A method and system for self-adaptive optimization of phosphogypsum concrete mix proportion

Abstract

This invention belongs to the field of concrete mix design technology, specifically relating to an adaptive optimization method and system for phosphogypsum concrete mix proportions. The method includes: constructing an identification framework and inputting mix proportion parameters into a prediction model group; generating a basic quality function based on a membership function weighted by a dual-factor weighted method of error dispersion and historical hit rate; calculating the distance and probability conflict coefficient between quality functions to construct a conflict matrix, extracting eigenvalues ​​to construct a consistency index, and adjusting the weights based on the hit rate; when the conflict measure exceeds a dynamic threshold, performing non-uniform quality redistribution based on the focal element cardinality and credibility; calculating the joint confidence degree and constructing a multi-objective fitness function with weights; running an optimization algorithm, dynamically perturbing the variation parameters using the characteristic variables of the conflict matrix, shrinking the search step size according to the uncertainty index, and outputting the optimal mix proportion. This invention achieves the synergy of uncertainty reasoning and swarm intelligence algorithms, accurately capturing solutions that balance performance and cost.

Description

Technical Field

[0001] This invention relates to the field of concrete mix design technology. More specifically, this invention relates to an adaptive optimization method and system for phosphogypsum concrete mix proportions. Background Technology

[0002] Phosphogypsum is a major industrial byproduct of the phosphate chemical industry, and using it as an admixture in concrete preparation is one of the core approaches to achieving large-scale resource utilization of phosphogypsum. Concrete mix design is a crucial step in determining the final mechanical properties, workability, and economic cost of concrete. Essentially, it involves finding the optimal combination of parameters that satisfy multiple objective constraints within a multi-dimensional parameter space, including phosphogypsum dosage, water-cement ratio, sand ratio, and admixture dosage. With the deep application of artificial intelligence technology in civil engineering, combining machine learning prediction models with uncertainty reasoning methods and heuristic optimization algorithms has become an important development direction for intelligent design and decision-making in concrete engineering.

[0003] However, phosphogypsum commonly contains harmful impurities such as soluble phosphorus and fluorine, which can interfere with the cement hydration process and alter the setting time and subsequent mechanical property development of concrete. This results in a more complex and nonlinear relationship between the performance of phosphogypsum concrete and its mix proportion parameters compared to ordinary concrete. Furthermore, there is a natural interdependence between the compressive strength, fluidity, and economic cost of concrete. Simply pursuing high strength leads to increased use of cementitious materials and admixtures, thus increasing project costs; conversely, excessive cost reduction may sacrifice the workability or mechanical properties of concrete, affecting project quality and safety. This places more stringent demands on the predictive accuracy, multi-source conflict information processing capabilities, and global optimization efficiency of mix proportion optimization methods.

[0004] Chinese patent document CN116595621A discloses a method and system for fusion diagnosis of concrete dam deformation. This scheme constructs a framework for identifying the anomaly levels of concrete dam deformation, directly maps the outputs of multiple monitoring models to basic probability assignments to generate a fundamental quality function, uses a fixed threshold to determine the degree of evidence conflict and performs evidence fusion, and finally achieves anomaly diagnosis of concrete dam deformation based on the fusion results.

[0005] This existing technology directly maps the prediction results of a single model to generate a basic quality function, without fully considering the dispersion of prediction errors and historical prediction hit rates of different models. This generation method makes the initial basic probability assignment lack a stable and objective basis for credibility. When the model prediction fluctuates significantly, it directly leads to a decrease in the reliability of the evidence source itself, thus affecting the accuracy of the final fusion result. At the same time, this method uses a fixed threshold to determine the degree of evidence conflict, which cannot accurately characterize the spatial distance and probability distribution differences between evidence. When there is a strong conflict in the multi-source prediction results, this fixed judgment method will produce a fusion result that deviates significantly from the actual situation, thus misleading the subsequent decision-making process. In addition, this method is only designed for the scenario of concrete dam deformation diagnosis, and does not construct a multi-objective identification framework and a dedicated evidence processing mechanism for the characteristics of phosphogypsum concrete mix design. It also does not deeply integrate the evidence fusion result with heuristic optimization algorithms, and cannot effectively solve the multi-objective conflict and global optimization problem in phosphogypsum concrete mix design. The above defects together make it difficult for the existing technology to obtain the optimal phosphogypsum concrete mix design that takes into account mechanical performance, construction performance and economy in a short time, and cannot meet the actual engineering needs for efficient and accurate mix design. Summary of the Invention

[0006] To address the technical problems of difficulty in predicting reliability assessment, inflexible evidence conflict handling mechanisms, and low efficiency in collaborative optimization of multiple performance indicators in existing phosphogypsum concrete mix design, this invention provides an adaptive optimization method and system for phosphogypsum concrete mix proportions.

[0007] In a first aspect, the present invention provides an adaptive optimization method for phosphogypsum concrete mix proportions, comprising: S1: constructing an identification framework with performance grades of compressive strength, fluidity, and cost as the hypothesis set, and collecting mix proportion parameters as inputs to the mix proportion prediction model group; using a membership function based on a dual-factor weighted factor of error dispersion and historical hit rate to map the outputs of each prediction model to generate a basic quality function; S2: calculating the distance and probability conflict coefficient between the basic quality functions to obtain a comprehensive conflict measure between pairwise evidence to construct a conflict matrix, and extracting the off-diagonal elements of the matrix to calculate the mean value to obtain the global scalar of the current comprehensive conflict measure, and extracting the eigenvalues ​​of the conflict matrix to construct a consistency index and adjust the hit rate weight; based on historical comprehensive... The system generates a dynamic threshold for the combined conflict measure. When the global scalar of the current combined conflict measure exceeds the dynamic threshold corrected by the consistency index, the weights are redistributed based on the focal element cardinality and evidence credibility. The conflict quality is then non-uniformly redistributed, and the remaining conflict quality is assigned to the entire set. S3: The quality functions are fused. The fused quality functions are used to calculate the joint confidence level within each identification framework. This confidence level is then converted into a reverse penalty weight to construct a global fitness function using a weighted multi-objective function. An optimization algorithm is run to iteratively search for the fit ratio. The feature variables of the conflict matrix are extracted during the iteration, and the mutation parameters are dynamically adjusted to perturb the particles. The search step size is dynamically reduced based on the uncertainty index of the probability distribution. After obtaining the optimal fit ratio, the historical hit rate is updated.

