A method for detecting the air tightness of a diaphragm accumulator

Abstract

The present application relates to the technical fields of air tightness detection, more particularly to a diaphragm accumulator air tightness detection method, which is used to solve the problem that the prior art cannot accurately depict the equipment operation characteristics, cannot identify typical defect patterns in combination with spectral clustering, and cannot provide quantitative basis for equipment failure trend analysis and intelligent maintenance; the present application depicts the equipment operation characteristics by constructing a low-dimensional state atlas, identifies typical defect patterns in combination with spectral clustering, establishes a nonlinear correlation model to reveal the defect evolution relationship, optimizes the classification model parameters by using an evolutionary algorithm to improve the identification accuracy, mines the most possible defect evolution path based on the transition probability matrix, and provides quantitative basis for equipment failure trend analysis and intelligent maintenance.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of air tightness detection, and more particularly to a diaphragm accumulator air tightness detection method. BACKGROUND

[0002] Air tightness of a diaphragm accumulator is a core guarantee for hydraulic system performance, but the existing detection technology has three bottlenecks: the traditional gas pressure decay method cannot quantify the pre-charged gas pressure loss caused by diaphragm fatigue, and it is difficult to evaluate long-term reliability; the water immersion method is low in efficiency and easy to damage the product, and cannot meet the batch production demand; the helium mass spectrometry leak detection method is high in cost and complex in operation, and is difficult to popularize in industrial scenes, and the quantitative correlation model of diaphragm fatigue characteristics and air tightness decay is missing, resulting in lack of data support for material selection and structure optimization.

[0003] The patent application with the publication number CN119164562A discloses a diaphragm accumulator air tightness detection method, system and storage medium, which comprises obtaining preparation completion trigger information of a diaphragm accumulator; obtaining test plan information of the diaphragm accumulator based on the preparation completion trigger information; controlling a pre-set hydraulic system to perform liquid charging test on the diaphragm accumulator according to the test plan information to determine gas pressure detection information; analyzing the gas pressure detection information and the test plan information to determine liquid gas pressure information of the hydraulic system; determining whether the gas pressure detection information meets the requirements of the liquid gas pressure information; if not, prompting according to pre-set unqualified prompt information; if yes, prompting according to pre-set qualified prompt information, the application has the effect of improving the air tightness detection accuracy of the diaphragm accumulator;

[0004] However, the above-mentioned reference patent controls the liquid charging test of the hydraulic system on the diaphragm accumulator, improves the air tightness and gas pressure detection accuracy by comparing the gas pressure and liquid charging time information respectively, and tests the liquid charging according to the method to improve the liquid charging convenience when the test rule does not conform to the ladder rule, but cannot accurately reconstruct the internal structure characteristics of the diaphragm accumulator, cannot accurately identify the crack propagation behavior, reduces the sensitivity to different scale defect characteristics, and cannot provide technical support for equipment health state evaluation and intelligent maintenance decision; at the same time, cannot accurately depict the equipment operation characteristics, cannot identify the typical defect mode combined with spectral clustering, and cannot provide quantitative basis for equipment fault trend analysis and intelligent maintenance.

[0005] Therefore, the application proposes a diaphragm accumulator air tightness detection method aiming at the above-mentioned problems. SUMMARY

[0006] The application aims to provide a diaphragm accumulator air tightness detection method, solve the problem that the prior art cannot accurately reconstruct the internal structure characteristics of the diaphragm accumulator, cannot accurately identify the crack propagation behavior, reduces the sensitivity to different scale defect characteristics, cannot provide technical support for equipment health state evaluation and intelligent maintenance decision, and cannot accurately depict the equipment operation characteristics, cannot identify the typical defect mode in combination with spectral clustering, and cannot provide quantitative basis for equipment failure trend analysis and intelligent maintenance.

[0007] The application aims to provide a diaphragm accumulator air tightness detection method, solve the problem that the prior art cannot accurately reconstruct the internal structure characteristics of the diaphragm accumulator, cannot accurately identify the crack propagation behavior, reduces the sensitivity to different scale defect characteristics, cannot provide technical support for equipment health state evaluation and intelligent maintenance decision, and cannot accurately depict the equipment operation characteristics, cannot identify the typical defect mode in combination with spectral clustering, and cannot provide quantitative basis for equipment failure trend analysis and intelligent maintenance.

[0008] A diaphragm accumulator air tightness detection method comprises the following steps:

[0009] Step one: collecting state monitoring data in the running process of the diaphragm accumulator, the state monitoring data comprising gas pressure data, temperature data, vibration data, acoustic emission signals and strain data, and performing pretreatment operation on the collected state monitoring data;

[0010] Step two: performing real-time monitoring and closed-loop control on the gas flow, gas pressure and gas path state in the inflation process, and automatically switching the gas path according to the detection requirement;

[0011] Step three: fusing the gas pressure, temperature, vibration and acoustic emission signals, and adopting acoustic wave propagation inversion, dynamic crack identification and multi-scale analysis method to preliminarily position and analyze the evolution trend of the internal defects of the accumulator;

[0012] Step four: constructing a high-dimensional graph based on the state monitoring data, depicting the nonlinear correlation between the defect modes, combining the evolutionary algorithm to optimize the model parameters, and realizing the defect identification and evolution path analysis.

