Method and system for collaborative evaluation of mechanical properties of concrete members based on machine learning

Abstract

This invention relates to the fields of civil engineering structural safety inspection and machine learning technology, and discloses a machine learning-based collaborative evaluation method and system for the mechanical properties of concrete components. The method constructs a coupled graph of mechanical indices with confidence levels, trains a missing-node-aware joint evaluation model, and suppresses the propagation of noise from missing nodes through joint attenuation of observability masks and edge confidence weights. Then, it performs weighted SHAP interaction effect analysis and hierarchical verification, and generates a supplementary detection scheme with comprehensive information gain ranking based on the D-optimal criterion. Through iterative closed-loop optimization, it outputs the collaborative evaluation results and evaluation confidence levels of the concrete component's compressive strength, splitting tensile strength, and elastic modulus.

Description

Technical Field

[0001] This invention relates to the field of civil engineering structural safety testing and machine learning technology, and more specifically, to a method and system for collaborative evaluation of the mechanical properties of concrete components based on machine learning. Background Technology

[0002] In the safety assessment of existing concrete structures, it is necessary to conduct a coordinated assessment of three mechanical properties—compressive strength, splitting tensile strength, and modulus of elasticity—for a large number of components. Due to differences in on-site testing methods and costs, the test data for a batch of components exhibits a severe imbalance and lack of information.

[0003] Existing machine learning-based evaluation methods typically utilize the coupling relationship between mechanical indicators to construct a graph structure model for joint prediction, and combine SHAP analysis to identify the interaction effect between the ratio parameters and mechanical indicators to help improve prediction accuracy.

[0004] However, such methods face the following technical drawbacks: First, the edge weights of the mechanical index coupling graph rely on the correlation coefficients and mutual information values ​​between the indices. A high missing rate results in extremely scarce samples with complete data for multiple indices, leading to significant bias in edge weight estimation. During graph attention message passing, missing index nodes lack effective supervisory signals, and their noise propagates to other index nodes through the graph structure, reducing the prediction accuracy of existing indices. Second, SHAP interaction effect analysis requires complete data for multiple indices. With a high missing rate, complete samples are severely insufficient, and direct analysis will produce statistically unreliable pseudo-interaction effects. Derived features constructed based on pseudo-interaction effects introduce additional noise and exacerbate overfitting. Furthermore, existing methods fail to integrate the need for interaction effect verification and index missing completion into the allocation decision of limited supplementary detection resources, resulting in the limited detection budget failing to effectively serve the improvement of overall evaluation accuracy. Summary of the Invention

[0005] This invention provides a machine learning-based collaborative evaluation method and system for the mechanical properties of concrete components, which solves the technical problems in related technologies such as insufficient accuracy of multi-index collaborative evaluation, inadequate utilization of the coupling relationship between indices, and lack of processing capability for missing data when existing mechanical index data of concrete structures are missing.

[0006] This invention discloses a collaborative evaluation method for the mechanical properties of concrete components based on machine learning, comprising the following steps: The mix proportion parameters and measured values ​​of three mechanical indices—compressive strength, splitting tensile strength, and elastic modulus—of the batch of components are obtained. Missing indicator vectors for each sample are generated. Paired Pearson correlation coefficients and mutual information values ​​are calculated for each pair of mechanical indices using a subset of samples that are not missing at the same time. The effective sample size of each pair is normalized to a confidence weight. The triplet of the correlation coefficient, mutual information value, and confidence weight is used as a side attribute to generate a coupling relationship graph of mechanical indices with confidence. The standardized proportioning parameter features are generated into a shared latent feature vector through a shared feature encoding layer, and then mapped to the initial node features of each index node in the mechanical index coupling graph through an independent linear mapping layer. Missing graph attention message passing is performed on the mechanical index coupling graph. The product of the observability mask value of the source node and the confidence weight of the corresponding edge is used as a decay factor to weight the message channel. After multiple rounds of message passing, the predicted values ​​of the three mechanical indices are generated through the independent regression output layer of each index. The model is trained with regression loss calculated only for non-missing indices to obtain the missing graph joint evaluation model. The missing information-aware joint evaluation model is applied to all components to be evaluated. After setting the observability mask of the missing indicators to zero, the missing information-aware message passing and regression output are executed to generate the collaborative evaluation results of all three mechanical indicators and the evaluation confidence of each indicator.

[0007] Furthermore, the confidence weight is calculated as follows: the number of simultaneously non-missing samples is divided by the total number of all samples to obtain the confidence weight normalized by the sample size ratio; or, based on the minimum sample size required for the Pearson correlation coefficient test at a given significance level, the number of simultaneously non-missing samples is divided by the minimum sample size and then the minimum value is taken with 1, which is used as the confidence weight normalized based on the statistical test power.

[0008] Furthermore, in the missing perceptual graph attention message passing, the updated features of the target node in round t are as follows: for each neighboring source node of the target node, the decay factor, attention weight, and the source node feature vector of round t-1 after linear transformation of the learnable weight matrix in round t are multiplied respectively. The sum of the product results of all neighboring source nodes is then processed by the activation function to obtain the node feature vector of the target node in round t. The attention weight is calculated using the standard graph attention mechanism, which is obtained by concatenating the features of the source node and the target node, performing linear transformation and LeakyReLU activation, and then normalizing by softmax.

[0009] Furthermore, the regression loss calculated only for non-missing indicators is as follows: for all mechanical indicators of all samples, the square of the difference between each standardized predicted value and the standardized measured value is weighted by the corresponding element in the missing indicator vector, the sum of the weighted squared differences is divided by the total number of all non-missing indicators as the loss function value, and the Adam optimization algorithm is used to iteratively minimize the loss function.

[0010] Furthermore, the confidence level is calculated as follows: for the k-th indicator of the i-th sample, the product of the decay factor and the attention weight received by the node from each neighbor node in all message passing rounds is accumulated and summed, and then divided by the product of the total number of message passing rounds and the number of neighbor nodes of the corresponding node to obtain the confidence level of the indicator.