[0008] By adopting the above technical solutions, this invention overcomes the model prediction instability under complex hydration environments of phosphogypsum by constructing a multi-index identification framework and introducing a two-factor weighting mechanism based on error and hit rate; by constructing a conflict matrix and combining it with a dynamic threshold to trigger non-uniform quality redistribution in real time, it effectively filters out high-conflict redundancy in multi-source information fusion and improves the logical stability under extremely complex environments; finally, by deeply feeding the underlying conflict features and fusion uncertainty index into the heuristic optimization algorithm, it achieves adaptive control of algorithm mutation perturbation and search step size, thereby accurately capturing the optimal mix ratio scheme that takes into account mechanical performance, construction performance and economic cost in a multi-dimensional parameter space, which has extremely high engineering application value.

[0009] Preferably, the distance between the basic quality functions is the Jousselme distance, and the probability conflict coefficient is the Pignistic probability conflict coefficient. The calculation of the distance between the basic quality functions and the probability conflict coefficient to construct a conflict matrix, and the calculation of the current comprehensive conflict measure, includes: obtaining the Jousselme distance between the i-th and j-th pieces of evidence; calculating the cosine similarity between the Pignistic probability distribution vectors of the i-th and j-th pieces of evidence, and subtracting the cosine similarity from 1 to obtain the Pignistic probability conflict coefficient; taking the square root of the product of the Jousselme distance and the Pignistic probability conflict coefficient as the comprehensive conflict measure between each pair of pieces of evidence, filling it into the corresponding position of the symmetric matrix, and constructing the conflict matrix; extracting the set of off-diagonal elements of the conflict matrix, and calculating its arithmetic mean as the global scalar of the current comprehensive conflict measure.

[0010] Preferably, the step of extracting the eigenvalues ​​of the conflict matrix to construct a consistency index and adjust the hit rate weight includes: performing eigenvalue decomposition on the conflict matrix, extracting the first eigenvalue with the largest absolute value and the second eigenvalue with the second largest absolute value; using the ratio of the absolute value of the second eigenvalue to the absolute value of the first eigenvalue as a conflict dispersion index, and subtracting this index from 1 to obtain the consistency index; taking the absolute value of the eigenvector corresponding to the first eigenvalue, normalizing it using the L1 norm as the conflict participation degree of each piece of evidence, and performing inverse normalization on the conflict participation degree to obtain the credibility weight, thereby adjusting the hit rate.

[0011] Preferably, the step of calculating the redistribution weight based on the focal element cardinality and evidence credibility, performing non-uniform redistribution of conflict quality, and assigning the remaining conflict quality to the entire set includes: when the intersection of the focal elements that cause conflict is an empty set, extracting the focal elements involved in the conflict, the union of the focal elements involved in the conflict, and their internal non-empty subsets as redistribution objects; taking the inverse of the focal element cardinality of each redistribution object, multiplying it by the mean of the evidence credibility that causes conflict, and normalizing it to obtain the redistribution weight; splitting the total conflict quality according to a set retention coefficient: multiplying the total conflict quality by 1 and subtracting the retention coefficient, then multiplying the difference by the redistribution weight, and adding it to the basic probability assignment of each non-empty subset; assigning the remaining part obtained by multiplying the total conflict quality by the retention coefficient to the entire set of focal elements of the identification frame to ensure that the total probability sum of the global quality function is equal to 1.

[0012] By adopting the above technical solution, the present invention employs a non-uniform redistribution mechanism based on the reciprocal of the focal element cardinality and the mean of credibility, which effectively solves the logical violation problem caused by the traditional Dempster rule when dealing with highly conflicting evidence. By accurately compensating the conflict quality proportionally to the relevant non-empty subsets and the entire set of focal elements, the useful subdivision information in the conflict source is preserved, and the uncertainty of the system is reasonably expressed, thereby improving the logical stability of the system in extremely complex data environments.

[0013] Preferably, the process of obtaining the dynamic threshold corrected by the consistency index and the state determination process include: constructing a fixed-time window cache queue, and sequentially storing the comprehensive conflict measure calculated in multiple rounds and the corresponding consistency index; calculating the arithmetic mean of the comprehensive conflict measure in the queue, correcting the arithmetic mean using the consistency index and an adjustment coefficient that decays with the number of iterations, and outputting the dynamic threshold; when the current comprehensive conflict measure is greater than the dynamic threshold, triggering the non-uniform redistribution; when the current comprehensive conflict measure is less than or equal to the dynamic threshold, determining it as a low-conflict state, skipping redistribution and directly entering the Dempster rule fusion step.

[0014] Preferably, the step of generating the basic quality function using a membership function based on a dual-factor weighting of error dispersion and historical hit rate includes: calculating the standard deviation of the prediction error of each prediction model on the historical validation set, normalizing and inverting it to obtain a dimensionless error stability index; linearly weighting it with the model's historical hit rate, and normalizing it using the L1 norm to obtain a comprehensive confidence weight; dynamically scaling the Gaussian kernel width of the Gaussian membership function based on the comprehensive confidence weight, thereby calculating the nonlinear probability mapping value of each performance level and generating the basic quality function.

