CN110427974A - A Health Status Detection Method of Hydraulic Components Based on Generalized Support Vector Machine - Google Patents

Abstract

The invention belongs to the related technical field of equipment state monitoring, and discloses a method for detecting the health status of hydraulic components based on generalized support vector machines. The signal is used as sample data, and the corresponding health status data is obtained at the same time; (2) the correlation between the obtained feature data and the health status data is calculated, and the training features are selected; (3) the stacking integration learning method is adopted Build a plurality of generalized support vector machine models, and use the data output in the generalized support vector machine model training process to train to obtain the random forest model, thus obtained the generalized support vector machine detection model; (4) will obtain The real-time sensor signal data of the measured hydraulic components is input to the generalized support vector machine detection model for health status detection. The invention improves detection precision and flexibility, and has strong applicability.

Description

A Health Status Detection Method of Hydraulic Components Based on Generalized Support Vector Machine

technical field

The invention belongs to the related technical field of equipment state monitoring, and more specifically relates to a method for detecting the health state of hydraulic components based on a generalized support vector machine.

Background technique

The hydraulic system is a key component of industrial electromechanical equipment, and its operating status directly affects the stability and reliability of the entire electromechanical equipment. If the hydraulic system fails suddenly, it will cause the equipment to stop, affect the production efficiency of the enterprise, and bring serious economic losses to the enterprise. Therefore, it is very necessary to perceive the health status of the hydraulic system.

The health status of the hydraulic system cannot be simply divided into normal operation and fault damage, and the damage process is a long-term and deepening process. Health status perception is to perceive any time node in the damage process, and the obtained fault states may be various. The staff need to make corresponding repairs according to the different fault states obtained from the health state perception of hydraulic system components Measures to prevent accidents, avoid economic losses and personal injuries.

Support vector machine (Support vector machine, SVM) is a supervised machine learning method, which can realize the classification of damaged and non-damaged eigenvalues, and has the advantages of small sample size requirements and fast diagnosis speed. However, the classic support vector machine is only suitable for the diagnosis of two classifications. The commonly used methods to help the support vector machine to perform multi-classification, such as OVO and OVA, will lead to a decline in the diagnosis effect, and the demand for samples of other classification methods is large, and the diagnosis speed slower. Correspondingly, there is a technical requirement in this field to develop a method for detecting the health status of hydraulic components based on generalized support vector machines with better accuracy, so as to realize accurate perception of the health status of the hydraulic system of electromechanical equipment.

Contents of the invention

Aiming at the above defects or improvement needs of the prior art, the present invention provides a generalized support vector machine-based method for detecting the health status of hydraulic components. A good health detection method for hydraulic components based on generalized support vector machines. The detection method uses a generalized support vector machine based on different kernel functions to perceive multiple health states of various hydraulic components of the hydraulic system, and uses an integrated learning method to integrate the health state perception results based on the generalized support vector machine. The final health status perception result is obtained, thereby realizing the health status perception of hydraulic components, improving detection accuracy and flexibility, and having strong applicability.

In order to achieve the above object, the present invention provides a method for detecting the state of health of hydraulic components based on generalized support vector machines, the detection method comprising the following steps:

(1) Obtain multiple sets of sensor signals of sample hydraulic components in different health states as sample data, and simultaneously obtain the health state data of the sample hydraulic components corresponding to each sample data;

(2) performing data cleaning on the sample data, and extracting statistical features from the cleaned sample data;

(3) Calculate the correlation between the characteristic data obtained based on the Pearson correlation coefficient and the health status data, and sort the characteristic data according to the corresponding correlation from large to small, and select the first predetermined one The features of numbers are used as training features;

(4) Based on the health status data corresponding to the training features and the training features, a plurality of generalized support vector machine models are constructed by stacking integrated learning methods, and the data output during the training process of the generalized support vector machine models are used To train the random forest model, and thus obtain the generalized support vector machine detection model integrated by the stacked ensemble learning method;

(5) Obtain the sensor signal data of the hydraulic component to be tested in real time online, and input the obtained sensor signal data into the generalized support vector machine detection model after processing, and then perform the generalized support vector machine detection model on the hydraulic component to be measured Health status detection.