[0011] Furthermore, after obtaining the missing perception joint evaluation model, the following steps are also included: The missing information perception joint evaluation model is used to perform forward inference on all samples, generate completion values ​​and corresponding completion confidence for the missing index positions, combine the completion values ​​with the original non-missing standardized measured values ​​to generate a pseudo-complete multi-index data matrix, and mark the observation confidence level for each data item according to the comparison results of the missing indicator vector and completion confidence with the preset high confidence threshold. Based on the pseudo-complete multi-index data matrix and the standardized matching parameter feature matrix, an independent gradient boosting regression model is trained for each of the three mechanical indices. During training, the geometric mean of the numerical values ​​of the observed confidence scores of each index for each sample is used as the sample weight to weight the loss contribution of each sample. After training, the SHAP interaction value is calculated. The confidence interval of each pair of feature interaction values ​​is calculated using weighted Bootstrap resampling. The stability index is calculated based on the probability that the confidence interval does not cross zero. Feature pairs with a stability index higher than a preset threshold are marked as high-confidence interaction effects, and feature pairs with a stability index lower than the preset threshold are marked as interaction hypotheses to be verified, thus generating a hierarchical interaction effect hypothesis set.

[0012] Furthermore, the numerical value rules for the observation confidence level are as follows: the numerical value corresponding to the index with a measured value is 1.0; the numerical value corresponding to the index that is missing and whose completion confidence is higher than the high confidence threshold is a preset intermediate value between the high confidence threshold and 1; the numerical value corresponding to the index that is missing and whose completion confidence is lower than the high confidence threshold is a preset low value that is lower than the preset intermediate value; in the weighted Bootstrap resampling, the sample weight is used as the sampling probability weight of each sample for sampling with replacement, and the distribution of each pair of feature interaction values ​​is obtained after multiple resampling.

[0013] Furthermore, after generating the set of hypotheses on hierarchical interaction effects, the following steps are also included: For each feature pair corresponding to the interaction hypothesis to be verified, based on the regression design matrix containing the intercept term, two main effect terms and the interaction term, the D-optimal criterion is used to select a set of candidate components from the existing component set that maximizes the determinant of the information matrix, wherein the information matrix is ​​formed by the product of the transpose of the regression design matrix and itself. For each component-index combination in the candidate component set, calculate the comprehensive information gain. The comprehensive information gain is the weighted sum of the expected increment of the determinant of the information matrix after completing the index and the complement of the confidence level of the index completion, after normalization. Sort the supplementary detection schemes in descending order of the comprehensive information gain to generate a list of supplementary detection schemes. The expected increment of the determinant of the information matrix is ​​calculated using the rank-one update formula of the matrix determinant.

[0014] Furthermore, it also includes an iterative closed-loop optimization step: According to the priority of the supplementary detection scheme list, new measured data are obtained, the completion values ​​at the corresponding positions in the pseudo-complete multi-index data matrix are replaced and the missing indicator vector is updated, and the missing perception joint evaluation model and gradient boosting regression model are retrained. Explicit interaction feature generation is performed only on feature pairs corresponding to high-confidence interaction effects in the set of hierarchical interaction effects, and the product term of the standardized features of the pair is concatenated as a derived feature to the standardized matching parameter feature matrix. Repeat the steps of generating a pseudo-complete multi-indicator data matrix, weighted SHAP interaction effect analysis, generating a supplementary detection scheme, and the iterative closed-loop optimization, until the supplementary detection budget is exhausted or all interaction hypotheses to be verified have been confirmed. Finally, apply the missing awareness joint evaluation model to all components to be evaluated to output the final collaborative evaluation result.

[0015] This invention provides a machine learning-based collaborative evaluation system for the mechanical properties of concrete components, comprising: The data acquisition and coupling relationship diagram generation module is used to acquire the proportioning parameter characteristics of batch components and the measured values ​​of three mechanical indicators, generate missing indicator vectors, calculate the pairwise correlation coefficients, mutual information values ​​and confidence weights of each pair of mechanical indicators, and generate a coupling relationship diagram of mechanical indicators with confidence. The missing information-aware joint evaluation model training module is used to generate initial node features for each index node by passing the standardized ratio parameter features through a shared feature encoding layer and an independent linear mapping layer. On the mechanical index coupling relationship graph, the missing information graph attention message passing is performed using the product of the observability mask value and the confidence weight as a decay factor. Predicted values ​​are generated through the independent regression output layer of each index. The missing information-aware joint evaluation model is obtained by training with regression loss calculated only for non-missing indicators. The collaborative evaluation output module is used to apply the missing information-aware joint evaluation model to all components to be evaluated, perform missing information-aware message passing and regression output, and generate collaborative evaluation results of all three mechanical indicators and the evaluation confidence of each indicator.

[0016] This invention addresses the technical problem of decreased cross-index prediction accuracy caused by the propagation of noise information from missing index nodes through the graph structure under high missing rates by introducing a joint attenuation mechanism of observability mask and edge confidence weights in the attention message passing of the missing index. It achieves the technical effect of maintaining the effectiveness and stability of message passing on the coupling graph of mechanical indices under high missing rates. Furthermore, this invention solves the technical problem of false interaction effect misjudgment leading to invalid derived features and exacerbating overfitting under high missing rates by combining weighted Bootstrap resampling with a stability index grading strategy. It achieves the technical effect of suppressing the interference of low-quality completion values ​​on interaction effect identification and reducing the risk of false interaction feature construction. Finally, this invention integrates the interaction effect verification requirement and the index missing completion requirement into the supplementary detection optimization objective by using comprehensive information gain based on the D-optimal criterion. This solves the technical problem that limited detection budgets fail to effectively serve the improvement of overall evaluation accuracy, achieving the technical effect of improving the marginal contribution of each supplementary detection to the overall evaluation accuracy under limited detection budget constraints. Attached Figure Description

[0017] Figure 1 This is a flowchart of the machine learning-based collaborative evaluation method for the mechanical properties of concrete components provided in this embodiment of the invention; Figure 2 This is a schematic diagram showing the detection coverage and missing distribution of the three mechanical indicators provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the comparison of the edge attributes of the coupling relationship of mechanical indicators provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the effective message strength of each round of message transmission for the elastic modulus node of component ID007 provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the mass weight heatmap of various mechanical properties of a typical component provided in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the weight comparison of some component samples provided in an embodiment of the present invention; Figure 7 This is a schematic diagram comparing the stability index of the SHAP interaction effect with the threshold provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of the comprehensive information gain ranking of candidate component supplementary detection provided in an embodiment of the present invention. Detailed Implementation

[0018] In the safety assessment of existing concrete structures, it is necessary to conduct a coordinated evaluation of three mechanical properties—compressive strength, splitting tensile strength, and modulus of elasticity—for a large number of components. Due to differences in on-site testing methods and costs, the test data of a batch of components exhibits a severe imbalance and lack of information. Existing machine learning-based assessment methods typically utilize the coupling relationships between mechanical indicators to construct graph structural models for joint prediction, and combine SHAP analysis to identify the interaction effects between mix proportion parameters and mechanical indicators to help improve prediction accuracy. However, such methods face two mutually restrictive technical bottlenecks.

