KR20220095631A - Gear pump failure classification method using Support Vector Machines - Google Patents

Abstract

A gear pump failure classification method using support vector machines includes: a step of separating vibration data of the gear pump into a certain section to analyze a spectrum for each section and extract a feature by using a mel-frequency cepstral coefficient (MFCC) algorithm; and a step of classifying the extracted feature into normality or failure by using a support vector machine technique.

Description

Gear pump failure classification method using Support Vector Machines

The present invention relates to a method for classifying a failure of a mechanical device, and more particularly, to a method for classifying a failure of a gear pump using a support vector machine (SVM).

Hydraulic systems are widely used in modern industries because they can obtain a large force even in a relatively small size. Among them, gear pumps have a simple structure, low cost, and high efficiency through high-speed motion, so they are widely used in various industrial fields such as forklifts, construction machinery, hydraulic brakes, and steering systems. A lot of effort is going on.

In particular, failure analysis tends to separate failure data by applying various machine learning techniques based on operational data, typically ANN (Artificial Neural Networks), SVM (Support Vector Machines), PCA (Principal Component Analysis), etc. There is this.

The dual SVM technique was developed from statistical learning theory in the 1960s, and it has strengths in low-dimensional data or noise data because high-dimensional mapping of input features using a kernel is possible. It is known that the SVM technique is suitable for fault diagnosis of rotating bodies.

KR 10-2019-0062739 A

The present invention has been proposed to solve the above technical problems, and provides a method for classifying a failure of a gear pump capable of predicting failure of a gear pump in advance using an MFCC algorithm and an SVM technique.

According to an embodiment of the present invention for solving the above problem, the vibration data of the gear pump is divided into predetermined sections using a Mel-Frequency Cepstral Coefficient (MFCC) algorithm, and the spectrum is analyzed for each section to extract the features. and classifying the extracted features as normal or faulty using a support vector machine (SVM) technique is provided.

In addition, in the step of extracting the features, after calculating the pre-emphasis for each frame of vibration data of the gear pump, the data is set to the level of 20-40 ms frame through a Hamming window. And, it is characterized in that the FFT (Fast Fourier Transform) is calculated for the divided frames and passed through a Mel filter bank as in Equation (1).

<Equation 1>

- f is the frequency

In addition, the signal data that has passed through the Mel Filter Bank is logarithm-processed, DCT (Discrete Cosine Transform) is calculated with the logarithm-processed data, and features are extracted by applying the MFCC algorithm. characterized in that

In addition, the SVM (Support Vector Machine) is characterized in that any one of Optimazed Linear Kernel, Linear Kernel, Gaussian Kernel, and Polynomial Kernel is applied as a kernel function.

The failure classification method of the gear pump according to an embodiment of the present invention can predict the failure of the gear pump in advance by using the MFCC algorithm and the SVM technique. That is, by predicting the failure of the gear pump in advance, it is expected that the operation rate of equipment can be improved and maintenance costs can be reduced.

An exemplary view of the gear pump of FIG. 1
2 is a view showing the severity and frequency of failure on the system when a failure occurs;
3 is a view showing FTA (Fault Tree Analysis);
4 is a configuration diagram of a vibration test device of a gear pump
5 is a hydraulic circuit diagram for FIG. 4
6 is a view showing the state before and after the test of the gear pump
7 is a photograph showing abrasion caused by foreign substances entering the hydraulic oil of a gear pump used in an actual forklift.
8 is an exemplary diagram of an MFCC (Mel-Frequency Cepstral Coefficient) algorithm applied to a failure classification method of a gear pump according to an embodiment of the present invention;
9 is a view showing the vibration of the gear pump
10 is a view showing a result of FFT (Fast Fourier Transform) processing the data of FIG. 9
11 is a diagram in which features are extracted from the data of FIG. 10 through an MFCC algorithm;
12 is a diagram showing the similarity of extracted features;
13 is a view showing the results of classification by SVM;
14 is a matrix showing classification data;

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings in order to describe in detail enough that a person of ordinary skill in the art to which the present invention pertains can easily implement the technical idea of the present invention.

The gear pump, which is the center of the present invention, is a rotating body that generates high pressure to obtain a large force, and at this time, the vibration characteristics of the pump are changed according to various parameters of the system. The increase in vibration in the gear pump is a direct factor that degrades the performance of the pump, and the magnitude of the measured acceleration of the pump is an indicator for measuring the state.