[0013] As a preferred embodiment of the application, the process of real-time monitoring and closed-loop control on the gas flow, gas pressure and gas path state in the inflation process in step two comprises:

[0014] Using a thermal flowmeter to collect the gas flow, and transmitting the collected gas flow data to a control unit;

[0015] Using a piezoelectric pressure sensor to collect the gas pressure, and transmitting the collected data to the control unit;

[0016] Using a switch state sensor to collect the gas path conduction state, using a valve position sensor to collect the valve opening degree, using a differential pressure flow detection module to collect the gas flow state, and inputting the collected data into the control unit;

[0017] The control system is composed of sensors, a control unit and an actuator, the control unit receives sensing signals from gas flow, gas pressure and gas path state, compares the received actual measurement values with preset target set values, calculates deviation, generates a control output signal according to the deviation, and drives the actuator to adjust the system state;

[0018] The control unit adopts a proportional-integral-derivative control algorithm to perform closed-loop adjustment on the gas flow and the gas pressure, and changes the gas flow size, adjusts the gas pressure or switches the gas path according to preset program instructions.

[0019] As a preferred embodiment of the present application, the process of automatically switching the gas path according to the detection requirement in step two comprises:

[0020] The control system determines whether to perform gas path switching according to the task identification and the operation parameters, and generates a switching command after identifying the switching condition;

[0021] The control system selects a target gas path and issues a switching command, and the gas path switching valve changes the gas flow direction in response to the command;

[0022] The components in the control system work cooperatively, and the gas path switching mechanism guides the gas flow into different paths according to the task identification;

[0023] The control system realizes path switching and parameter adjustment of the gas path system.

[0024] As a preferred embodiment of the present application, the process of fusing the gas pressure, temperature, vibration and acoustic emission signals and performing acoustic wave propagation inversion analysis in step three comprises:

[0025] The preprocessed gas pressure data, temperature data, vibration data and acoustic emission signals are obtained, and features are extracted therefrom, the extracted features are normalized and weighted to be fused into a comprehensive feature vector F;

[0026] The acoustic wave propagation inversion method is used to calculate the material internal physical parameters or structure information according to the boundary measured acoustic wave signals;

[0027] In an isotropic linear elastic medium, the acoustic wave satisfies a three-dimensional second-order wave equation;

[0028] When the acoustic source position and medium parameters are known, the finite difference method is used to solve the wave equation to obtain a theoretical wave field u(x, t);

[0029] The inversion target is to optimize the medium parameters m(x) by minimizing the difference between the observed wave field u gc (x i ,t) and the calculated wave field u(x i ,t; m);

[0030] introducing an adjoint field λ(x, t), calculating the gradient of the objective function to the medium parameter by the adjoint state equation;

[0031] Using the gradient, the medium parameter m is iteratively updated by using the L-BFGS algorithm until convergence.

[0032] As a preferred embodiment of the present application, the process of using dynamic crack identification and multi-scale analysis method in step three includes:

[0033] Dynamic crack propagation causes the stiffness of the material to decrease, which is manifested as a decrease in modal frequency and an increase in damping ratio of the structure, and the crack location can be identified by combining the local energy density change;

[0034] The Hilbert-Huang transform is used to process the response signal, and the instantaneous frequency f n (t) and the damping ratio ζ n (t) of the nth order are extracted;

[0035] The local energy density function is introduced, and the energy concentration degree increases when the crack propagates, and the crack location is determined by combining the change of wave propagation path;

[0036] The multi-scale analysis method is used to process the structural characteristics at different scales, and the wavelet decomposition is used to realize the cross-scale feature extraction;

[0037] The micro variable is defined, where ε is a small parameter much smaller than 1, and the field quantity u has a multi-scale expansion form, the expansion is substituted into the control equation and separated by ε, and the variable is solved step by step;

[0038] The wavelet basis function ψ j,k (x) is used for multi-scale decomposition of the signal.

[0039] As a preferred embodiment of the present application, the process of preliminary positioning and evolution trend analysis of the internal defects of the accumulator in step three includes:

[0040] The defects in the accumulator can be analyzed by multi-source data fusion to analyze the spatio-temporal evolution characteristics;

[0041] The internal reflection coefficient distribution r(x) of the accumulator is reconstructed by inversion, the measured echo s(t) is calculated by inverse calculation to obtain the r(x) distribution map, and the defect location is determined;

[0042] The defect parameter D(t) is defined to represent the damage factor, and its evolution is defined to comply with the exponential growth model;

[0043] When the defect parameter D(t) reaches or exceeds the set threshold D sd , it is determined that the defect is in a dangerous state, and repair or replacement measures need to be taken.

[0044] As a preferred embodiment of the present application, the process of constructing high-dimensional map based on state monitoring data and depicting the nonlinear correlation between defect patterns in step four comprises:

[0045] Obtain the pre-processed state monitoring data, form time series state vectors, and construct a state data matrix for the total observation time length T;

[0046] Extract the state information of consecutive time periods in turn using a fixed window length and sliding step to form a series of trajectory segments, calculate the total number of trajectories according to the set window length and sliding step, and generate a trajectory set;

[0047] Use the diffusion mapping algorithm to reduce the dimension of the high-dimensional trajectory, measure the similarity between trajectories by constructing a Gaussian kernel function, and normalize to obtain a transition probability matrix;

[0048] Extract the first several non-trivial eigenvectors and their corresponding eigenvalues of the matrix to construct the node representation in the low-dimensional embedding space, and finally obtain a set of low-dimensional nodes, which constitute the map structure of the device state;

[0049] For each pair of low-dimensional nodes, if one node is the nearest neighbor of the other node, an edge is established between the two nodes, and finally a weighted graph is constructed.

