Variable speed-based two-dimensional piston pump health state evaluation method

Abstract

The application provides a health state evaluation method of a variable-speed two-dimensional (2D) piston pump, which comprises the following steps: data acquisition: collecting vibration signals reflecting normal and fault states of the pump; data preprocessing: extracting instantaneous frequency by using GLCT to perform equal-angle resampling to eliminate the influence of the rotating speed, then performing order analysis on the data to obtain order domain characteristic parameters; time-frequency analysis method is used to construct a time-frequency fault sample set; model training and optimization: constructing a fault diagnosis model, taking the training sample set as input and the state label as output, and training the model. The vibration signal of the 2D piston pump is input into the trained and optimized fault diagnosis model, and the state label value is output by the fault diagnosis model, so that the final health state classification and recognition result can be directly obtained, and the method is simple and reliable, and the accuracy of the diagnosis result is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hydraulic element health state evaluation, and particularly relates to a two-dimensional piston pump health state evaluation method based on variable rotating speed. BACKGROUND

[0002] The micro two-dimensional piston pump is a new type of hydraulic pump, which utilizes the principle of two degrees of freedom of the piston to form a new distribution mode, has the characteristics of realizing high power-to-weight ratio while considering the miniaturization of the structure, and has the advantages of simple structure, small size, light weight, few friction pairs and fast response compared with the traditional hydraulic pump. As a key basic part of aircraft, robots and high-end equipment, the health state of the two-dimensional piston pump is closely related to the reliability and stability of the overall operation of the equipment. Therefore, accurately evaluating the health state of the two-dimensional piston pump can effectively solve the problem of difficult discovery of hydraulic system faults, and has great practical significance for formulating a reasonable maintenance plan and ensuring the stability and safety of the equipment.

[0003] With the rapid development of signal processing technology and artificial intelligence, applying artificial intelligence to the health state evaluation and analysis of the two-dimensional piston pump provides a new idea for realizing real-time and accurate evaluation of the current health state of the system, predicting the performance degradation degree and development trend of the system, discovering early faults, and realizing condition-based maintenance.

[0004] The health state recognition of the two-dimensional piston pump is actually a process of classifying and recognizing the characteristics of the fault and damage degree according to the vibration signal. The vibration signal collected according to the damage degree of the test object has vibration characteristics corresponding to the damage degree in theory. Therefore, data processing and analysis based on the collected vibration signal is a hot topic in the health diagnosis and recognition of mechanical equipment based on data-driven health state research. However, most of the current domestic and foreign researches are on the steady state constant rotating speed condition, and the research on the variable rotating speed condition is still rare.

[0005] The vibration response of the variable speed condition is very complex in time domain, frequency domain and time-frequency domain, and has a greater impact on the signal essential characteristics. In order to eliminate the influence of speed change, by exploring the relationship between the speed and the vibration signal, finding a mapping relationship between the speed and the vibration signal is a way to eliminate the influence of speed. An important method for variable speed health state recognition is order analysis, and the common order analysis mainly has two methods: hardware order analysis method and order analysis method based on instantaneous frequency estimation. The hardware order analysis method needs to realize synchronous sampling by installing hardware, and the phase reference signal collected from the encoder or tachometer is essential, but due to budget cost or technical reasons, many equipment cannot always install the encoder and tachometer. In some important devices such as aircraft and robots, the sensor is not convenient to install. Therefore, the core of the order analysis method of the instantaneous frequency estimation is to accurately estimate the instantaneous frequency of the vibration signal by the parameterized time-frequency analysis method, so as to classify and identify the fault.

[0006] By obtaining the vibration signal, eliminating the influence of speed change on the vibration signal, mining the effective information inside the signal, and establishing a perfect fault classification system, the health state classification accuracy of the two-dimensional piston pump is greatly affected, so the research on the classification method is also very key to obtain the health state of the two-dimensional piston pump. With the increasingly optimized development of the classification algorithm, the classification methods are endless, and for a specific research object, selecting a suitable classification algorithm is the research direction at this stage. The health state recognition of the two-dimensional piston pump is a small sample data classification test, so the support vector machine is used to classify the small sample data and optimize the parameters, which can ensure higher accuracy of the health state recognition of the two-dimensional piston pump. SUMMARY

[0007] In order to solve the above problems of the prior art, the present application provides a two-dimensional piston pump health state evaluation method based on variable speed, which can monitor and identify the health state of the two-dimensional piston pump based on variable speed, and intelligently identify the health state of the two-dimensional piston pump with high precision. The evaluation method of the present application has higher classification accuracy than other traditional methods CPA-SVM, and the population-based algorithm CPA initializes multiple solutions and iteratively enhances, although the algorithm needs more function evaluations, but due to information sharing, local optimum can be avoided, so that the optimization of the objective function can be effectively performed, and the optimal health monitoring model can be ensured.