Further, the sensor signal includes temperature, flow, pressure and power.

Further, the statistical features include absolute mean value, effective value, square root amplitude, skewness, kurtosis, kurtosis index, skewness index and waveform index.

Further, median filtering is used to clean the sample data.

Further, when using the Pearson correlation coefficient for feature selection, the calculation formula of the correlation coefficient between the sample data and the health status label data is:

In the formula, cov(X, Y) represents the covariance; σ _X represents the standard deviation of the extracted feature data set X; σ _Y represents the standard deviation of Y; X[i] represents the sample data of the i-th feature; The health status label data corresponding to the sample data.

Further, the predetermined number is 3.

Further, the generalized support vector machine model includes multiple different kernel functions.

Further, the signals of two sensors associated with the sample hydraulic components in different health states are acquired as sample data.

Generally speaking, compared with the prior art through the above technical solutions conceived by the present invention, the generalized support vector machine-based hydraulic component health status detection method provided by the present invention mainly has the following beneficial effects:

1. Based on the correlation between the characteristic data calculated by the Pearson correlation coefficient and the health status data, using the Pearson correlation coefficient can intuitively reflect the degree of correlation between the extracted feature value and the health status of the component for selection The eigenvalue data with the highest correlation and the best effect can improve the accuracy of the final classification effect, thereby improving the detection accuracy.

2. Use the stacked ensemble learning method to construct multiple generalized support vector machine models, and use the data output in the generalized support vector machine model training process to train the random forest model, thus obtaining the integrated by the stacked ensemble learning method The generalized support vector machine detection model uses the Stacking integrated learning method to integrate multiple classifiers to improve the accuracy rate. The Stacking method also includes K-fold cross-validation, which can effectively reduce the overfitting of the classifier training. combine.

3. Using the generalized support vector machine method can make up for the problem that the traditional SVM can only be divided into two categories. Compared with other multi-classification methods, the generalized support vector machine has a shorter training time and better classification effect.

4. The detection method can sense the health state of hydraulic components in time, and improves the accuracy and effectiveness of hydraulic component health state perception, and can also be used for the perception of hydraulic system health state, in order to solve the problem of hydraulic system health state perception A new way of thinking is provided.

Description of drawings

Fig. 1 is the schematic flow chart of the method for detecting the state of health of a hydraulic component based on a generalized support vector machine provided by the present invention;

Fig. 2 is a schematic flowchart of model building and online sample testing of the stacked integrated generalized multi-classification support vector machine involved in the generalized support vector machine-based hydraulic component health state detection method in Fig. 1 .

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not constitute a conflict with each other.

Please refer to Fig. 1 and Fig. 2, the generalized support vector machine based hydraulic component health status detection method provided by the present invention, the detection method can realize the precision of the health status diagnosis of electromechanical equipment, and it mainly includes the following steps:

Step 1: Obtain multiple sets of sensor signals of the sample hydraulic components in different health states as sample data, and simultaneously obtain the health state data of the sample hydraulic components corresponding to each sample data.

Specifically, the signals of two sensors associated with the sample hydraulic components in different health states, such as temperature, flow, pressure, power and other sensor signals, are obtained as sample data, and the health of the sample hydraulic components corresponding to each sensor signal is obtained status data. In this embodiment, the two associated sensor signals of the sample hydraulic components in Y health states are obtained, and the state corresponding to each signal sample can also be obtained; of course, when the object is a hydraulic system, multiple Each hydraulic component performs health status detection respectively, and then the health status of the hydraulic system is determined by combining multiple health status detection results obtained. Here, the associated sensor is determined by the hydraulic system. A certain hydraulic system in different hydraulic systems The sensors corresponding to each hydraulic component are different, and the number of sensors selected is also different.

Step 2, performing data cleaning on the sample data, and extracting statistical features from the cleaned sample data.

