Method for predicting performance of recycled brick aggregate concrete based on rule-based reasoning network

Abstract

The application discloses a recycled brick aggregate concrete performance prediction method based on a rule reasoning network, relates to the technical field of concrete performance prediction, and comprises the following steps: collecting recycled brick aggregate concrete sample data and calculating the attribute reliability of each attribute; a confidence rule base is constructed, a rule-specific attribute weight matrix is constructed, attribute importance factors under each rule are generated through an attribute reliability fusion algorithm with monotonicity and constraint according to the attribute reliability and the rule-specific attribute weight matrix; the activation degree and the rule activation weight are calculated according to the attribute importance factors; the consequent confidence of all rules is aggregated according to the rule activation weight to obtain a fusion confidence, and the brick aggregate influence factor prediction value vector of each sample is obtained by defuzzification; the rule-specific attribute weight matrix and the attribute reliability fusion algorithm with monotonicity and constraint are realized, and the technical effect of effectively enhancing the knowledge expression ability and the robustness of the model is achieved.

Description

Technical Field

[0001] This invention relates to the field of concrete performance prediction technology, and in particular to a method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks. Background Technology

[0002] Recycled brick aggregate concrete is a green and environmentally friendly building material. Its mechanical properties (such as brick aggregate influencing factors) are affected by the nonlinear coupling of multiple factors such as glue-water ratio, compressive strength, apparent density, water absorption, crushing value, and brick aggregate replacement rate. When using traditional artificial intelligence machine learning algorithms to predict the performance of recycled brick aggregate concrete, there are often problems of data dependence or insufficient interpretability. Among the existing technologies, the rule reasoning model of Belief Rule Base (BRB) provides a solution because it combines the interpretability of expert systems with small sample learning ability.

[0003] However, existing BRB models typically assume that all rules share a uniform set of attribute weights, meaning that the importance of the same attribute across all rules is set to a fixed value. This globally shared weight setting ignores the objective law that attribute importance dynamically changes with system state (i.e., different rules are activated). In complex systems, the degree of influence of the same attribute on the output often differs significantly under different states; this limits the model's ability to effectively characterize the differences in attribute importance under different rules and makes it difficult to accurately reflect the actual operating mechanism of complex systems.

[0004] Secondly, existing BRB models do not adequately consider the impact of interference factors in observational data. Sample data collected in actual engineering projects is inevitably affected by interference factors such as measurement errors and environmental noise, causing some attribute values ​​to deviate from their true values ​​and become unreliable. Although some scholars have proposed confidence rule base models with attribute reliability to quantify the credibility of input data, their attribute reliability fusion algorithms have significant flaws. Traditional fusion algorithms may paradoxically increase the final attribute weight when attribute reliability is uncertain—that is, attributes with unreliable information are given higher importance, which contradicts objective intuition. Furthermore, this fusion algorithm does not fully consider the rationality under extreme conditions. When attribute reliability approaches zero, it cannot completely shield the influence of unreliable information on the inference results; unreliable information may still participate in the inference process, leading to reduced model accuracy and insufficient robustness in noisy scenarios. Summary of the Invention

[0005] This application provides a method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks, which solves the problems of insufficient interpretability and data dependence in the prior art. It realizes a fusion algorithm based on rule-specific attribute weight matrix and attribute reliability with monotonicity and constraint, which effectively enhances the knowledge representation ability and robustness of the model.

[0006] This application provides a method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks, including: S1: Collect sample data of recycled brick aggregate concrete. Each sample contains several input attributes and a true target performance value. Calculate the attribute reliability of each attribute. Construct a confidence rule base and initialize rule weights, attribute weights, and consequent confidence. S2: Construct a rule-specific attribute weight matrix. Based on the attribute reliability and the rule-specific attribute weight matrix, generate attribute importance factors for each rule using an attribute reliability fusion algorithm that is monotonic and constrained. S3: Calculate the activation level and rule activation weight based on the attribute importance factor; S4: Activate weights according to rules, aggregate the consequent confidence of all rules to obtain fusion confidence, and defuzzify to obtain the predicted value vector of brick aggregate impact factor for each sample. S5: Based on the true target performance value vector and the predicted value vector, generate a single-sample prediction error vector and an overall mean square error; compare the mean square error with a preset expected error threshold. If the mean square error is not greater than the expected error threshold, the model converges and the final model is directly output for predicting the performance of recycled brick aggregate concrete.