[0008] Specifically, the present application provides a two-dimensional piston pump health state evaluation method based on variable speed, which comprises the following steps:

[0009] S1, data acquisition: collect the vibration signal reflecting the normal state and fault state of the two-dimensional piston pump as the original state data;

[0010] S2, data preprocessing analysis: GLCT time-frequency analysis is performed on the original state data of the two-dimensional piston pump vibration signal collected in step S1 to extract the instantaneous rotational frequency of the vibration signal;

[0011] S3, extracting angular domain vibration signal and order domain characteristic parameters, including the following sub-steps:

[0012] S31, data fitting processing is performed on the instantaneous rotational frequency extracted in step S21, and a segmented multi-order polynomial fitting method is used to intercept the data of the frequency rising stage of the instantaneous frequency for data fitting to obtain a data fitting curve, and an order spectrum is obtained according to the fitting curve;

[0013] S32, for the angular domain vibration signal s(l) of the two-dimensional piston pump, the data length of the equal-angle interval sampling is L, then the peak index Cf, the pulse index If and the kurtosis index K V of the angular domain vibration signal s(l) are calculated according to the following formulas respectively:

[0014]

[0015]

[0016]

[0017] Wherein, S rms is the root mean square value, S max is the peak value, is the absolute average amplitude, and β is the kurtosis;

[0018] The calculation formula of the root mean square value S rms is:

[0019]

[0020] The calculation formula of the peak value S max is:

[0021] S max = max(|s(l)|)

[0022] The calculation formula of the absolute average amplitude is:

[0023]

[0024] The calculation formula of the kurtosis β is:

[0025]

[0026] Let the order spectrum of the angular domain vibration signal s(l) be Sl(m), where l=1, 2, …, L, and m is the order data length variable, taking m=1, 2, …, Dmax where D max is the maximum order, then the order spectrum root mean square value Y1 is defined as follows:

[0027]

[0028] and the size of Y1 is used to measure the vibration energy intensity of the entire order spectrum;

[0029] S33, using the amplitudes corresponding to the 11th, 22nd, 33rd and 44th orders of the order spectrum as the order domain characteristic parameters for analysis:

[0030]

[0031] In the formula, n is the number of order spectrum data in the interval, Y2, Y3, Y4 and Y5 are the amplitudes corresponding to the 11th, 22nd, 33rd and 44th orders of the order spectrum, respectively;

[0032] S4, fault diagnosis model training and optimization, specifically including the following sub-steps:

[0033] S41, constructing a fault diagnosis model: taking Cf, If, K V , Y1, Y2, Y3, Y4 and Y5 as the training characteristic parameters of the CPA-SVM optimized classification algorithm to construct a fault diagnosis model, and dividing the sample library into a training sample set and a test sample set according to a proportion, then taking the training sample set as the input of the fault diagnosis model and the state label as the output of the fault diagnosis model;

[0034] S42, under the classification accuracy target function of SVM, the penalty factor c and the radial basis number g in a large range are accurately optimized by using the CPA optimization algorithm to optimize the model parameters;

[0035] S43, using the optimized CPA-SVM to train the fault classification model, obtaining the optimal model parameters best c and best g, thereby obtaining the trained and optimized fault diagnosis model;

[0036] S5, model verification: calling the trained and optimized fault diagnosis model, taking the test sample set as the input of the trained and optimized fault diagnosis model and the state label as the output of the trained and optimized fault diagnosis model, and verifying the comprehensive performance of the diagnosis model;

[0037] S6, model diagnosis: inputting the vibration signal of the two-dimensional piston pump into the trained and optimized fault diagnosis model, the fault diagnosis model outputs the state label value, and the health state classification recognition result is obtained.

[0038] Preferably, step S2 specifically includes the following sub-steps:

[0039] Step S21, the original STFT formula is constructed as follows:

[0040]

[0041] where w(u-t') is a certain window;

[0042] Step S22, on the basis of the STFT formula, a series of discrete demodulation operators are used to approximate the optimal demodulation operator The STFT formula based on the discrete demodulation operator is as follows:

[0043]

[0044] Step S23, for each TF point (t', w), if the discrete demodulation operator is close to the modulation component of the signal, the TF representation around its IF has a higher energy concentration, and its amplitude |S(t', w, c)| reaches a maximum among all values, and then for each TF point, according to the amplitude of |S(t', w, c)|, the optimal parameter c' of the parameter c is obtained:

[0045]

[0046] The time-frequency representation of the proposed time-frequency analysis method is:

[0047] GS(t', w) = S(t', w, c')

[0048] The spectrum calculation formula is:

[0049] Spec(t', w) = |GS(t', w)| 2

[0050] Step S24, determine the discrete demodulation operator A rotation parameter a is introduced to realize rotation on the time-frequency plane:

[0051]

[0052] where T s is the sampling time, and F s is the sampling frequency;

[0053] Let the rotation parameter a have N values, and divide the time-frequency plane into N+1 segments:

[0054] a = -π / 2 + π / (N+1), -π / 2 + 2·π / (N+1)... -π / 2 + N·π / (N+1);

[0055] Step S25, ensure that the demodulation operator The STFT formula of the discrete demodulation operator can be rewritten as the GLCT formula by describing all possible modulation components in the signal:

[0056]

[0057] Step S26, GLCT time-frequency analysis is performed according to the GLCT formula of step S25, and the instantaneous rotational frequency S(t', w, a) of the vibration signal is extracted.

[0058] Preferably, step S31 specifically comprises the following sub-steps:

[0059] S311, the relationship between the rotational speed and the first-order instantaneous frequency is represented as:

[0060]

[0061] Where f(t) is the shaft rotational frequency, and P(t) is the reference shaft speed.