Specifically, data cleaning is performed on the obtained sample data to remove singular values, and statistical features are extracted from the cleaned sample data, and the statistical features include absolute mean value, effective value, square root amplitude, skewness, Kurtosis, kurtosis index, skewness index and waveform index.

The median filtering is involved in cleaning the sample data, which specifically includes the following steps:

(21) The number of groups of sensor signal data collected by one of the corresponding sensors selected based on the sample data is m, and the number of data points included in each group is n, that is, the sensor has collected m groups of data, and each group of data has n sample points.

(22) Determine the length k of the median filter window.

(23) Take the first k data in the first set of data, then:

N _(k+i-1)/2 ＝median(Ni,N _i+1 ,...,N _i+k-1 )

In the formula, N _i represents the i-th data; i is initially 1; median() represents the median value, and 1 is added when (k+i-1) is an odd number.

(24) i=i+1, repeat step (23) until i+k=n, so all m sets of data are processed to obtain a median-filtered data set.

(25) Select m sets of sensor data (each set includes n data points) obtained by another sensor in the corresponding sensor based on the sample data, and perform the processing from step (22) to step (24).

See Table 1 for the 8 statistical feature values used in this embodiment.

Table 1 Statistical characteristic indicators

In the table, x _i is the signal sequence of the i-th sampling; i=1, 2, ..., N, N is the total sampling times; σ is the standard deviation of the data. The proposed statistical feature set is F=[F ₁ , F ₂ ,…,F ₈ ], the statistical feature parameters include dimensioned feature parameters and dimensionless feature parameters, [F ₁ , F ₂ ,…,F ₅ ] is Dimensioned statistical indicators, [F ₆ , F ₇ , F ₈ ] are dimensionless statistical characteristic indicators.

Step 3: Calculate the correlation between the characteristic data obtained based on the Pearson correlation coefficient and the health status data, and sort the characteristic data according to the corresponding correlation from large to small, and select the top predetermined Number features are used as training features.

Specifically, the predetermined number is N, which can be changed according to the final effect. Feature selection using the Pearson correlation coefficient mainly includes the following steps:

(31) Determine the real health status label data set Y corresponding to the extracted feature data, the feature data set X of one of the selected sensors, X contains 8 kinds of feature data extracted, and X[i] represents the i-th sample data of this feature.

(32) Use the following formula to calculate the correlation coefficient between X[i] and Y:

Get 8 sets of correlation coefficients for 8 features, where cov(X, Y) means covariance; σ _X means the standard deviation of X, and σ _Y means the standard deviation of Y.

(33) According to the correlation coefficient obtained by each feature, take the three features with the largest correlation coefficient as the input data set of the stacking integrated generalized multi-class support vector machine model.

(34) Perform steps (31) to (33) on the features extracted by another selected sensor, and also select three types of features as the input data set of the stacked and integrated generalized multi-classification support vector machine model.

Step 4, based on the training features and the health status data corresponding to the training features, use the stacked ensemble learning method to construct multiple generalized support vector machine models, and use the data output during the generalized support vector machine model training process To train the random forest model, and thus obtain the generalized support vector machine detection model integrated by the stacked ensemble learning method.

Specifically, use the obtained N features and the health status corresponding to each training feature to construct K×M generalized support vector machine models and obtain the output data during model training, and use the output of K×M generalized support vector machines The data is then used to train a random forest model, so far a generalized support vector machine detection model integrated by the stacking method is obtained.

In this embodiment, the construction of the generalized support vector machine detection model mainly includes the following steps:

(41) Using the selected eigenvalue data of the two sensors obtained, each sensor has m sets of eigenvalue data, and merges the m sets of eigenvalue data of each sensor into a large data set X, assuming that each sensor There are 4 types of features, one sensor has 100 sets of data, then the data in X is 2*4=8 columns, 100 rows.

(42) Use the health status label data Y corresponding to each row of eigenvalue data in the large data set X, and Y is m rows and 1 column.