Therefore, in the present invention, based on the results of FMEA (Failure Mode and Effects Analysis) to acquire realistic data, a gear pump wear test was performed in a form similar to the actual situation, and continuous data was acquired until a failure occurred in a steady state. did The vibration characteristic data obtained through these experiments was classified into a normal pump and a malfunctioning pump using the SVM (Support Vector Machines) method, a classification algorithm. Finally, by comparing the performance evaluation parameter of the classified data, such as accuracy The optimal kernel function was selected.

1 is an exemplary view of the gear pump.

Referring to FIG. 1, the gear pump is divided into an external gear pump and an internal gear pump as shown in FIG. 1, and the gear pump used in the experiment is an external gear pump in which two gears are in external contact with each other.

External gear pumps are generally composed of an inlet, outlet, housing, and two gears. The gear is classified into a drive gear that receives rotational power from a power source and a driven gear that rotates in mesh with the drive gear, and the drive gear is fastened to the drive shaft with a key. has been

The principle of operation of the external gear pump is that hydraulic oil is sucked and compressed by two gears that rotate while meshing with each other in the case, and the fluid trapped between the inner surface of the case and the gear is compressed and discharged from the inlet to the outlet as the gear rotates. At the same time, hydraulic oil is easily sucked in from the suction port because the pressure is lower than atmospheric pressure as the hydraulic oil is discharged toward the discharge port.

- Perform FMECA of gear pump

FMECA is a concept that includes criticality analysis in FMEA (Failure Mode and Effects Analysis). By early identification of vulnerabilities, such as the possibility of fatal failure, and reflecting them in the design, the potential failure potential in the design stage is eliminated or minimized in advance. is the method used.

FMECA finds all possible failure types, classifies failure types according to failure mechanisms, understands the impact of the identified failure types on the part itself and the system as a whole, and identifies failure types with high criticality by calculating the criticality value for each failure type. It is a process.

<Table 1>

In order to analyze the failure mechanism of the gear pump, major components were divided into gear case, cover, drive shaft gear, driven shaft gear, side plate, seal, bearing and fastening bolt as shown in <Table 1>. The failure mechanism and failure identification number according to the type are indicated.

2 is a view showing the severity and frequency of failure on the system when a failure occurs.

Referring to FIG. 2, when a failure occurs, the failure mechanism is classified by the severity and frequency of failure on the system, and is shown as a matrix as shown in FIG. 2, and the fatality (2-1) of the uneven load caused by wear is the most high can be seen.

<Table 2>

3 is a diagram illustrating a fault tree analysis (FTA).

FMECA results using the classified matrix and failure mechanism are shown in <Table 2>, and the failure mechanisms according to failure types such as noise, internal leakage, external leakage and component damage, which are the main failure modes, are expressed as FTAs as shown in FIG. Failure (②) due to contaminants in the gear pump caused bearing wear, increased gear backlash, and seal damage, and is related to the most part among the causes of failure.

According to the results of FMECA/FTA (Fault Tree Analysis), the gear pump experiment was conducted in the form of a wear test due to the inflow of foreign substances, and the experiment was designed to determine the failure of the pump by measuring the vibration according to the test time.

4 is a configuration diagram of a vibration test apparatus of a gear pump, and FIG. 5 is a hydraulic circuit diagram of FIG. 4 .

As shown in FIG. 4, the configuration of the experimental apparatus includes a hydraulic tank for storing hydraulic oil, a gear pump as an experiment target, a motor for driving the gear pump, a relief valve for generating hydraulic pressure by controlling the flow rate, a flow meter for measuring the flow rate, and hydraulic pressure It consists of a hydraulic meter for measurement, an accelerometer for vibration signal measurement, a power supply that supplies power to the sensor, and DAQ that reads the sensor’s current value and transmits it to a computer. The data is processed using the computer software LabView. did

In order to clarify the relationship between the type, structure, and operation sequence of the hydraulic equipment used in the experimental apparatus, a hydraulic circuit diagram was prepared as shown in FIG. 5, and when the motor is turned on, the gear pump is driven, After suction, compress and discharge. The discharged flow passes through the flow meter and hydraulic system to generate high pressure at the relief valve, and returns the hydraulic oil to the hydraulic tank. During this process, the data acquired from each sensor is transmitted and saved to the computer through the data acquisition board (DAQ). The data acquisition board (DAQ) digitizes the received analog signal so that it can be read by the computer, and serves as an interface between the computer and the signal coming from the external environment.