[0062] S312, a segmented multi-order polynomial fitting method is used to obtain a fitting curve:

[0063] R k (t)=a k +b k t+c k t 2

[0064] Where Rk(t) is the function formula of each segment of the rotational speed curve, k is the serial number of the segment, a k , b k , and c k are polynomial coefficients.

[0065] S313, the rotational speed fitting curve is integrated to obtain an equiangular interval interpolation phase detection time sequence, and the solving process is:

[0066]

[0067] Where T n is the key phase time mark, n is the time mark serial number, and T0 is the initial fitting time; the above formula is substituted into the fitting curve formula, and an effective algebraic solution is obtained to obtain a set of equiangular interval interpolation phase detection time sequences T n :

[0068]

[0069] S314, based on the equiangular interval interpolation phase detection time sequence obtained in step S313, the Lagrange value method is used to complete smoothing processing to obtain an angle domain signal.

[0070] S315, performing Fourier transform on the angle domain signal data obtained in step S314 to obtain an order spectrum.

[0071] Preferably, 200 groups of data are obtained to divide the training sample set and the test sample set in a ratio of 4:1, and the iteration number in a group group_iter=20, the attraction rate attraction_rate=0.8, the growth rate growth_rate=2, the reproduction rate reproduction_rate=1.8, the number of carnivorous plants nCPlant=10, and the number of prey nPrey=20 are defined.

[0072] Preferably, step S42 specifically comprises: taking the eight characteristic vectors obtained through time-frequency domain and order analysis as training characteristic parameters of the CPA-SVM optimized classification algorithm, and taking the target labels 1, 2, 3, and 4 in the last column as four different wear states determined by the optimized classification algorithm according to the characteristic vectors, then determining the ranges of c and g, and under the target function of the classification accuracy of the SVM, the c and g in a large range are further optimized by the CPA optimization algorithm to accurately optimize the parameters.

[0073] Preferably, the vibration signal is collected by a vibration sensor installed on the two-dimensional piston pump in step S1.

[0074] Preferably, the optimal parameters are best c=224.6036 and best g=31.5104.

[0075] Preferably, step S42 specifically comprises the following steps:

[0076] The carnivorous algorithm continuously iterates and evolves to find suitable SVM parameter values, thereby improving the accuracy of SVM classification. The mathematical model can be simplified to find the optimal decision vector [c, g], where c and g represent the penalty parameter vector and the kernel function vector of the SVM, respectively, to maximize the following objective:

[0077] max F=SVM_Acc

[0078]

[0079]

[0080] In the formula, F represents the objective function, SVM_Acc represents the classification accuracy of the SVM, M T represents the number of samples correctly classified into the corresponding category, M FThe number of individuals not correctly divided into the corresponding category is represented; first define the number of iterations within the group group_iter=20, the attraction rate attraction_rate=0.8, the growth rate growth_rate=2, the reproduction rate reproduction_rate=1.8, the number of carnivorous plants nCPlant=10 and the number of prey nPrey=20, in the objective function SVM_Acc, the maximum number of iterations T=40, for the optimization algorithm SVM parameters c and g, first through the grid search method to roughly optimize the parameters, determine the range of c and g is 2 -2 -2 8 , i.e. 0.25-256 and 2 -5 -2 5 , i.e. 0.03-32.

[0081] Compared with the prior art, the effects of the present application are as follows:

[0082] (1) The present application takes the original vibration signal of the two-dimensional (2D) piston pump as the data source, can non-destructively monitor the piston pump, and adopts the keyless order analysis method to eliminate the influence of the two-dimensional (2D) piston pump speed, so that the monitoring result is more accurate.

[0083] (2) Since the traditional instantaneous frequency extraction method has the disadvantages of low time-frequency resolution and large instantaneous frequency error, the GLCT time-frequency analysis method is adopted in the present application to more accurately extract the instantaneous frequency, and then perform angle domain resampling to further complete the order analysis in the angle domain.

[0084] (3) The two-dimensional (2D) piston pump health state evaluation method based on vibration signal, using some dimensionless parameters in the angle domain vibration signal which are not sensitive to speed change, can more accurately reflect the failure of the two-dimensional (2D) piston pump, such as the angle domain vibration signal peak value index, pulse index and kurtosis index. For the order domain, it is found that the order energy and other parameters show obvious failure, and the order spectrum root mean square value and the order amplitude in the order domain are used as characteristic parameters to construct a high-dimensional and complex overall fault sample set.

[0085] (4) The two-dimensional (2D) piston pump health state evaluation method based on vibration signal fully utilizes the CPA method to optimize the two parameters, and under the judgment of the fitness objective function, the parameters of the SVM multi-class prediction penalty factor c and the radial basis g are searched multiple times to improve the health state recognition ability of the two-dimensional (2D) piston pump.