[0007] Furthermore, the method also includes: S51: When the mean square error is greater than the preset expected error threshold, a backpropagation parameter update mechanism is executed, including: calculating the partial derivatives of consequent confidence, rule activation weight, rule weight and attribute weight based on the single sample prediction error vector and the overall mean square error, and updating the parameters; Specifically, a specific rule is used to calculate the partial derivatives of the attribute weights: for the j-th attribute weight of the k-th rule, the calculation of its partial derivative is based solely on the derivative of the k-th rule's own activation level with respect to the attribute weight, while the partial derivatives of any other rule besides the k-th rule with respect to the attribute weight are set to zero; thus, the j-th attribute weight of the k-th rule can only affect the activation level of the k-th rule itself. During parameter updates, parameters are trained on all samples one by one. Once the input is complete, the current round of parameter updates ends.

[0008] Furthermore, constructing the rule-specific attribute weight matrix includes: constructing a two-dimensional matrix based on the total number of rules in the confidence rule base and the total number of input attributes, and denoting the rule-specific attribute weight matrix as matrix. Its dimensions are Where L is the total number of rules in the confidence rule base, and M is the total number of input attributes; the matrix elements represent the specificity weight of the j-th attribute in the k-th rule. The initial value is randomly generated within the interval [0.1, 0.9].

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages: By constructing an attribute reliability fusion algorithm with monotonicity and constraints, the attribute reliability of each attribute is aggregated with the rule-specific attribute weight matrix to generate an attribute importance factor. This fusion method ensures that the attribute importance factor monotonically increases with attribute reliability, and when the reliability approaches zero, the attribute importance factor is strictly zero, thus completely shielding the attribute from its influence on rule activation and final output. When the reliability approaches one, the attribute importance factor converges to the rule-specific attribute weights themselves, degenerating into a conventional confidence rule base model. This fundamentally solves the paradox that unreliable information may still participate in reasoning, ensuring the model's robustness in extreme cases and accuracy in normal cases, enabling the model to effectively cope with engineering scenarios where some attribute data is abnormal or unreliable in actual measurements. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the process for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks in an embodiment of the present invention. Detailed Implementation

[0011] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0013] Example 1: As Figure 1As shown, a method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks is described, the method comprising: S1: Collect sample data of recycled brick aggregate concrete. Each sample contains several input attributes and a true target performance value. Calculate the attribute reliability of each attribute. Construct a confidence rule base and initialize rule weights, attribute weights, and consequent confidence.

[0014] In some embodiments, sample data is collected and consequent confidence levels and model parameters are initialized. Several (N) samples are collected from recycled brick aggregate concrete production sites or laboratories. Each sample contains M input attributes and a target performance value. In this embodiment, M=6, and the six input attributes are: glue-to-water ratio, compressive strength test value (MPa), apparent density (kg / m³), water absorption rate (%), crushing value (%), and brick aggregate replacement rate (dimensionless). The target performance value is the brick aggregate influence factor (dimensionless), which directly reflects the mechanical properties of the brick aggregate; a larger value indicates better mechanical properties.

[0015] After data collection is complete, set the j-th input attribute value of the i-th sample to... ,in ; N is the total number of samples, and M is the total number of input attributes; the true target performance value of the i-th sample is... Construct a vector of real target performance values ,in, Let be the true brick aggregate influence factor for the i-th sample (obtained from laboratory mechanical tests). The input data of all samples constitute an N-row, M-column matrix, and the target value constitutes a vector of length N.

