A Method for Predicting the Mix Proportion of Solid Waste Cementitious Materials Based on Prior Knowledge and Reinforcement Learning
By combining prior knowledge and reinforcement learning with weighted similarity retrieval and multi-source information fusion, the stability and accuracy issues of solid waste cementitious material mix ratio prediction were solved, achieving efficient and accurate performance prediction to meet engineering needs at different stages.
Patent Information
- Application Number
- CN202611374983.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-09-07
AI Technical Summary
Existing technologies struggle to achieve efficient, accurate, and stable prediction of solid waste cementitious material proportions. Traditional experimental methods are costly and inefficient, machine learning methods rely on large-scale data and have weak generalization capabilities, and pure large language models lack mechanistic constraints, making it difficult to meet the needs of rapid research and development and iteration.
We employ a method based on prior knowledge and reinforcement learning, using weighted similarity retrieval, inverse distance weighting, multi-source information fusion, multi-temperature gradient sampling, and dynamic experience pool updating to construct a stable prediction model, which is then combined with material mechanism knowledge to predict performance.
It improves the consistency, interpretability, robustness and adaptability of prediction results, reduces the bias and noise impact of model generation, adapts to the prediction needs at different stages, and improves prediction accuracy and generalization ability.
Smart Images

Figure CN122867917B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of interdisciplinary technology of multi-source solid waste cementitious materials and artificial intelligence, and in particular to a method for predicting the proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning. Background Technology
[0002] Currently, the performance prediction of solid waste cementitious materials mainly includes traditional experimental methods based on repeated mixing and mechanical performance testing, machine learning data-driven methods that rely on large-scale labeled data to build prediction models, improved data-driven methods that introduce manually set rule constraints, and preliminary technical paths that directly apply large language models to performance prediction. Each of these methods is supported by experimental verification, data statistical regularity, manual empirical rules, or general language models to achieve the performance prediction of solid waste cementitious materials.
[0003] However, existing technologies have several significant drawbacks: traditional experimental methods rely on extensive trial fittings and mechanical performance testing, which are inherently time-consuming, costly, and inefficient, making it difficult to meet the engineering needs of rapid material research and development and iteration; machine learning data-driven methods heavily rely on large-scale, high-quality labeled datasets, resulting in complex model training processes, high parameter tuning costs, and significantly reduced generalization ability when there are insufficient samples, data distribution deviations, or high noise levels. They are also susceptible to data quality issues such as abnormal distribution of target variables and inconsistent labels, limiting stability and prediction accuracy; improved methods that introduce artificial rule constraints rely excessively on domain experience, and when the data distribution does not match the preset rules, it can actually reduce the model's prediction accuracy and robustness; the technical path of directly applying large language models suffers from problems such as numerical regression drift, prediction results going out of bounds, lack of material mechanism constraints, and difficulty in multi-source information fusion. Overall, it is difficult to achieve the integrated prediction goal of avoiding complex training, high accuracy, strong generalization, and adaptive optimization, and it cannot meet the actual needs of solid waste cementitious material formulation optimization design, efficient utilization of bulk solid waste resources, and engineering applications. Summary of the Invention
[0004] The purpose of this invention is to provide a method for predicting the proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning, thereby solving the above-mentioned technical problems.
[0005] To achieve the above objectives, this invention provides a method for predicting the mix proportions of solid waste cementitious materials based on prior knowledge and reinforcement learning, comprising the following steps: S1. Obtain sample data of cementitious materials from multiple sources of solid waste, and obtain a standardized feature input vector containing 17 dimensions by optimizing the structured feature selection and introducing key component features. S2. Based on the standardized feature input vector, the performance prediction baseline is obtained through weighted similarity retrieval and inverse distance weighted calculation. S3. Based on the standardized feature input vector, high-quality few-shot example samples are obtained through experience pool similarity matching retrieval and sample quantity supplementation and adjustment. S4. Based on the standardized feature input vector, the material mechanism knowledge description text is obtained through targeted retrieval from the material knowledge base and knowledge text transformation; S5. Based on baseline, high-quality few-shot example samples, material mechanism knowledge description text and standardized feature input vector, structured prompts are obtained by splicing and fusing multi-source information; S6. Based on structured prompts, multiple sets of candidate performance prediction results are obtained through multi-gradient temperature parameter control and multiple sampling of large models. S7. Based on multiple sets of candidate performance prediction results and the baseline, the optimal performance prediction result is obtained through relative deviation calculation and weighted comprehensive scoring. S8. Based on the optimal performance prediction results, preset evaluation criteria, and obtain the updated experience pool through dynamic entry threshold adjustment and low-quality sample pruning. S9. Based on the optimal performance prediction results, the final performance prediction values and related statistical information corresponding to the proportion of multi-source solid waste cementitious materials are obtained through result statistical processing and information integration output.