[0019] First, the edge weights of the mechanical index coupling graph rely on the correlation coefficients and mutual information values ​​between the indices. However, a high missing rate results in extremely scarce samples with complete data for multiple indices, leading to significant biases in the estimation of edge weights. Furthermore, during graph attention message passing, missing index nodes lack effective supervisory signals, and their noise propagates to other index nodes through the graph structure, actually reducing the prediction accuracy of existing indices. Second, SHAP interaction effect analysis requires complete data for multiple indices to reliably identify the interaction between the ratio parameters and mechanical indices. With a high missing rate, complete samples are severely insufficient, and direct analysis will produce statistically unreliable pseudo-interaction effects. Derived features constructed based on pseudo-interaction effects introduce additional noise and exacerbate overfitting. Moreover, existing methods fail to integrate the needs for interaction effect verification and index missing completion into the allocation decision of limited supplementary detection resources, resulting in the limited detection budget failing to effectively serve the improvement of overall evaluation accuracy.

[0020] According to an embodiment of this invention, a collaborative evaluation method for the mechanical properties of concrete components based on machine learning is provided. It should be understood that the hardware environment for implementing this method includes: a database server for storing batch component inspection data, and a computing device with a graph processor for performing model training and inference calculations.

[0021] At least one embodiment of the present invention discloses a collaborative evaluation method for the mechanical properties of concrete components based on machine learning, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain multi-source test data of batch components and generate a mechanical index coupling relationship diagram with confidence level; Acquire the mix proportion parameters and on-site testing conditions of a batch of components to be evaluated in existing concrete structures, as well as the measured values ​​of three mechanical properties for each component: compressive strength, splitting tensile strength, and modulus of elasticity, allowing for the inclusion of some missing properties. Generate a three-dimensional binary missing property indicator vector for each sample. ,in Indicates the first The first sample The mechanical properties have measured values. This indicates that the indicator is missing. This serves as an index for mechanical parameters.

[0022] The proportioning parameter features are preprocessed using Z-score standardization to eliminate dimensional differences between different proportioning parameters, resulting in a standardized proportioning parameter feature matrix. ,in The total number of all samples. This represents the dimension of the proportioning parameter characteristics. The measured values ​​of the three mechanical properties were also preprocessed using Z-score standardization to eliminate the dimensional differences between compressive strength, splitting tensile strength, and elastic modulus, resulting in standardized measured values ​​of the mechanical properties. This is to enable subsequent calculation of correlation coefficients, mutual information, and regression loss on a uniform scale.

[0023] For each pair of mechanical indices Using only a subset of samples where both pairs of indicators are not missing, the pairwise Pearson correlation coefficient is calculated based on the standardized measured values ​​of the mechanical indicators. and nonlinear mutual information value At the same time, the effective sample size of this subset Normalized values ​​are used as confidence weights for estimating the correlation between the indicators. : in The total number of all samples. As an indicator With indicators The number of non-missing samples. (Based on correlation coefficient) Mutual information value and confidence weight triples As an edge attribute, a confidence-based mechanical index coupling graph is generated with the three mechanical indices as nodes. ,in These correspond to the three index nodes: compressive strength, splitting tensile strength, and elastic modulus, respectively. Let be the set of edges between each pair of indicators.

[0024] It should be noted that the above confidence weights It is a measure of the paired effective sample size after normalization. Specifically, sample size ratio normalization can be used, that is... Divide by total sample size Alternatively, a normalization method based on statistical power can be used, which determines the minimum sample size required at a given significance level based on the Pearson correlation coefficient test. The confidence weight is defined as Thus, when the effective sample size reaches the minimum sample size required for statistical testing, the confidence weight saturates to 1.

[0025] Step 2: Train the missing-aware joint evaluation model; Standardized proportioning parameter feature matrix Input the shared feature encoding layer to generate a shared latent feature vector for each sample. ,in For the dimensions of the matching parameter characteristics, Let be the dimension of the latent features. The above shared feature encoding layer is implemented using a multilayer perceptron, and its input is a dimensional . Standardized proportioning parameter feature vector The output is a dimension of Shared latent feature vectors During training, the entire missing-aware joint evaluation model undergoes end-to-end joint training. A shared coding layer is used instead of an independent coding layer because all three mechanical indices are jointly affected by the concrete mix proportion parameters, and the shared coding layer enables latent features to capture the general characterization of mechanical properties by the mix proportion parameters.

[0026] Step 201: Share the latent feature vector The results are mapped to a confidence-based mechanical index coupling graph through three independent linear mapping layers. Initial node characteristics of the three indicator nodes ( ), and at the same time, the missing indicator vector for each sample Mapped as an observability mask for three nodes .

[0027] Step 202: In the coupling relationship diagram of mechanical indices with confidence level The upper part performs attention message passing on the missing perceptual graph. For the first... Round-robin message passing, from the source node To the target node When transmitting messages, the observability mask value of the source node is used. confidence weights of corresponding edges Multiply, as the decay factor of this message channel. : in For the first Each sample source node The observability mask value, For the edge Confidence weighting. When the measured values ​​of the source indicator are missing. The messages transmitted from this node to the target node are completely masked; when the number of effective paired samples is small. The smaller the value, the less the impact of the message channel. Target node In the The update characteristics of the wheel are: in For nodes The set of neighboring nodes, For the first Wheel from source node To the target node Attention weights For the first The learnable weight matrix of the round, For activation function, For the source node In the The node feature vectors of the wheel. The attention weights mentioned above. The standard graph attention mechanism is used for computation, which involves concatenating the features of the source and target nodes, performing a linear transformation and LeakyReLU activation, and then normalizing using softmax. After passing the missing information message, the output is a fused feature vector of the three indicator nodes. ( ),in To determine the total number of message passing rounds, standardized predicted values ​​for three mechanical indices are generated through separate regression output layers (fully connected layers with an output dimension of 1). ,Right now ,in and The first The learnable weight vector and bias scalar of the regression output layer. This indicates transpose.