[0015] By adopting the above technical solution, this invention establishes a two-factor evaluation mechanism based on the standard deviation of prediction error and the historical hit rate of the model, thereby achieving real-time evaluation of the input quality of multi-source models. This mechanism can automatically identify model deviations caused by material batch fluctuations or interference from harmful impurities, and by dynamically adjusting the membership function, it weakens the negative impact of low-quality data on the optimization results from the source, ensuring the reliability of the input end of the adaptive optimization system.

[0016] Preferably, the Gaussian kernel width satisfies the following relationship:

[0017] In the formula, Characterizes the Gaussian kernel width; Characterizing the first The confidence weight of individual prediction models; A preset lower bound characterizing the width of the Gaussian kernel; A preset upper limit characterizing the width of the Gaussian kernel.

[0018] Preferably, the step of constructing a global fitness function using a weighted multi-objective function and running an optimization algorithm to iteratively search for the mix proportion includes: normalizing the joint confidence level of achieving the target levels corresponding to compressive strength, flowability, and cost, and using 1 minus the normalized value or its reciprocal as the inverse penalty weight of the corresponding sub-objective function; performing directional processing on each sub-objective function, converting compressive strength into a minimization term, and flowability into an interval deviation penalty term, and directionalizing it with the cost target; summing the product of the inverse penalty weight and the corresponding directional sub-objective function to obtain the global fitness function for single-objective optimization; initializing a multi-dimensional position vector representing the mix proportion parameters, and running a particle swarm optimization algorithm within the set mix proportion constraint boundary, with the goal of minimizing the global fitness function, and iteratively updating the individual and global optimal positions of the particles.

[0019] By adopting the above technical solution, the present invention ensures that the search process always evolves in the direction with the highest consensus of model prediction, balances the contradiction between compressive strength, flow performance and economic cost, and improves the feasibility of the optimization results in practical engineering applications.

[0020] Preferably, the feature variable is the spectral radius of the contemporary conflict matrix, and the optimization algorithm is a particle swarm optimization algorithm. The step of extracting the feature variable of the conflict matrix, dynamically adjusting the mutation parameter to perturb the particles, and dynamically shrinking the search step size according to the uncertainty index of the probability distribution includes: linearly mapping the spectral radius to a specific numerical range [1.5,2] of the Tent chaotic map to generate the mutation parameter; when a particle is trapped in a local optimum, substituting its position variable into the Tent mapping equation containing the mutation parameter for perturbation and mutation; calculating the Shannon entropy of the fused output Pignistic probability distribution as the uncertainty index, normalizing the Shannon entropy, and synchronously adjusting the learning factor and inertia weight in the particle swarm velocity update equation with its monotonically increasing function to achieve adaptive shrinkage of the search step size.

[0021] By adopting the above technical solution, the present invention shortens the convergence algebra and improves the capture accuracy of the optimal solution, realizing the underlying collaboration between uncertain reasoning and swarm intelligence algorithms.

[0022] Secondly, the present invention provides an adaptive optimization system for phosphogypsum concrete mix proportions, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned adaptive optimization method for phosphogypsum concrete mix proportions is implemented.

[0023] By adopting the above technical solution, a computer program is generated from the above-mentioned adaptive optimization method for phosphogypsum concrete mix proportions and stored in a memory so that it can be loaded and executed by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0024] The technical solution of the present invention has the following beneficial technical effects: This invention constructs an identification framework with different performance levels as the hypothesis set, and uses a two-factor weighted membership function based on the model prediction error dispersion and the model's historical hit rate to map the output of each sub-prediction model to the basic quality function. This scheme can objectively characterize the complex influence of various mix proportion parameters on multiple performance indicators of phosphogypsum concrete, and reduce the bias of a single prediction model caused by the interference of harmful impurities in phosphogypsum.

[0025] Based on this, this solution establishes a dynamic evidence conflict handling mechanism, enabling real-time dynamic updates of credibility weights and conflict determination criteria. This mechanism can automatically adjust the fusion strategy according to the consistency of evidence, ensuring that the system's response to multi-source information conflicts is more targeted. By performing a non-uniform redistribution of conflict quality within the focal elements involved in the conflict and their related focal elements, this invention filters out contradictory redundancies in the information flow, improving the logical rationality and decision robustness of the fusion results in highly conflicted environments.

[0026] In subsequent optimization iterations, this invention introduces a chaotic mapping mutation strategy controlled by the spectral radius of the conflict matrix, and dynamically updates the particle search step size based on the shrinkage rate of the probability distribution entropy value of the fused output, thus deeply improving the particle swarm optimization algorithm. Furthermore, this invention introduces a fitness evaluation mechanism that transforms joint trust into reverse penalty weights, reversing the algorithmic black hole of traditional weighted methods that easily misclassifies low-trust, inferior solutions as optimal solutions, enhancing the penalty strength of the global fitness function for substandard particles, and ensuring the correctness of the optimization physical vector. Simultaneously, by implementing off-diagonal mean dimensionality reduction on the multidimensional conflict matrix, the curse of dimensionality in measure scalarization under multi-model architectures is solved. This strategy enhances the algorithm's ability to escape local optima in complex solution spaces, improving overall convergence efficiency. Ultimately, this invention can efficiently solve for the optimal mix design that simultaneously considers compressive strength, flow performance, and economic cost-effectiveness, enhancing the comprehensive utilization value of phosphogypsum resources in the building materials field. Attached Figure Description

[0027] Figure 1 This is a flowchart of an adaptive optimization method for the mix proportion of phosphogypsum concrete in this invention; Figure 2 This is a schematic diagram illustrating how the conflict determination threshold changes with the number of iterations; Figure 3 This is a schematic diagram illustrating how the speed update step size parameter changes with the normalized Pignistic probability entropy; Figure 4 This is a comparative diagram showing how the global joint trust level changes with the number of iterations under different schemes. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0029] This invention discloses an adaptive optimization method for the mix proportion of phosphogypsum concrete, referring to... Figure 1 This includes steps S1-S3: S1: Construct a multi-target identification framework and generate a basic quality function.