[0006] Preferably, the specific steps of S1 include: S11. Collect raw sample data of coal-based multi-source solid waste cementitious materials, and obtain basic data of raw samples by extracting three core information: raw material composition, proportioning parameters and process conditions. S12. Based on the original sample data, the core feature set is obtained by screening four key performance correlation features: water-cement ratio, alkali equivalent, Si / Al ratio, and Na / Al ratio. S13. Based on the core feature set, feature vectors are constructed by integrating feature dimensions and unifying feature format to obtain the original material feature vector with 17 dimensions; S14. Based on the original material feature vector, the calculation is completed by following the Z-score standardization calculation logic and substituting each original feature parameter to obtain the standardized feature input vector of 17-dimensional features. The Z-score standardization formula is: ; in, For the first The feature values of each original feature; For the first The mean of the original features; For the first The standard deviation corresponding to each original feature.
[0007] Preferably, the multi-source solid waste cementing material is a coal-based multi-source solid waste cementing material, which is composed of a mixture of five types of solid waste components: coal gangue, fly ash, slag, desulfurization gypsum, and carbide slag.
[0008] Preferably, step S2 specifically includes: S21. Based on the standardized feature input vector, by following the weighted Euclidean distance calculation logic and substituting the feature weights and feature vector parameters, the distance value between the sample to be predicted and each sample in the historical sample library is obtained. S22. Based on the distance values between the sample to be predicted and each sample in the historical sample database, sort the distance values in ascending order and select the top samples after sorting. The samples corresponding to the minimum distances are used to obtain the Top-k historical similar sample set; S23. Based on the Top-k historical similar sample set, the performance prediction baseline is obtained by following the inverse distance weighted calculation logic and substituting the sample distance and the actual performance value parameters. The weighted Euclidean distance formula is as follows: ; in, For the sample to be predicted and the historical sample database, the first... The weighted Euclidean distance between samples; For feature weights; The standardized feature vector of the sample to be predicted; For the first in the historical sample database Standardized feature vectors of each sample; For the first in the historical sample database The first sample One standardized feature value; For the prediction sample The One standardized feature value; The inverse distance weighted formula is: ; in, Based on the sample to be predicted The obtained performance prediction baseline; This represents the number of Top-k historical similar samples selected. For the sample to be predicted and the first Weighted Euclidean distance between historically similar samples; For the first The true performance values of historically similar samples; For the sample to be predicted and the first The weighted Euclidean distance between historically similar samples, and .
[0009] Preferably, the specific steps of S3 include: S31. Based on a pre-built experience pool, the experience pool is arranged according to... The format stores the feature vectors, prediction results, and evaluation scores, among which... For the prediction results of the sample, To obtain the evaluation score corresponding to the prediction result, and to retrieve similar samples that match the standardized feature input vector, an initial high-quality candidate sample set is obtained; S32. Based on the initial high-quality candidate sample set, the result of whether the number of samples is sufficient or insufficient is obtained by verifying whether the number of samples meets the preset few-shot example number requirement. S33. Based on the judgment result of insufficient sample quantity, the Top-k historical similar samples of S2 are introduced to supplement the sample quantity and obtain a high-quality candidate sample set with sufficient quantity. S34. Based on the high-quality candidate sample set that meets the quantity requirements, the sample evaluation scores are verified twice and low-quality samples are removed to obtain high-quality few-shot example samples.
[0010] Preferably, step S4 specifically includes: S41. Based on the standardized feature input vector, an initial set of mechanism knowledge is obtained by retrieval of mechanism information, including reactivity law, hydration mechanism, and strength development trend, from the material knowledge base. S42. Based on the initial set of mechanism knowledge, by converting professional mechanism terms into natural language and simplifying redundant expressions, the material mechanism knowledge description text is obtained.