[0028] Step 203: During training, the loss function calculates the regression loss only for non-missing metrics: in The total number of all samples. For the first The first sample The standardized measured values ​​of the indicators, For the corresponding standardized predicted value, For the corresponding element in the missing indicator vector, the denominator This represents the total number of all non-missing indicators. Because... and All parameters have been Z-score standardized, and the three mechanical indices are on a uniform dimensionless scale, with consistent dimensions in the loss function. The Adam optimization algorithm is used to iteratively minimize the loss function to obtain the missing-aware joint evaluation model.

[0029] Step 3: Perform a collaborative evaluation on all components to be evaluated and output the evaluation results of multiple mechanical indices; The finally trained missing-aware joint evaluation model is applied to all components to be evaluated. For components with only partial measured values ​​of indicators, the existing measured values ​​are assigned to the corresponding indicator nodes as strong prior features. After setting the observability mask of the missing indicators to 0, missing-aware message passing and regression output are performed to generate standardized collaborative evaluation results for all three mechanical indicators. ( Then, through the inverse transformation of the Z-score standardization of each index in step 1, the mechanical index evaluation values ​​under the original dimensions are restored, as well as the evaluation confidence level of each index.

[0030] It should be noted that the confidence level of each of the above indicators is calculated based on the cumulative value of the effective decay factor converging to that indicator node during message transmission. Specifically, for the first... The first sample The confidence level of the indicators. Defined as this node in The cumulative strength of valid messages received in round-robin messaging: in For the total number of rounds of message passing, For nodes The number of neighboring nodes, As the attenuation factor, For the first The attention weight of the wheel. When this metric itself has measured values... When the observability and edge confidence of its neighbor indicators are both high, the assessment confidence tends to be high; when the indicator is missing and the information of the neighbor indicators is insufficient, the assessment confidence is low.

[0031] In this embodiment of the application, in order to further improve the accuracy of collaborative evaluation, in addition to step 2, the following steps are also included, which are used to gradually enhance the performance of the missing perception joint evaluation model through an iterative closed loop of pseudo-complete data completion, interaction effect analysis and supplementary detection optimization.

[0032] Step 4: Generate a pseudo-complete multi-indicator data matrix; Using a missing-aware joint evaluation model, forward inference is performed on all samples to generate graph message passing completion values ​​for the locations of missing indicators. This completion value is the output under standardized scaling. The completion confidence of each completion value is generated based on the cumulative calculation of the effective decay factor in message passing. Complete the confidence level The calculation method is the same as the confidence level assessment in step 3. The calculation method is the same. The padded value will be... Compared with the original non-missing standardized measured values Combine to generate a pseudo-complete multi-indicator data matrix At the same time, the observation confidence level is labeled for each data item. :like Then it is marked as the highest level; if and If the confidence level is higher than the preset high confidence threshold, it is marked as medium level; if and If the confidence level is below the preset high confidence threshold, it is marked as low level.

[0033] Furthermore, the aforementioned preset high confidence threshold is used to distinguish the reliability of the completed value, and to determine the confidence level of the completed value. Missing indicator completion values ​​above the high confidence threshold are considered highly reliable estimates, while completion values ​​below the high confidence threshold are considered less reliable estimates. This high confidence threshold can be set according to the accuracy requirements of the specific engineering assessment. A higher high confidence threshold means a more stringent assessment of the reliability of the completion values, thus resulting in a higher proportion of high-quality data in subsequent weighted analyses.

[0034] Step 5: Perform weighted SHAP interaction effect analysis to generate a set of hierarchical interaction effect hypotheses; Based on pseudo-complete multi-indicator data matrix and standardized proportioning parameter characteristic matrix Each of the three mechanical indices is trained with an independent gradient boosting regression model, and the feature matrix of the standardized ratio parameter is used. As input, a pseudo-complete multi-indicator data matrix Each mechanical index is listed as a prediction target, and the sample weights defined in step 501 are used during training. The loss contribution of each sample is weighted. After training, the SHAP interaction value is calculated based on each gradient boost regression model for subsequent interaction effect analysis.

[0035] Step 501: Calculate the weight of each sample. The geometric mean of the observed confidence levels of each indicator in the sample is used as the sample weight: in For the first The first sample Numericalization of the observation confidence level of each indicator. As an index for mechanical indicators, this allows high-confidence samples dominated by measured values ​​to contribute more significantly to the estimation of interaction effects. Specifically, the numerical value rules for the observation confidence levels are as follows: highest level ( (i.e., there are measured values) Medium level ( and (Higher than the preset high confidence threshold) Take a preset midpoint between the high confidence threshold and 1, for example, take... Low level ( and (Below the preset high confidence threshold) Take a preset low value that is less than the value of the medium-level numerical value, for example, take The above numerical values ​​can be adjusted according to the accuracy requirements of specific engineering assessments. The core principle is to make the weight of the measured value sample significantly higher than the weight of the supplementary value sample.

[0036] Step 502: Calculate the confidence interval for each pair of feature interaction values ​​using weighted Bootstrap resampling. For each resampling, ... Sampling with replacement is performed on the sampling probability weights of each sample. The SHAP interaction value on the resampled dataset is calculated. After multiple resampling, the distribution of each pair of feature interaction values ​​is obtained, and then the confidence interval of each pair of feature interaction values ​​is calculated.

[0037] Step 503: Calculate the stability index of each pair of feature interactions. , defined as the probability that the confidence interval does not cross zero: in and These are the two feature indices in the feature pair. The stability index... Higher than the preset threshold The features are labeled as high-confidence interaction effects, and the stability index is... Below the preset threshold The feature pairs are labeled as interaction hypotheses to be verified, generating a set of hierarchical interaction effect hypotheses. The aforementioned preset threshold... The preset threshold can be set according to the accuracy requirements of specific engineering assessments, with a higher preset threshold. This means that the confirmation of interaction effects is more rigorous.

[0038] Furthermore, the aforementioned stability index The physical meaning is: the probability that the interaction effect direction of this feature pair remains consistent (i.e., the confidence interval is always positive or always negative) in the SHAP interaction value distribution obtained from multiple Bootstrap resampling. When the stability index... When the stability index approaches 1, it indicates that the interaction effect of this feature pair can be stably reproduced on different subsets of data, and has high statistical reliability; when the stability index... A lower value indicates that the interaction effect is inconsistent in direction across different subsets of data. The possibility of pseudo-interaction caused by data noise or completion error cannot be ruled out. It should be included in the set of interaction hypotheses to be verified and re-evaluated after subsequent supplementary testing.