(43) Determine the value of the parameter K in the K-fold cross-validation, and divide X into K sets by row, and each set has Row data, combine K sets K times, each combination contains training set and verification set, ensure that each set of K sets is used as a verification set in K combinations, and the training set of each combination It is the combination of the remaining K-1 sets. Specifically, if K=3, then divide X into three sets a, b, and c by row; combine the three sets three times, in which a is the verification set in the first combination, and b and c are jointly used as training set, b in the second combination is the verification set, a and c are used as the training set together, and the third combination is similar.

(44) Use the generalized multi-classification support vector machine with K linear kernels to train and classify K combinations, and predict the verification set after each training to obtain the prediction data for the verification set. K trained The linear kernel generalized support vector machine model predicts the verification set to obtain K prediction data sets, and merges the K sets of verification set prediction data to obtain data with the same number of rows as the original X data set m. Named as H, H is used as the training input data of the random forest model.

(45) Using K polynomial (Poly) kernels and K radial basis (RBF) kernels generalized multi-classification support vector machine, repeat step (44) to obtain another two sets H ₂ , H ₃ . The support vector machine includes a plurality of different kernel functions.

(46) Use the random forest model, train a random forest model with the data sets H, H ₂ , H ₃ , and finally obtain a generalized multi-class support vector machine detection model integrated by stacking.

Step 5: Obtain the sensor signal data of the hydraulic component to be tested in real time online, and input the obtained sensor signal data into the generalized support vector machine detection model, and then the generalized support vector machine detection model performs health status detection of the hydraulic component to be tested . Wherein, the features corresponding to the sensor signal data are consistent with the training features.

In order to further describe the present invention in detail, the health state data of the hydraulic system is used to verify the method. The experimental device simulates four types of faults, and the fault modes and control parameters are shown in Table 2.

Table 2 Hydraulic system failure modes and their control parameters

Each component that performs health status awareness has a different health status, as shown in Table 3:

Table 3 Health Status Labels and Meanings

The experimental data set of the hydraulic system platform contains 18 kinds of sensor data, including pressure, power, flow, temperature, vibration, efficiency sensors, etc., and each sensor is arranged at the corresponding component position. The frequency of the pressure and power sensors is 100 Hz, and the frequency of the flow sensor is 10 Hz. Temperature, vibration, efficiency sensor frequency is 1Hz. A total of 2205 sets of data were collected, one set of data was 60s, and the data points in each set of data of each sensor were different according to the frequency of the sensor, and each set of data had a corresponding health status label.

The temperature sensors TS3 and TS4 are selected for the health status perception of the cooler; the pressure sensors PS1 and PS2 are selected for the health status perception of the valve; the flow sensor FS1 and the system efficiency sensor data SE are selected for the health status perception of the pump; for energy storage Select the power sensor EPS1 and pressure sensor PS2 for the health status perception of the device. Perform median filtering to obtain the filtered data, and then extract 8 statistical features for each set of data of each sensor selected for each component.

Correlation analysis is performed using the Pearson coefficient to obtain the eigenvalue data with the strongest correlation with the health status data of the component. The features most correlated with cooling system health status are effective value, absolute mean value, and kurtosis index; the features most correlated with valve health status are square root amplitude, skewness, waveform index; and pump health status correlation The strongest features are square root magnitude, absolute mean, and skewness index; since the correlation coefficients of the relevant eigenvalues of the accumulator are very small, all eigenvalues are used.

Merge the eigenvalue data corresponding to each component into a matrix, that is, each component has 2205 sets of data, and the number of columns obtained is different due to the number of selected eigenvalues and the number of sensors. At the same time, four corresponding label data sets with 2205 rows and one column are obtained. The label types of each data set are shown in Table 3.

Randomly select 70% of the eigenvalue matrix of each component as the training set and 30% as the test set. Use Stacking to integrate the generalized support vector machine method, in which 3-fold cross-validation is adopted, and the generalized support vector machine adopts linear kernel, poly kernel and RBF kernel respectively.