[0011] Preferably, the specific steps of S5 include: S51. Based on the baseline, extract baseline values, label baseline source information and organize the content to obtain the baseline prediction value and source description text; S52. Based on high-quality few-shot sample texts, standard few-shot sample texts are obtained by unifying sample formats and standardizing sample content layout. S53. Based on baseline predicted values and source description text, standard few-shot example text, material mechanism knowledge description text, and standardized feature input vector, structured prompts are obtained by following structured splicing logic and orderly integrating the four types of information. The structured suggestion concatenation formula is as follows: ; in, Provide structured hints; This is a function for structured splicing operations; This is a collection of high-quality few-shot example samples selected from the experience pool; Based on the sample to be predicted The retrieved text describing the material mechanism.
[0012] Preferably, the specific steps of S6 include: S61. Preset a set of gradient temperature parameters, and obtain the single-sample temperature parameters by retrieving the temperature parameters in the set one by one; S62. Based on structured prompts and single-sampled temperature parameters, a generation operation is performed by inputting a large language model to obtain a single set of candidate performance prediction results; S63. Based on multiple sets of single-sample temperature parameters, multiple sets of candidate performance prediction results are obtained by repeatedly performing generation operations and summarizing all generation results. The formula for generating multiple temperature samples is: ; in, For the first Candidate performance prediction results generated from the second sampling; For large language models; For the first The temperature parameters were sampled once.
[0013] Preferably, the specific steps of S7 include: S71. Based on multiple sets of candidate performance prediction results and the baseline, by following the relative deviation calculation logic and substituting the candidate results and baseline values, the relative deviation values of each candidate result for each performance index are obtained. The formula for relative deviation is: ; in, For the candidate prediction results in the 1st Predicted values for each performance metric; This is a relative deviation; S72. Based on the relative deviation value, and with preset index weight coefficients and error tolerance thresholds, the comprehensive evaluation score of each candidate result is obtained by following the weighted comprehensive scoring calculation logic. The weighted comprehensive scoring formula is as follows: ; in, The comprehensive evaluation score of the candidate prediction results; For the first Weighting coefficients for each performance indicator; This represents the error tolerance threshold for the corresponding indicator. S73. Based on the comprehensive evaluation score of each candidate result, sort the scores in descending order and select the candidate result corresponding to the highest score to obtain the optimal performance prediction result.
[0014] Preferably, the specific steps of S8 include: S81. Based on the optimal performance prediction results, extract the comprehensive evaluation score and compare it with the preset dynamic entry threshold to obtain the entry qualification result. S82. Based on the qualified entry judgment result, the standardized feature input vector, the optimal performance prediction result and the corresponding comprehensive evaluation score of the sample are integrated and written into the experience pool to obtain the experience pool after the addition of the sample. The experience pool update formula is as follows: ; in, For the updated experience pool; This is the experience pool before the update; The standardized feature input vector is the sample to be added to the experience pool; The optimal performance prediction result for the samples to be added to the experience pool; The overall evaluation score for samples to be added to the experience pool; The entry threshold is dynamic; S83. Based on the experience pool after the addition of new samples, the updated experience pool is obtained by periodically screening and deleting samples whose comprehensive evaluation scores are lower than the preset standard.
[0015] Therefore, the present invention employs the above-mentioned method for predicting the proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning, which has the following beneficial effects: 1. By introducing a weighted Euclidean distance similarity retrieval and an inverse distance weighted baseline construction mechanism, different feature weights can be assigned according to the degree of influence of material components on performance. This allows for accurate matching of historical samples with similar key components, weakens the interference of outliers on prediction results, provides stable and reliable reference anchors for subsequent large model predictions, effectively constrains the bias of model generation, reduces performance fluctuations caused by random sampling of pure large models, and improves the consistency and interpretability of prediction results.
[0016] 2. By constructing an experience pool similarity matching and dynamic supplementation of Few-Shot examples, it is possible to prioritize the retrieval of examples that highly match the current sample feature distribution from the validated high-quality prediction samples. When there are insufficient samples, historical similar samples are introduced to ensure that the examples fit the ratio-performance correlation law of solid waste cementitious materials, improve the learning efficiency of the large model on the material performance prediction logic, reduce unfounded inferences in the model generation process, and enhance the consistency and adaptability of the prediction results.
[0017] 3. Based on the multi-source information fusion structured prompting mechanism, the performance prediction baseline, high-quality Few-Shot examples, material mechanism knowledge and sample characteristics can be integrated in an orderly manner. This allows the large model to take into account both data statistical regularities and material science mechanisms when generating predictions, avoiding the "black box" type of erroneous predictions of pure data-driven models. It ensures that the prediction results conform to the basic physicochemical laws such as hydration reaction and activity activation of solid waste cementitious materials. At the same time, the structured format also improves the stability of model generation and reduces the probability of irrelevant outputs or format chaos.