[0030] It should be noted that, in order to avoid the nonlinear fluctuations in data caused by impurities in phosphogypsum interfering with the cement hydration process and the failure of the physical prediction model due to strong coupling of multiple objectives, this step establishes an initial evidence distribution source with the ability to resist disturbances by extracting the core proportion feature vector, inputting it into the prediction model group, and combining it with the error dispersion to map the membership degree.

[0031] Preferably, as an example, an identification framework is constructed with performance levels as the hypothesis set for compressive strength, fluidity, and cost, respectively. The phosphogypsum content, water-cement ratio, sand ratio, and admixture dosage are collected as inputs to the mix proportion prediction model group. A two-factor weighted Gaussian membership degree based on the model prediction error dispersion and the model's historical hit rate is used to map the output of each sub-prediction model to the basic probability assignment of the focal element, generating the basic quality function, including: 1. Identification Framework Construction and Feature Vector Acquisition. First, the processor constructs an identification framework with performance levels as the set of hypotheses for compressive strength, flowability, and cost, establishing a complete set of independent hypotheses. Next, it reads historical mix proportion data files from external storage and extracts the proportion values ​​to form the input feature vector. For example, in specific engineering design applications, the aforementioned performance level hypothesis sets are divided into three state levels: low, medium, and high. Their boundary scales are anchored based on historical sample quantiles, engineering design requirements, or preset thresholds. The carrier used to read historical data is a comma-separated value file, and the extracted feature vector is composed of the specific values ​​of phosphogypsum content, water-cement ratio, sand ratio, and admixture dosage.

[0032] 2. Prediction Model Configuration and Two-Factor Credibility Analysis. Subsequently, the processor configures the mix proportion prediction model group, injecting the input feature vector into the architecture containing multiple sub-prediction models to output the corresponding continuous prediction values. Model parameter training is completed using historical phosphogypsum concrete mix proportion samples and their corresponding measured indicators. Next, the standard deviation of the prediction error of each sub-prediction model on the historical validation set is calculated as the model prediction error dispersion. Simultaneously, the percentage of times each sub-prediction model's prediction level matches the actual level is read as the model's historical hit rate. Finally, the model prediction error dispersion is normalized, and the inverse value is taken to obtain the dimensionless error stability index. This index is linearly weighted with the model's historical hit rate, and then L1 norm normalization is performed to obtain the comprehensive credibility weight. In the actual model architecture selection and configuration scheme, the sub-prediction models within the mix proportion prediction model group adopt a hybrid network configuration of regression tree models, support vector regression models, neural network regression models, or a combination of these models. Alternatively, multi-output regression models can be directly deployed to achieve simultaneous parallel derivation of the three performance indicators.

[0033] 3. Membership mapping and basic probability assignment reconstruction. After obtaining the weight parameters, the processor extracts the first... The predicted continuous values ​​of the sub-prediction model and the first The cluster centers corresponding to each level are normalized to the same dimensional scale according to the historical sample value range; at the same time, the Gaussian kernel width is dynamically scaled and adjusted based on the confidence weights obtained from the previous calculation.

[0034] Specifically, the Gaussian kernel width satisfies the following relationship:

[0035] In the formula, Characterizes the Gaussian kernel width; Characterizing the first The confidence weight of individual prediction models; A preset lower bound characterizing the width of the Gaussian kernel; The preset upper limit represents the width of the Gaussian kernel. The preset lower limit is defined as being greater than zero and strictly less than the preset upper limit.

[0036] For the evolution control of the Gaussian kernel width, the subtraction operator represents the limit search span defined by the preset upper and lower limits, while the multiplication operator represents the hindering compensation game process based on the confidence weight. If the multiplication operation compensation of the above inverse mapping is cancelled in the underlying logic, the sensor feature streams of high and low reliability will be in an equal weight state. Under the physical condition of extreme fluctuations in phosphogypsum content, the microcontroller will directly accept the low reliability boundary data, causing the subsequent probability distribution map to exhibit an unstable and divergent state.

[0037] The above relationship belongs to the category of weighted linear boundary interpolation, which transforms the confidence contraction principle in expert systems into a data dispersion suppression strategy in physical space. This makes the Gaussian kernel width narrower when the confidence is higher, ensuring that the probability density around the core predicted value exhibits a sharpening and clustering effect.

[0038] Based on the obtained Gaussian kernel width, the processor further calculates the nonlinear probability mapping values ​​for each level. The membership values ​​of each level satisfy the following relation:

[0039] In the formula, Characterizes the membership degree value; The normalized model predictions are represented; Characterizes the cluster centers after normalization; Characterizing the first Gaussian kernel width of individual prediction models; This represents an exponential function with the natural constant e as its base.

[0040] 4. Basic Probability Reconstruction and Allocation of Combined Focal Elements. Finally, the processor normalizes the membership values ​​of each level output by the same sub-prediction model for the same performance index, obtaining the initial allocation ratio of each single-level focal element. Then, based on the proximity of adjacent level membership values, the processor proportionally subtracts the corresponding quality from the initial quality of the single-level focal elements and assigns it to adjacent combined focal elements as an uncertain quality proportion. Simultaneously, based on the confidence weight, the processor again proportionally subtracts from the remaining quality and assigns it to the entire set of focal elements in the identification framework as retained quality, ensuring that the sum of the basic probability assignments of all focal elements equals one, ultimately generating the basic quality function corresponding to each sub-prediction model. In the specific quality transfer rules, the closer the values ​​of adjacent level membership values, the greater the proportion of uncertain quality transferred to adjacent combined focal elements; conversely, the lower the confidence weight of the sub-prediction model, the more retained quality is allocated to the entire set of focal elements, thus achieving a global distribution of uncertain information.