[0050] As a preferred embodiment of the present application, the process of depicting the nonlinear correlation between defect patterns in step four comprises:

[0051] Classify the low-dimensional node set using the spectral clustering method, divide nodes with similar characteristics into the same category, and each category represents a defect pattern. Calculate the center point of each category as the representative state of the defect in that category;

[0052] Calculate the squared Euclidean distance between the centers of each defect category as a measure of their differences, and finally obtain a nonlinear correlation matrix reflecting the mutual relationship between each defect pattern;

[0053] Set the threshold to the top 20% quantile of the similarity between all defect pairs. If the squared Euclidean distance between two defect categories is less than the threshold, a directed edge is established between the corresponding nodes, and a directed association graph between defects is constructed, which represents the possible evolution path of the defects.

[0054] As a preferred embodiment of the present application, the process of optimizing model parameters combined with evolutionary algorithms in step four comprises:

[0055] Design an input-output model with device state vectors as input and defect category numbers as output, and model parameters as optimization variables;

[0056] Define a comprehensive evaluation function, including the following three: classification accuracy, loss function value and parameter norm, each item is combined into a whole fitness index according to the set weight;

[0057] The evolutionary search steps are as follows:

[0058] T1: initialize a batch of random parameter individuals, calculate the fitness value of each individual;

[0059] T2: select the highest fitness of part of individuals as the parent, generate new individuals by crossing operation, and add a small variation disturbance:

[0060] T3: repeat the above process until the preset iteration number or convergence condition is reached;

[0061] Finally output the optimal model parameters.

[0062] As a preferred embodiment of the application, the process of realizing defect identification and evolution path analysis in step four includes:

[0063] Input each state vector collected during the operation of the equipment into the trained model, output the corresponding defect category label, and generate a complete defect label sequence;

[0064] Statistical adjacent time defect label transition frequency, calculate the frequency ratio of each defect to another defect, and form a defect transition probability matrix;

[0065] Define a path sequence composed of multiple defect categories, and the probability of the path is the product of all adjacent defect transition probabilities in the path;

[0066] Among all possible paths, find the path with the maximum probability as the most likely defect evolution path.

[0067] Compared with the prior art, the application has the following advantages:

[0068] (1) In the present application, by fusing gas pressure, vibration, temperature and acoustic emission signals, key features are extracted and a comprehensive feature vector is constructed, the internal structure characteristics of the diaphragm accumulator are reconstructed by combining the acoustic wave propagation inversion method, the modal frequency, damping ratio and energy distribution changes are analyzed based on the dynamic crack identification model, the crack propagation behavior is identified, the multi-scale wavelet analysis method is introduced, the sensitivity to defect features of different scales is enhanced, the defect positioning accuracy and evolution trend prediction ability are improved, and the whole process integrates physical models and data-driven methods, which provides technical support for equipment health state evaluation and intelligent maintenance decision;

[0069] (2) In the present application, the device running characteristics are described by constructing a low-dimensional state map, the typical defect modes are identified by combining spectral clustering, the non-linear correlation model is established to reveal the defect evolution relationship, the classification model parameters are optimized by using the evolutionary algorithm to improve the recognition accuracy, the most possible defect evolution path is mined based on the transition probability matrix, and quantitative basis is provided for device fault trend analysis and intelligent maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 The flow chart of the air tightness detection method of the diaphragm accumulator in the present application is shown.

[0071] Figure 2 The step flow chart of the evolutionary search in the present application is shown. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings of the embodiments of the present application; obviously, the described embodiments are only part of the embodiments of the present application, but not all the embodiments; based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0073] Embodiment one: as shown in the present application, a diaphragm accumulator air tightness detection method is proposed, which comprises the following steps: Figure 1

[0074] Step one: collecting the state monitoring data of the diaphragm accumulator during operation, the state monitoring data including gas pressure data, temperature data, vibration data, acoustic emission signal and strain data, and performing pretreatment operation on the collected state monitoring data, the pretreatment operation including signal denoising, outlier rejection and time synchronization;

[0075] By collecting the multi-dimensional state monitoring data of the diaphragm accumulator during operation, such as gas pressure, temperature, vibration, acoustic emission and strain, and performing pretreatment operation such as signal denoising, outlier rejection and time synchronization, the device running state information is comprehensively obtained, and the data quality and consistency are effectively improved; high-quality data not only provides a reliable basis for subsequent state recognition, fault diagnosis and life prediction, but also helps to capture early and small abnormal changes of the device, and realizes early warning of faults.

[0076] Step two: real-time monitoring and closed-loop control of the gas flow, gas pressure and gas path state during the inflation process, and automatically switching the gas path according to the detection requirements;

[0077] The process of real-time monitoring and closed-loop control of the gas flow, gas pressure and gas path state during the inflation process in step two comprises:

[0078] ​Gas flow is collected by using thermal flow meter, sensor outputs electric signal through heat change, collected gas flow data is transmitted to control unit, control unit adjusts regulating valve opening degree according to current flow value;

[0079] Gas pressure is collected by using piezoelectric pressure sensor, sensor outputs electric signal through voltage change, collected data is transmitted to control unit, control unit compares current gas pressure value with set value and executes adjusting operation;

[0080] Gas path conduction state is collected by using switch state sensor, sensor outputs binary signal to indicate whether gas path is conducted, valve opening degree is collected by using valve position sensor, sensor outputs continuous signal to indicate valve position, gas flow state is collected by using differential pressure flow detection module, sensor outputs signal to judge whether gas flows, collected data is input into control unit to judge whether gas path is abnormal;

