A numerical-real fusion test evaluation method based on result fusion

Abstract

The application discloses a kind of based on result fusion's digital-real fusion test evaluation method, comprising: through digital-real test result data fusion module, digital test result and installation test result are integrated;Step two, through digital-real fusion test requirement dynamic adaptation module, the multi-objective test requirement model of equipment is constructed;Step three, through digital-real fusion test result evaluation module, to realize the depth evaluation of digital-real fusion test result, the performance of bayesian neural network model is optimized using back propagation mechanism and iterative training of model, ensure that bayesian neural network model can dynamically adapt the multi-objective test requirement model of construction.This application fuses the data of both by removing redundancy and filling in the gaps, and then adjusts the fusion degree of the multi-objective test requirement model and the index factor model for the digital-real test result by establishing a test requirement dynamic adaptation module, so that the fusion advantages of both are better possessed.

Description

Technical Field

[0001] This invention belongs to the fields of control engineering and computer science, and specifically relates to a data-real fusion test and evaluation method based on result fusion. Background Technology

[0002] When testing complex industrial manufacturing lines, special-purpose vehicles, and other systems or equipment, traditional methods typically involve physical testing. With advancements in technology and technology, digital testing has emerged and developed rapidly. Compared to physical testing, digital testing offers advantages such as comprehensive testing scenarios, rich testing conditions, and abundant test samples; however, its reliability remains low due to technological limitations. Physical testing, while facing limited testing conditions and high iterative testing costs, boasts advantages such as high accuracy and reliability of test results. Each method has its strengths and weaknesses. The challenge in integrating these two sets of limited and inconsistent test data—balancing the advantages and disadvantages of both methods while simultaneously achieving high reliability and comprehensive testing conditions—is an unavoidable challenge for the future. Summary of the Invention

[0003] The technical problem this invention aims to solve is to provide a data-real fusion test evaluation method based on result fusion. This method encompasses a data fusion module for data from both digital and physical test results, a dynamic adaptation module for data-real fusion test requirements, and a data-real fusion test result evaluation module. It effectively leverages the advantages of abundant digital test results and high reliability of physical test results while mitigating their inherent disadvantages, thus achieving a reliable and effective fusion evaluation of both data-real test results. The method first designs a data fusion module that integrates digital and physical test results, addressing the issues of redundancy removal and data supplementation. Second, it designs a dynamic adaptation module for data-real fusion test requirements, dynamically adjusting the target of the test evaluation. Finally, it designs a data-real fusion test result evaluation module that adaptively improves a Bayesian neural network and iteratively trains the model based on backpropagation to form a data-real fusion test result evaluation model, achieving a reliable and effective fusion evaluation of the data-real test results.

[0004] This invention provides a data-real fusion testing and evaluation method based on result fusion, comprising:

[0005] Step 1: The digital test results and the actual test results are integrated into one by the data fusion module, which at the same time makes up for missing test data and eliminates redundant and inefficient data.

[0006] Step 2: Construct a multi-objective test requirement model for the equipment through the data-real fusion test requirement dynamic adaptation module, and then iteratively optimize and verify the model based on the key primary test requirement data under ideal operating conditions and the applicable compliance requirement data under widely used operating conditions.

[0007] Step 3: Perform the following operations through the data-real fusion test result evaluation module:

[0008] a. Construct an adaptively improved Bayesian neural network model to achieve in-depth evaluation of the results of the data-real fusion test. The Bayesian neural network model includes an input layer, a hidden layer, and an output layer, where the input layer... Receive the output from step one, which integrates the digital test results and the actual test results:

[0009] ;

[0010] The hidden layer uses variational inference to optimize network parameters. As weight, For bias, It is an activation function:

[0011] ;

[0012] The output layer is responsible for generating the final evaluation results. As weight, For output bias, It is the output activation function:

[0013] ;

[0014] b. Optimize the performance of the Bayesian neural network model by using backpropagation mechanism and iterative training of the model to ensure that the Bayesian neural network model can dynamically adapt to the multi-objective testing requirement model constructed in step two.

[0015] This method is suitable for evaluating the results of multi-stage, multi-round iterative testing of equipment or systems.

[0016] The advantages of this invention compared to the prior art are as follows:

[0017] (1) When equipment is in a complex system such as having multiple sub-components and requiring multiple rounds of iterative testing, a huge amount of manpower and resources are needed to conduct limited physical testing, while digital testing cannot provide good credibility. Regarding the evaluation methods of these two testing methods and test results, traditional methods have a low level of understanding of digital testing and a poor understanding of the degree of combination and integration of data and physical testing. This invention integrates the data of both by removing redundancy and filling gaps, and then adjusts the degree of integration of data and physical test results by establishing a dynamic adaptation module for test requirements and constructing a multi-objective test requirement model and an indicator factor model, so that it can better possess the advantages of the integration of the two.