[0086] (5) The two-dimensional (2D) piston pump health state evaluation method based on vibration signals, based on the training and testing of the angle domain and order domain characteristic parameters, can successfully solve the problems of high-dimensional design variables, the existence of various constraints and the search space with multiple local optimal solutions after optimizing SVM by CPA, improve the ability of two-dimensional (2D) piston pump health state recognition, and ensure the accuracy of model output. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 is a flowchart of the present application;

[0088] Figure 2 is an implementation process of the two-dimensional piston pump health state recognition method based on vibration signals according to the present application;

[0089] Fig. 3 is a comparison diagram of the instantaneous frequency of the vibration signals of the four states of the two-dimensional piston pump actually measured in the embodiment of the present application, wherein Figure 3a is a health state diagram, Figure 3b is a mild wear fault diagram, Figure 3c is a moderate wear fault diagram, Figure 3d is a severe wear fault diagram;

[0090] Fig. 4 is a comparison diagram of the angle domain resampling of the vibration signals of the four states of the two-dimensional piston pump actually measured in the embodiment of the present application, wherein Figure 4a is a health state diagram, Figure 4b is a mild wear fault diagram, Figure 4c is a moderate wear fault diagram, Figure 4d is a severe wear fault diagram;

[0091] Figure 5 is a schematic diagram of the best fitness of 93.5% of the fault classification model of the SVM optimized by the CPA-SVM optimization algorithm in the embodiment of the present application;

[0092] Fig. 6 is a two-dimensional piston pump health state classification result diagram based on SVM in the embodiment of the present application, wherein Figure 6a is a health state diagnosis result diagram, Figure 6b is a diagnosis result confusion matrix diagram.DETAILED DESCRIPTION

[0093] Hereinafter, the embodiments of the present application will be described with reference to the accompanying drawings.

[0094] The application will be further described below in conjunction with the drawings and specific embodiments, but the scope of protection of the application is not limited thereto. The embodiments of the application are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the application, and cannot be understood as a limitation of the application.

[0095] In the variable speed condition, the change of speed will cover the sensitive information of the relevant degradation indicators, and the changes of the vibration signal in the time domain and frequency domain are very complex and intense, which leads to the method of positioning the health status of the piston pump through sensitive features becoming less applicable. Therefore, how to eliminate the influence of variable speed and locate the degradation state indicator of the piston pump is an important problem of the application. The application builds a non-constant speed piston pump wear fault test bench, collects data of piston pumps with different wear degrees, uses the encoder-free test method, extracts the instantaneous frequency based on GLCT for equal-angle resampling to eliminate the influence of speed, and then analyzes the order of the data to explore the health status of the piston pump.

[0096] Specifically, the application collects data of two-dimensional piston pumps with different piston wear degrees through a vibration sensor installed on the two-dimensional piston pump, collects vibration signals reflecting normal state and health state evaluation state of the pump as original state data, and then uses the GLCT method to accurately extract the instantaneous frequency of strong nonlinear signals, and obtains the rotational speed corresponding to the rotational frequency from the mapping of the instantaneous frequency and the instantaneous rotational speed, laying a foundation for further equal-angle domain resampling and eliminating the influence of variable speed. The GLCT time-frequency analysis method can more accurately extract the instantaneous frequency, and then perform angle domain resampling, and then perform Fourier transform on the angle domain resampled data to obtain the order spectrum for order analysis; thereafter, some dimensionless indexes in the angle domain which are basically irrelevant to the operating condition and are not sensitive to the change of the rotational speed are extracted as the time domain characteristic parameters and the root mean square value of the order spectrum as the order domain characteristic parameters as the total characteristic parameters; finally, the one-to-one method of the support vector machine (SVM) multi-classification is used to classify the characteristic parameters, and a training model is constructed, and then the Carnivorous Plant Algorithm (CPA) is used to optimize the radial basis kernel parameter g and the support vector machine penalty factor c which affect the sample training and testing process, to establish a CPA-SVM health state evaluation classification training model to shorten the time of establishing the health state evaluation classification model and improve the accuracy of SVM classification. The test result analysis shows that the classification accuracy of the test results obtained by the SVM and ELM methods is compared. The results show that the health state evaluation recognition accuracy of the CPA-SVM model based on the Carnivorous Plant Algorithm optimized support vector machine is 93.75%, which is improved by 11.25% compared with the SVM without optimization, and the classification accuracy is also higher than that of ELM, which reflects the advantages of the Carnivorous Plant Algorithm optimized support vector machine method in health state recognition.

[0097] The specific steps of the health state evaluation method of the variable speed two-dimensional piston pump are as follows:

[0098] The application provides a health state evaluation method of a variable speed two-dimensional piston pump based on CPA-SVM, which mainly has two steps: firstly, the GLCT-based time-frequency analysis method is used to extract the instantaneous frequency of the vibration signal, and angle domain resampling is performed, and then Fourier transform is performed on the angle domain resampled data to obtain the order spectrum for order analysis. Then, four groups of data corresponding to four health state evaluation states are established according to the obtained characteristic parameters, and the classification optimization method based on CPA-SVM is used for health state evaluation classification model training, which reflects the advantages of CPA-SVM in two-dimensional piston pump health state recognition.

[0099] As shown in Figure 1 and Figure 2 , which specifically includes the following steps:

[0100] S1, the different health states of two-dimensional piston pump are embodied by replacing pistons with different degrees of wear. The vibration signals of two-dimensional piston pumps with four different degrees of wear are collected. The test sampling frequency is 20000hz, the data processing down-sampling is 5120hz, and the sampling time is 35s. 100 groups of data are collected for each of the four health state evaluation states, and a total of 400 groups of data are collected.

