A locomotive vehicle brake valve detection method based on CNN

Abstract

The present application belongs to the technical field of railway locomotive and vehicle maintenance, and particularly relates to a locomotive and vehicle brake valve detection method based on CNN, comprising the following steps: taking the brake valve as a carrier and various factor data as neurons to construct a neural network system of the brake valve body, performing correlation analysis on the data of each neuron after data import to form a predictive diagnosis of the brake valve installation and use effect, judging whether to install and use or re-maintain according to the diagnosis result, and finally achieving the effects of reducing locomotive and vehicle brake system failures and improving railway vehicle transportation safety. Through the brake valve detection using the method, the influence of test environment, use environment and maintenance factors on brake valve performance can be fully considered, the brake valve can be put into use in the best state after maintenance, and the hidden danger of railway locomotive and vehicle driving safety can be reduced.

Description

Technical Field

[0001] This invention belongs to the field of railway locomotive and rolling stock maintenance technology, specifically relating to a CNN-based method for detecting brake valves in locomotives and rolling stock. Background Technology

[0002] Railway locomotives and rolling stock are the means of transporting goods by rail, and the brake valve is one of the core components of railway freight cars. It is a key actuator for deceleration and stopping, and an essential device for ensuring train operation safety. For modern railways, the importance of braking is not merely a safety issue; it has become a significant factor limiting further improvements in train speed and traction quality. Traditional methods for testing railway locomotive and rolling stock brake valves involve disassembling, repairing, assembling, and testing the valves before installation. Because the post-repair testing environment differs significantly from the actual operating environment of the train, brake failures can occur under various operating conditions, such as emergency ventilation, exhaust leakage, and spontaneous release, posing safety hazards to railway transportation. Therefore, relying solely on the results of a single test of the brake valve under the brake chamber environment as the basis for installation, especially considering the impact of edge-range test data on post-installation performance, lacks basic fault prediction methods and poses a risk to train operation safety. Therefore, to reduce the probability of malfunctions during locomotive and rolling stock operation and ensure driving safety, it is necessary to comprehensively analyze the brake valve data after testing. This analysis should identify key factors that may affect the performance of the brake valves and analyze the effects of these factors on performance, leading to guidelines for operation and maintenance, and ultimately improving the safety performance of the brake valves. Because of the uncertainty of the operating environment of brake valves and the correlation between influencing factors, and because data on these correlated factors can be extracted through railway vehicle operation-related information systems or manually maintained. Summary of the Invention

[0003] To address the technical problem that the lack of basic fault prediction methods for the impact of edge test data within the aforementioned qualified range on the post-installation application effect poses a potential threat to vehicle operation safety, this invention provides a CNN-based method for detecting brake valves in locomotives and rolling stock. This method uses the brake valve as a carrier and various factor data as neurons to construct a neural network system for the brake valve body. Through correlation analysis of the data imported from each neuron, a predictive diagnosis of the brake valve's post-installation application effect is formed. Based on the diagnostic results, it is determined whether the valve should be installed and used or re-inspected, ultimately achieving the effect of reducing locomotive and rolling stock braking system failures and improving railway vehicle transportation safety.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A CNN-based method for detecting brake valves in rolling stock includes the following steps:

[0006] S1. Classify and organize the test data of the brake valve, and construct it as the input layer of a two-dimensional convolutional neural network;

[0007] S2. Construct convolution kernels for experimental numerical factors, station environmental factors, key component measurement data factors, application interval environmental factors, and personnel skill level factors. Define correlation functions based on the correlation between experimental data and influencing factors, and dynamically determine the convolution kernel parameters according to specific values.

[0008] S3. After the numerical values ​​of the two-dimensional matrix of the convolution kernel are determined, convolution feature extraction is performed on the input layer of the two-dimensional convolutional neural network.