[0039] Furthermore, the stability index in step 502 above The calculation method is as follows: In all Bootstrap resampling iterations, the feature pairs are statistically analyzed. The stability index is obtained by dividing the number of resampling operations where the lower bound of the SHAP interaction value confidence interval is greater than zero or the upper bound of the confidence interval is less than zero by the total number of resampling operations. The estimated value is obtained. When the total number of resampling attempts is sufficient, this estimated value converges to the true value of the stability index.

[0040] Step 6: Generate a list of supplementary detection schemes sorted by overall information gain; Step 601: For each feature pair corresponding to the interaction hypothesis to be verified Based on a regression design matrix containing main effects and interaction terms, the D-optimal criterion is used to select components from the existing set that enable the information matrix to be optimized. determinant The set of candidate components that is maximized.

[0041] Furthermore, the aforementioned information matrix It consists of the product of the transpose of the regression design matrix and itself, i.e. ,in Indicates transpose. To target feature pairs The constructed regression design matrix has each row corresponding to a component sample, and each column contains the intercept term, features, and other parameters. Main effect terms, characteristics Main effect terms and characteristics With features The interaction term (i.e., the product of their standardized eigenvalues). Information matrix. determinant This reflects the current component set in feature pairs The larger the determinant, the stronger the ability of the current design point to identify the interaction effect. The D-optimal criterion maximizes the information matrix. determinant The candidate components that are most representative in the feature space and contribute the most to the estimation of interaction effects are selected from the existing component set, so as to maximize the statistical power of the newly measured data obtained by subsequent supplementary detection for the verification of interaction effects.

[0042] Step 602: For each component in the candidate component set, evaluate its current missing status and completion confidence for each metric. For component-indicator combinations that are in key areas of interaction effect verification but have low confidence in completing missing indicators, [further details needed]. Calculate the overall information gain : in For the components Complete the indicators Post-information matrix The expected increment of the determinant For components No. The confidence level of the completed indicators, and To separately control for interaction effects, a trade-off coefficient was used to assess the relative importance of the verification requirements and the requirements for completing missing indicators. It should be noted that... and The numerical ranges differ, and in practical applications, both should be normalized to the mean based on their ranges to scale them to the mean. Then perform a weighted summation after excluding the range to ensure and This allows for effective control over the relative importance of the two. When there are many interaction hypotheses to be verified at the beginning of the iteration, the importance can be appropriately increased. The focus is on verifying the interaction effect; when the need for missing completion becomes more prominent in the later stages of iteration, the size can be appropriately increased. Based on comprehensive information gain Sort the results from largest to smallest to generate a list of supplementary detection schemes.

[0043] Furthermore, the above The calculation method is as follows: assuming that for the component The After supplementary testing of the indicator, the corresponding row of the component is added to the regression design matrix. The updated information matrix is ​​calculated using the rank-one update formula for the matrix determinant (Sherman-Morrison formula). The increment of the determinant relative to the current determinant is obtained. The expected value. This calculation does not require reconstructing the complete information matrix. It can be completed in constant time and is suitable for batch calculation of all component-index combinations in the candidate component set.

[0044] Step 7: Perform iterative closed-loop optimization and output the final collaborative evaluation results; Obtain the newly added measured data for supplementary testing according to the priority of the supplementary testing plan list, and replace the pseudo-complete multi-index data matrix with the newly added measured values ​​after Z-score standardization. The missing indicator vector is updated by filling in the missing values ​​at the corresponding positions. Using the updated data, retrain the missing-aware joint evaluation model from step 2, and retrain the gradient boosting regression model from step 5.

[0045] Step 701: Perform explicit interaction feature generation only on feature pairs corresponding to high-confidence interaction effects in the set of hierarchical interaction effects, that is, concatenate the product term of the standardized features of the pair as derived features to the standardized matching parameter feature matrix. .

[0046] Step 702: Repeat the iterative closed loop from Step 4 to Step 7 until the supplemental detection budget is exhausted or all pending interaction hypotheses have been confirmed.

[0047] Step 703: Apply the final iteration of the missing information joint evaluation model to all components to be evaluated in the manner of Step 3, generate the final standardized collaborative evaluation results of all three mechanical indicators, and then restore the mechanical indicator evaluation values ​​under the original dimensions and the evaluation confidence of each indicator through the inverse transformation of Z-score standardization of each indicator.

[0048] In this embodiment, the specific problem addressed by the iterative closed loop in steps 4 to 7 is that single-cycle model training cannot fully utilize the coupling relationship between mechanical indicators due to excessively high missing rates, and one-time interaction effect analysis is prone to producing spurious interaction judgments under incomplete data conditions. Through iterative closed loop, the newly added measured values ​​introduced in each round of supplementary detection simultaneously serve two objectives—correcting the edge weight estimation of the mechanical indicator coupling relationship graph with confidence and confirming or denying the interaction hypothesis to be verified, enabling the missing-aware joint evaluation model to continuously correct evaluation biases as data is gradually enriched.

[0049] This implementation dynamically masks the message output of missing indicator nodes during message passing by combining an observability mask and edge confidence weights introduced in the missing graph attention message passing process, and correspondingly attenuates the edge weights based on small sample estimates. Therefore, when the measured value of an indicator node is missing, that node will not output unsupervised, noisy messages to its neighbors, thus avoiding the decrease in cross-indicator accuracy caused by the propagation of missing information to existing indicator nodes through the graph structure. Simultaneously, when the number of effective samples between a pair of indicators is insufficient, the message channel weights on that edge are adaptively attenuated, reducing the impact of unreliable edge weight estimation on the message passing results. Thus, this implementation can maintain the effectiveness and stability of message passing on the confidence-based mechanical indicator coupling graph even with a high missing rate.

[0050] Furthermore, in the weighted SHAP interaction effect analysis, the geometric mean of the completion confidence scores is used as the sample weight for weighted Bootstrap resampling. This ensures that high-confidence samples, which are mainly composed of measured values, contribute more to the interaction effect estimation and suppresses the interference of low-quality completion values ​​on interaction identification. Combined with a tiered strategy using the stability index, only feature pairs whose confidence intervals do not cross zero after Bootstrap validation are identified as high-confidence interaction effects and derived features are generated accordingly. Feature pairs that are not yet statistically significant are retained as hypotheses to be verified. This avoids the risk of misclassifying noise-induced pseudo-interactions as valid interactions and constructing invalid derived features, thus exacerbating the risk of overfitting.