[0016] In some embodiments, the reliability of each input attribute is calculated. The specific calculation process is as follows: For the j-th attribute, the mean of all its samples is calculated. and standard deviation Specifically: ; ; Based on the fluctuation characteristics of this attribute during actual measurement, engineering experts assign a tolerance coefficient t. The tolerance coefficient reflects the expert's tolerance for measurement errors; if the measurement of this attribute is easily affected by environmental interference, t takes a larger value, 3; if the measurement is stable, t takes a smaller value, 1.5; in this embodiment, all attributes are uniformly set to 2. Based on the tolerance coefficient, mean, and standard deviation, the reliability interval of attribute j is defined as follows: If the attribute value of a sample falls outside the reliable range, then the attribute value is considered unreliable.

[0017] Count the number of unreliable samples for the j-th attribute, and denote the proportion of reliable samples as the attribute reliability. Then, the attribute reliability is: ;in, It is the number of unreliable samples for the j-th attribute; Let be the total number of samples for the j-th attribute; the attribute reliability is 1 when all samples fall within the reliable interval; and the attribute reliability is 0 when all samples are unreliable.

[0018] The attribute reliability of all attributes constitutes an attribute reliability vector, which reflects the overall reliability of each attribute in the original dataset.

[0019] The core of a rule-based reasoning network is a confidence rule base containing L rules. The number of rules is determined by dividing each input attribute into three reference levels (low, medium, and high), theoretically allowing a maximum of L rules. However, since M=6, that is, 729 rules.

[0020] The rule weight vector is denoted as: in, This represents the weight of the k-th rule. The larger the rule weight, the greater its influence on the reasoning. The initial value is randomly generated within the interval [0,1].

[0021] In this embodiment, the brick aggregate influencing factors are divided into three evaluation results: low (L), medium (M), and high (H), i.e. Matrix elements This indicates the confidence level that the output of the k-th rule belongs to the c-th evaluation result, which must satisfy... A sum of 1 indicates a complete rule, while a sum less than 1 indicates uncertainty. Initially, the confidence level of each rule is randomly generated across three levels.

[0022] The above content yielded six raw attribute values ​​for each recycled brick aggregate concrete sample. These raw values ​​have different physical dimensions and numerical ranges, making them unsuitable for direct input into the rule-based reasoning network for algebraic operations. Each specific value must be converted into a pre-defined confidence level belonging to each reference level.

[0023] S2: Construct a rule-specific attribute weight matrix. Based on the attribute reliability and the rule-specific attribute weight matrix, generate attribute importance factors for each rule using an attribute reliability fusion algorithm that has monotonicity and constraint.

[0024] Constructing the rule-specific attribute weight matrix includes denoting the rule-specific attribute weight matrix as a matrix. Its dimensions are , where L is the total number of rules in the confidence rule base, and M is the total number of input attributes. The matrix elements represent the specificity weight of the j-th attribute in the k-th rule, with initial values ​​randomly generated within the interval [0.1, 0.9]. The purpose of the matrix is ​​to give different importance to the same attribute under different rules.

[0025] In some embodiments, for the k-th rule and the j-th attribute, their specificity weights and attribute reliability are combined into a comprehensive index, known as the attribute importance factor. When an attribute has high attribute reliability (i.e., reliable data) and a large specificity weight in the current rule, its contribution to rule activation should be amplified; conversely, if the consequent confidence is low, its contribution should be suppressed even if its specificity weight is large.

[0026] The formula for calculating the importance factor of an attribute is: ;in, It is an attribute importance factor; Let the reliability of the j-th attribute be denoted as . Let J be the rule-specific attribute weight of the j-th attribute in the k-th rule. .

[0027] For a fixed rule base, the importance factors of all attributes constitute an importance factor matrix. , dimension Its element in the k-th row and j-th column is .