[0018] (2) During the testing process, the test results of sub-links / steps differ significantly from the test results of the equipment or system as a whole. Furthermore, the mature conditions for final assembly testing are often not met on a large scale due to various reasons. Traditional methods typically establish complex feature models and combine them with historical data to predict the performance of the equipment or system under test. However, due to the low cost of actual assembly test data, effective results are often difficult to obtain. This method, through separate numerical tests of each sub-link, establishes a complex Bayesian model and dynamically adjusts the demand factors to train the iterative model as needed. This significantly improves the grasp of the overall system test evaluation and enhances the accuracy of the test results assessment. Attached Figure Description

[0019] Figure 1 This is a system structure block diagram of a result fusion-based numerical-real fusion test and evaluation method according to the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings.

[0021] This invention relates to a data-real fusion test evaluation method based on result fusion, including the design of a data fusion module for data of data-real fusion test results, a dynamic adaptation module for data-real fusion test requirements, and a data-real fusion test result evaluation module. It can effectively absorb the advantages of rich sample data of digital test results and high credibility of physical test results, while weakening the inherent disadvantages of both, and achieve a credible, reliable and effective fusion evaluation of data-real test results.

[0022] Figure 1 This is a system structure block diagram of a data-real fusion test and evaluation method based on result fusion according to the present invention. Figure 1 As shown, the system implementing this method includes a data fusion module 1 for data and real-data fusion test results, a dynamic adaptation module 2 for data and real-data fusion test requirements, and a data and real-data fusion test result evaluation module 3. These modules work in coordination to execute the method according to the present invention. The following refers to... Figure 1 Describe the method. The method includes the following steps:

[0023] Step one involves integrating digital test results with actual test results using a data fusion module. This process can compensate for missing test data and eliminate redundant and inefficient data. The specific implementation is as follows:

[0024] a. In view of the fact that it is difficult to conduct destructive, extreme and dangerous tests on equipment in actual practice, resulting in incomplete test data, and that it is convenient to conduct tests in digital simulation but it is difficult to conduct tests that are completely realistic and have high credibility and reliability, resulting in redundant test data with low credibility, a data redundancy removal and missing data filling mechanism is designed based on the results data of digital test and actual equipment test.

[0025] b. After conducting multi-stage (n) and multi-round (m) practical and digital tests on the equipment, the following results are obtained through a data redundancy removal and missing data filling mechanism:

[0026] ,

[0027] ,

[0028] in, These are the results of actual testing. This is the result of the m-th round of the n-th stage in the actual test results. These are the results of a numerical test.

[0029] Redundant data with low credibility is removed or its weight is reduced. Redundant data with low credibility refers to data where the results of that stage already exist in the actual installation test, but whose numerical test results differ significantly from the actual installation test results. For some stage result data missing in the actual installation test, the missing data is supplemented using numerical test result data obtained through multiple weighted averages.

[0030] ,

[0031] ,

[0032] ,

[0033] in, The weighting coefficients determine the degree to which redundant data is reduced. In the experiment, a real-world test is performed. Not incomplete, with Therefore Right now There is redundancy. It is a very small constant. Indicates to The output value after redundancy removal. Further testing will be conducted using a real-world setup. Incomplete, in formula Indicates to The output values ​​are then supplemented and incomplete. Finally, after redundancy removal and supplementation, an output that integrates the digital test results and the actual installation test results is constructed. ,in This matrix represents The i-th row, This matrix represents The j-th line.

[0034] Step two involves constructing a multi-objective test requirement model for the equipment using a data-real fusion test requirement dynamic adaptation module. Then, based on the key primary test requirements under ideal operating conditions and the applicable compliance requirements under widely applicable operating conditions, the main station functional indicators and auxiliary station operational indicators are analyzed. The model is then iteratively optimized and validated. The specific implementation is as follows:

[0035] a. Construct a multi-objective test requirement model for the equipment based on multiple indicators, including primary combat function indicators and secondary combat operation indicators. The primary combat function indicators include the equipment's payload, range, and maximum power. The secondary combat operation indicators include the equipment's rapid component replacement capability, harsh operating condition tolerance level, and fault disturbance level. Finally, a multi-objective test requirement model covering both primary and secondary combat function indicators is constructed. The model's inputs include payload, range, maximum power, rapid replacement capability, harsh operating condition tolerance level, fault disturbance level, and the requirement index for each indicator, defined as: TRI is the total input. ~ There are six indicators. ~ Output: Demand index for each indicator. , ~ These are test indices representing maximum power output, fuel efficiency, transmission system efficiency, maximum fuel storage capacity, equipment frame materials, consistency level of key components, load-bearing capacity of parts, suspension reliability, robustness of the electronic control system, and dustproof / waterproof / corrosion-proof / high-temperature / low-temperature / sealing performance of the structure. A training layer is set up within the model. Due to the correlation between the various indicators, the relationship between each indicator and the ten output indices is multivariate. Therefore, a multivariate regression method is used to determine the correlation between six indicators, and then the test indices for each indicator in the TRO are generated, as shown in the following formula:

[0036] ,

[0037] in, These are regression coefficients, representing the correlation between output and input indicators. Each This represents the corresponding equipment performance testing requirements indicators, specifically... Determining the values ​​requires training the model. After collecting sufficient samples, a multiple regression method is used to fit the sample data and solve for the regression coefficients. During training, the least squares method is used to minimize the error between the predicted and actual values.

