8), rotor shaft crack (C9), bearing housing loosening (C10) and coupling defect (C11).

The optimal feature is determined in such a way that the difference between the feature values of 11 classes is the maximum and the difference between the feature values in the same class is the minimum.

4 illustrates another embodiment of an SVM classification model for classifying one of eleven classes (C1 to C11) from a total of six SVM classifiers for eleven classes (C1 to C11) including a steady state.

The SVM1 classifier 121 classifies the steady state C1 and the abnormal state C2 to C11, and is classified into one of four defect type groups by the SVM2 classifier 122 classified as an abnormal state. That is, classified into one of the defect types of {C2, C10, C11}, {C3, C4, C5}, {C6, C7}, {C8, C9}, and according to the classification result SVM3 classifier 123, SVM4 classifier 124, the SVM5 classifier 125, and the SVM6 classifier 126 are selected and finally diagnosed as one of the defect types belonging to the defect type group.

Here, the four defect type groups are presented as an example, and the SVM classification model 100 may be configured by dividing into various methods by applying the aforementioned method.

In the SVM classification model 100, the optimal feature is the same as described above, the difference in the feature values of the class of the defect type class of each SVM classifier is the maximum, the difference in the feature value in the same class is the minimum Are each determined.

5 is an example of a vibration signal obtained from an experimental apparatus that simulates a defect of a rotating machine having a steady state C1 and four defect types C2, C3, C4, and C8.

Each defect type was measured 5 times (marked as D1, D2, D3, D4, D5), and the defect types simulated by the experimental apparatus in addition to the normal state were described above as mass unbalance (C2) and bearing inner ring defect (C3). , Bearing outer ring defect (C4), bearing ball defect (C5), rotor contact wear type 1 (C6), rotor contact wear type 2 (C7), rotor axle bending (C8), rotor shaft crack (C9), bearing There are 10 types of housing loosening C10 and coupling defect C11.

6 is a graph showing one example of feature data for eleven classes with the acquired vibration signal as an input. In Example 1 according to the present invention, the average value of the measurement signal, RMS, skewness, kurtosis, maximum peak value, 5th to 9th order central moment, absolute mean, crest factor, shape factor Signals with 18 characteristics were extracted: impulse factor, clearance factor, entropy, modified entropy, and mean value divided by standard deviation.

In Fig. 6, the measurement data (denoted as D1 to D5) of the total number of times of measurement is divided into 20 signals each consisting of 9,500 sample data, each having 18 characteristics from the equalized signal (feature number 1 in Fig. 6). To 18) was extracted.

In Fig. 6, the number on the horizontal axis means each defect type, and 1 is the steady state C1. Referring to FIG. 6, features 1, 2, 11, 17, and 18 relatively distinguish each defect type, and features 3 to 9 of FIG. 6 distinguish defect 3 from other defects relatively well. It can be seen that there is no difference in characteristics with respect to the other defects except the third defect C3. Therefore, it is difficult to expect good SVM classifier performance when only 3 to 9 features are used to classify 11 defect types.

Table 1 shows five optimal feature numbers for the SVM classifier of FIG. 3 and the six SVM classifiers of FIG. 4. Where D1 to D5 are feature data sets.

Looking at Table 1, it can be seen that the five optimal feature numbers for each SVM classifier are different, and it can be seen that there are also differences from the five optimal feature numbers of the SVM classifier of FIG. 3.

Classifier D1 D2 D3 D4 D5 SVM1 (Figure 4) 18,1,13,12,14 18,1,13,12,14 18,1,12,14,15 18,1,13,12,14 18,1,13,12,14 SVM2 (Figure 4) 17,18,10,16,2 17,16,11,2,10 17,10,2,11,16 17,18,10,16,2 17,16,11,2,10 SVM3 (Figure 4) 1,18,16,2,11 1,18,16,11,2 1,18,11,16,2 1,18,16,2,11 1,18,16,11,2 SVM4 (Figure 4) 11,17,2,13,1 11,2,17,13,1 17,11,2,13,1 11,17,2,13,1 11,2,17,13,1 SVM5 (Figure 4) 1,13,17,4,15 1,18,13,17,16 1,18,17,16,13 1,13,17,4,15 1,18,13,17,16 SVM6 (Figure 4) 17,11,16,2,1 11,17,2,16,1 17,11,2,16,1 17,11,16,2,1 11,17,2,16,1 SVM (Figure 3) 1,17,18,11,2 1,11,18,17,2 1,17,18,11,2 1,18,11,2,17 1,18,17,11,2

Defect diagnosis was performed according to the SVM classification model shown in FIGS. 3 and 4 by using five data (D1 to D5) measured for each defect type. The SVM classifiers were trained using the 220 × 18 feature data sets for each of the five signals (D1 to D5), and the 220 × 18 feature data sets were verified using four signals that were not used for learning for D1 to D5. It was.