[0041] It should be added that the calibration procedure for the boundary of each state level in the above set of performance levels of compressive strength, fluidity and cost includes: extracting the historical quality inspection records of phosphogypsum concrete mix proportions produced continuously by the target mixing plant in the past three years; using the measured strength range of the servo press, the slump loss rate over time at the construction site and the fluctuating unit price of raw materials as joint physical feature vectors to perform statistical normal distribution fitting; and extracting the absolute dimensional values ​​corresponding to the 33rd and 67th percentiles of the cumulative probability density distribution respectively, which are used as objective hard calibration boundaries to strictly define the low, medium and high state levels.

[0042] It should be added that the calibration procedure for the aforementioned lower and upper limits of the Gaussian kernel width includes: extracting the background fluctuation variance data of compressive strength of multiple batches of phosphogypsum concrete standard specimens in constant temperature and humidity curing chambers, combining it with the standard deviation of static measurement noise of the dry powder feeding weighing sensor at the on-site mixing plant for benchmark summation, and estimating the lower limit through statistical confidence intervals; at the same time, extracting and calibrating the extreme values ​​of prediction errors induced by the full-range historical test samples under the physical limit saturation state of phosphogypsum material moisture content as the upper limit, thereby locking the safe operation boundary of the microcontroller through physical measurement data.

[0043] It should be added that the physical transformation procedure for the above-mentioned proportion of uncertain mass transferred to adjacent combined focal elements and the proportion of reserved mass allocated to the entire set of focal elements includes: obtaining the Euclidean distance calibration benchmark of the cluster centers of adjacent state levels in the normalized feature space, substituting the ratio of the absolute difference of membership degree output by the real-time feature vector to the distance benchmark into the preset negative exponential decay hardware lookup table function, and physically mapping to obtain the proportion of uncertain mass transferred to the combined focal elements; at the same time, directly using the numerical complement of the dimensionless error stability index obtained by the bottom sensor as the rigid allocation coefficient, and calculating and retaining the final reserved mass allocated to the entire set of focal elements accordingly.

[0044] Thus, by using two-factor confidence analysis and dynamic adaptive scaling of Gaussian kernel width, the risk of output distortion caused by local sample oscillation in the single mix ratio prediction model was suppressed, laying a solid foundation for high-fidelity underlying data for subsequent handling of multi-source conflict information fusion.

[0045] S2: Extract eigenvalues ​​of the conflict matrix and redistribute non-uniform quality.

[0046] It should be noted that, in order to avoid the strong physical index deviation caused by the spatial heterogeneity of material impurities among the multi-objective prediction outputs of phosphogypsum, as well as the serious submergence and failure of the underlying true features due to the fixed threshold fusion strategy, this step implements local conflict quality redistribution by constructing a multi-source probability conflict matrix and extracting feature spectra to build a consistency index and combining it with dynamic time-series thresholds.

[0047] Preferably, as an example, the Jousselme distance and the Pengnistic probability conflict coefficient of the evidence pair constituting the basic quality function are calculated. The geometric mean of the two is used to construct a conflict matrix and perform eigenvalue decomposition. A consistency index is constructed based on the second largest and largest eigenvalues, and the weight of the hit rate factor is adjusted. When the comprehensive conflict measure exceeds the conflict judgment threshold obtained by combining the historical comprehensive conflict measure moving average with the consistency index, the conflict quality is redistributed according to the normalized product of the cardinality of the focal elements participating in the conflict, the union of the focal elements participating in the conflict, the inverse of the cardinality of the focal elements of the relevant non-empty subsets within the union, and the confidence of the reverse conflict. The remaining conflict quality is assigned to the entire set, including: 1. Conflict Coefficient Analysis and Matrix Construction. First, for the evidence pairs generated by the pairwise combination of the three sub-prediction models mentioned above, the processor calculates the Jousselme distance, which measures the degree of allocation difference. This is calculated by taking the square root of half of the inner product of the difference between the column vectors of the two basic quality functions and the interaction matrix. The elements of the interaction matrix are composed of the ratio of the number of elements in the intersection of each pair of focal elements to the number of elements in the union. Next, according to the probability transformation formula, the basic probability of each focal element is divided by the number of levels of states it contains and then evenly distributed to each single-element hypothesis to obtain the corresponding Pignistic probability distribution vector. Subsequently, the cosine similarity between the two Pignistic probability distribution vectors is calculated, and the Pignistic probability conflict coefficient is derived by subtracting the cosine similarity from one. Finally, the geometric mean of the product of the Jousselme distance and the Pignistic probability conflict coefficient is calculated as the comprehensive conflict measure. This geometric mean is symmetrically assigned to the corresponding off-diagonal positions of the matrix, and all diagonal elements are set to zero, thus completing the construction of the symmetric conflict matrix. In specific data extrapolation, for the ternary identification framework formed by low, medium, and high levels of compressive strength, assuming the measured Jousselme distance is 0.2 and the Pengistic probability conflict coefficient is 0.15, the system calculates the product of these two and takes the square root to obtain a comprehensive conflict measure of 0.1732. This comprehensive conflict measure is then used as the off-diagonal element to construct a 3×3 real symmetric conflict matrix. For example, three independent conflict measure values ​​outside the diagonal of this 3×3 real symmetric matrix are extracted, such as 0.1732, 0.5218, and 0.655, respectively. The arithmetic mean of these three values ​​is calculated to obtain 0.45. This value of 0.45 is used as the global scalar of this comprehensive conflict measure and pushed into the subsequent cache queue for threshold comparison. Subsequently, all off-diagonal elements of the conflict matrix are extracted, and their arithmetic mean or maximum value is calculated. This arithmetic mean or maximum value is used as the global scalar representing the overall conflict level of the system, i.e., the comprehensive conflict measure, for subsequent state determination and queue updates.