[0018] 4. By adopting a multi-temperature gradient sampling and weighted comprehensive scoring screening mechanism, multiple sets of diverse candidate prediction results can be generated through different temperature parameters. At the same time, a comprehensive evaluation system is constructed based on the relative deviation from the baseline to screen out the optimal result that not only conforms to the data distribution pattern but also has controllable error. This not only utilizes the diversity advantage generated by the large model but also effectively reduces the random error of a single sampling through deviation constraints, thereby improving the robustness and reliability of the prediction results.
[0019] 5. By introducing a dynamic entry threshold and a periodic pruning experience pool update mechanism, effective samples can be quickly accumulated with a low threshold in the early stage of system operation. Subsequently, the entry standard is gradually increased as the prediction accuracy improves. At the same time, low-quality samples are cleaned up regularly to avoid the experience pool being polluted by noisy data. This also continuously improves the prediction accuracy and generalization ability of the system, adapting to the prediction needs at different stages.
[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0021] Figure 1 A flowchart of the solid waste cementitious material mix ratio prediction method based on prior knowledge and reinforcement learning provided by the present invention; Figure 2 This is a schematic diagram of the core architecture of the corresponding training stage of the solid waste cementitious material proportion prediction method based on prior knowledge and reinforcement learning provided by the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.
[0023] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.
[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0025] Existing technologies for predicting the performance of solid waste cementitious materials have significant shortcomings: traditional experimental methods are time-consuming, costly, and inefficient; machine learning methods rely on large-scale, high-quality data and have weak generalization ability; artificial rule constraints have poor adaptability; and direct application of large language models is prone to numerical drift and lacks mechanistic constraints, all of which make it difficult to achieve efficient, accurate, and stable predictions.
[0026] Based on the above analysis, this invention is designed. (See appendix.) Figures 1-2 A method for predicting the mix proportions of solid waste cementitious materials based on prior knowledge and reinforcement learning includes the following steps: S1. Obtain sample data of multi-source solid waste cementitious materials, and obtain a standardized feature input vector containing 17 dimensions by optimizing structured feature screening and introducing key component features.
[0027] The specific steps of S1 include: S11. Collect raw sample data of coal-based multi-source solid waste cementitious materials, and obtain basic data of raw samples by extracting three core information: raw material composition, proportioning parameters and process conditions. S12. Based on the original sample data, the core feature set is obtained by screening four key performance correlation features: water-cement ratio, alkali equivalent, Si / Al ratio, and Na / Al ratio. S13. Based on the core feature set, feature vectors are constructed by integrating feature dimensions and unifying feature format to obtain the original material feature vector with 17 dimensions; S14. Based on the original material feature vector, the calculation is completed by following the Z-score standardization calculation logic and substituting each original feature parameter to obtain the standardized feature input vector of 17-dimensional features. The Z-score standardization formula is: ; in, For the first The feature values of each original feature; For the first The mean of the original features; For the first The standard deviation corresponding to each original feature.
[0028] This process ensures that the distribution has a mean of 0 and a standard deviation of 1, thereby eliminating the influence of different units on subsequent calculations and providing a basis for subsequent similarity calculations.
[0029] Among them, the multi-source solid waste cementitious material is a coal-based multi-source solid waste cementitious material, which is composed of five types of solid waste components: coal gangue, fly ash, slag, desulfurization gypsum, and carbide slag.
[0030] S2. Based on the standardized feature input vector, the performance prediction baseline is obtained through weighted similarity retrieval and inverse distance weighted calculation.
[0031] The specific steps in S2 include: S21. Based on the standardized feature input vector, by following the weighted Euclidean distance calculation logic and substituting the feature weights and feature vector parameters, the distance value between the sample to be predicted and each sample in the historical sample library is obtained. S22. Based on the distance values between the sample to be predicted and each sample in the historical sample database, sort the distance values in ascending order and select the top samples after sorting. The samples corresponding to the minimum distances are used to obtain the Top-k historical similar sample set; S23. Based on the Top-k historical similar sample set, the performance prediction baseline is obtained by following the inverse distance weighted calculation logic and substituting the sample distance and the actual performance value parameters. The weighted Euclidean distance formula is as follows: ; in, For the sample to be predicted and the historical sample database, the first... The weighted Euclidean distance between samples; For feature weights; The standardized feature vector of the sample to be predicted; For the first in the historical sample database Standardized feature vectors of each sample; For the first in the historical sample database The first sample One standardized feature value; For the prediction sample The One standardized feature value; Specifically, higher weights are given to the water-cement ratio and the content of alkali activator to highlight their dominant role in material performance.