For example, in Table 1, when D3 is used as training input data, the verification data is D1, D2, D4, and D5. The classification results for 11 classes according to the SVM classifier of FIG. 3 are shown in Table 2, and the results using the SVM classifiers of FIG. 4 are shown in Table 3. FIG.

In Tables 2 and 3, when the number of features is 1 and the learning feature data set is D1, the feature 1 of FIG. 6 is used when the SVM classifier of FIG. 3 is used and the six SVM classifiers of FIG. 4 are used. It means that the features 18, 17, 1, 11, 1, 17 of FIG. 6 are used, respectively. Up to five optimal feature numbers are shown in Table 1.

Number of features D1 D2 D3 D4 D5 One 39.3 59.2 64.6 61.9 60.7 3 73.8 91.3 88.8 83.6 83.5 5 78 91 88.4 85.8 83.3 10 84.7 93.5 91.1 87 87.2

Number of features D1 D2 D3 D4 D5 One 66.6 63.8 69.2 72.3 77.6 3 78.6 83.6 77.7 77.4 78.6 5 86.6 84.6 78.9 77.9 79.1 10 84.3 94 91.7 87.6 88

In Tables 2 and 3, when only one optimal feature data for each defect type is used, the SVM classifier of FIG. 3 shows the lowest classification performance of 59.2% with the exception of D1. However, the SVM classifier of FIG. When used, it shows more than 63.8% classification performance.

When the number of input feature data is 3 or 5, except that D1 is used as learning data, the classification performance of the SVM classifier of FIG. 3 is better. However, when the number of input feature data is 10, FIG. Using the SVM classifier of 4 gives better results.

In comparison with the learning time measured by the SVM classification model of FIG. 4 with respect to the learning time measured by the SVM classification model of FIG. 3 shown in Table 4, in both cases, there was almost no difference in learning time according to the number of feature data.

Since the learning time measured by the SVM classification model of FIG. 4 is short as 60 to 70%, the learning time measured by the SVM classification model of FIG. 3 is short. It can be seen that similar classification performance can be obtained.

Number of features D1 D2 D3 D4 D5 One 0.58 0.59 0.66 0.63 0.65 3 0.66 0.61 0.65 0.6 0.71 5 0.69 0.63 0.64 0.62 0.64 10 0.68 0.64 0.62 0.69 0.65

FIG. 7 illustrates an embodiment of a screen for diagnosing a defect in accordance with the present invention, and is configured to display brief information on an SVM classifier as in ① of FIG. From the features and the number of features extracted from the display, such as (3) of Figure 7 can be visually represented by the defect diagnosis results for each defect type.

In addition, as shown in ④ of FIG. 7, an input signal for diagnosis can be read, a diagnosis can be performed, and a control such as storing a diagnosis result can be configured.

<Example 2>

It looks at Example 2 according to the present invention. Embodiment 2 relates to a defect diagnosis apparatus for a rotating machine according to the present invention.

The defect diagnosis apparatus of the rotary machine according to the present invention is implemented to perform the defect diagnosis method of the rotary machine of the first embodiment, and a program for carrying out the defect diagnosis method of the rotary machine is mounted.

More specifically, the defect diagnosis apparatus of the rotary machine according to the present invention is a plurality of acceleration sensors and displacement sensors, one of the acceleration sensor and displacement sensors installed on one side of the rotary machine to obtain a signal from the components of the rotary machine to be diagnosed Means for preprocessing the signals measured by the sensor through frequency conversion or wavelet conversion, and the like.

In addition, Example 2 includes a means for obtaining a feature vector consisting of feature values of the preprocessed signals, and the normal or multiple abnormal defects of the rotating machine from the SVM classification model previously learned using the obtained feature vector and the SVM algorithm. It consists of means for diagnosing.

For the technical configuration described in Embodiment 1, description thereof is omitted in Embodiment 2 in order to avoid repeated description.

Embodiments of the present invention described above have been given specific examples for clarity of understanding, and the scope of protection of the present invention is not limited to the above embodiments, and various other modifications based on the technical spirit of the present invention are possible. It will be apparent to those skilled in the art to which the present invention pertains.

According to the present invention, signals detected by a plurality of acceleration sensors and displacement sensors installed in components of a rotating machine to be diagnosed are input, and the input signals are preprocessed through frequency conversion or wavelet conversion, and the like. A method and apparatus for diagnosing a defect of a rotating machine capable of classifying a feature vector with values and classifying the normal or multiple abnormal defects of the rotating machine from a pre-trained SVM classification model using the feature vector and the support vector machine algorithm. Therefore, even if you are not an expert in diagnosis of rotating machine, it is easy to determine the steady state and abnormal type of defect information of rotating machine.