[0081] Control system is composed of sensor, control unit and actuator, control unit receives sensing signals from gas flow, gas pressure and gas path state, compares received actual measurement value with preset target set value, calculates deviation, generates control output signal according to the deviation, drives actuator to adjust system state, and adjustment process is completed based on feedback mechanism;

[0082] Control unit adopts proportional-integral-derivative control algorithm to perform closed-loop adjustment on gas flow and gas pressure, changes gas flow size, adjusts gas pressure or switches gas path according to preset program instruction;

[0083] Proportional term: control amount is generated according to current error;

[0084] Integral term: correction amount is generated according to error accumulation;

[0085] Derivative term: dynamic compensation amount is generated according to error change rate;

[0086] Controller parameters are set according to preset rules;

[0087] Actuator includes gas flow regulating valve, pressure regulating valve and gas path switching valve;

[0088] The process of automatically switching gas path according to detection requirement in step two includes:

[0089] Control system judges whether to execute gas path switching according to task identification and running parameters, task identification is provided by external input, running parameters include gas pressure, flow and time stamp, control system generates switching command after identifying switching condition;

[0090] The control system selects the target gas path and issues a switching command, the gas path switching valve responds to the command to change the direction of gas flow, after switching is completed, continue to collect gas flow, gas pressure and gas path state data;

[0091] Each component in the control system works cooperatively, the sensor is responsible for data collection, the control unit is responsible for data processing and generating control signals, the actuator adjusts operation according to the control signals, and the gas path switching mechanism guides the gas flow into different paths according to the task identification;

[0092] The control system realizes path switching and parameter adjustment of the gas path system, and the control unit executes subsequent operations according to the collected data;

[0093] Through real-time collection of gas flow, pressure and gas path state information by sensors, combined with closed-loop control mechanism based on PID algorithm, dynamic adjustment and stable control of gas parameters are realized to ensure the safety and precision of the inflation process; at the same time, the system judges and executes the automatic switching of the gas path path through the task identification and operating parameters, improving the flexibility and adaptability of the system; each component works cooperatively to form a "perception-decision-execution" closed loop, which not only improves the control response speed and adjustment precision, but also supports automatic operation under multiple working conditions, suitable for gas path management and intelligent control in complex detection tasks.

[0094] Step three: fuse gas pressure, temperature, vibration and acoustic emission signals, use acoustic wave propagation inversion, dynamic crack identification and multi-scale analysis methods to preliminarily locate and analyze the evolution trend of internal defects of the accumulator;

[0095] The process of fusing gas pressure, temperature, vibration and acoustic emission signals and performing acoustic wave propagation inversion analysis in step three includes:

[0096] Obtain the preprocessed gas pressure data, temperature data, vibration data and acoustic emission signals, and perform feature extraction on them, the feature extraction process is as follows:

[0097] Extract the mean, variance, and average absolute value of the difference between adjacent time points from the gas pressure data;

[0098] Extract the maximum value, minimum value, standard deviation and temperature change rate from the temperature data;

[0099] Extract the peak value, effective value, peak-to-peak value, kurtosis, main frequency and its amplitude from the vibration data;

[0100] Extract the pulse count, impact number, amplitude maximum value, main frequency and its energy distribution from the acoustic emission signal;

[0101] Normalize and weight the extracted features into a comprehensive feature vector F, the weight coefficients are obtained by historical data training, which is expressed as: F=[f1,f2,…fn ] T where f i represents the i-th feature component in the comprehensive feature vector, n represents the feature dimension, and F represents the overall physical characteristics of the current system state;

[0102] The physical property parameters or structural information inside the material are calculated according to the boundary measured acoustic wave signals by the acoustic wave propagation inversion method;

[0103] In the isotropic linear elastic medium, the acoustic wave satisfies the three-dimensional second-order wave equation, which is expressed as follows: ;

[0104] Where u(x,t) represents the displacement field of the acoustic wave at position x and time t, c is the wave velocity in the medium, and ▽ 2 is the Laplace operator;

[0105] When the acoustic source position and medium parameters are known, the finite difference method is used to solve the wave equation to obtain the theoretical wave field u(x,t);

[0106] The objective of inversion is to optimize the medium parameter m(x) by minimizing the difference between the observed wave field u gc (x i ,t) and the calculated wave field u(x i ,t; m), and the comprehensive feature vector F is used to set the value range of the medium parameter m(x) in the wave equation solving process as the initial constraint condition of inversion;

[0107] The objective function is expressed as follows:

[0108] Where J(m) is the objective function, representing the energy difference between the observation and the calculated wave field, m(x) is the medium parameter distribution, N is the number of sensors, and T is the total observation time;

[0109] The adjoint field λ(x,t) is introduced, and the gradient of the objective function with respect to the medium parameter is calculated by the adjoint state equation:

[0110] ;

[0111] Using the gradient, the L-BFGS algorithm is used to iteratively update the medium parameter m until convergence;

[0112] The process of using the dynamic crack identification and multi-scale analysis method in step three includes:

[0113] Dynamic crack propagation causes the stiffness of the material to decrease, which is manifested as a decrease in modal frequency and an increase in damping ratio of the structure, and the crack position can be identified by combining the local energy density change, and the structural dynamics equation is expressed as follows:

[0114] where M is the mass matrix, C is the damping matrix, K(t) is the stiffness matrix which varies with crack propagation, f(t) is the external excitation, u(t) is the displacement vector, are the first and second time derivatives of displacement;

[0115] When the integrated eigenvector F numerically mutates or deviates from the historical mean value beyond a set threshold, it is determined that the system is in an abnormal operating state, triggering the dynamic crack identification module to start. This vector does not participate in the calculation of modal frequency or damping ratio, and only serves as a trigger condition for crack identification.