[0032] The inverse distance weighted formula is: ; in, Based on the sample to be predicted The obtained performance prediction baseline; This represents the number of Top-k historical similar samples selected. For the sample to be predicted and the first Weighted Euclidean distance between historically similar samples; For the first The true performance values of historically similar samples; For the sample to be predicted and the first The weighted Euclidean distance between historically similar samples, and .
[0033] Samples that are closer in distance contribute more to the baseline, which is used to constrain the output range of large language models and avoid results that deviate significantly from engineering experience.
[0034] S3. Based on the standardized feature input vector, high-quality few-shot example samples are obtained through experience pool similarity matching retrieval and sample quantity supplementation and adjustment.
[0035] The specific steps of S3 include: S31. Based on a pre-built experience pool, the experience pool is arranged according to... The format stores the feature vectors, prediction results, and evaluation scores, among which... For the prediction results of the sample, To obtain the evaluation score corresponding to the prediction result, and to retrieve similar samples that match the standardized feature input vector, an initial high-quality candidate sample set is obtained; S32. Based on the initial high-quality candidate sample set, the result of whether the number of samples is sufficient or insufficient is obtained by verifying whether the number of samples meets the preset few-shot example number requirement. S33. Based on the judgment result of insufficient sample quantity, the Top-k historical similar samples of S2 are introduced to supplement the sample quantity and obtain a high-quality candidate sample set with sufficient quantity. S34. Based on the high-quality candidate sample set that meets the quantity requirements, the sample evaluation scores are verified twice and low-quality samples are removed to obtain high-quality few-shot example samples.
[0036] This allows for the reuse of high-quality reasoning experience, thereby improving the predictive stability of the model under complex formulation conditions.
[0037] S4. Based on the standardized feature input vector, the material mechanism knowledge description text is obtained through targeted retrieval from the material knowledge base and knowledge text transformation.
[0038] The steps in S4 specifically include: S41. Based on the standardized feature input vector, the system obtains an initial set of mechanism knowledge by retrieval of mechanism information in the material knowledge base, including reactivity rules, hydration mechanisms, and strength development trends. Specifically, when the water-cement ratio is high, the system retrieves knowledge such as "increased water-cement ratio leads to increased porosity and decreased strength"; when the alkali activator content is high, the system retrieves relevant mechanism descriptions of "enhanced alkali activation reaction". S42. Based on the initial set of mechanism knowledge, by converting professional mechanism terms into natural language and simplifying redundant expressions, the material mechanism knowledge description text is obtained.
[0039] S5. Based on baseline, high-quality few-shot example samples, material mechanism knowledge description text, and standardized feature input vectors, structured prompts are obtained by splicing and fusing multi-source information.
[0040] The specific steps of S5 include: S51. Based on the baseline, extract baseline values, label baseline source information and organize the content to obtain the baseline prediction value and source description text; S52. Based on high-quality few-shot sample texts, standard few-shot sample texts are obtained by unifying sample formats and standardizing sample content layout. S53. Based on baseline predicted values and source description text, standard few-shot example text, material mechanism knowledge description text, and standardized feature input vector, structured prompts are obtained by following structured splicing logic and orderly integrating the four types of information. The structured suggestion concatenation formula is as follows: ; in, Provide structured hints; This is a function for structured splicing operations; This is a collection of high-quality few-shot example samples selected from the experience pool; Based on the sample to be predicted The retrieved text describing the material mechanism.
[0041] Specifically, the structured prompts include: Baseline predicted values and their sources; several high-quality historical prediction examples; material mechanism knowledge related to the current formulation; characteristic information of the current sample.
[0042] By combining information inputs, multi-source constraints are provided to large language models, enabling their prediction results to simultaneously satisfy data consistency and mechanistic rationality.
[0043] S6. Based on structured prompts, multiple sets of candidate performance prediction results are obtained through multi-gradient temperature parameter control and multiple sampling of a large model.