[0009] S4. The conclusions of the analysis of the causes of brake valve failure, brake valve application and failure outside the station are classified and organized by the transfer learning model for the parameter adjustment of the excitation function for weight calculation. The excitation function for weight calculation of influencing factors is designed and the parameters generated by the transfer model are substituted into the excitation function to calculate the weight of the influencing factors.

[0010] S5. Construct a weighted feature pooling layer to extract features that affect the effect and form corresponding numerical analysis weights;

[0011] S6. To assess the overall performance of the brake valve, determine the quality of brake valve maintenance, and make diagnostic predictions for its application, thereby completing a comprehensive test of the brake valve.

[0012] The method for constructing the input layer of the two-dimensional convolutional neural network in S1 is as follows: In the two-dimensional matrix, the rows are defined according to the classification of brake valve detection data, namely leakage Boolean type, leakage amount numerical type, pressure numerical type, and time numerical type; the columns are defined as test items, namely brake valve inflation position, emergency braking position, braking and release sensitivity, partial reduction valve, stability, emergency pressurization, full release position, and equalization sensitivity; a 4×8 two-dimensional matrix is ​​constructed, and the data in the two-dimensional matrix are the values ​​converted from the original test records. Items not involved are filled with 0, and Boolean types are filled with 0 or 1 from the test records.

[0013] The method for constructing the experimental numerical factor convolution kernel in S2 is as follows: Based on the maintenance standards, set the leakage deviation coefficient m and the time deviation coefficient n. The convolution kernel is set to extract a 2×2 two-dimensional matrix of numerical interval features, storing the maximum and minimum values ​​of the leakage standard and the maximum and minimum values ​​of the time standard, respectively. By multiplying the deviation coefficients corresponding to the input two-dimensional matrix data type, the values ​​exceeding the range of the dynamic convolution kernel are counted. The time value is counted using the median value of the range, and deviation coefficients less than the median value are used with their negative absolute values ​​in the calculation. This is extracted as a 2×2 two-dimensional matrix with counters as features, and recorded as the leakage over-limit count and the time over-limit count, respectively.

[0014] The method for constructing the convolution kernel of station environmental factors in S2 is as follows: based on the influence of temperature and air source pressure of the brake valve test station on the test data results, a dynamic convolution kernel is set with the total air source pressure of the station and the standard test air source pressure value. Since temperature and pressure mainly affect the time value, the extracted feature is a 2×2 two-dimensional matrix with the trend feature of time performance. The extracted feature is a 2×2 two-dimensional matrix with counter as the feature, and the deviation trend count is recorded respectively.

[0015] The method for constructing the convolution kernel of key component measurement data factors in S2 is as follows: based on the deviation of the measurement data of the brake valve components from the standard value, a 2×2 two-dimensional matrix of deviation coefficient influence feature extraction is set. The two-dimensional matrix takes the values ​​of brake valve spring test deviation coefficient, rubber template deviation coefficient, slide valve roughness deviation coefficient, and slide valve stop valve wear deviation coefficient. A 2×2 two-dimensional matrix of feature extraction based on the influence of component test data deviation on brake valve data is extracted, and the influence degree values ​​are recorded respectively.

[0016] The method for constructing the convolution kernel for environmental factors in the application interval in S2 is as follows: based on the temperature and humidity difference of the application interval and the application time point, a 2×3 two-dimensional matrix is ​​extracted, which corresponds to the highest and lowest values ​​of the temperature and humidity difference and the estimated temperature and humidity values ​​at the running time point. The matrix is ​​extracted into a 2×3 two-dimensional matrix with counters as features, and the deviation trend count is recorded respectively.

[0017] The method for constructing the convolution kernel of personnel skill factors in S2 is as follows: based on personnel skill level, rework rate, key maintenance failure ratio, and years of service, a 2×2 two-dimensional matrix for feature extraction is set up, and a 2×2 two-dimensional matrix for feature extraction of the influence of personnel skill factors on brake valve data is extracted, and the influence degree values ​​are recorded respectively.