[0048] 2. Feature Space Decomposition and Credibility Correction. After obtaining the conflict matrix, the processor uses the Jacobi method numerical algorithm to decompose its features, extracting a sequence of feature values ​​arranged in descending order of absolute value. Next, the second largest absolute value of the extracted feature is divided by the largest absolute value of the first feature to obtain the conflict dispersion index. The consistency index is generated by subtracting the conflict dispersion index, which is used to perform dimensionless correction on the weight of the model's historical hit rate. Subsequently, the absolute value of the eigenvector corresponding to the first feature value is extracted and normalized using the L1 norm to obtain the conflict participation degree of each piece of evidence. Finally, the conflict participation degree is subjected to inverse normalization to obtain the conflict-corrected credibility weight, which is then merged with the corrected hit rate weight as the final credibility weight. For example, if the absolute value of the largest eigenvalue obtained from the decomposition is 1.56 and the absolute value of the second largest eigenvalue is 0.34, then the calculated value of the conflict dispersion index is 0.2179, and the corresponding consistency index is 0.7821; simultaneously, the absolute value of the eigenvector corresponding to the first eigenvalue is reflected as a transpose matrix composed of 0.4, 0.5 and 0.6.

[0049] Specifically, the credibility weights satisfy the following relationship:

[0050] In the formula, Characterizing the first The credibility weight of each piece of evidence; Characterizing the first The degree of conflict of interest in each piece of evidence; Characterizes the zero-bias parameter; In the representation of the summation operation, the first The degree of conflict in the evidence.

[0051] 3. Timing Threshold Update and Fusion State Adjudication. Next, the processor establishes a fixed-length cache queue time window based on a first-in, first-out (FIFO) mechanism. The comprehensive conflict measure output from multiple rounds of data acquisition is sequentially pushed into the queue, and the corresponding consistency index is recorded synchronously. Simultaneously, the oldest value with the furthest time interval is removed in real time. Subsequently, the arithmetic mean of the comprehensive conflict measure of the resident sequences in the cache queue is calculated and corrected by combining the current consistency index with an adjustment coefficient that decays with the number of iterations. The latest conflict determination threshold is then output after being truncated to preset upper and lower limits. Figure 2As shown; finally, the processor monitors the current comprehensive conflict metric in real time. If it exceeds the conflict determination threshold, a non-uniform redistribution mechanism with cardinality constraints is triggered. If it is less than or equal to the determination threshold, a low-conflict state is set and the system directly enters the standard Dempster rule fusion process. In the system memory queue configuration environment, the fixed time window capacity is preferably configured as an integer value in the range of 5 to 10. Assuming this length is set to 8, the arithmetic mean of the queue resident data is 0.24. In the actual threshold comparison and determination stage, if the real-time detected comprehensive conflict metric is 0.45, redistribution is immediately initiated. If the metric is 0.28, it is determined to fall into the safe zone, thus skipping the redistribution calculation.

[0052] 4. Dynamic Cardinality Constraints and Non-Uniform Quality Flow. As a branch processing method for triggering the redistribution security strategy, when the intersection of the focal elements that cause conflict quality is an empty set, the processor extracts the participating conflict focal elements, the union of the participating conflict focal elements, and the relevant non-empty subsets within the union to serve as redistribution objects. Subsequently, the number of hierarchical subsets contained in each redistribution object is calculated, i.e., the focal element cardinality. The reciprocal of this cardinality is taken as the local trust support degree and multiplied by the mean of the evidence credibility weight. The product result of all involved non-empty focal element sets is normalized to obtain the redistribution weight. Finally, the processor sets a retention coefficient according to the current measure, obtains the total conflict quality through the unnormalized conjunction rule, multiplies it by one minus the reduction ratio of the retention coefficient, and then multiplies it by the normalized redistribution weight. The result is superimposed on the basic probability assignment of the corresponding non-empty focal elements, and the remaining part of the total conflict quality multiplied by the retention coefficient is assigned as global uncertainty information to the full set of focal elements of the identification frame. In real-world computational allocation scenarios with highly conflicting multi-source evidence, if the total conflict quality generated by the empty set is as high as 0.45, then the cardinality of a single participating conflict focal element is one, and the cardinality of the non-empty subset related to the combination type is two. Assuming the extracted confidence weights have an average value of 0.305, and the retention coefficient, which is positively correlated with the comprehensive conflict measure and limited to between 0.1 and 0.5, is 0.25, the system extracts 75% of the conflict quality and performs targeted non-empty allocation based on the weights. The remaining 25% is directly injected into the focal elements of the entire set to ensure that the global probability sum is constant at one.

[0053] It should be added that the physical calibration procedure for the above-mentioned fixed time window capacity and the adjustment coefficient that decays with the number of iterations includes: extracting convergence cycle data of multiple batches of historical phosphogypsum concrete mix proportion optimization tasks, calculating the average number of iterations required for the system to transition from a high-conflict state to a steady state, and directly calibrating it as the time window capacity after rounding it up; at the same time, extracting the descent slope of the natural envelope of the convergence trajectory measured by the sensor in multiple batches of historical annealing optimization experiments, and using its negative exponential fitting parameter on the iteration time axis as the adjustment coefficient for this decay, thereby anchoring the pure mathematical time window sliding process to the actual convergence rhythm of the algorithm.

[0054] It should be added that the engineering calibration procedure for the retention coefficient mentioned above includes: extracting the gradation variation coefficient and the measured fluctuation range of moisture content of coarse and fine phosphogypsum aggregates from previous production lines, inputting the product of the two into a preset rheological empirical lookup table function, obtaining the physical equivalent ratio representing the unmeasurable noise floor of the batching system, and strictly truncating it to the physical limit range of 10% to 50% as the objective benchmark for dynamically calculating the retention coefficient.

[0055] Thus, by using conflict spectrum radius feature decomposition in the multidimensional feature space and dynamic quality transfer mechanism based on focal element basis topology, the risk of severe information conflict in the early stage of optimizing the mix proportion of highly heterogeneous materials is successfully eliminated in the underlying control domain, providing a high-fidelity convergence data path free from false maxima for subsequent heuristic iterative optimization.