[0044] The specific steps of S6 include: S61. Preset a set of gradient temperature parameters, and obtain the single-sample temperature parameters by retrieving the temperature parameters in the set one by one; S62. Based on structured prompts and single-sampled temperature parameters, a generation operation is performed by inputting a large language model to obtain a single set of candidate performance prediction results; S63. Based on multiple sets of single-sample temperature parameters, multiple sets of candidate performance prediction results are obtained by repeatedly performing generation operations and summarizing all generation results. The formula for generating multiple temperature samples is: ; in, For the first Candidate performance prediction results generated from the second sampling; For large language models; For the first The sampling temperature parameters are 0.3, 0.5, 0.7, 0.9 and 1.0. The results generated by low temperature sampling are more stable and closer to historical experience; high temperature sampling has stronger exploratory capabilities and can discover potential optimization space. Through the multi-temperature sampling mechanism, a set of candidate prediction results with both stability and diversity can be obtained.
[0045] S7. Based on multiple sets of candidate performance prediction results and the baseline, the optimal performance prediction result is obtained through relative deviation calculation and weighted comprehensive scoring.
[0046] The specific steps of S7 include: S71. Based on multiple sets of candidate performance prediction results and the baseline, by following the relative deviation calculation logic and substituting the candidate results and baseline values, the relative deviation values of each candidate result for each performance index are obtained. The formula for relative deviation is: ; in, For the candidate prediction results in the 1st Predicted values for each performance metric; This is a relative deviation; S72. Based on the relative deviation value, and with preset index weight coefficients and error tolerance thresholds, the comprehensive evaluation score of each candidate result is obtained by following the weighted comprehensive scoring calculation logic. The weighted comprehensive scoring formula is as follows: ; in, The comprehensive evaluation score of the candidate prediction results; For the first Weighting coefficients for each performance indicator; This represents the error tolerance threshold for the corresponding indicator. When the deviation of a certain indicator exceeds the threshold, its corresponding score is recorded as 0, thereby suppressing prediction results that deviate significantly from engineering experience.
[0047] S73. Based on the comprehensive evaluation score of each candidate result, sort the scores in descending order and select the candidate result corresponding to the highest score to obtain the optimal performance prediction result.
[0048] S8. Based on the optimal performance prediction results, preset evaluation criteria are used to obtain an updated experience pool through dynamic entry threshold adjustment and low-quality sample pruning.
[0049] The specific steps of S8 include: S81. Based on the optimal performance prediction results, extract the comprehensive evaluation score and compare it with the preset dynamic entry threshold to obtain the entry qualification result. S82. Based on the qualified entry judgment result, the standardized feature input vector, the optimal performance prediction result and the corresponding comprehensive evaluation score of the sample are integrated and written into the experience pool to obtain the experience pool after the addition of the sample. The experience pool update formula is as follows: ; in, For the updated experience pool; This is the experience pool before the update; The standardized feature input vector is the sample to be added to the experience pool; The optimal performance prediction result for the samples to be added to the experience pool; The overall evaluation score for samples to be added to the experience pool; The entry threshold is dynamic; S83. Based on the experience pool after the addition of new samples, the updated experience pool is obtained by periodically screening and deleting samples whose comprehensive evaluation scores are lower than the preset standard.
[0050] S9. Based on the optimal performance prediction results, the final performance prediction values and related statistical information corresponding to the proportion of multi-source solid waste cementitious materials are obtained through result statistical processing and information integration output, which are used to assist engineering decision-making.