In accordance with KS A 0006, the laboratory environment was maintained at room temperature 20±15° C. and humidity 65±20%, and the motor rotation speed was set to 1750 rpm and pressure 100 bar.

<Table 3>

Impurities used for experiments are classified according to their components and uses in KS A0090, a related standard. As shown in <Table 3>, particles for equipment wear testing include GBL, GBM, and White Fused Alumina.

In the case of white fused alumina, more than 99% of it consists of Al ₂ O ₃ , so it is suitable for use as an impurity in the gear pump wear test _{. 3} ) was replaced.

6 is a view showing the state before and after the test of the gear pump, and FIG. 7 is a photograph showing abrasion caused by the inflow of foreign substances into the hydraulic oil of the gear pump actually used in the forklift.

The gear pump wear test using iron oxide powder was stopped as soon as it did not satisfy the minimum required pressure of 50 bar to drive the pump tester. it can be confirmed that It can be seen that the iron oxide powder abraded the housing while the gear rotated at high speed, and the photo (refer to Fig. 7) where foreign substances were introduced into the hydraulic oil of the gear pump used in the forklift and the wear occurred and the gear worn by the foreign material input experiment The wear shape of the pump is similar.

- Feature extraction algorithm

The vibration data of the gear pump is data obtained by experimenting for a long time, and the amount of data is very large, and it is difficult to perform machine learning with this data as a whole. Therefore, in general, machine learning is performed with feature data equivalent to the original data by extracting the features of the data.

Algorithms for extracting features of data are very diverse, and new algorithms are being developed through continuous research. The vibration data of the gear pump obtained in the experiment consists of waveforms with amplitude and period, and these waveforms have the same characteristics as the sound waveform. Therefore, the algorithm used to extract features from the vibration data of the gear pump was selected as MFCC (Mel-Frequency Cepstral Coefficient), an algorithm for extracting sound features.

8 is an exemplary diagram of an MFCC (Mel-Frequency Cepstral Coefficient) algorithm applied to a method for classifying a failure of a gear pump according to an embodiment of the present invention.

MFCC (Mel-Frequency Cepstral Coefficient) does not target the entire input sound, but divides it into a certain section and analyzes the spectrum for this section to extract features. It shows higher accuracy than other algorithms, and the analysis process is As shown in FIG. 8, it can be divided into six steps.

First, after calculating the pre-emphasis for each frame, the data is set to an appropriate level of 20-40ms frames through a Hamming window, and FFT (Fast Fourier Transform) is applied to the divided frames. ) and pass it through the Mel Filter Bank as in Equation (1). In this case, f means frequency.

<Equation 1>

Signal data that has passed through the Mel Filter Bank is subjected to logarithm processing to preserve the signal amplitude component and discard the low-importance phase component. Finally, after calculating DCT (Discrete Cosine Transform) with logarithm-processed data, only 12 are left in the calculation result. change, and this is because the performance of voice recognition is lowered.

- Feature extraction of vibration data

9 is a view showing the vibration of the gear pump, FIG. 10 is a view showing the result of FFT (Fast Fourier Transform) processing on the data of FIG. 9, and FIG. 11 is a view in which features are extracted from the data of FIG. to be.

As a result of the gear pump experiment, the normal pump generated small-scale vibration with an amplitude of 4 g or less, but the malfunctioning pump showed an increase in amplitude up to 60 g, which is shown in FIG. 9 .

In addition, as a result of performing FFT (Fast Fourier Transform) on the vibration data of the pump, in the case of a normal pump, as shown in FIG. As shown in Fig. 10(b), it can be seen that the vibration is reduced in the low frequency band and the vibration is irregularly generated over the entire frequency band. These data were extracted with the MFCC algorithm and shown in FIG. 11, and it can be seen that the characteristic spectra of the normal pump and the malfunctioning pump are different.

12 is a diagram showing the similarity of extracted features.

If the extracted features overlap with other features, there is confusion in recognizing that they are the same data even though different data are included when classifying the data, and the accuracy can be greatly reduced. .

Therefore, in order to compare the extracted features with each other, the similarity of the extracted features is shown in Fig. 12, and the same features are put on the horizontal and vertical axes so that the same features are completely matched and come out in a bright color, and the other parts are expressed in a dark color. .