[0051] Furthermore, the supplementary detection scheme based on the D-optimal criterion integrates the requirements for verifying interaction effects and filling missing indicators into the same optimization objective function by comprehensively considering information gain. This allows the component-indicator combination selected for each supplementary detection to simultaneously serve the confirmation or denial of interaction effects and the correction of the edge weights of the coupling relationship graph of mechanical indicators with confidence. Thus, under the constraint of a limited detection budget, the marginal contribution of each supplementary detection to the overall evaluation accuracy is maximized.

[0052] like Figure 2-8 As shown, the method in this embodiment operates as follows: A city conducted a safety survey of a batch of existing concrete frame structures built in the 20XX, selecting 120 beam and column components for evaluation. Due to limited on-site testing resources, only the rebound compressive strength test was completed for some components, while the ultrasonic testing coverage for splitting tensile strength and modulus of elasticity was insufficient, resulting in a severe imbalance and missing data for the batch of components. The mix proportion parameters of each component include six items: water-cement ratio, sand ratio, cement content, fly ash content, water-reducing agent content, and curing age. Three mechanical indicators, namely compressive strength (index 1), splitting tensile strength (index 2), and modulus of elasticity (index 3), need to be evaluated simultaneously.

[0053] Acquire multi-source test data of batch components and generate a coupling relationship diagram of mechanical indices with confidence level; Data was processed from 120 components, with 85% coverage of measured compressive strength, 52% coverage of splitting tensile strength, and 43% coverage of elastic modulus. A three-dimensional missing indicator vector was generated for each component; for example, the missing indicator vector for component ID007 was... This indicates that the compressive strength and splitting tensile strength have measured values, but the elastic modulus is missing.

[0054] After Z-score standardization of the proportioning parameters and the measured values ​​of the three mechanical indices, the effective sample size for each pair of mechanical indices that are simultaneously non-missing was calculated, and the pairwise Pearson correlation coefficient, nonlinear mutual information value, and confidence weight were calculated. Taking the compressive strength and elastic modulus pair as an example, the effective sample size for simultaneously non-missing values ​​was calculated. Total number of components The confidence weights are: Table 1. Edge attributes of the coupling relationship graph of mechanical indices: As shown in Table 1, the pairwise valid sample size for splitting tensile strength and elastic modulus is the smallest (38 samples), and its confidence weight is... The lowest value means that the message channel on that edge will experience the strongest attenuation during subsequent graph attention transmission, which aligns with the actual distribution of data quality.

[0055] Training the missing-aware joint evaluation model; The standardized proportion parameter feature matrix of 120 components is input into the shared feature encoding layer, and the output latent feature dimension is set to 32 to generate a shared latent feature vector for each component.

[0056] In step 201, component ID007 (missing indicator vector) is used. Taking (e.g.,) as an example, the observability mask for its three indicator nodes is: compressive strength node Splitting tensile strength node Elastic modulus node This indicates that the elastic modulus node does not output messages to neighboring nodes during message passing.

[0057] In step 202, taking the elastic modulus node (index 3) receiving a message from the compressive strength node (index 1) in the first round of message passing for component ID007 as an example, the attenuation factor is: The attenuation factor for receiving messages from the splitting tensile strength node (index 2) is: Conversely, if the missing indicator vector of component ID015 is The attenuation factor when the splitting tensile strength node transmits a message to the compressive strength node is: The messaging channel was completely blocked.

[0058] In step 203, the calculation logic of weighted regression loss is illustrated using the loss contribution of 5 components as an example: Table 2 Examples of component loss calculations (standardized scale): The denominator of the loss function is the total number of non-missing metrics among all 120 components. Assuming the total number of non-missing metrics is 214, the loss function only includes the loss for rows with a missing '1' label; rows with a missing '0' label do not participate in gradient updates. Perform a collaborative evaluation on all components and output evaluation results with multiple mechanical indices; Component ID007 (elastic modulus missing) Taking the example of a node with a modulus of elasticity, let the total number of message passing rounds be 3, and the number of neighbors of the modulus of elasticity node be 2 (compressive strength node and splitting tensile strength node). The attention weights for each round are as follows: Table 3. Calculation of confidence level for the elastic modulus of component ID007 at the node: The confidence level is assessed as the average strength of valid messages from each source across all rounds, for a total of 3 rounds × 2 source nodes = 6 items: After the inverse Z-score transformation, the elastic modulus assessment value of component ID007 was restored to its original dimension (GPa), with an assessment confidence level of 0.353, which is at a medium confidence level. This reflects the objective limitations of missing elastic modulus and low confidence of its neighboring edges.

[0059] Generate a pseudo-complete multi-indicator data matrix; A pre-trained missing-aware joint evaluation model is used to generate completion values ​​for the missing indicators of 120 components, with a preset high-confidence threshold of 0.65. The construction process of the pseudo-complete matrix is ​​demonstrated using four typical components as examples: Table 4 shows the pseudo-complete multi-index data matrix and observation confidence levels for some components: The confidence level of the elastic modulus completion for component ID003 is 0.712 > 0.65, and it is marked as medium level with a mass weight of 0.6; the confidence level of the elastic modulus completion for component ID007 is 0.353 < 0.65, and it is marked as low level with a mass weight of 0.2.

[0060] Perform a weighted SHAP interaction effect analysis to generate a set of hierarchical interaction effect hypotheses. In step 501, taking the sample weight calculation of components ID003 and ID007 as an example, the sample weight is taken as the geometric mean of the quality weights of the three indicators: Component ID003 (all three indicators are at high or medium confidence level): Component ID007 (Low confidence level in elastic modulus): Component ID058 (both splitting tensile strength and modulus of elasticity are at low confidence levels): In steps 502 and 503, weighted Bootstrap resampling (500 times) is performed on 15 pairs of features, combining each of the 6 matching parameters, and the stability index of each feature pair is calculated with a preset threshold of 0.80. Table 5 shows the results of the SHAP interaction effect analysis of some features: The stability index of (water-cement ratio, cement dosage) is 0.934 > 0.80, which is identified as a high-confidence interaction effect, and the corresponding product-derived feature will be generated in step 701; the stability index of (water-cement ratio, water-reducing agent dosage) is 0.762 < 0.80, which is included in the set of hypotheses to be verified and needs to be re-evaluated after supplementary testing.