[0051] In summary, this invention effectively solves the core problems in the prediction of solid waste cementitious materials performance: long experimental cycles, high costs, and low efficiency in traditional methods; machine learning relies on large-scale, high-quality data and has weak generalization ability; and pure large language models are prone to numerical drift and lack mechanistic constraints. Through historical sample baseline constraints, experience pool enhancement, material knowledge fusion, multi-temperature sampling, and result evaluation optimization, it significantly improves prediction stability, accuracy, and generalization ability, suppresses numerical drift, and aligns with material mechanism laws. Table 1, through a horizontal comparison of multiple indicators, quantitatively verifies the comprehensive advantages of this invention compared to traditional machine learning models, intuitively demonstrating the significant improvement in prediction stability and accuracy. See Table 1 for details. Table 1 Comparison between machine learning and this method:
[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for predicting the mix proportions of solid waste cementitious materials based on prior knowledge and reinforcement learning, characterized in that, The specific steps include: S1. Obtain sample data of cementitious materials from multiple sources of solid waste, and obtain a standardized feature input vector containing 17 dimensions by optimizing the structured feature selection and introducing key component features. S2. Based on the standardized feature input vector, the performance prediction baseline is obtained through weighted similarity retrieval and inverse distance weighted calculation. S3. Based on the standardized feature input vector, high-quality few-shot example samples are obtained through experience pool similarity matching retrieval and sample quantity supplementation and adjustment. S4. Based on the standardized feature input vector, the material mechanism knowledge description text is obtained through targeted retrieval from the material knowledge base and knowledge text transformation; S5. Based on baseline, high-quality few-shot example samples, material mechanism knowledge description text and standardized feature input vector, structured prompts are obtained by splicing and fusing multi-source information; S6. Based on structured prompts, multiple sets of candidate performance prediction results are obtained through multi-gradient temperature parameter control and multiple sampling of large models. S7. Based on multiple sets of candidate performance prediction results and the baseline, the optimal performance prediction result is obtained through relative deviation calculation and weighted comprehensive scoring; the specific steps of S7 include: S71. Based on multiple sets of candidate performance prediction results and the baseline, by following the relative deviation calculation logic and substituting the candidate results and baseline values, the relative deviation values of each candidate result for each performance index are obtained. The formula for relative deviation is: ; in, For the candidate prediction results in the 1st Predicted values for each performance metric; This is a relative deviation; S72. Based on the relative deviation value, and with preset index weight coefficients and error tolerance thresholds, the comprehensive evaluation score of each candidate result is obtained by following the weighted comprehensive scoring calculation logic. The weighted comprehensive scoring formula is as follows: ; in, The comprehensive evaluation score of the candidate prediction results; For the first Weighting coefficients for each performance indicator; This represents the error tolerance threshold for the corresponding indicator. S73. Based on the comprehensive evaluation score of each candidate result, sort the scores in descending order and select the candidate result corresponding to the highest score to obtain the optimal performance prediction result; S8. Based on the optimal performance prediction result, preset evaluation criteria, and obtain the updated experience pool through dynamic pool entry threshold adjustment and low-quality sample pruning. S9. Based on the optimal performance prediction results, the final performance prediction values and related statistical information corresponding to the proportion of multi-source solid waste cementitious materials are obtained through result statistical processing and information integration output.
2. The method for predicting the mix proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning according to claim 1, characterized in that: The specific steps of S1 include: S11. Collect raw sample data of coal-based multi-source solid waste cementitious materials, and obtain the basic data of the raw samples by extracting three core information: raw material composition, proportioning parameters and process conditions. S12. Based on the original sample data, the core feature set is obtained by screening four key performance correlation features: water-cement ratio, alkali equivalent, Si / Al ratio, and Na / Al ratio. S13. Based on the core feature set, feature vectors are constructed by integrating feature dimensions and unifying feature format to obtain the original material feature vector with 17 dimensions. S14. Based on the original material feature vector, the calculation is completed by following the Z-score standardization calculation logic and substituting each original feature parameter to obtain the standardized feature input vector of 17-dimensional features. The Z-score standardization formula is: ; in, For the first The feature values of each original feature; For the first The mean of the original features; For the first The standard deviation corresponding to each original feature.
3. The method for predicting the mix proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning according to claim 2, characterized in that: The multi-source solid waste cementitious material is a coal-based multi-source solid waste cementitious material, which is composed of five types of solid waste components: coal gangue, fly ash, slag, desulfurization gypsum, and carbide slag.
4. The method for predicting the mix proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning according to claim 3, characterized in that: The specific steps in S2 include: S21. Based on the standardized feature input vector, by following the weighted Euclidean distance calculation logic and substituting the feature weights and feature vector parameters, the distance value between the sample to be predicted and each sample in the historical sample library is obtained. S22. Based on the distance values between the sample to be predicted and each sample in the historical sample database, sort the distance values in ascending order and select the top samples after sorting. The samples corresponding to the minimum distances are used to obtain the Top-k historical similar sample set; S23. Based on the Top-k historical similar sample set, the performance prediction baseline is obtained by following the inverse distance weighted calculation logic and substituting the sample distance and the actual performance value parameters. The weighted Euclidean distance formula is as follows: ; in, For the sample to be predicted and the historical sample database, the first... The weighted Euclidean distance between samples; For feature weights; The standardized feature vector of the sample to be predicted; For the first in the historical sample database Standardized feature vectors of each sample; For the first in the historical sample database The first sample One standardized feature value; For the prediction sample The One standardized feature value; The inverse distance weighted formula is: ; in, Based on the sample to be predicted The obtained performance prediction baseline; This represents the number of Top-k historical similar samples selected. For the sample to be predicted and the first Weighted Euclidean distance between historically similar samples; For the first The true performance values of historically similar samples; For the sample to be predicted and the first The weighted Euclidean distance between historically similar samples, and .