[0101] S2, GLCT time-frequency analysis is performed on the 4x100 groups of data of the collected four groups of variable speed two-dimensional piston pump vibration signals to extract the instantaneous speed frequency. The method for extracting the instantaneous frequency based on GLCT time-frequency analysis is as follows:

[0102] A series of discrete demodulation operators are used to approximate the best demodulation operator The STFT formula considering the discrete demodulation operator is as follows:

[0103]

[0104] For each TF point (t’, w), if the discrete demodulation operator Approaches the modulation component of the signal, and the TF representation around its IF has a higher energy concentration, and its amplitude |S(t’, w, c)| reaches a maximum among all values. Then for each TF point, the best parameter c is obtained according to the amplitude of |S(t’, w, c)|:

[0105]

[0106] The time-frequency representation of the proposed time-frequency analysis method is:

[0107] GS(t’, w) = S(t’, w, c’)

[0108] The spectrum can be defined as:

[0109] Spec(t’, w) = |GS(t’, w)| 2

[0110] Next, the discrete demodulation operator According to the existing research results, by introducing A rotation effect will be generated on the time-frequency result, and the rotation degree is arctan(-c). For a signal, if the sampling time (Ts) and the sampling frequency (Fs) are determined, a time-frequency plane of t∈(0, Ts), f∈(0, Fs / 2) will be determined. Therefore, by introducing a rotation parameter a, rotation on the time-frequency plane is realized, as follows:

[0111]

[0112] Ensure that the demodulation operator The above equation can be rewritten as:

[0113]

[0114] If parameter a has N values, the time-frequency plane can be divided into N+1 segments:

[0115] a = -π / 2 + π / (N+1), -π / 2 + 2·π / (N+1)... -π / 2 + N·π / (N+1)

[0116] From the above equation, it can be seen that the method GLCT used by the application introduces one more parameter a than STFT.

[0117] S3, angular domain resampling is performed on the measured device without speed estimation, to obtain an instantaneous frequency and perform data fitting processing on the obtained instantaneous frequency, and specifically, a segmented multi-order polynomial fitting method is used to intercept the frequency rising stage of the instantaneous frequency for data fitting, to obtain a fitting curve.

[0118] For the measured device without speed estimation, key phase time stamping needs to be calculated. In order to perform data fitting processing on the obtained instantaneous frequency, the frequency rising stage of the instantaneous frequency is intercepted for data fitting, and first, a speed curve is obtained, and the relationship between speed and first-order instantaneous frequency is:

[0119]

[0120] Where f(t) is the shaft rotation frequency, and P(t) is the reference shaft speed.

[0121] The obtained discrete speed information is converted into a smooth curve through a fitting method, and then the speed model can be further simulated. A segmented multi-order polynomial fitting method is used to achieve high-precision fitting effect:

[0122] R k (t)=a k +b k t+c k t 2

[0123] Where Rk(t) is the function formula of each segment of the speed curve, k is the serial number of the segment, a k , b k , c k are polynomial coefficients. The result obtained by integrating the speed fitting curve is the phase demodulation time stamp sequence, and the solving process is:

[0124]

[0125] Where T nis the initial fitting time. Substitute the above formula into the previous formula of Rk(t), and get the effective algebraic solution, to get a set of equal angle interval interpolation phase time series T n :

[0126]

[0127] The fitting curve and angle resampling data obtained by the above method can solve the frequency spectrum ambiguity problem of directly Fourier transforming the collected vibration signal, and obtain the order spectrum.

[0128] For the angular domain vibration signal s(l) of the two-dimensional piston pump, the data length of equal angle interval sampling is L, and its peak index Cf, pulse index If and kurtosis index K V are defined as follows, respectively:

[0129]

[0130]

[0131]

[0132] Where S rms is the root mean square value, defined as:

[0133]

[0134] The peak S max is defined as:

[0135] S max = max(|s(l)|)

[0136] The absolute average amplitude is defined as:

[0137]

[0138] The kurtosis β is defined as:

[0139]

[0140] Let the order spectrum of the angular domain vibration signal s(l) (l=1, 2, …, L) be Sl(m), where m represents the order data length variable, and take m=1, 2, …, D max , where Dmax is the maximum order, and the order spectrum root mean square value is defined as:

[0141]

[0142] And use Y1 size to measure the vibration energy intensity of the entire order spectrum.

[0143] In order to better represent the feature difference under different degrees of wear, the amplitudes corresponding to the 11th, 22nd, 33rd and 44th orders are used as characteristic parameters for analysis. Since the maximum amplitude may not occur exactly at the integer order for each signal segment during the calculation of the order spectrum, the average value of the 11th, 22nd, 33rd and 44th order intervals ± 0.15 is selected as the characteristic value for selection, that is:

[0144]

[0145] In the formula, n is the number of order spectrum data in the interval.