[0018] The method for calculating the weights of influencing factors in S4, and then substituting the parameters generated by the migration model into the activation function to calculate the weights of the influencing factors, is as follows:

[0019] S4.1 Classify vehicle control valve malfunctions into three main categories: single-vehicle malfunctions, centralized air test malfunctions, and operational malfunctions. Then, classify the main causes of the malfunctions into secondary categories. Next, classify the causes of the malfunctions in the secondary categories into tertiary categories of component factors, including templates, springs, spool valves, stop valves, mounting seats, and orifices. Extract the probability of malfunction m based on the proportion of malfunctions generated by the components in the tertiary categories. For brake valves with similar working principles, dynamically increase the probability value based on the probability n of the same type of malfunction. Take the median value (mn) / 2 of the increased part as the malfunction pre-occurrence judgment.

[0020] S4.2 Classify the operational failures caused by temperature and humidity, the number of events n, the total number of failures m, and calculate the probability n / m of failures caused by temperature and humidity. Based on the open parabolic relationship between temperature and humidity during the operation of the brake valve, calculate the predicted probability factor by multiplying the calculated values ​​based on the temperature and humidity difference at the application site with the probability.

[0021] S4.3 Classify operational failures caused by personnel skill issues, with the number of events n and the total number of failures m. Calculate the probability n / m of failures caused by personnel skill levels. Based on the negative correlation between brake valve failures and personnel experience years, multiply the calculated values ​​of repair rate and years of service with the probability to calculate the prediction probability factor.

[0022] The method for constructing the weighted feature pooling layer in S5 is as follows: The numerical summation of the convolutional results of the experimental data is S = ∑i=1m∑j=1na. ij The deviation between the result and the expected numerical parameter n is calculated as f(n) = (a ij -(an 2 +bn-c)) 2 Taking the derivative of f(n) yields a and b, which are then used to apply the numerical weight activation function f(x) = ax. 2 + bx + c, obtain the weight value A; numerical summation of site environmental factors through convolution operation S = ∑i=1m∑j=1na ij The deviation between the result and the expected numerical parameter n is calculated as f(n) = (a ij -(an 2 +bn-c)) 2 Taking the derivative of f(n) yields a and b, which are then used to apply the numerical weight activation function f(x) = ax. 2 + bx + c, obtain the weight value B; perform interval convolution operation and sum the numerical values ​​S = ∑i=1m∑j=1na ij The deviation between the result and the expected numerical parameter n is calculated as f(n) = (a ij -(an 2 +bn-c)) 2 Taking the derivative of f(n) yields a and b, which are then used to apply the numerical weight activation function f(x) = ax. 2 + bx+c, to obtain the weight value C; after the key component factor calculation, the activation function f(K)=K*∑i=1m∑j=1na is used. ij The weight value D is obtained; the personnel skill factor is obtained through the incentive function f(K)=K*∑i=1m∑j=1na ij The weight value E is then obtained.

[0023] The method for judging the overall performance of the brake valve and predicting its maintenance quality in S6 is as follows: Data is normalized based on weighted influencing factors to form a comprehensive weight value M = A + B*a + C*b + D*c + E. Here, coefficient a represents the influence of external air source pressure and temperature difference on the failure of the centralized test air; coefficient b represents the influence of temperature and humidity in the operating area on operational failures; and coefficient c represents the influence of personnel factors on failures occurring in individual vehicles, external areas, and operating areas. Correspondingly, the fault diagnosis is categorized according to the transfer learning model, ultimately leading to the judgment of the overall performance of the brake valve and the prediction of its maintenance quality, thus completing the comprehensive testing of the brake valve.