[0056] S3: Multi-objective fitness construction and adaptive particle swarm optimization.

[0057] It should be noted that, in order to avoid the objective failure consequences of the natural physical constraints between the multi-objective properties of phosphogypsum causing the optimization trajectory to fall into local false extremes and the static evolution step size causing the bus control to overshoot the boundary oscillation in the later stage, this step establishes an adaptive optimization control link with global jump-out and local precise anchoring capabilities by constructing a trust-weighted global fitness combined with spectral radius chaotic variation and entropy contraction step size control mechanism.

[0058] Preferably, as an example, the modified mass function after conflict redistribution, or the mass function obtained by Dempster rule fusion under low-conflict conditions, is used to calculate the joint trust degree within each identification framework; using the trust degree-weighted multi-objective function as the fitness, the particle swarm optimization algorithm is run to search for fit ratios. During iteration, the Tent mapping, whose parameters are adjusted by the spectral radius of the contemporary conflict matrix, is used to mutate the particles, and the step size is updated according to the shrinkage rate of the pignistic probability entropy. After obtaining the actual validation results, the historical hit rate of the model is updated, and the optimal fit ratio is output, including: 1. Joint Trust Resolution and Global Fitness Construction. First, the processor extracts the joint trust values ​​corresponding to three performance directions: compressive strength reaching the target strength level, flow rate falling within the target construction range, and cost being at the low-cost level. After normalization, since the global fitness optimization direction is minimization, the system adopts a reverse weighting mechanism. This involves subtracting the normalized value of each joint trust value from 1, or taking the reciprocal after adding a minimum bias, and attaching this reciprocal as the reverse penalty weight for the corresponding sub-objective function. This design ensures that inferior solutions with lower target trust values ​​receive larger penalty multipliers, preventing erroneous convergence. Next, the three sub-objective functions are simultaneously normalized, converting the compressive strength sub-objective function into a revenue minimization term and the flow rate sub-objective function into a target range deviation penalty term, thus aligning it with the cost sub-objective function representing overhead. Finally, the normalized weights are multiplied by their corresponding aligned sub-objective functions and summed to construct the global fitness function for single-objective optimization in the particle swarm optimization algorithm. In the specific system data transformation process, the aforementioned joint trust degree is extracted by the processor through the execution of the Pignistic probability transformation on the fused quality function within the identification framework. At the same time, before being fed into the fitness calculation, the values ​​of each sub-objective function are uniformly processed by minimax normalization and strictly mapped to the bounded range of zero to one, thereby satisfying the requirement of consistent lower bound convergence form of the minimization solution.

[0059] 2. Optimal Topology Establishment and Iterative Search Advancement. After global fitness is achieved, the processor initializes multidimensional position vectors representing phosphogypsum content, water-cement ratio, sand ratio, and admixture dosage in the particle swarm, constructing a search space within pre-defined material ratio constraints. Subsequently, a cyclic evaluation system is initiated, substituting each multidimensional position vector into the global fitness function for minimization target decoding. As the computational iteration loop continues, the processor continuously detects and updates the records of the individual optimal positions traversed by each particle, while extracting and sharing the best global optimal position within the current full cluster search coverage, continuously providing feedback and guiding the particle's movement. In the actual instantiated model configuration parameters, the processor backbone execution flow opens a four-dimensional population space with a scale ranging from 50 to 100; and the lower and upper bounds of specific physical parameters are fixed as follows: phosphogypsum content 20% to 50%, water-cement ratio 0.3 to 0.6, sand ratio 35% to 50%, and admixture dosage range 0.5% to 2.5%.

[0060] 3. Spectral radius mutation induction and population information entropy step size constraint. Upon entering the iterative loop, the processor extracts the absolute value of the largest eigenvalue of the current conflict matrix as the spectral radius, linearly mapping it to the effective closed interval of the chaotic parameter to generate the current mutation parameter. In this embodiment, the effective closed interval is defined as [1.5,2]. Next, the processor generates uniformly distributed random numbers and substitutes them into the Tent mapping iterative equation with this mutation parameter to generate a chaotic sequence. If the current value of the chaotic sequence is greater than a random threshold, then the worst-performing particles in the bottom 10% of the current population's fitness ranking are selected, and their position variables are normalized to the [0,1] interval according to the corresponding upper and lower limits before perturbation. When the processor detects that a particle is trapped in a local optimum, its position variables are also normalized and substituted into the Tent mapping equation for iterative perturbation. The results of the aforementioned perturbation are then back-normalized to the material mix design constraint space to generate new positions after mutation. Finally, at the end of each iteration, the processor calculates the Shannon information entropy of the fused Pignistic probability distribution vector corresponding to the candidate mix design of each particle in the current swarm. The average Shannon information entropy of all particles is taken as the current swarm Shannon information entropy. This current swarm Shannon information entropy is divided by the theoretical maximum entropy of the ternary identification framework to obtain the normalized entropy scale. Then, based on the monotonically increasing function corresponding to this scale, the individual learning factor, swarm learning factor, and inertia weight in the algorithm's speed update equation are synchronously adjusted. When the entropy decreases, the inertia weight and learning factor decrease synchronously to achieve adaptive contraction of the speed update step size. Figure 3 As shown. For example, if the measured value of the extracted spectral radius is 0.85, the processor transforms it into a mutation parameter of 1.72 through affine mapping; at the same time, it is set that if a particle fails to achieve a minimum value improvement for more than 5 consecutive times, it is judged as prematurely locked into a local optimum, thereby triggering the reassignment replacement mechanism of the mutation position.

[0061] The Tent mapping used when triggering a disturbance satisfies the following relationship:

[0062] In the formula, Characterizes the next generation of position coordinates resulting from the current dimensional variation. Represents the original position coordinates of the current dimension. Characterizes the variation parameters obtained from contemporary extraction mapping.