5. The method for predicting the mix proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning according to claim 4, characterized in that: The specific steps of S3 include: S31. Based on a pre-built experience pool, the experience pool is arranged according to... The format stores the feature vectors, prediction results, and evaluation scores, among which... For the prediction results of the sample, To obtain the evaluation score corresponding to the prediction result, and to retrieve similar samples that match the standardized feature input vector, an initial high-quality candidate sample set is obtained; S32. Based on the initial high-quality candidate sample set, the result of whether the number of samples is sufficient or insufficient is obtained by verifying whether the number of samples meets the preset few-shot example number requirement. S33. Based on the judgment result of insufficient sample quantity, the Top-k historical similar samples of S2 are introduced to supplement the sample quantity and obtain a high-quality candidate sample set with sufficient quantity. S34. Based on the high-quality candidate sample set that meets the quantity requirements, the sample evaluation scores are verified twice and low-quality samples are removed to obtain high-quality few-shot example samples.
6. The method for predicting the mix proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning according to claim 5, characterized in that: The steps in S4 specifically include: S41. Based on the standardized feature input vector, an initial set of mechanism knowledge is obtained by retrieval of mechanism information, including reactivity law, hydration mechanism, and strength development trend, from the material knowledge base. S42. Based on the initial set of mechanism knowledge, by converting professional mechanism terms into natural language and simplifying redundant expressions, the material mechanism knowledge description text is obtained.
7. The method for predicting the mix proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning according to claim 6, characterized in that: The specific steps of S5 include: S51. Based on the baseline, extract baseline values, label baseline source information and organize the content to obtain the baseline prediction value and source description text; S52. Based on high-quality few-shot sample texts, standard few-shot sample texts are obtained by unifying sample formats and standardizing sample content layout. S53. Based on baseline predicted values and source description text, standard few-shot example text, material mechanism knowledge description text, and standardized feature input vector, structured prompts are obtained by following structured splicing logic and orderly integrating the four types of information. The structured suggestion concatenation formula is as follows: ; in, For structured hints; This is a function for structured splicing operations; This is a collection of high-quality few-shot example samples selected from the experience pool; Based on the sample to be predicted The retrieved text describing the material mechanism.
8. The method for predicting the mix proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning according to claim 7, characterized in that: The specific steps of S6 include: S61. Preset a set of gradient temperature parameters, and obtain the single-sample temperature parameters by retrieving the temperature parameters in the set one by one; S62. Based on structured prompts and single-sampled temperature parameters, a generation operation is performed by inputting a large language model to obtain a single set of candidate performance prediction results; S63. Based on multiple sets of single-sample temperature parameters, multiple sets of candidate performance prediction results are obtained by repeatedly performing generation operations and summarizing all generation results. The formula for generating multiple temperature samples is: ; in, For the first Candidate performance prediction results generated from the second sampling; For large language models; For the first The temperature parameters were sampled once.
9. The method for predicting the mix proportion of solid waste cementitious materials based on prior knowledge and reinforcement learning according to claim 8, characterized in that: The specific steps of S8 include: S81. Based on the optimal performance prediction results, extract the comprehensive evaluation score and compare it with the preset dynamic entry threshold to obtain the entry qualification result. S82. Based on the qualified entry judgment result, the standardized feature input vector, the optimal performance prediction result and the corresponding comprehensive evaluation score of the sample are integrated and written into the experience pool to obtain the experience pool after the addition of the sample. The experience pool update formula is as follows: ; in, For the updated experience pool; This is the experience pool before the update; The standardized feature input vector is the sample to be added to the experience pool; The optimal performance prediction result for the samples to be added to the experience pool; The overall evaluation score for samples to be added to the experience pool; The entry threshold is dynamic; S83. Based on the experience pool after the addition of new samples, the updated experience pool is obtained by periodically screening and deleting samples whose comprehensive evaluation scores are lower than the preset standard.
Citation Information
Patent Citations
Polymer composite material design method based on LLM knowledge database construction and reasoning optimization
CN121148531A
New material performance prediction method and system based on artificial intelligence
CN121964013A