[0024] Compared with the prior art, the beneficial effects of this invention are:

[0025] This invention addresses the safety hazards of railway operations caused by relying solely on one-time test results for the maintenance of railway locomotives and rolling stock, neglecting the comprehensive influence of factors such as operating environment, component replacement and break-in, and personnel maintenance skills on brake valve performance. By constructing a convolutional neural network for detecting railway locomotive and rolling stock brake valves, this invention uses test data factors, station environment factors, vehicle operating section factors, component maintenance and replacement factors, and maintenance personnel skill factors as neurons in the neural network. It summarizes and categorizes faults and causes occurring during the operation of railway locomotives and rolling stock to form a fault diagnosis library. Fault phenomena and causes of brake valves with similar functions are categorized and supplemented into the brake valve fault diagnosis library through a transfer learning model. The invention identifies the influence of the interrelationships of neuronal factors on the actual performance of the brake valve during operation and designs activation functions accordingly. When one-time test data from a test bench is input into the convolutional neural network and formatted as a two-dimensional array, the convolutional kernels of each neuron complete parameter configuration based on the actual values ​​of different factors. Convolutional operations are then performed, and the influencing factors are derived through activation functions. These influencing factors are fully linked and normalized to form a comprehensive diagnostic conclusion for the brake valve. By using this method to test brake valves, the impact of the test environment, operating environment, and maintenance factors on brake valve performance can be fully considered, ensuring that the brake valves are put into use in the best condition after maintenance, thereby reducing potential safety hazards for railway locomotives and rolling stock. Attached Figure Description

[0026] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0027] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0028] Figure 1 This is a schematic diagram of the detection method of the present invention.

[0029] Among them: 1 is the test data input, 2 is the test numerical factor convolution kernel, 3 is the station environmental factor convolution kernel, 4 is the operational interval environmental factor convolution kernel, 5 is the key component factor convolution kernel, 6 is the personnel skill factor convolution kernel, 7 is the numerical factor feature extraction, 8 is the station factor feature extraction, 9 is the interval environmental factor feature extraction, 10 is the component factor feature extraction, 11 is the skill factor feature extraction, 12 is the comprehensive weight, and 13 is the brake valve pre-diagnosis result. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. These descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the claims of the present invention. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0032] In this embodiment, as Figure 1As shown, the neural network consists of an input layer, an influencing factor convolutional layer, and its internal components including influencing factor convolutional kernels, correlation functions, transfer learning models, weight calculation activation functions, feature extraction pooling, and comprehensive weight functions. Experimental data input 1 is the experimental data result transformed into a numerical matrix, serving as the raw data input layer for the neural network. This matrix is ​​then processed with corresponding convolutional kernels to extract corresponding features. Experimental numerical factor convolutional kernel 2, along with station environmental factor convolutional kernel 3, interval environmental factor convolutional kernel 4, key component factor convolutional kernel 5, and personnel skill factor convolutional kernel 6, act as network neurons, influencing numerical factor feature 7. Station environmental factor convolutional kernel 3 acts as a network neuron, influencing numerical factor feature 7, station factor feature extraction 8, and component factor feature extraction 10. Interval environmental factor convolutional kernel 4 acts as a network neuron, influencing numerical factor feature 7, interval environmental factor feature extraction 9, and component factor feature extraction 10. Key component factor convolutional kernel 5 acts as a network neuron, influencing numerical factor feature 7 and component factor feature extraction 10. Personnel skill factor convolution kernel 6 acts as a network neuron, influencing numerical factor feature 7 and skill factor feature extraction 11. After forming a comprehensive weight value from numerical factor feature 7, station factor feature extraction 8, interval environmental factor feature extraction 9, component factor feature extraction 10, and skill factor feature extraction 11, the comprehensive weight value of each extracted feature is calculated. The calculated pre-diagnosis result 13 for the brake valve is then obtained, and based on the diagnosis result, corresponding diagnostic conclusions, maintenance guidance, and installation and operation precautions are given.