[0116] The Hilbert-Huang transform is used to process the response signal, and the instantaneous frequency f n (t) and the damping ratio ζ n (t) are extracted:

[0117] where m n is the nth order mass coefficient, c n is the nth order damping coefficient, and k n (t) is the nth order stiffness coefficient;

[0118] Crack development causes the stiffness coefficient k n (t) to decrease, resulting in a decrease in modal frequency and an increase in damping ratio;

[0119] The local energy density function is introduced:

[0120] where is the displacement velocity vector;

[0121] The energy concentration degree increases during crack propagation, and the crack position is determined by combining the change in wave propagation path;

[0122] Discrete wavelet transform is used to decompose the signal into three levels, and the Daubechies db4 wavelet basis function is selected to extract structural features at macro and micro scales. The micro variable is defined, where ε is a small parameter much smaller than 1, and the field quantity u has a multi-scale expansion form:

[0123] u(x, t) = u0(x, ξ, t) + εu1(x, ξ, t) + ε 2 u2(x, ξ, t) + …;

[0124] Substitute the expansion into the control equation and separate the variables according to ε, and solve step by step;

[0125] The wavelet basis function ψ j,k (x) is used for multi-scale decomposition of the signal:

[0126] where j controls the scale, k controls the translation, c j,k are wavelet coefficients, reflecting local features at different scales;

[0127] The comprehensive feature vector F reflects the trend of the overall state of the system. During the wavelet decomposition process, the analysis weight at different scales is set according to the change of the feature vector. For example, when the feature vector shows that the high-frequency energy is enhanced, the wavelet coefficients in the high-frequency band are preferentially analyzed.

[0128] The process of preliminary positioning and evolution trend analysis of internal defects of the accumulator in step three includes:

[0129] The defects (such as cavitation, fatigue cracks or corrosion points) in the accumulator can be analyzed for their spatio-temporal evolution characteristics through multi-source data fusion.

[0130] The internal reflectivity distribution r(x) of the accumulator is reconstructed using inversion:

[0131] where G(x s ,x,t) is the Green's function from the sound source position x s to point x, G(x,x r ,t) is the Green's function from point x to the receiver position x r , s(t) is the received echo signal, and r(x) is the reflectivity distribution representing the defect intensity.

[0132] The measured echo s(t) is used to calculate the r(x) distribution through inversion to determine the defect location.

[0133] The defect parameter D(t) is defined to represent the damage factor, and its evolution is defined to conform to the exponential growth model:

[0134] D(t)=D0e θt , where D0 is the initial defect degree and θ is the growth coefficient, which is obtained by fitting the experiment. In practical applications, the value of θ is affected by the current working condition, and it can be obtained through linear mapping of the comprehensive feature vector.

[0135] When the defect parameter D(t) reaches or exceeds the set threshold D sd , it is determined that the defect is in a dangerous state, and repair or replacement measures need to be taken.

[0136] An example of defect positioning based on the content of step three is given:

[0137] Consider a scenario where an accumulator is used in a wind turbine system to store and release energy to balance the demand of the power grid; due to long-term high pressure and frequent temperature changes, the accumulator is prone to internal defects such as fatigue cracks or corrosion points; in order to ensure the safe operation of the system, it is necessary to regularly monitor the health of the accumulator;

[0138] First, install multiple acoustic sensors on the outside of the accumulator to collect acoustic signals on the boundary; at the same time, use built-in pressure, temperature and vibration sensors to obtain real-time data;

[0139] After preprocessing the collected data, feature extraction is performed according to the method described in step three, and these features are fused into a comprehensive feature vector F;

[0140] According to the known sound source position and medium parameters (such as wave speed c), the finite difference method is used to solve the wave equation;

[0141] Using the difference between the observed wave field and the calculated wave field, the medium parameters m(x) are optimized through the target function minimization process, and the L-BFGS algorithm is used for iterative update until convergence;

[0142] Finally, according to the reconstructed internal reflectance distribution r(x) of the accumulator, the location of potential defects is determined;

[0143] When the comprehensive feature vector F is detected to have a sudden change, start the dynamic crack identification module to analyze the changes in frequency, mode and energy characteristics;

[0144] Use wavelet decomposition method to analyze the structural characteristics at different scales to determine the specific location and development trend of the crack;

[0145] After the above steps, the positioning results of the internal defects of the accumulator and the defect parameter D(t) curve with time are obtained, and Table 1 shows some key moments of defect parameter values and their corresponding dangerous state judgments:

[0146] Table 1 Defect parameter conditions

[0147]

[0148] Among them, the set dangerous threshold is 0.08, from the table we can see that at the 100th day, the defect parameter D(t) of the accumulator exceeds the set dangerous threshold, indicating that immediate repair or replacement measures need to be taken to avoid accidents;

[0149] By fusing multi-source heterogeneous signals such as gas pressure, temperature, vibration and acoustic emission, key features are extracted and a comprehensive feature vector is constructed, and the internal structure characteristics of the diaphragm accumulator are reconstructed by combining the acoustic wave propagation inversion method; the dynamic crack identification model is used to analyze the changes of modal frequency, damping ratio and energy distribution, and the crack propagation behavior is identified; at the same time, the multi-scale wavelet analysis method is introduced to enhance the sensitivity to defect features of different scales and improve the defect positioning accuracy and evolution trend prediction ability; the whole process is based on the combination of physical model and data driving, has strong robustness and engineering applicability, and can effectively support equipment health state evaluation and intelligent maintenance decision.