[0028] Based on the importance factor matrix, for the first training sample, the degree of activation of the k-th rule is calculated. The degree of activation reflects the matching strength between the sample and the rule's preconditions. The larger the value, the more suitable the rule is for describing the current sample.

[0029] S3: Calculate the activation level and rule activation weight based on the attribute importance factor.

[0030] For the j-th attribute and the k-th rule, the activation level is calculated using the following formula: ;in, It represents the activation level of the j-th attribute and the k-th rule; Let be the individual matching degree between the j-th attribute and the k-th rule; This is the attribute importance factor. The product symbol represents the weighted product of the matching degrees of all attributes according to the importance factor index. The individual matching degree represents the matching degree of the antecedent attribute relative to the reference value in the k-th rule, which is pre-set by experts based on the antecedent attribute reference value.

[0031] For all N samples and L rules, an N×L dimensional rule activation matrix is ​​obtained, where the element in the i-th row and k-th column represents the activation level, quantitatively characterizing the matching strength between each sample and each rule. Since each rule is assigned different attribute weights, the activation level of the k-th rule is determined by the matching degree of its antecedent attributes and the rule-specific attribute weights, rather than all rules sharing a single set of attribute weights in the traditional rule-based inference network (RIN).

[0032] Different rules have varying degrees of importance in an expert system, which is characterized by the initial rule weight vector. This embodiment requires combining the rule's weight with its activation level to calculate the weight of each rule actually invoked in this inference process, i.e., the rule activation weight.

[0033] The calculation of rule activation weights follows the principle of importance weighting: the higher the inherent weight of a rule and the higher the degree of matching between the current sample and the preconditions of the rule, the greater the contribution of the rule to the final output.

[0034] Specifically, the activation weight calculation formula is as follows: ;in, It is the activation weight of the j-th attribute activating the k-th rule; Let be the rule weight of the k-th rule, with a value range of (0,1). L represents the total number of rules in the confidence rule base.

[0035] S4: Activate the weights according to the rules, aggregate the consequent confidence of all rules to obtain the fused confidence, and defuzzify to obtain the predicted value vector of brick aggregate influence factor for each sample.

[0036] In this application, the Evidence-Based Reasoning (ER) algorithm is used to aggregate the consequent confidence of all rules to obtain the fusion confidence that the output of the i-th sample belongs to the c-th evaluation result. The ER parsing algorithm expression is as follows: ; in, It is the fusion confidence score belonging to the c-th evaluation result; The confidence level of the consequent of the k-th rule belonging to the c-th evaluation result is the confidence level of the consequent. is the activation weight of the j-th attribute activating the k-th rule; L is the total number of rules in the confidence rule base; It is the sum of the consequent confidence scores of the k-th rule. If the sum is 1, it means the rule is complete; if it is less than 1, it means the rule is incomplete and there are unknown results. u is the conflict factor, which represents the degree of conflict during fusion. .

[0037] Defuzzing yields the predicted vector of brick aggregate impact factors for each sample, which is: The fused confidence distribution is then transformed into specific predicted values. Let the standard utility values ​​of the three evaluation results (low, medium, and high) be respectively... , , Based on engineering experience, experts determined the predicted value of the brick aggregate influence factor for the i-th sample as follows: ;in, This is the predicted value of the impact factor on brick aggregate; It is the standard utility value of the c-th evaluation result; the confidence level of each evaluation result is used as the weight to perform a weighted summation of the utility values ​​of each level.

[0038] For each training sample i, a specific predicted value is obtained. The predicted values ​​of all samples are combined to obtain the predicted value vector. .

[0039] S5: Based on the true target performance value vector and the predicted value vector, generate a single-sample prediction error vector and an overall mean square error; compare the mean square error with a preset expected error threshold. If the mean square error is not greater than the expected error threshold, the model converges and the final model is directly output for predicting the performance of recycled brick aggregate concrete.