[0038] b. Design, training, iteration, and validation of the multi-objective testing requirement model. First, the Analytic Hierarchy Process (AHP) is used to perform hierarchical analysis of each indicator. Then, an initial indicator factor model is established by combining the Delphi Method and expert experience evaluation methods. Next, game theory is used to iteratively evaluate and adjust the indicator factor model through reverse positioning. Finally, the TOPSIS method is used to analyze the distance between the current model parameters and the negative ideal model to verify the model's merits. The model parameters include all input indicators. and demand index The negative ideal model refers to the scheme that corresponds to the worst value among all indicators.

[0039] In the process of finding the negative ideal model, the negative ideal solution for each indicator is first determined, that is, the worst-case value (most unfavorable value) of the indicator. For example, for the range indicator, its worst-case value is the minimum demand. Then, the negative ideal solutions for all indicators are listed, forming a negative ideal solution vector: Similarly, we can obtain .in, Let the solution be a negative ideal solution. Then, calculate the distances between various solutions and the ideal and negative ideal solutions using the Euclidean distance formula.

[0040] ,

[0041] ,

[0042] in, It is the value of the p-th scheme on the q-th index. and These are the ideal solution and the negative ideal solution, respectively. Finally, the relative similarity is calculated. To evaluate the merits of each solution:

[0043] ,

[0044] The larger the value, the closer the solution is to the ideal, and the better the solution is. According to... The values ​​are sorted from largest to smallest.

[0045] Step 3: Perform the following operations through the data-real fusion test result evaluation module:

[0046] a. Construct an adaptively improved Bayesian neural network model to achieve in-depth evaluation of the results of the data-real fusion test. The Bayesian neural network model includes an input layer, hidden layers, and an output layer, where the input layer... Receive the output of fusion digital test results and actual test results :

[0047] ;

[0048] The hidden layer uses variational inference to optimize network parameters (weights and biases). As weight, For bias, It is an activation function. Output for hidden layer:

[0049] ,

[0050] The output layer is responsible for generating the final evaluation results. As weight, For output bias, It is the output activation function:

[0051] .

[0052] b. Optimize the performance of the Bayesian neural network model using backpropagation and iterative training to ensure that the model can dynamically adapt to the multi-objective testing requirements constructed in step two. Dynamic adaptation refers to adjusting weights and bias parameters as needed. After obtaining the optimized network parameters in step a, store the suboptimal network parameters. During iterative training to optimize the Bayesian neural network model's performance, allocate a small amount of computing power to train the suboptimal network parameters. If the best training effect of the suboptimal parameters is better than that of the optimal network parameters, train the suboptimal network parameters as well. Similarly, expand the range of suboptimal network parameters sequentially. The loss function L of the backpropagation mechanism is expressed as:

[0053] ,

[0054] in, For the true value, Here, M represents the predicted value, and M is the number of samples. After calculating the gradient of the loss function with respect to the parameters of the Bayesian neural network model using the backpropagation algorithm, the network weights are updated using this gradient. Learning rate:

[0055] ,

[0056] ,

[0057] Finally, the output solution is combined with the multi-objective test requirement model constructed in step two. and Regularly evaluate and update the factor weights of the multi-objective testing requirements model:

[0058] ,

[0059] Where F is the current factor weight, and k is the adjustment coefficient. It refers to the degree of improvement in the evaluation results.

[0060] The multi-objective testing requirement model constructed in step two not only provides the initial conditions for the iterative training in step three, but also serves as an evaluation criterion during the training process, continuously promoting model optimization. This closed-loop system ensures the flexibility and adaptability of the data-real fusion testing evaluation method, enabling it to respond in real time to different testing requirements and environmental changes.

[0061] In summary, this method includes: designing a data fusion module for numerical and physical test results, which integrates numerical testing and physical testing, and can identify and address the problem of missing data samples in both numerical and physical tests, as well as the redundancy of data samples; designing a dynamic adaptation module for fusion test requirements, which can dynamically adjust the test evaluation objectives based on the accuracy of physical testing and the comprehensiveness of numerical test samples according to dynamic requirements; and a fusion test result evaluation module, which integrates the aforementioned fused sample data with test requirements, uses an adaptively improved Bayesian neural network to process the input data, forms a fusion test result evaluation model, and then performs iterative training of the model based on a backpropagation mechanism to achieve a reliable and effective fusion evaluation of the numerical and physical test results.