[0150] Embodiment two: the technical scheme of the embodiment of the application is different from that of embodiment one in that

[0151] As shown in Figure 1 and Figure 2 , step four: based on the state monitoring data, a high-dimensional graph is constructed to depict the nonlinear correlation between defect modes, and an evolutionary algorithm is used to optimize the model parameters to realize defect identification and evolution path analysis;

[0152] The process of constructing a high-dimensional graph based on state monitoring data and depicting the nonlinear correlation between defect modes in step four includes:

[0153] The preprocessed state monitoring data is obtained and is formed into a time series state vector:

[0154] x(t k )=[x (1) (t k ),x (2) (t k ),…,x (n) (t k )] T ∈R n ;

[0155] Where t k is the kth sampling time, n is the number of collected data, and each component represents the real-time measurement value of a type of physical quantity;

[0156] For the total observation time length T, a state data matrix is constructed:

[0157] X=[x(t1),x(t2),…,x(t T )]∈R n×T ;

[0158] A fixed window length and a sliding step are used to extract state information in continuous time periods in turn, forming a series of trajectory segments, each trajectory being composed of data at multiple time points, and according to the set window length and sliding step, the total number of trajectories is calculated and a trajectory set is generated;

[0159] The diffusion map algorithm is used to reduce the dimensionality of the high-dimensional trajectory, the similarity between trajectories is measured by constructing a Gaussian kernel function, and the transition probability matrix is obtained by normalization. The first several non-trivial eigenvectors and their corresponding eigenvalues of the matrix are extracted to construct the node representation in the low-dimensional embedding space:

[0160] z i =[λ1φ1(i),λ2φ2(i),…,λ d φ d (i)] T ∈R d , where λ j is the jth eigenvalue, and d is the dimension of the embedding space.

[0161] Finally, a set of low-dimensional nodes is obtained, which constitutes the graph structure of the device state.

[0162] For each pair of low-dimensional nodes, if one node is the nearest neighbor of the other node, an edge is established between the two nodes, and the weight of the edge is determined according to the distance between the nodes. The smaller the distance, the higher the weight. Finally, a weighted graph is constructed:

[0163] , where Z is the node set, each node corresponds to a low-dimensional embedding state, β is the edge set, representing the connection relationship between states, and W is the weighted adjacency matrix.

[0164] The process of depicting the nonlinear association between defect patterns in step four includes:

[0165] The spectral clustering method is used on the low-dimensional node set to classify, and nodes with similar characteristics are divided into the same category. Each class represents a defect pattern, and the center point of each class is calculated as the representative state of the defect in that class.

[0166] The Euclidean distance square between the centers of each defect class is calculated as a measure of their difference, and a nonlinear function is introduced to convert the distance into association strength. The smaller the distance, the stronger the association. Finally, a nonlinear association matrix reflecting the mutual relationship between each defect pattern is obtained.

[0167] Set the threshold to the top 20% quantile of the similarity between all defect pairs. If the Euclidean distance square between two defect categories is less than the threshold, a directed edge is established between the corresponding nodes, and a directed association graph between defects is constructed:

[0168] , where C is the total number of defect categories, β c is the directed edge set, representing the possible transition direction of the defect category, and Q is the nonlinear association strength matrix. The graph represents the possible evolution path of the defect.

[0169] The process of optimizing model parameters in step four by combining evolutionary algorithms includes:

[0170] An input-output model is designed, with the input being the device state vector and the output being the defect category number, and the model parameters being the variables to be optimized;

[0171] A comprehensive evaluation function is defined, including the following three items:

[0172] Classification accuracy: the degree of consistency between the model prediction result and the true label;

[0173] Loss function value: reflecting the size of prediction error;

[0174] Parameter norm: control model complexity to prevent overfitting;

[0175] Each item is combined into a whole fitness index according to the set weight:

[0176] ;

[0177] Where η is the model parameter to be optimized, Acc(η) is the classification accuracy, L(η) is the loss function value, are weight coefficients;

[0178] The evolutionary search steps are as follows:

[0179] T1: Initialize a batch of random parameter individuals, and calculate the fitness value of each individual;

[0180] T2: Select the highest fitness of some individuals as the parent, generate new individuals by crossover operation, and add a small variation disturbance:

[0181] T3: Repeat the above process until the preset iteration number or convergence condition is reached;

[0182] When the population converges or reaches the maximum iteration number, select the individual with the highest fitness as the final model parameter set and save it;

[0183] The process of realizing defect recognition and evolutionary path analysis in step four includes:

[0184] Each state vector collected during the operation of the device is input into the trained model, and the corresponding defect category label is output:

[0185] , where is the predicted defect category label at time t k , f(·) is the defect recognition model function, x(t k ) is the state vector at time t k , is the optimal model parameter after training;

[0186] Generate a complete defect label sequence to record the changes of defect types during the operation of the equipment;

[0187] Calculate the transition frequency between adjacent defect labels at adjacent time points, calculate the frequency ratio of each defect to another defect, and form a defect transition probability matrix, which describes the possible transition trend between defects;