[0033] A method for detecting brake valves in locomotives and rolling stock based on CNN, characterized by comprising the following steps:

[0034] Step 1: Classify and organize the brake valve test data to construct a two-dimensional convolutional neural network input layer. In this two-dimensional matrix, rows are defined according to the brake valve test data classification: leakage Boolean type, leakage quantity numerical type, pressure numerical type, and time numerical type. Columns are defined as test items: brake valve inflation position, emergency braking position, braking and release sensitivity, partial depressurization valve, stability, emergency pressurization, full release position, and equalization sensitivity. A 4×8 two-dimensional matrix is ​​constructed, where the data are converted values ​​from the original test records. Items not involved are filled with 0, and Boolean types are filled with 0s or 1s from the test records.

[0035] Step 2: Construct convolution kernels for experimental numerical factors, station environmental factors, key component measurement data factors, application interval environmental factors, and personnel skill level factors. Define correlation functions based on the correlation between experimental data and influencing factors to dynamically determine the convolution kernel parameters according to specific numerical values.

[0036] Experimental numerical factor convolution kernel: Based on the maintenance standards, the leakage deviation coefficient *m* and the time deviation coefficient *n* are set. The convolution kernel is set to extract a 2×2 two-dimensional matrix of numerical range features, which stores the standard maximum and minimum values ​​of leakage and time, respectively. By multiplying the deviation coefficients corresponding to the input two-dimensional matrix data types, the values ​​exceeding the range of the dynamic convolution kernel are counted. The time value is counted based on the median value of the range. For deviation coefficients less than the median value, the negative absolute value is used in the calculation. The result is extracted as a 2×2 two-dimensional matrix with counters as features, which is used to record the leakage exceedance count and the time exceedance count, respectively.

[0037] Site environmental factor convolution kernel: Based on the influence of temperature and air source pressure at the brake valve test site on the test data results, a dynamic convolution kernel is set with the total air source pressure at the site and the standard test air source pressure value. Since temperature and pressure mainly affect the time value, the extracted feature is a 2×2 two-dimensional matrix with the trend feature of time performance. It is extracted into a 2×2 two-dimensional matrix with counter as the feature, and the deviation trend count is recorded respectively.

[0038] Using the convolution kernel of environmental factors in the application interval: Based on the temperature and humidity difference in the application interval and the application time point, a 2×3 two-dimensional matrix is ​​extracted, which corresponds to the highest and lowest values ​​of the temperature and humidity difference and the estimated temperature and humidity values ​​at the running time point. The extracted matrix is ​​a 2×3 two-dimensional matrix with counters as features, and the deviation trend count is recorded respectively.

[0039] Key component factor convolution kernel: Based on the deviation of the measured data of the brake valve components from the standard values, a 2×2 two-dimensional matrix of deviation coefficient influence feature extraction is set. The two-dimensional matrix takes the values ​​of brake valve spring test deviation coefficient, rubber template deviation coefficient, slide valve roughness deviation coefficient, and slide valve stop valve wear deviation coefficient. The feature extraction 2×2 two-dimensional matrix of the influence of component test data deviation on brake valve data is extracted, and the influence degree value is recorded respectively.

[0040] Personnel skill factor convolution kernel: Based on personnel skill level, rework rate, key maintenance failure ratio, and years of service, a 2×2 two-dimensional matrix for feature extraction is set to extract the influence of personnel skill factors on brake valve data, and the influence degree values ​​are recorded respectively.

[0041] Step 3: After determining the numerical values ​​of the two-dimensional matrix of the convolution kernel, perform convolution feature extraction on the input layer of the two-dimensional convolutional neural network.

[0042] Step 4: Using the transfer learning model, classify and organize the conclusions of the analysis of the causes of brake valve failures, other similar brake valve applications, and external station failures for parameter adjustment of the excitation function for weight calculation. Design the excitation function for calculating the weights of influencing factors, and after substituting the parameters generated by the transfer model into the excitation function, calculate the weights of the influencing factors.