- Failure classification of gear pump

Features extracted through MFCC were classified into normal and faulty using SVM (Support Vector Machine). The SVM technique is a technique that creates a hyperplane that classifies data using a kernel function based on characteristics in the data.

13 is a diagram showing the results of classification by SVM.

SVM (Support Vector Machine) has various kernel functions, can make decision boundaries with complex data characteristics, and shows high accuracy regardless of dimension. 13 is a result of classification by SVM (Support Vector Machine) by changing the kernel function by adding the previously extracted features, the upper left is the Optimazed Linear Kernel, the upper right is the Linear Kernel, the lower left is the Gaussian Kernel, and the lower right is Polynomial Kernel. At this time, the data were classified by setting the default values of Auto and 1 for curvature (Gamma) and allowable error (Cost) of the decision boundary, which are variables of the support vector machine (SVM) technique, respectively.

14 is a matrix showing classification data, and Table 4 shows classification performance evaluation indexes.

<Table 4>

Data classified by machine learning can evaluate classification performance using various indicators, and Accuracy, Precision, Recall, and F-1 score are generally used to enable cross-validation. These indicators use classified data as shown in FIG. 14. As shown in Table 4, Accuracy is the ratio of True to True and False to False in all cases, and Precision is the ratio of the model to True The ratio of the ones that are actually True among those classified as , Recall is the ratio of those that the model predicts as True among those that are actually True, and the F-1 score is the harmonic average of Precision and Recall.

<Table 5>

As for the classification performance evaluation index for the learned algorithm, as shown in Table 5, the Accuracy of the Optimized Linear Kernel was the highest at 92.5%, and at this time, the Precision was 94.3%, the Recall was 91.0%, and the F-1 score was 92.6%.

In conclusion, as a result of the FMECA performance of the gear pump, it was found that the most fatal failure was the increase in noise and vibration due to wear.

In the wear test of the gear pump, the equipment was stopped as soon as the minimum operating pressure fell below 50 bar, the vibration data obtained through the experiment was applied to the MFCC algorithm to extract features, and the SVM technique was used to Data of worn gear pumps were classified. These results are expected to improve the operation rate of equipment and reduce maintenance costs by predicting the failure of the gear pump in advance. The results obtained are summarized as follows.

- The wear phenomenon of the gear pump housing due to the inflow of iron oxide powder is similar to the failure mode of the housing that is actually occurring in the gear pump for forklifts.

- Among the machine learning methods, the SVM method can classify gear pump failures, and the Optimized Linear Kernel shows the best performance.

- Optimized Linear Kernel has precision 94.3%, recall 91%, F-1 score 92.6%, and accuracy 92.5%.

The failure classification method of a gear pump according to an embodiment of the present invention can predict the failure of the gear pump in advance using the MFCC algorithm and the SVM technique.

can That is, by predicting the failure of the gear pump in advance, it is expected that the operation rate of equipment can be improved and maintenance costs can be reduced.

As such, those skilled in the art to which the present invention pertains will understand that the present invention may be embodied in other specific forms without changing the technical spirit or essential characteristics thereof. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. The scope of the present invention is indicated by the following claims rather than the above detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

Claims (4)

extracting features by dividing the vibration data of the gear pump into predetermined sections using a Mel-Frequency Cepstral Coefficient (MFCC) algorithm and analyzing the spectrum for each section; and
classifying the extracted features as normal or faulty using a support vector machine (SVM) technique;
A method of classifying a failure of a gear pump comprising a.
According to claim 1,
The step of extracting the feature is
After calculating the pre-emphasis for each frame of the vibration data of the gear pump, the data is set to a level of 20-40 ms frame through a Hamming window, and the divided frame is A failure classification method of a gear pump, characterized in that it passes through a Mel filter bank as in Equation (1) by calculating FFT (Fast Fourier Transform).
<Equation 1>

- f is the frequency
3. The method of claim 2,
Signal data that has passed through the Mel Filter Bank is logarithm-processed, DCT (Discrete Cosine Transform) is calculated with logarithm-processed data, and then the MFCC algorithm is applied to extract features. A failure classification method of a gear pump, characterized in that.
According to claim 1,
The SVM (Support Vector Machine) is,
A failure classification method of a gear pump, characterized in that any one of Optimazed Linear Kernel, Linear Kernel, Gaussian Kernel and Polynomial Kernel is applied as a kernel function.

Images