[0061] Generate a list of supplementary detection schemes sorted by overall information gain. In step 601, for the interaction hypothesis to be verified (water-cement ratio, water-reducing agent dosage), a corresponding regression design matrix is ​​constructed. Each row contains an intercept term, a main effect term of water-cement ratio, a main effect term of water-reducing agent dosage, and an interaction term of their product. The determinant of the current information matrix is ​​calculated, and a set of 18 candidate components (total 18) that contribute the most to the ability to identify the interaction effect are selected from 120 components.

[0062] In step 602, missing indicators in the candidate component set are filled in, and component-indicator combinations with low confidence are calculated to obtain the comprehensive information gain. The weighting coefficients are taken as follows: , (In the early stages of iteration, the focus was on verifying interaction effects.) The results after normalization are as follows: Table 6 Ranking of Candidate Component Supplementary Detection Comprehensive Information Gain (Partial): Component ID007 has the lowest confidence level for elastic modulus completion (0.353) and the largest contribution to the incremental determinant of the interaction effect information matrix (normalized value 0.831). Its comprehensive information gain of 0.803 ranks first, so it should be given priority for supplementary testing.

[0063] Perform iterative closed-loop optimization and output the final collaborative evaluation result; Based on the priority of the supplementary testing plan list, the first round of supplementary testing will be conducted on the top 8 component-index combinations. After obtaining the newly measured values, the corresponding values ​​in the pseudo-complete matrix will be replaced using Z-score standardization, and the missing indicator vector will be updated. For example, after the elastic modulus supplementary testing of component ID007 is completed, its missing indicator vector will be updated to... The effective sample size for the component's indices (splitting tensile strength, elastic modulus) increased from 38 to 39, and the confidence weights were updated as follows: In step 701, the product-derived features corresponding to the two pairs of high-confidence interaction effects (water-cement ratio, cement dosage) and (sand ratio, fly ash content) are concatenated to the proportioning parameter feature matrix, expanding the feature dimension from 6 to 8, and the missing information perception joint evaluation model is retrained.

[0064] After three rounds of closed-loop iteration, the stability index of (water-ash ratio, water-reducing agent dosage) increased to 0.847, exceeding the threshold of 0.80, and turned into a high-confidence interaction effect and generated corresponding derived features; the stability indices of (curing age, fly ash dosage) and (sand ratio, water-reducing agent dosage) were 0.783 and 0.611, respectively, and were still in the state of pending verification, but the supplementary testing budget had been exhausted, and the iteration was terminated.

[0065] In step 703, the final iteration of the missing-aware joint evaluation model is applied to all 120 components to generate the final collaborative evaluation results of the three mechanical indices and inversely transform them back to the original dimensions: Table 7: Final Collaborative Evaluation Results of Some Components After supplementary testing, the confidence level of the elastic modulus of component ID007 increased from the initial 0.353 to 0.883, significantly enhancing the reliability of the assessment results. The confidence levels of the splitting tensile strength and elastic modulus of component ID058 were 0.654 and 0.521, respectively, which are still relatively low and should be given attention in engineering decision-making.

[0066] The data flow logic of the entire implementation process is as follows: Step 1 generates missing indicator vectors and edge attributes of the coupling relationship graph of mechanical indicators with confidence scores from the original detection data of 120 components, where the confidence score weights directly reflect the differences in data coverage; Step 2 calculates the attenuation factor using these edge attributes and mask values, dynamically adjusting the influence of each message channel during message passing, and the loss function is only calculated for non-missing indicators during training; Step 3 outputs the evaluation confidence score, which inherits the accumulated information of the attenuation factor in Step 2; Step 4 generates the observation confidence score, etc., based on the evaluation confidence score. In step 5, the geometric mean of each index level is used as the sample weight to ensure that high-confidence samples dominated by measured values ​​contribute more to the estimation of interaction effects. In step 6, the integrated information gain unifies the interaction effect verification requirements (normalized determinant increment) and the missing information completion requirements (normalized uncertainty) to the same ranking index. In step 7, the iterative closed loop feeds the newly added measured data back to steps 2 and 5, simultaneously corrects the side weight estimation and updates the stability judgment of the interaction effect until the supplementary detection budget is exhausted, and finally outputs a complete collaborative evaluation result that combines evaluation value and confidence level.

[0067] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A collaborative evaluation method for the mechanical properties of concrete components based on machine learning, characterized in that, Includes the following steps: The mix proportion parameters and measured values ​​of three mechanical indices—compressive strength, splitting tensile strength, and elastic modulus—of the batch of components are obtained. Missing indicator vectors for each sample are generated. Paired Pearson correlation coefficients and mutual information values ​​are calculated for each pair of mechanical indices using a subset of samples that are not missing at the same time. The effective sample size of each pair is normalized to a confidence weight. The triplet of the correlation coefficient, mutual information value, and confidence weight is used as a side attribute to generate a coupling relationship graph of mechanical indices with confidence. The standardized proportioning parameter features are generated into a shared latent feature vector through a shared feature encoding layer, and then mapped to the initial node features of each index node in the mechanical index coupling graph through an independent linear mapping layer. Missing graph attention message passing is performed on the mechanical index coupling graph. The product of the observability mask value of the source node and the confidence weight of the corresponding edge is used as a decay factor to weight the message channel. After multiple rounds of message passing, the predicted values ​​of the three mechanical indices are generated through the independent regression output layer of each index. The model is trained with regression loss calculated only for non-missing indices to obtain the missing graph joint evaluation model. The missing information-aware joint evaluation model is applied to all components to be evaluated. After setting the observability mask of the missing indicators to zero, the missing information-aware message passing and regression output are executed to generate the collaborative evaluation results of all three mechanical indicators and the evaluation confidence of each indicator.

2. The machine learning-based collaborative evaluation method for the mechanical properties of concrete components according to claim 1, characterized in that, The confidence weight is calculated as follows: the number of simultaneously non-missing samples is divided by the total number of all samples to obtain the confidence weight normalized by the sample size ratio; or, based on the minimum sample size required for the Pearson correlation coefficient test at a given significance level, the number of simultaneously non-missing samples is divided by the minimum sample size and then the result is minimized by 1, which is used as the confidence weight normalized based on the statistical test power.