[0188] Define a path sequence composed of multiple defect categories, and the probability of the path is the product of the transition probabilities of all adjacent defects in the path:

[0189] ;

[0190] Where The probability of transition from the i k th defect to the i k+1 th defect, K is the total number of defect categories in the path;

[0191] Among all possible paths, find the path with the maximum probability:

[0192] , where is the optimal evolution path;

[0193] As the most likely defect evolution path;

[0194] Give an example of high-dimensional graph modeling and defect evolution path prediction simulation according to the content of step four:

[0195] Consider an industrial manufacturing environment, continuously monitor the state of a key production equipment, which may have various types of defects such as mechanical wear and tear, fatigue cracks, etc. due to long-term operation; In order to ensure the safe operation of the equipment and prolong its service life, it is necessary to construct a high-dimensional graph through state monitoring data and analyze the nonlinear correlation between these defect patterns to predict the potential defect evolution path;

[0196] Collect state monitoring data of multiple physical quantities (such as vibration, temperature, pressure, etc.) from the equipment to form a time series state vector;

[0197] According to the total observation time length T, construct a state data matrix X, which contains all the data at the sampling time;

[0198] Use a fixed window length and sliding step to extract a series of trajectory segments from X, and apply diffusion mapping algorithm to reduce the dimension of these high-dimensional trajectories to obtain low-dimensional node representation z i ;

[0199] A weighted graph is established based on the distance between low-dimensional nodes to reflect the graph structure of the device state;

[0200] Spectral clustering method is used for classification on the low-dimensional node set to determine each defect mode, calculate the Euclidean distance square between defect centers, and convert it into a nonlinear correlation strength to construct a directed correlation graph and show the transfer possibility between different defect modes;

[0201] An input-output model is designed, a comprehensive evaluation function is defined, and an evolutionary algorithm is used to search for optimal model parameters to maximize the classification accuracy and minimize the loss function value;

[0202] The real-time collected state vector is input into the trained model to obtain the corresponding defect category label, the transfer frequency between adjacent time defect labels is analyzed to form a defect transfer probability matrix, the probability of all possible paths is calculated according to the transfer probability, and the most likely defect evolution path is found;

[0203] After the above steps, the defect evolution path analysis result of the device in a period of time is obtained, and Table 2 below is a simplified defect transfer probability matrix example:

[0204] Table 2 Defect transfer

[0205]

[0206] Each cell represents the probability of transferring from one defect type to another defect type;

[0207] By mapping multi-dimensional state monitoring data to low-dimensional graph structure, the global evolution characteristics of the device running state are described, the spectral clustering method is used to extract typical defect modes, and the nonlinear correlation modeling is used to reveal the potential evolution relationship between defects, the evolutionary algorithm is introduced to optimize the input-output model parameters, the classification accuracy is improved, and the model generalization ability is enhanced, and finally the most possible evolution path is mined based on the defect transfer probability matrix, providing an interpretable quantitative analysis tool for the device fault evolution trend.