For the generalized support vector machine of each kernel, the parameters are continuously debugged in the experiment, and the final parameters are as follows:

Linear kernel: k=0, p=1, lambda=1*10 ^-8 , epsilon=1*10 ^-6 ;

Poly core: k=0, p=2, lambda=1*10 ^-8 , epsilon=1*10 ^-6 ;

RBF kernel: k=0, p=1, lambda=1*10 ^-8 , epsilon=1*10 ^-6 .

The final results are shown in Table 4:

Table 4 Classification results and accuracy

fault location Number of training sets Number of test sets first layer accuracy Test set prediction accuracy cooling system 1506 649 100% 100% Valve V10 1506 649 100% 100% MP1 pump 1335 580 100% 100% accumulator leak 1335 580 82% 76%

In order to illustrate the accuracy of this method, this method was compared with the LDA method, ANN method, linear kernel SVM, and RBF kernel SVM. The results are shown in Table 5. The health status perception accuracy of this method is better than other methods.

Table 5 Comparison between different methods

The health state detection method of hydraulic components based on the generalized support vector machine provided by the present invention, the detection method combines the Pearson correlation coefficient, the stacking ensemble learning method and the generalized support vector machine, improves the accuracy of the final classification effect, and then improves the detection accuracy , and has good flexibility and strong applicability.

Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (8)

1. a kind of hydraulic unit method for detecting health status based on Generalized Support Vector Machine, which is characterized in that this method includes Following steps:

(1) it obtains multiple groups sample hydraulic unit and be in the sensor signal under different health status using as sample data, and together When obtain state of health data locating for the corresponding sample hydraulic unit of each sample data；

(2) data cleansing is carried out to the sample data, and statistical nature is extracted to the sample data after cleaning；

(3) correlation between obtained characteristic and the state of health data is calculated based on Pearson correlation coefficients, And be ranked up the characteristic from large to small according to corresponding correlation, choose the feature conduct of predetermined number before coming Training characteristics；

(4) based on state of health data corresponding to the training characteristics and the training characteristics, using stacking integrated study side Method constructs multiple Generalized Support Vector Machine models, and using the number exported during the Generalized Support Vector Machine model training According to training to obtain Random Forest model, the Generalized Support Vector Machine detection integrated by stacking integrated learning approach is resulted in Model；

(5) it obtains the sensor signal data of hydraulic unit to be measured in real time online, and obtained sensor signal data is inputted To the Generalized Support Vector Machine detection model, and then the Generalized Support Vector Machine detection model carries out hydraulic unit to be measured Health status detection.

2. the hydraulic unit method for detecting health status based on Generalized Support Vector Machine, feature exist as described in claim 1 In: the sensor signal includes temperature, flow, pressure and power.

3. the hydraulic unit method for detecting health status based on Generalized Support Vector Machine, feature exist as described in claim 1 In: the statistical nature includes absolute mean, virtual value, root amplitude, flexure, kurtosis, kurtosis index, flexure index and waveform Index.

4. the hydraulic unit method for detecting health status based on Generalized Support Vector Machine, feature exist as described in claim 1 In: the sample data is cleaned using median filtering.

5. the hydraulic unit method for detecting health status based on Generalized Support Vector Machine, feature exist as described in claim 1 In: when carrying out feature selecting using Pearson correlation coefficients, the relative coefficient of sample data and health status label data Calculation formula are as follows:

In formula, cov (X, Y) indicates covariance；σ_XIndicate the standard deviation of extracted characteristic data set X；σ_YIndicate the standard deviation of Y； X [i] indicates the sample data of i-th kind of feature；Y indicates the corresponding health status label data of sample data.

6. the hydraulic unit method for detecting health status as described in any one in claim 1-5 based on Generalized Support Vector Machine, It is characterized by: the predetermined number is 3.

7. the hydraulic unit method for detecting health status as described in any one in claim 1-5 based on Generalized Support Vector Machine, It is characterized by: the Generalized Support Vector Machine model includes multiple and different kernel functions.

8. the hydraulic unit method for detecting health status as described in any one in claim 1-5 based on Generalized Support Vector Machine, It is characterized by: obtaining the signal of two sensors associated by the sample hydraulic unit under different health status as sample number According to.