[0146] S4, select the peak index Cf, pulse index If and kurtosis index KV in the dimensionless index as the degradation characteristic parameters in the angle domain, denoted as X = [Cf, If, KV], select the average value of the 11th, 22nd, 33rd and 44th order intervals ± 0.15 as the health state evaluation characteristic parameters in the order domain, denoted as Y = [Y1, Y2, Y3, Y4, Y5], and combine the two into the total characteristic parameters denoted as Z = [Cf, If, KV, Y1, Y2, Y3, Y4, Y5]. V V

[0147] S5, the Gaussian radial basis function is used to map the health state evaluation characteristic parameter samples to a higher dimensional space. By solving the quadratic optimization problem, the normal vector w and the intercept b on the hyperplane are obtained using the mathematical expression of the optimal hyperplane. The Lagrange multiplier method is used to convert the above optimization problem into a dual problem to obtain the optimal classification function.

[0148] S6, 80 of the 100 groups of data are trained and the remaining 20 groups are tested. Then the grid search method is used to roughly optimize the parameters to determine the range of c and g. Under the target function of the classification accuracy of SVM, c and g in a large range are further optimized using the CPA optimization algorithm. Further, the CPA-SVM is used for health state evaluation classification model training, and the optimal parameters obtained are best c = 224.6036 and best g = 31.5104. At this time, the target function score of the model is the best, and the remaining parameters are the default values.

[0149] The specific method is as follows: the first 80 groups of 4x100 signals, i.e. a total of 320 signals, are trained, and the remaining 4x20 groups, i.e. a total of 80 groups, are tested. The Cf, If, KV V ​​, Y1, Y2, Y3, Y4, Y5 are training feature parameters of the CPA-SVM optimization classification algorithm, and the target labels 1, 2, 3, and 4 in the last column are four different wear states corresponding to the feature vectors, and then according to the determined penalty factor c and the range of radial basis g, the c and g in a large range are further optimized by the CPA optimization algorithm under the classification accuracy target function of the SVM. Further, the CPA-SVM is used for health state evaluation classification model training, and the optimal parameters obtained are best c = 224.6036 and best g = 31.5104, at which time the model has the best target function score, and the remaining parameters are default values. DETAILED EMBODIMENTS

[0151] A two-dimensional piston pump test bench was built. In this test, different wear degrees of pistons were replaced to simulate different health state evaluation states of the two-dimensional (2D) piston pump, vibration signals of the two-dimensional (2D) piston pump were collected, and finally the signals were processed by the method proposed in the application. The test classified the pistons with different wear degrees from the two-dimensional (2D) piston pump in the health state evaluation, which were in turn healthy state, mild wear, moderate wear, and severe wear. As shown in Figures 3a-3d , the Figure 3a is a health state vibration signal instantaneous frequency diagram, Figure 3b is a mild wear state vibration signal instantaneous frequency diagram, Figure 3c is a moderate wear state vibration signal instantaneous frequency diagram, Figure 3d is a severe wear state vibration signal instantaneous frequency distribution diagram. As Figures 3a-3d can be seen, it is difficult to determine the different health states of the two-dimensional (2D) piston pump corresponding to the pressure signals by observing the differences in the instantaneous waveform diagram.

[0152] The measured two-dimensional (2D) piston pump generalized linear frequency modulation transform (GLCT) extracted vibration signal angular frequency domain diagram in the health state is shown in Figures 4a-4d , wherein Figure 4a is a health state vibration signal angular frequency domain diagram, Figure 4b is a mild wear vibration signal angular frequency domain diagram, Figure 4c is a moderate wear vibration signal angular frequency domain diagram, Figure 4d is a severe wear health state evaluation state vibration signal angular frequency domain diagram.

[0153] The four wear states collected by the test bench are all under the same working condition. The pulse signal of the stepper motor goes through two stages. The first stage is the rising stage of the pulse signal from 5000 to 15000 at a rate of 500 per second. The second stage is the falling stage from 15000 to 2500 at a rate of 1000 pulses. The whole signal collection time is from 1 second to 33 seconds, a total of 32 seconds. It is known that the stepper motor is 8 subdivided values. Through conversion, the speed change is 200 r / min to 600 r / min.

[0154] The peak indicators, pulse indicators and kurtosis indicators extracted in the time-frequency domain and the 11, 22, 33 and 44 order data extracted in the order domain constitute eight characteristic vectors to identify the health state of the two-dimensional (2D) piston pump under variable speed. 180 groups of data out of the 200 groups of data are trained and the remaining 20 groups are tested. The group iteration number is defined as group_iter=20, the attraction rate is defined as attraction_rate=0.8, the growth rate is defined as growth_rate=2, the reproduction rate is defined as reproduction_rate=1.8, the number of carnivorous plants is defined as nCPlant=10, and the number of prey is defined as nPrey=20.

[0155] The eight characteristic vectors obtained by time-frequency domain and order analysis are used as the training characteristic parameters of the CPA-SVM optimization classification algorithm. The target labels 1, 2, 3 and 4 in the last column are the four different wear states corresponding to the characteristic vectors, and then the range of the penalty factor c and the radial basis g is determined. Under the target function of the classification accuracy of SVM, the c and g in a large range are further optimized by the CPA optimization algorithm. Further, the CPA-SVM is used to train the health state evaluation classification model, and the optimal parameters obtained are best c=224.6036 and best g=31.5104. At this time, the model has the best target function score.