[0043] 1) Vehicle control valve malfunctions are categorized into three main types: single-vehicle malfunctions, centralized test malfunctions, and operational malfunctions. The primary causes of these malfunctions are then classified into secondary categories, such as external leakage, abnormal release, poor braking, and frequent emergency starts. These secondary categories are further subdivided into tertiary categories based on component factors, such as templates, springs, spool valves, stop valves, mounting seats, and orifices. The probability of malfunction (m) is extracted based on the proportion of malfunctions occurring in each of the tertiary categories of components. For brake valves with similar operating principles, the probability of occurrence (n) is dynamically increased, and the median (mn) / 2 of the increased probability is used as the pre-malfunction prediction.

[0044] 2) Classify the operational failures caused by temperature and humidity, with the number of events n and the total number of failures m. Calculate the probability n / m of failures caused by temperature and humidity. Based on the open parabolic relationship between temperature and humidity during the operation of the brake valve, calculate the predicted probability factor by multiplying the calculated values ​​based on the temperature and humidity differences at the application site with the probability.

[0045] 3) Classify operational failures caused by personnel skill issues, with the number of events n and the total number of failures m. Calculate the probability n / m of failures caused by personnel skill levels. Based on the negative correlation between brake valve failures and personnel experience years, calculate the return rate and years of service, multiply the calculated values ​​by the probability, and calculate the prediction probability factor.

[0046] Step 5: Construct a weighted feature pooling layer to extract features affecting the effect and form corresponding numerical analysis weights. The numerical summation of the convolution operation results of the experimental data is S = ∑i=1m∑j=1na. ij The deviation between the result and the expected numerical parameter n is calculated as f(n) = (a ij -(an 2 +bn-c)) 2 Taking the derivative of f(n) yields a and b, which are then used to apply the numerical weight activation function f(x) = ax. 2 +bx+c, yielding the weight value A. The numerical summation of the convolution operation for site environmental factors is S=∑i=1m∑j=1na. ij The deviation between the result and the expected numerical parameter n is calculated as f(n) = (a ij -(an 2 +bn-c)) 2 Taking the derivative of f(n) yields a and b, which are then used to apply the numerical weight activation function f(x) = ax. 2 + bx + c, yielding the weight value B. Performing interval convolution operations and numerical summation S = ∑i=1m∑j=1na ij The deviation between the result and the expected numerical parameter n is calculated as f(n) = (a ij -(an 2+bn-c)) 2 Taking the derivative of f(n) yields a and b, which are then used to apply the numerical weight activation function f(x) = ax. 2 + bx + c, yielding the weight value C. After calculating the key component factors, the activation function f(K) = K * ∑i=1m∑j=1na is applied. ij The weight value D is obtained. Personnel skill factors are determined through the incentive function f(K) = K * ∑i=1m∑j=1na. ij The weight value E is then obtained.

[0047] Step Six: Finally, the data is normalized according to the weighted influencing factors to form a comprehensive weight value M = A + B*a + C*b + D*c + E. Here, coefficient a mainly represents the influence of the external air source pressure and temperature difference on the failure of the centralized test air, coefficient b represents the influence of the temperature and humidity in the operating area on the failure of operation, and c represents the influence of personnel factors on the failure of individual vehicles, external areas, and operating areas. Correspondingly, the fault diagnosis classification in the transfer learning model is used to make a judgment on the overall performance of the brake valve, the quality of brake valve maintenance, and the diagnosis of operation prediction, thus completing the comprehensive testing of the brake valve.

[0048] The above description only illustrates the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and all such changes should be included within the protection scope of the present invention.

Claims

1. A CNN-based rolling stock brake valve detection method, characterized in that, Comprise the following steps: S1, the test data of brake valve classification arrangement, build for two-dimensional convolution neural network input layer; S2, build test numerical factor convolution kernel, station environment factor convolution kernel, key component measurement data factor convolution kernel, use interval environment factor convolution kernel, personnel skill level factor convolution kernel, through the correlation of test data and influence factor definition correlation function, thus according to the specific numerical dynamic determination convolution kernel parameter; S3, the two-dimensional matrix value of convolution kernel is determined, and the two-dimensional convolution neural network input layer is extracted; S4, the brake valve itself fault, brake valve use and station external fault reason analysis conclusion are classified and arranged by using the transfer learning model for weight calculation excitation function parameter adjustment, and the transfer model generation parameter is substituted into the excitation function, and the influence factor weight is calculated; S5, build weight feature pooling layer, used for extracting influence effect characteristic and forming corresponding numerical analysis weight; S6, the judgment and use prediction diagnosis of brake valve maintenance quality is formed according to the overall performance of brake valve, and the comprehensive detection of brake valve is completed.