[0040] In the above content, the true target value of the i-th sample is Construct a vector of real target performance values ,in, The true brick aggregate influence factor for the i-th sample (obtained from laboratory mechanical tests).

[0041] In some embodiments, for the i-th sample, the single-sample prediction error is defined as the difference between the true value and the predicted value: ; This is the single-sample prediction error. It can be positive or negative; a positive value indicates an underestimation, and a negative value indicates an overestimation. Its absolute value reflects the degree of deviation in the single-sample prediction. All single-sample prediction errors are grouped into a column vector. .

[0042] For a batch of N samples, the mean squared error (MSE) is used as the overall performance metric. The mean squared error over the entire training set is calculated as follows: ; This is the single-sample prediction error; MSE is a non-negative value, and the closer it is to 0, the higher the model's prediction accuracy. This step compares the mean squared error with a preset expected error threshold. (In this embodiment, 0.001 is used for comparison.) If If the model has converged, we can proceed to output the final model; otherwise, continue with the subsequent steps.

[0043] MSE is one of the most commonly used overall performance metrics in regression prediction problems. It squares the error, which eliminates the problem of positive and negative cancellation and amplifies the impact of large error samples, thereby guiding the model to prioritize correcting those predictions that are seriously off track.

[0044] Repeat steps two through five until any of the following convergence conditions are met: Condition 1: Current mean square error ,in The preset expected error threshold is 0.001 in this embodiment. Condition 2: In K consecutive iterations (k is 10 in this example), the absolute value of the change in MSE is less than the tolerance value. (In this embodiment, 0.0001 is used); Condition 3: The number of iterations reaches the preset maximum number. (This example uses 500 times).

[0045] If any convergence condition is met, the iteration stops, and all current optimal parameters are saved: the final rule-specific weight matrix, the rule weight vector, and the consequent confidence vector. These parameters together constitute the trained prediction model. For a new recycled brick aggregate sample, only the above optimal parameters are needed to quickly predict its brick aggregate influence factor.

[0046] The method further includes: S51: When the mean square error is greater than the preset expected error threshold, a backpropagation parameter update mechanism is executed, including: calculating the partial derivatives of consequent confidence, rule activation weight, rule weight and attribute weight based on the single sample prediction error vector and the overall mean square error, and updating the parameters.

[0047] Specifically, a specific rule is used to calculate the partial derivatives of the attribute weights: for the j-th attribute weight of the k-th rule, its partial derivative is calculated only based on the derivative of the k-th rule's own activation level with respect to the attribute weight, while the partial derivatives of any other rule with respect to the attribute weight are set to zero; thus, the j-th attribute weight of the k-th rule can only affect the activation level of the k-th rule itself; during parameter updates, all samples are trained one by one, and the current round of parameter updates ends after the input is completed.

[0048] This invention completely shields the influence of attributes on rule activation and final output by calculating attribute importance factors and setting boundedness, fundamentally resolving the paradox of unreliable information still participating in reasoning in existing technologies. Simultaneously, because the confidence level automatically adapts to changes in data distribution during iteration, when a systematic shift occurs in the measurement value of a certain attribute in a new batch of samples, the model can automatically reduce the confidence level of that attribute through error feedback without retraining, enhancing the model's robustness to noisy data and thus maintaining prediction stability. This is particularly crucial for engineering scenarios such as recycled brick aggregate concrete, where raw material sources are diverse and on-site measurement conditions are harsh.

[0049] In this application, information is transmitted through ER algorithm inference and parameters are transmitted using a gradient descent model. This achieves decoupled independent updating of rule-specific attribute weights. In traditional BRB models, all rules share a set of attribute weights. When updating a weight of a certain attribute, the activation levels of all rules contribute gradients, requiring the summation of gradients from all rules. This leads to mutual constraints and interference between the update directions of different rules for the same attribute weight. This invention employs a rule-specific differentiation method. The weight of the j-th attribute of the k-th rule is calculated and updated independently based solely on the activation level of the k-th rule itself. Error signals from other rules do not affect this weight. Through this decoupled update, the attribute weights of each rule can independently evolve along the optimal direction of their respective error gradients, avoiding gradient interference and mutual cancellation between different rules. This allows each rule to accurately learn the optimal attribute importance configuration for the state corresponding to that rule, significantly improving the model's ability to characterize the dynamic changes in attribute importance with state in complex systems.