[0062] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0063] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A data-real fusion testing and evaluation method based on result fusion, characterized in that, include: Step 1: The digital test results and the actual test results are integrated into one by the data fusion module, which at the same time makes up for missing test data and eliminates redundant and inefficient data. Step 2: Through the data-real fusion test requirement dynamic adaptation module, based on the key primary test requirements under ideal operating conditions and the applicable compliance requirements under widely common operating conditions, the main station functional indicators and auxiliary station operation indicators are analyzed. Based on the main station functional indicators and auxiliary station operation indicators, a multi-objective test requirement model of the equipment is constructed; then the model is iteratively optimized and verified. The inputs to the multi-objective test requirement model include load capacity, driving range, maximum power, rapid repair capability, harsh operating condition tolerance level, and fault disturbance level, as well as the requirement index for each indicator, defined as: TRI is the total input. ~ There are six indicators. ~ Demand index for each indicator; Output , ~ These are test indices representing the maximum power output, fuel efficiency, transmission system efficiency, maximum fuel storage capacity, equipment frame materials, consistency level of key components, load-bearing level of parts, suspension reliability, robustness of electronic control system, and dustproof / waterproof / corrosion-proof / high temperature resistance / low temperature resistance / sealing performance of the structure, respectively. Step 3: Perform the following operations through the data-real fusion test result evaluation module: a. Construct an adaptively improved Bayesian neural network model to achieve in-depth evaluation of the results of the data-real fusion test. The Bayesian neural network model includes an input layer, a hidden layer, and an output layer, where the input layer... Receive the output from step one, which integrates the digital test results and the actual test results: ； The hidden layer uses variational inference to optimize network parameters. As weight, For bias, It is an activation function: ； The output layer is responsible for generating the final evaluation results. As weight, For output bias, It is the output activation function: ； b. Optimize the performance of the Bayesian neural network model by using backpropagation mechanism and iterative training of the model to ensure that the Bayesian neural network model can dynamically adapt to the multi-objective test requirement model constructed in step two. Step one includes: When conducting multi-stage, multi-round practical and digital tests on the equipment, where the number of stages is n and the number of rounds is m, the practical test results are obtained. and numerical test results : ， ， in, These are the results of actual testing. This is the result of the m-th round of the n-th stage in the actual test results. These are numerical test results; Redundant data with low credibility is removed or its weight is reduced. Redundant data with low credibility refers to data where the results of that stage already exist in the actual installation test, but whose numerical test results differ significantly from the actual installation test results. For some stage results missing in the actual installation test, the missing data is supplemented using numerical test results obtained through multiple weighted averages. This process is expressed as follows: ， ， ， in, This represents the weighting coefficient, which determines the degree of reduction of redundant data. Let's assume the actual test results in the experiment... Not incomplete, with This indicates the results of the digital test. Right now There is redundancy. It is a very small constant. Indicates the results of the numerical test. The output value after redundancy removal Indicates the results of actual testing. Fill in the missing output values; The final output, which integrates digital test results and actual installation test results, is as follows: ; Step two includes: a. Construct a multi-objective test requirement model for the equipment based on the main combat function indicators and auxiliary combat operation indicators. The main combat function indicators include the equipment's payload, range, and maximum power. The auxiliary combat operation indicators include the equipment's rapid component replacement and repair capabilities, the equipment's tolerance to harsh operating conditions, and the level of disturbance to the system caused by faults. Finally, construct a multi-objective test requirement model that covers both the main combat function indicators and auxiliary combat operation indicators of the equipment. b. For the multi-objective test requirement model, design training, iteration and verification methods, including: first, use the analytic hierarchy process (AHP) to perform hierarchical analysis of each indicator; then, combine the Delphi method and expert experience evaluation method to establish an initial indicator factor model; then, use game theory to evaluate repeatedly from a reverse perspective and iteratively adjust the indicator factor model; finally, use the TOPSIS method to analyze the distance between the current parameters of the model and the negative ideal model to verify the merits of the model scheme. Step three includes: The loss function L of the backpropagation mechanism is expressed as: ， in, For the true value, Here, M is the number of samples, and the gradient of the loss function with respect to the network parameters is calculated using backpropagation. The network weights are then updated using this gradient. Learning rate: ， ， Finally, based on the multi-objective test requirement model constructed in step two, the factor weights of the multi-objective test requirement model are periodically evaluated and updated: ， Where F is the current factor weight, and k is the adjustment coefficient. It refers to the degree of improvement in the evaluation results.

Images