2. The CNN-based rolling stock brake valve detection method according to claim 1, characterized in that, The method for building two-dimensional convolution neural network input layer in S1 is that in the two-dimensional matrix, the row is defined according to the brake valve detection data classification, which is respectively leakage boolean type, leakage quantity numerical type, pressure numerical type and time numerical type; the column is defined as test project, which is respectively brake valve inflation position, emergency brake position, brake and relief sensitivity, local reduction valve, stability, emergency pressure increase, full relief position and balance sensitivity; a 4*8 two-dimensional matrix is constructed, the data in the two-dimensional matrix is the numerical value converted from the original test record, the items not involved are filled with 0, and the boolean type is filled with 0 or 1 in the test record.

3. The CNN-based rolling stock brake valve detection method according to claim 1, characterized in that, The method for building test numerical factor convolution kernel in S2 is that according to the maintenance standard, leakage quantity deviation coefficient m and time quantity deviation coefficient n are set, a two-dimensional matrix of 2*2 is set for numerical interval feature extraction, which respectively stores the maximum and minimum values of leakage quantity standard and time standard; the product operation is carried out on the input two-dimensional matrix data type corresponding to the deviation coefficient, and the count is carried out on the range interval outside the dynamic convolution kernel, wherein the time quantity is counted by the range median, and the absolute value of the negative deviation coefficient is taken for operation when it is less than the median, and a two-dimensional matrix of 2*2 is extracted as the feature of the counter, which is recorded as leakage overrun count and time quantity overrun count respectively.

4. The CNN-based rolling stock brake valve detection method according to claim 1, characterized in that, The method for building station environment factor convolution kernel in S2 is that according to the influence degree of brake valve test site temperature and air source pressure on test data result, dynamic convolution kernel is set according to the total air source and standard test air source pressure value of the site, and since temperature and pressure mainly affect the time value extraction feature, a two-dimensional matrix of 2*2 is extracted as the feature of the counter, which is recorded as deviation trend count respectively.

5. The CNN-based rolling stock brake valve detection method according to claim 1, characterized in that, The method for constructing the key component measurement data factor convolution kernel in S2 is: according to the deviation degree of the brake valve component measurement data from the standard value, setting the deviation coefficient influence feature extraction 2*2 two-dimensional matrix, the two-dimensional matrix values are the brake valve spring test deviation coefficient, the rubber template deviation coefficient, the slide valve roughness deviation coefficient, and the slide valve stop valve wear deviation coefficient, extracting the feature extraction 2*2 two-dimensional matrix influenced by the component test data deviation on the brake valve data, and recording the influence degree values respectively.

6. The CNN-based rolling stock brake valve detection method according to claim 1, characterized in that, The method for constructing the interval environmental factor convolution kernel in S2 is: according to the temperature and humidity difference of the use interval and the use time point, setting the feature extraction 2*3 two-dimensional matrix, corresponding to the highest and lowest values of the temperature and humidity difference and the calculated temperature and humidity values at the running time point, extracting the 2*3 two-dimensional matrix with the counter as the feature, and recording the deviation trend count respectively.

7. The CNN-based rolling stock brake valve detection method according to claim 1, characterized in that, The method for constructing the personnel skill factor convolution kernel in S2 is: according to the personnel skill level, the repair rate, the key maintenance fault ratio, and the working years, setting the feature extraction 2*2 two-dimensional matrix, extracting the feature extraction 2*2 two-dimensional matrix of the influence of the personnel skill factor on the brake valve data, and recording the influence degree values respectively.