[0050] Since the weight of the j-th attribute of the k-th rule is determined solely by the activation level of the k-th rule, this weight parameter always strictly corresponds to the importance of that attribute in the antecedent of the rule in a physical sense, with clear semantics and a well-defined direction. Simultaneously, the attribute weights of each rule can be independently adjusted based on its own error feedback, unaffected by error signals from other rules. This ensures that the rule weight parameters maintain a stable physical meaning during iteration, facilitating domain experts to individually review, understand, and correct each rule after training, further enhancing the model's interpretability and engineering reliability. This maintains the semantic purity and interpretability of the knowledge representation of each rule.

[0051] By inputting each training sample into the model one by one for parameter training, and ending the update cycle uniformly after all samples have been processed, the random perturbation of the convergence direction caused by the sample order during online updates is avoided, thus ensuring the stability of the gradient descent direction.

[0052] To make the technical solution of the present invention clearer, the above specific embodiments may involve several specific numerical values, thresholds, parameters, formulas, and examples. Those skilled in the art should understand that these specific contents are merely exemplary descriptions given for ease of implementation and are not essential technical features necessary for realizing the present invention, nor do they constitute a limitation on the scope of protection of the present invention. The mathematical formulas, parameter values, model structures, calculation rules, constraints, and data acquisition methods described in the specific embodiments section of this application are preferred examples for realizing the technical concept of the present invention, not the only implementation methods. Based on the core logic and inventive concept disclosed in this application, those skilled in the art can make reasonable equivalent substitutions, parameter adjustments, or logical optimizations to the parameters, formulas, and algorithm models according to actual conditions, industry standards, or conventional experimental methods, without producing substantial differences. Any changes that do not depart from the inventive concept of this application fall within the scope of this application. All mathematical expressions in this application should ensure that their physical meaning is clear when applied; parameters such as weights and thresholds can be calibrated using statistical or machine learning methods known in the art.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks, characterized in that, include: S1: Collect sample data of recycled brick aggregate concrete. Each sample contains several input attributes and a true target performance value. Calculate the attribute reliability of each attribute. Construct a confidence rule base and initialize rule weights, attribute weights, and consequent confidence. S2: Construct a rule-specific attribute weight matrix. Based on the attribute reliability and the rule-specific attribute weight matrix, generate attribute importance factors for each rule using an attribute reliability fusion algorithm that is monotonic and constrained. S3: Calculate the activation level and rule activation weight based on the attribute importance factor; S4: Activate weights according to rules, aggregate the consequent confidence of all rules to obtain fusion confidence, and defuzzify to obtain the predicted value vector of brick aggregate impact factor for each sample. S5: Based on the true target performance value vector and the predicted value vector, generate a single-sample prediction error vector and an overall mean square error; compare the mean square error with a preset expected error threshold. If the mean square error is not greater than the expected error threshold, the model converges and the final model is directly output for predicting the performance of recycled brick aggregate concrete.

2. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, The method further includes: S51: When the mean square error is greater than the preset expected error threshold, a backpropagation parameter update mechanism is executed, including: calculating the partial derivatives of the consequent confidence, rule activation weight, rule weight and attribute weight based on the single sample prediction error vector and the overall mean square error, and updating the parameters; Specifically, a specific rule is used to calculate the partial derivatives of the attribute weights: for the j-th attribute weight of the k-th rule, the calculation of its partial derivative is based solely on the derivative of the k-th rule's own activation level with respect to the attribute weight, while the partial derivatives of any other rule besides the k-th rule with respect to the attribute weight are set to zero; thus, the j-th attribute weight of the k-th rule can only affect the activation level of the k-th rule itself. During parameter updates, parameters are trained on all samples one by one. Once the input is complete, the current round of parameter updates ends.

3. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, Constructing the rule-specific attribute weight matrix includes: building a two-dimensional matrix based on the total number of rules in the confidence rule base and the total number of input attributes, and denoting the rule-specific attribute weight matrix as a matrix. Its dimensions are Where L is the total number of rules in the confidence rule base, and M is the total number of input attributes; the matrix elements represent the specificity weight of the j-th attribute in the k-th rule. The initial value is randomly generated within the interval [0.1, 0.9].

4. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, In step S1, the j-th input attribute value of the i-th sample is set to... For the j-th attribute, calculate the mean of all its samples. and standard deviation Based on the preset tolerance coefficient, the reliability interval for each attribute j is defined as follows: ; If the attribute value of a sample falls outside the reliability interval, then the attribute value is considered unreliable. The reliability of the j-th attribute is calculated by counting the number of unreliable samples that fall outside the reliability interval. ;in, It is the number of unreliable samples for the j-th attribute; Let be the total number of samples for the j-th attribute.

5. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, In step S2, the formula for calculating the attribute importance factor is: ;in, It is an attribute importance factor; Let the reliability of the j-th attribute be denoted as . Let J be the rule-specific attribute weight of the j-th attribute in the k-th rule. .

6. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, In step S3, the activation level is calculated using the following formula for the j-th attribute and the k-th rule: ;in, It represents the activation level of the j-th attribute and the k-th rule; Let be the individual matching degree between the j-th attribute and the k-th rule; It is an attribute importance factor; The activation weight calculation formula is: ;in, It is the activation weight of the j-th attribute activating the k-th rule; Let be the rule weight of the k-th rule, with a value range of (0,1). L represents the total number of rules in the confidence rule base.

7. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, In step S4, the consequent confidence scores of all rules are aggregated to obtain the fusion confidence score belonging to the c-th evaluation result, expressed as: ; in, It is the fusion confidence score belonging to the c-th evaluation result; The confidence level of the consequent of the k-th rule belonging to the c-th evaluation result is the confidence level of the consequent. is the activation weight of the j-th attribute activating the k-th rule; L is the total number of rules in the confidence rule base; is the sum of the consequent confidence scores of the k-th rule. If the sum is 1, it means the rule is complete; if it is less than 1, it means the rule is incomplete and there are unknown results. u is the conflict factor, which represents the degree of conflict during fusion.

8. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, In step S4, defuzzification yields the predicted value vector of the brick aggregate impact factor for each sample, which is: the fused confidence distribution is transformed into specific predicted values, and the standard utility values ​​of the three evaluation results are respectively... , , Then the predicted value of the brick aggregate impact factor for the i-th sample is: ;in, This is the predicted value of the impact factor on brick aggregate; This is the standard utility value of the c-th evaluation result; the confidence level of each evaluation result is used as a weight to perform a weighted sum of the utility values ​​of each level; a specific predicted value is obtained for each training sample i. Combine the predicted values ​​of all samples to obtain the predicted value vector. .

9. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, In step S5, the single-sample prediction error is defined as the difference between the true value and the predicted value. The mean squared error over the entire training set is calculated as follows: Where MSE is the mean squared error; It is the single-sample prediction error; the mean squared error is a non-negative value, and the closer it is to 0, the higher the model prediction accuracy.

10. The method for predicting the performance of recycled brick aggregate concrete based on rule-based reasoning networks according to claim 1, characterized in that, The convergence conditions include: the current mean square error is not greater than the expected error threshold; the absolute value of the change in mean square error is less than a preset tolerance value in K consecutive iterations; the number of iterations reaches a preset maximum number; if any convergence condition is met, the iteration stops and all current optimal parameters are saved.

Images