[0208] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and improvement concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A method for testing the airtightness of a diaphragm accumulator, characterized in that, Includes the following steps: Step 1: Collect status monitoring data during the operation of the diaphragm accumulator. The status monitoring data includes gas pressure data, temperature data, vibration data, acoustic emission signals, and strain data. Perform preprocessing operations on the collected status monitoring data. Step 2: Real-time monitoring and closed-loop control of gas flow rate, gas pressure, and gas path status during the inflation process, and automatic switching of gas path according to detection requirements; Step 3: By integrating gas pressure, temperature, vibration and acoustic emission signals, and using acoustic wave propagation inversion, dynamic crack identification and multi-scale analysis methods, the internal defects of the accumulator are preliminarily located and their evolution trend is analyzed. The process of fusing gas pressure, temperature, vibration, and acoustic emission signals and performing acoustic wave propagation inversion analysis in step three includes: Preprocessed gas pressure data, temperature data, vibration data, and acoustic emission signals are acquired, and their features are extracted. The extracted features are then normalized and weighted and fused into a comprehensive feature vector F. The material's internal physical properties or structural information can be deduced from the acoustic wave signals measured at the boundary using the acoustic wave propagation inversion method. In an isotropic linear elastic medium, sound waves satisfy a three-dimensional second-order wave equation; When the location of the sound source and the parameters of the medium are known, the wave equation is solved using the finite difference method to obtain the theoretical wave field u(x,t); The goal of the inversion is to minimize the observed wavefield u gc (x i ,t) and the calculation of wave field u(x) i The difference between ,t;m) is used to optimize the medium parameter m(x); Introducing the adjoint field λ(x,t), the gradient of the objective function with respect to the medium parameters is calculated through the adjoint state equation; Using this gradient, the medium parameter m is iteratively updated using the L-BFGS algorithm until convergence; The process of using dynamic crack identification and multi-scale analysis methods in step three includes: Dynamic crack propagation causes a decrease in material stiffness, manifested as a decrease in structural modal frequency and an increase in damping ratio. The location of the crack can be identified by combining local energy density changes. The response signal is processed using the Hilbert-Huang transform to extract the nth instantaneous frequency f. n (t) and damping ratio ζ n (t); By introducing a local energy density function, the energy concentration increases during crack propagation, and the crack location is determined by combining the change in wave propagation path. Structural features at different scales are processed by multi-scale analysis methods, and cross-scale feature extraction is achieved by wavelet decomposition. Define micro variables ε is a small parameter that is much smaller than 1. The field quantity u has a multi-scale expansion form. Substitute the expansion into the control equation and separate the variables according to ε, and solve it step by step. Using wavelet basis function ψ j,k (x) Perform multi-scale decomposition on the signal; The process of preliminary location and evolution trend analysis of internal defects in the accumulator in step three includes: Defects in energy storage devices can be analyzed for their spatiotemporal evolution characteristics through multi-source data fusion. By inverting and reconstructing the internal reflection coefficient distribution r(x) of the accumulator, the distribution map of r(x) is obtained by inverse calculation of the measured echo s(t) to determine the location of the defect; The defect parameter D(t) is defined as the damage factor, and its evolution is defined as conforming to an exponential growth model. When the defect parameter D(t) reaches or exceeds the set threshold D sd When a defect is deemed to be in a dangerous state, repair or replacement measures must be taken. Step 4: Construct a high-dimensional map based on condition monitoring data to characterize the nonlinear correlation between defect patterns, and optimize model parameters by combining evolutionary algorithms to achieve defect identification and evolution path analysis; The process of constructing a high-dimensional map based on state monitoring data in step four includes: Obtain the preprocessed state monitoring data, construct a time-series state vector from it, and build a state data matrix for the total observation time length T. Using a fixed window length and sliding step size, state information of consecutive time periods is extracted sequentially to form a series of trajectory segments. The total number of trajectories is calculated based on the set window length and sliding step size, and a trajectory set is generated. The high-dimensional trajectory is reduced using the diffusion mapping algorithm. The similarity between trajectories is measured by constructing a Gaussian kernel function and then normalized to obtain the transition probability matrix. The first few non-trivial eigenvectors of the matrix and their corresponding eigenvalues ​​are extracted to construct the node representation in the low-dimensional embedding space, and finally a set of low-dimensional nodes are obtained, which constitute the graph structure of the device state. For each pair of low-dimensional nodes, if one node is the nearest neighbor of the other node, then an edge is established between the two nodes, eventually constructing a weighted graph. The process of characterizing the nonlinear correlation between defect patterns in step four includes: The low-dimensional node set is classified using spectral clustering. Nodes with similar characteristics are grouped into the same category. Each category represents a defect pattern. The center point of each category is calculated as the representative state of the defect of that category. The squared Euclidean distance between the centers of various defects is calculated as a measure of their differences, and finally a nonlinear correlation matrix reflecting the relationship between the various defect modes is obtained. The threshold is set as the top 20 percentile of the similarity between all defect pairs. If the squared Euclidean distance between two defect categories is less than the threshold, a directed edge is established between the corresponding nodes, thereby constructing a directed association graph between defects, which represents the possible evolution path of the defects. The process of optimizing model parameters using evolutionary algorithms in step four includes: Design an input-output model with the device state vector as input and the defect category number as output. The model parameters are variables to be optimized. Define a comprehensive evaluation function, which includes the following three items: classification accuracy, loss function value and parameter norm. Each item is combined into an overall fitness index according to a set weight. The evolutionary search steps are as follows: T1: Initialize a batch of individuals with random parameters and calculate the fitness value of each individual; T2: Select the individuals with the highest fitness as parents, perform crossover to generate new individuals, and add a small mutation perturbation: T3: Repeat the above process until the preset number of iterations or the convergence condition is reached; The final output is the optimal model parameters; The process of defect identification and evolution path analysis in step four includes: Each state vector collected during equipment operation is input into the trained model, and the corresponding defect category label is output to generate a complete defect label sequence. The frequency of defect transfer between adjacent defect labels is statistically analyzed, and the frequency ratio of each defect to another defect is calculated to form a defect transfer probability matrix. Define a path sequence consisting of multiple defect categories, where the probability of a path is the product of the transition probabilities of all its adjacent defects. Among all possible paths, identify the one with the highest probability as the most likely defect evolution path.

2. The method for testing the airtightness of a diaphragm accumulator according to claim 1, characterized in that, The process of real-time monitoring and closed-loop control of gas flow rate, gas pressure, and gas path status during the inflation process in step two includes: Gas flow rate is collected using a thermal flow meter, and the collected gas flow rate data is transmitted to the control unit. A piezoelectric pressure sensor is used to collect gas pressure, and the collected data is transmitted to the control unit. The gas path conduction status is collected using a switch status sensor, the valve opening is collected using a valve position sensor, and the gas flow status is collected using a differential pressure flow detection module. The collected data is then connected to the control unit. The control system consists of sensors, a control unit, and an actuator. The control unit receives sensor signals from gas flow, gas pressure, and gas path status, compares the received actual measured values ​​with preset target values, calculates the deviation, and generates a control output signal based on the deviation to drive the actuator to adjust the system state. The control unit uses a proportional-integral-derivative control algorithm to perform closed-loop regulation of gas flow and gas pressure, and changes the gas flow rate, adjusts the gas pressure, or switches the gas path according to preset program instructions.

3. The method for testing the airtightness of a diaphragm accumulator according to claim 2, characterized in that, The process of automatically switching the gas path according to the detection requirements in step two includes: The control system determines whether to perform a gas path switch based on the task identifier and operating parameters. After identifying the switching conditions, the control system generates a switching command. The control system selects the target air path and issues a switching command, and the air path switching valve responds to the command and changes the airflow direction. The various components of the control system work together, and the air path switching mechanism guides the airflow into different paths according to the task identifier; The control system enables path switching and parameter adjustment of the gas path system.

Images