8. The CNN-based locomotive vehicle brake valve detection method according to claim 1, characterized in that, The method for designing the influence factor weight calculation incentive function and substituting the migration model generation parameters into the incentive function to calculate the influence factor weight in S4 is: S4.1, classify the vehicle control valve faults, the main categories are single vehicle fault, centralized test wind fault, and driving operation fault; then classify the main causes of the faults into two levels; then classify the causes of the two-level abnormal occurrences to form the three-level classification of the component factors, including the template, the spring, the slide valve stop valve, the mounting seat and the aperture; extract the fault occurrence probability m corresponding to the fault proportion of the three-level classification components, and dynamically improve the probability value of the similar brake valve based on the similar working principle and the occurrence probability n, and take the median value (m-n) / 2 as the fault prediction occurrence judgment; S4.2, classify the operation faults caused by temperature and humidity, the event quantity n, the total fault quantity m, calculate the probability n / m of the faults caused by temperature and humidity, and according to the open parabolic relationship between the temperature and humidity in the working process of the brake valve, multiply the calculated value of the temperature difference and the humidity difference in the operation site with the probability to calculate the prediction probability factor; S4.3, classify the operation faults caused by personnel skill, the event quantity n, the total fault quantity m, calculate the probability n / m of the faults caused by the personnel skill level, and according to the negative correlation between the brake valve fault and the personnel experience years, multiply the repair rate and the working years with the probability to calculate the prediction probability factor.

9. The CNN-based rolling stock brake valve detection method according to claim 1, characterized in that, The method for constructing the weight feature pooling layer in S5 is: summing the convolution operation result values of the test data S = ∑i=1m∑j=1na ij The result is deviated from the expected value parameter n to calculate f(n)=(a ij -(an 2 +bn-c)) 2 , and the derivative of f(n) is obtained, a, b, and the weight value A are obtained through the numerical weight incentive function f(x)=ax 2 + bx+c, the weight value B is obtained by summing the convolution operation values of the site environment factors S=∑i=1m∑j=1na ij The result is deviated from the expected value parameter n to calculate f(n)=(a ij -(an 2 +bn-c)) 2 , and the derivative of f(n) is obtained, a, b, and the weight value A are obtained through the numerical weight incentive function f(x)=ax 2 + bx+c, the weight value B is obtained by summing the convolution operation values of the site environment factors S=∑i=1m∑j=1na ij The result is deviated from the expected value parameter n to calculate f(n)=(a ij -(an 2 +bn-c)) 2 , and the derivative of f(n) is obtained, a, b, and the weight value A are obtained through the numerical weight incentive function f(x)=ax 2 + bx+c, the weight value B is obtained by summing the convolution operation values of the site environment factors S=∑i=1m∑j=1na ij , the weight value D is obtained by the incentive function f(K)=K*∑i=1m∑j=1na The personnel skill factor passes through the incentive function f(K) = K *∑i=1m∑j=1na ij , to obtain the weight value E.

10. The CNN-based rolling stock brake valve detection method according to claim 1, characterized in that, The method for judging the brake valve maintenance quality of the overall performance of the brake valve in the S6 and the application prediction diagnosis is: data normalization processing is performed according to the weight influence factor to form a comprehensive weight M=A+B*a+C*b+D*c+E, wherein the a coefficient is the influence of the off-site wind source pressure and the temperature difference on the failure of the concentrated part of the test wind, the b coefficient is the influence of the temperature and humidity of the operation interval on the operation failure, and the c is the influence of the personnel factor on the failure of the single vehicle, the off-site and the operation interval. According to the fault diagnosis classification in the transfer learning model, the judgment of the brake valve maintenance quality of the overall performance of the brake valve and the application prediction diagnosis are finally formed, and the comprehensive detection of the brake valve is completed.

Images