[0156] The optimization process is as follows:

[0157] The specific steps of step S42 are as follows:

[0158] The carnivorous algorithm continuously iterates and evolves to find suitable SVM parameter values, thereby improving the accuracy of SVM classification. This mathematical model can be simplified to find the optimal decision vector [c, g], where c and g represent the penalty parameter vector and the kernel function vector of SVM, respectively, to maximize the following objective:

[0159] max F=SVM_Acc

[0160]

[0161]

[0162] In the formula, F represents the objective function, SVM_Acc represents the classification accuracy of the SVM, M T represents the number of samples correctly classified into the corresponding class, M F represents the number of samples not correctly classified into the corresponding class; first define the number of group iterations group_iter = 20, the attraction rate attraction_rate = 0.8, the growth rate growth_rate = 2, the reproduction rate reproduction_rate = 1.8, the number of carnivorous plants nCPlant = 10 and the number of prey nPrey = 20, in the objective function SVM_Acc, the maximum number of iterations T = 40, for the optimization algorithm SVM parameters c and g, first perform rough optimization of the parameters by grid search method, and determine the ranges of c and g as 2 -2 -2 8 , i.e. 0.25-256 and 2 -5 -2 5 , i.e. 0.03-32.

[0163] Table 1 Test samples and corresponding target feature parameters

[0164]

[0165]

[0166] Table 2 Comparison of classification methods

[0167] Classification method Test accuracy (%) CPA-SVM 93 SVM 82.5 ELM 85

[0168] As can be seen from Table 1, the average classification accuracy of the health status evaluation obtained by using the unoptimized SVM and ELM methods is 82.5% and 85%, respectively, combined with Figure 5 and Figure 6, it can be seen that the classification accuracy obtained by the CPA optimized SVM algorithm is improved more, which verifies the effectiveness and accuracy of the method.

[0169] It should be understood that although the present specification is described in terms of various embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that those skilled in the art can understand. Equivalent embodiments or changes made without departing from the spirit of the present application shall be included within the scope of protection of the present application.

Claims

1. A variable speed based two-dimensional piston pump health condition assessment method, characterized in that: It comprises the following steps: S1, data acquisition: collect vibration signals reflecting the normal state and fault state of the two-dimensional piston pump as the original state data; S2, data preprocessing analysis: GLCT time-frequency analysis is performed on the original state data of the two-dimensional piston pump vibration signal collected in step S1, and the instantaneous rotating frequency of the vibration signal is extracted; S3, extracting angular domain vibration signal and order domain characteristic parameter, which comprises the following sub-steps: S31, data fitting processing is performed on the instantaneous rotating frequency extracted in step S21, and the data fitting method is adopted to intercept the data of the frequency rising stage of the instantaneous frequency for data fitting, so as to obtain the data fitting curve, and the order spectrum is obtained according to the fitting curve; S32、For the angular domain vibration signal s(l) of the two-dimensional piston pump, the data length of the equiangular interval sampling is L, then the peak index Cf, the pulse index If and the kurtosis index K of the angular domain vibration signal s(l) are calculated according to the following formulas respectively: V ​ where S rms is the root mean square value, S max is the peak value, is the absolute mean amplitude, and β is the kurtosis. Root mean square value S rms The formula for calculating the root mean square value S is: Peak S max The formula for calculating is: S max = max(|s(l)|) absolute average amplitude The formula for calculating the absolute average amplitude is: The calculation formula of kurtosis β is: Let the order spectrum of the angular domain vibration signal s(l) be Sl(m), where l = 1, 2, …, L, m is an order data length variable, and m = 1, 2, …, D max , where D max is the maximum order, and the order spectrum root mean square value Y1 is defined as follows: And the size of Y1 is used to measure the vibration energy intensity of the whole order spectrum; S33, the amplitudes corresponding to the 11th, 22nd, 33rd and 44th orders of the order spectrum are used as the order domain characteristic parameters for analysis: In the formula, n is the number of order spectrum data in the interval, Y2, Y3, Y4 and Y5 are the amplitudes corresponding to the 11th, 22nd, 33rd and 44th orders of the order spectrum respectively; S4, fault diagnosis model training and optimization, specifically comprising the following sub-steps: S41, constructing a fault diagnosis model: taking Cf, If, K V , Y1, Y2, Y3, Y4, Y5 as the training feature parameters of the CPA-SVM optimized classification algorithm to construct a fault diagnosis model, and dividing the sample library into a training sample set and a test sample set according to a proportion, then taking the training sample set as the input of the fault diagnosis model and the state label as the output of the fault diagnosis model; S42, under the target function of the classification accuracy of SVM, the penalty factor c and the radial basis g in a large range are accurately optimized by using the CPA optimization algorithm to optimize the model parameters; S43, the optimized CPA-SVM is used for fault classification model training, and the optimal model parameters best c and best g are obtained, so as to obtain the trained and optimized fault diagnosis model; S5, model verification: calling the trained and optimized fault diagnosis model, taking the test sample set as the input of the trained and optimized fault diagnosis model, and taking the state label as the output of the trained and optimized fault diagnosis model, the comprehensive performance of the diagnosis model is verified; S6, model diagnosis: inputting the vibration signal of the two-dimensional piston pump into the trained and optimized fault diagnosis model, the fault diagnosis model outputs the state label value, and the health state classification recognition result is obtained.