3. The machine learning-based collaborative evaluation method for the mechanical properties of concrete components according to claim 1, characterized in that, In the missing graph attention message passing, the updated features of the target node in round t are as follows: for each neighboring source node of the target node, the decay factor, attention weight, and the source node feature vector of round t-1 after linear transformation of the learnable weight matrix in round t are multiplied respectively. The sum of the product results of all neighboring source nodes is then processed by the activation function to obtain the node feature vector of the target node in round t. The attention weight is calculated using the standard graph attention mechanism, which is obtained by concatenating the features of the source node and the target node, performing linear transformation and LeakyReLU activation, and then normalizing by softmax.

4. The machine learning-based collaborative evaluation method for the mechanical properties of concrete components according to claim 1, characterized in that, The regression loss calculated only for non-missing indicators is as follows: for all mechanical indicators of all samples, the square of the difference between each standardized predicted value and the standardized measured value is weighted by the corresponding element in the missing indicator vector, and the sum of the weighted squared differences is divided by the total number of all non-missing indicators as the loss function value. The Adam optimization algorithm is used to iteratively minimize the loss function.

5. The machine learning-based collaborative evaluation method for the mechanical properties of concrete components according to claim 1, characterized in that, The confidence level is calculated as follows: for the k-th indicator of the i-th sample, the product of the decay factor and the attention weight received by the node from each neighbor node in all message passing rounds is accumulated and summed, and then divided by the product of the total number of message passing rounds and the number of neighbor nodes of the corresponding node to obtain the confidence level of the indicator.

6. The machine learning-based collaborative evaluation method for the mechanical properties of concrete components according to any one of claims 1 to 5, characterized in that, After obtaining the missing information joint evaluation model, the following steps are also included: The missing information perception joint evaluation model is used to perform forward inference on all samples, generate completion values ​​and corresponding completion confidence for the missing index positions, combine the completion values ​​with the original non-missing standardized measured values ​​to generate a pseudo-complete multi-index data matrix, and mark the observation confidence level for each data item according to the comparison results of the missing indicator vector and completion confidence with the preset high confidence threshold. Based on the pseudo-complete multi-index data matrix and the standardized matching parameter feature matrix, an independent gradient boosting regression model is trained for each of the three mechanical indices. During training, the geometric mean of the numerical values ​​of the observed confidence scores of each index for each sample is used as the sample weight to weight the loss contribution of each sample. After training, the SHAP interaction value is calculated. The confidence interval of each pair of feature interaction values ​​is calculated using weighted Bootstrap resampling. The stability index is calculated based on the probability that the confidence interval does not cross zero. Feature pairs with a stability index higher than a preset threshold are marked as high-confidence interaction effects, and feature pairs with a stability index lower than the preset threshold are marked as interaction hypotheses to be verified, thus generating a hierarchical interaction effect hypothesis set.

7. The machine learning-based collaborative evaluation method for the mechanical properties of concrete components according to claim 6, characterized in that, The numerical value rules for the observation confidence level are as follows: the numerical value corresponding to the index with a measured value is 1.0; the numerical value corresponding to the index that is missing and whose completion confidence level is higher than the high confidence threshold is a preset intermediate value between the high confidence threshold and 1; the numerical value corresponding to the index that is missing and whose completion confidence level is lower than the high confidence threshold is a preset low value that is lower than the preset intermediate value. In weighted Bootstrap resampling, the sample weights are used as the sampling probability weights for each sample to perform sampling with replacement, and the distribution of each pair of feature interaction values ​​is obtained after multiple resampling.

8. The machine learning-based collaborative evaluation method for the mechanical properties of concrete components according to claim 6, characterized in that, After generating the set of hypotheses on hierarchical interaction effects, the following steps are also included: For each feature pair corresponding to the interaction hypothesis to be verified, based on the regression design matrix containing the intercept term, two main effect terms and the interaction term, the D-optimal criterion is used to select a set of candidate components from the existing component set that maximizes the determinant of the information matrix, wherein the information matrix is ​​formed by the product of the transpose of the regression design matrix and itself. For each component-index combination in the candidate component set, calculate the comprehensive information gain. The comprehensive information gain is the weighted sum of the expected increment of the determinant of the information matrix after completing the index and the complement of the confidence level of the index completion, after normalization. Sort the supplementary detection schemes in descending order of the comprehensive information gain to generate a list of supplementary detection schemes. The expected increment of the determinant of the information matrix is ​​calculated using the rank-one update formula of the matrix determinant.

9. The machine learning-based collaborative evaluation method for the mechanical properties of concrete components according to claim 8, characterized in that, It also includes iterative closed-loop optimization steps: According to the priority of the supplementary detection scheme list, new measured data are obtained, the completion values ​​at the corresponding positions in the pseudo-complete multi-index data matrix are replaced and the missing indicator vector is updated, and the missing perception joint evaluation model and gradient boosting regression model are retrained. Explicit interaction feature generation is performed only on feature pairs corresponding to high-confidence interaction effects in the set of hierarchical interaction effects, and the product term of the standardized features of the pair is concatenated as a derived feature to the standardized matching parameter feature matrix. Repeat the steps of generating a pseudo-complete multi-indicator data matrix, weighted SHAP interaction effect analysis, generating a supplementary detection scheme, and the iterative closed-loop optimization, until the supplementary detection budget is exhausted or all interaction hypotheses to be verified have been confirmed. Finally, apply the missing awareness joint evaluation model to all components to be evaluated to output the final collaborative evaluation result.

10. A machine learning-based collaborative evaluation system for the mechanical properties of concrete components, used to execute the machine learning-based collaborative evaluation method for the mechanical properties of concrete components as described in any one of claims 1 to 9, characterized in that, include: The data acquisition and coupling relationship diagram generation module is used to acquire the proportioning parameter characteristics of batch components and the measured values ​​of three mechanical indicators, generate missing indicator vectors, calculate the pairwise correlation coefficients, mutual information values ​​and confidence weights of each pair of mechanical indicators, and generate a coupling relationship diagram of mechanical indicators with confidence. The missing information-aware joint evaluation model training module is used to generate initial node features for each index node by passing the standardized ratio parameter features through a shared feature encoding layer and an independent linear mapping layer. On the mechanical index coupling relationship graph, the missing information graph attention message passing is performed using the product of the observability mask value and the confidence weight as a decay factor. Predicted values ​​are generated through the independent regression output layer of each index. The missing information-aware joint evaluation model is obtained by training with regression loss calculated only for non-missing indicators. The collaborative evaluation output module is used to apply the missing information-aware joint evaluation model to all components to be evaluated, perform missing information-aware message passing and regression output, and generate collaborative evaluation results of all three mechanical indicators and the evaluation confidence of each indicator.

Images