Claims (14)

In the fault diagnosis method of the rotating machine,
Measuring a signal by a plurality of acceleration sensors and displacement sensors installed on the components of the rotating machine to be diagnosed;
Preprocessing the measured signals to obtain a feature value;
Generating a feature vector from feature values of the preprocessed signals; And
And diagnosing a steady state or a plurality of abnormal defects of the rotating machine from the previously learned SVM classification model using the feature vector and the SVM algorithm. The method according to claim 1,
The pre-learned SVM classification model is a fault diagnosis method of a rotating machine, characterized in that it is composed of one SVM classifier that is classified as one of the normal state or abnormal defect type. The method according to claim 1,
The abnormal abnormality includes rotor mass unbalance (C2), bearing inner ring defect (C3), bearing outer ring defect (C4), bearing ball defect (C5), rotor contact wear type 1 (C6), rotor contact wear type 2 (C7), rotary axle bending (C8), rotor shaft crack (C9), bearing housing loosening (C10) and coupling defect (C11). The method according to claim 3,
The C2 to C11 are classified into four types of defect types by {C2, C10, C11}, {C3, C4, C5}, {C6, C7}, and {C8, C9} according to defect types, and among the classified defect types. Method for diagnosing faults of rotating machines, characterized by diagnosing one. The method according to any one of claims 1 to 3,
The feature value extraction of the signal includes the average value of the measured signal, RMS, skewness, kurtosis, maximum peak value, center moments from 5th to 9th order, absolute mean, crest element, shape factor, impulse element, clearance element, entropy, A method for diagnosing defects of a rotating machine, characterized in that it extracts by selecting one of the modified entropy and the mean divided by the standard deviation. The method according to any one of claims 1 to 3,
Wherein the pre-processing step is a fault diagnosis method of a rotating machine, characterized in that through the frequency conversion or wavelet conversion. The method according to any one of claims 1 to 3,
The optimal feature vector used in the SVM classification model is determined in such a way that the difference in feature values for each classification of defect types or types of defects of a rotating machine is maximum, and the difference in feature values in the same type of defect type or defect type group is minimized. Defect diagnosis method of the rotating machine, characterized in that. In the defect diagnosis device of the rotating machine,
A plurality of acceleration sensors and displacement sensors for obtaining signals from components of the rotating machine to be diagnosed;
Means for preprocessing signals measured by the plurality of acceleration sensors and displacement sensors;
Means for obtaining a feature vector consisting of feature values of the preprocessed signals; And
And a means for diagnosing a normal or a plurality of abnormal defects of the rotating machine from the SVM classification model previously learned using the obtained feature vector and the SVM algorithm. The method according to claim 8,
The pre-learned SVM classification model is a fault diagnosis apparatus of a rotating machine, characterized in that it consists of one SVM classifier for classifying the classification as one of the steady state or abnormal defect types. The method according to claim 8,
The abnormal abnormality includes rotor mass unbalance (C2), bearing inner ring defect (C3), bearing outer ring defect (C4), bearing ball defect (C5), rotor contact wear type 1 (C6), rotor contact wear type 2 (C7), rotary axle bending (C8), rotor shaft crack (C9), bearing housing loosening (C10) and coupling defect (C11). The method according to claim 10,
The C2 to C11 are classified into four types of defect types by {C2, C10, C11}, {C3, C4, C5}, {C6, C7}, and {C8, C9} according to defect types, and among the classified defect types. A fault diagnosis device for a rotating machine, characterized in that configured to diagnose as one. The method according to any one of claims 8 to 11,
The feature value extraction of the signal includes the average value of the measured signal, RMS, skewness, kurtosis, maximum peak value, center moments from 5th to 9th order, absolute mean, crest element, shape factor, impulse element, clearance element, entropy, A device for diagnosing defects of rotating machines, characterized in that it is extracted by selecting one of the modified entropy and the mean value divided by the standard deviation. The method according to any one of claims 8 to 11,
The pre-processing step is a fault diagnosis apparatus for a rotating machine, characterized in that the frequency conversion or through the wavelet conversion. The method according to any one of claims 8 to 11,
The optimal feature vector used in the SVM classification model is determined in such a way that the difference in feature values for each classification of defect types or types of defects of a rotating machine is maximum, and the difference in feature values in the same type of defect type or defect type group is minimized. Defect diagnosis device of the rotating machine, characterized in that.

Images