2. The variable speed-based two-dimensional piston pump health condition assessment method according to claim 1, characterized in that: Step S2 specifically comprises the following sub-steps: Step S21, the original STFT formula is constructed as follows: Wherein, w(u-t') is a certain window; u is a certain short time; Step S22, using a series of discrete demodulation operators based on the STFT formula to approximate the best demodulation operator The STFT formula based on the discrete demodulation operator is as follows: Step S23, for each TF point (t', w), if the discrete demodulation operator The modulation component of the proximity signal, whose TF representation around the IF, has a higher concentration of energy, whose amplitude |S(t', w, c)| reaches a maximum among all values, after which, for each TF point, the best parameter c' of the parameter c is obtained from the amplitude |S(t', w, c)|: The time-frequency representation of the proposed time-frequency analysis method is: GS(t',w)=S(t',w,c') The spectrum calculation formula is: Spec(t',w) = | GS(t',w) | 2 Step S24, determining a discrete demodulation operator Introducing a rotation parameter a, to achieve rotation on the time-frequency plane: where T s is the sampling time, F s is the sampling frequency; Suppose that the rotating parameter a has N values, and the time-frequency plane is divided into N+1 segments: a=-π / 2+π / (N+1),-π / 2+2·π / (N+1)...-π / 2+N·π / (N+1); Step S25, ensuring demodulation operators The STFT formulation of the discrete demodulation operator can be re-written to the GLCT formulation, describing all possible modulation components in the signal: Step S26, GLCT time-frequency analysis is performed according to the GLCT formula of step S25, and the instantaneous rotating frequency S(t',w,a) of the vibration signal is extracted.

3. The variable speed-based two-dimensional piston pump health condition assessment method of claim 1, wherein: Step S31 specifically comprises the following sub-steps: S311, the relationship between the rotating speed and the first-order instantaneous frequency is represented as: Wherein, f(t) is the shaft rotating frequency, and P(t) is the reference shaft rotating speed; S312, the fitting curve is obtained by using the piecewise multi-order polynomial fitting method: R k (t) = a k +b k t+c k t 2 wherein the function of the speed curve of each segment Rk(t) is k, the serial number of the segment, a k , b k , c k are polynomial coefficients; S313, the integral equation of the rotating speed fitting curve is solved to obtain an equal-angle interval interpolation phase discrimination time sequence, and the solving process is as follows: Wherein, T n is the key phase time, n is the time sequence number, T0 is the initial fitting time; the above formula is substituted into the fitting curve formula, and a set of equal angle interval interpolation phase time sequence T n is obtained by effective algebraic solution. S314, based on the equal-angle interval interpolation phase discrimination time sequence obtained in step S313, the Lagrange value method is used to complete smoothing processing to obtain an angle domain signal; S315, the angle domain signal data obtained in step S314 is subjected to Fourier transform to obtain an order spectrum.

4. The variable speed-based two-dimensional piston pump health condition assessment method of claim 1, wherein: 200 groups of data are acquired to divide the training sample set and the test sample set in a ratio of 4:1, the in-group iteration number group_iter is defined as 20, the attraction rate attraction_rate is defined as 0.8, the growth rate growth_rate is defined as 2, the reproduction rate reproduction_rate is defined as 1.8, the number of carnivorous plants nCPlant is defined as 10, and the number of prey nPrey is defined as 20.

5. The variable speed-based two-dimensional piston pump health condition assessment method of claim 1, wherein: Step S42 is specifically: 8 characteristic vectors obtained through time-frequency domain and order analysis are used as training characteristic parameters of the CPA-SVM optimized classification algorithm, and target labels 1, 2, 3 and 4 in the last column are four different wear states determined by the optimized classification algorithm according to the characteristic vectors, then the range of c and g is determined, and under the target function of the classification accuracy of the SVM, c and g in a large range are further optimized by using the CPA optimization algorithm to accurately optimize the parameters.

6. The variable speed-based two-dimensional piston pump health condition assessment method of claim 1, wherein: In step S1, the vibration signal is collected by the vibration sensor installed on the two-dimensional piston pump.

7. The variable speed-based two-dimensional piston pump health condition assessment method of claim 1, wherein: The optimal parameters are best c=224.6036 and best g=31.5104.

8. The variable speed-based two-dimensional piston pump health condition assessment method of claim 1, wherein: The specific steps of step S42 are: The carnivorous algorithm continuously iterates and evolves to find suitable SVM parameter values, thereby improving the accuracy of SVM classification. The mathematical model can be simplified to find the optimal decision vector [c, g], wherein c and g represent the penalty parameter vector and the kernel function vector of the SVM respectively, and the maximum is as follows: max F=SVM_Acc In the formula, F represents the objective function, SVM_Acc represents the classification accuracy of SVM, M T represents the number of samples correctly classified into the corresponding class, M F represents the number of samples not correctly classified into the corresponding class; first, define the number of group iterations group_iter=20, the attraction rate attraction_rate=0.8, the growth rate growth_rate=2, the reproduction rate reproduction_rate=1.8, the number of carnivorous plants nCPlant=10 and the number of prey nPrey=20, in the objective function SVM_Acc, the maximum number of iterations T=40, for the optimization algorithm SVM parameters c and g, first, through the grid search method, the parameters are roughly optimized, and the ranges of c and g are determined as 2 -2 -2 8 , that is, 0.25-256 and 2 -5 -2 5 , that is, 0.03-32.

Images