[0063] 4. Entity Mix Proportion Closed-Loop Verification and Historical Hit Rate Refresh. The particle swarm optimization algorithm terminates after reaching the preset maximum number of iterations, extracts the global optimal position coordinates, and obtains the actual verification results according to the mix proportion corresponding to these coordinates. Subsequently, the experimental evaluation level is compared with the initial prediction level of each sub-prediction model, and the model's historical hit rate is updated accordingly, outputting the optimal mix proportion. Specifically, according to the mix proportion corresponding to the extracted coordinates, concrete samples are prepared using a batching and mixing system. After molding and curing, the compressive strength is tested using a servo pressure testing machine, and the flowability index is measured. Simultaneously, the actual cost index is calculated based on the actual material usage and unit price. Next, the total number of verifications for each sub-prediction model participating in this verification is accumulated, and the historical hit rate of correctly predicted sub-prediction models is accumulated, updating the model's historical hit rate for use in the next optimization task. Finally, the experimentally verified optimal phosphogypsum dosage, water-cement ratio, sand ratio, and admixture dosage values ​​are returned to the end user through a standard output stream.

[0064] It should be added that the physical calibration procedure for the above-mentioned global maximum number of shards constant includes: during the overall deployment phase of the distributed storage cluster, based on the available memory addressing space of a single node processor and the upper limit of the number of concurrent thread pool cores, combined with the number of bytes of memory occupied by the particle position matrix in a single operation, the maximum population tolerance under the condition of no memory overflow is calculated, and after being decremented and rounded down, it is directly mapped and solidified into the global maximum number of shards constant to ensure that the particle swarm size is within the safe boundary of the system's computing power.

[0065] It should be added that the above-mentioned threshold for the number of consecutive stagnation detections that are judged as premature convergence lock-in state when the continuous improvement fails to achieve the minimum value, and the calibration procedure for the monotonically increasing function corresponding to the normalized entropy value, include: extracting the rated rotation speed benchmark of the mechanical spindle of the forced concrete mixer of the field model to complete a single tangential homogenization mixing cycle, and directly calibrating its physical cycle constant as the tolerance threshold for the optimization control flow to fall into premature convergence stagnation; at the same time, for the learning factor and inertia weight adjustment and control mechanism, based on the measured envelope data of rheological shear stress dissipation in the previous standard cement paste thixotropic flow experiment, fitting and constructing a negative exponential decay hardware lookup table array characterizing the viscous deceleration characteristics of real fluid, and solidifying it as the only objective mapping benchmark for micro-optimization step size shrinkage control.

[0066] In this way, the risk of pseudo-extreme value calculation deadlock caused by abrupt changes in the impurity properties of multi-source materials is eliminated, and the out-of-bounds oscillation of parameters issued by the system is effectively suppressed at the end of the control sequence. Finally, the lean mix proportion material control instructions that are highly consistent with the actual mechanical and rheological design expectations are delivered to the automatic batching equipment on site at high frequency.

[0067] In the experimental verification phase, the experimental environment uniformly adopted the boundary condition constrained by the proportion of phosphogypsum material, and used compressive strength, flowability, and cost as evaluation sub-objectives. The maximum number of iterations was set to 100, and the population size to 50. For comprehensive comparative evaluation, three comparison schemes were constructed. The first scheme served as the baseline model, its internal operating logic directly employing the standard Dempster combinatorial rule for fusion and executing the standard particle swarm optimization algorithm. The second scheme, based on the baseline model, introduced a mechanism based on the moving average of historical comprehensive conflict measures combined with a consistency index to correct the conflict determination threshold, thereby performing non-uniform redistribution of conflict quality. The third scheme, as the complete execution logic of this invention, further added a strategy of mutating particles using a Tent mapping with parameters adjusted by the spectral radius of the contemporary conflict matrix, and a control mechanism for updating the step size based on the shrinkage rate of the Pignistic probability entropy, based on the second scheme. The three comparison schemes were each run independently 50 times to obtain the statistically optimal index. The convergence effect of the multi-objective global joint trust degree corresponding to each scheme, i.e., the comparison of the change of global joint trust degree with the number of iterations under different schemes, was combined with... Figure 4 As shown.

[0068] Comparative test results show that the first scheme has an average global convergence iteration count of 85, achieving a compressive strength of 32.5 MPa and a flowability of 185 mm for the optimized phosphogypsum concrete, with a corresponding multi-objective global joint confidence score of only 0.72. The second scheme reduces the average global convergence iteration count to 62, increases the compressive strength to 38.2 MPa, and achieves a flowability of 205 mm, with a corresponding multi-objective global joint confidence score of 0.85. Under the same experimental conditions, the third scheme requires an average of only about 41 iterations to reach the convergence criteria, and its final output of the optimal mix proportion further increases the compressive strength to 42.6 MPa, achieves a flowability of 220 mm, and boasts a multi-objective global joint confidence score as high as 0.89.

[0069] A deeper analysis of the aforementioned technical effects reveals that the improvement in joint trust level of the second scheme compared to the first scheme is mainly due to the dynamically adjusted conflict determination threshold and the non-uniform quality redistribution mechanism executed according to the reciprocal of the focal element cardinality. This mechanism effectively filters out contradictory data that causes deviations, thereby reducing the risk of result deviation caused by the direct fusion of highly conflicting evidence and minimizing the adverse impact of contradictory evidence on the final fusion result. The leap in convergence speed and optimization upper limit of the third scheme compared to the second scheme is attributed to the synergistic effect of the Tent mapping mutation and the Pignistic probability entropy step size contraction mechanism. Specifically, when the algorithm falls into a premature convergence stagnation state, the chaotic mutation guided by the spectral radius of the conflict matrix can effectively perturb the particles to jump out of the local optimum position; at the same time, the convergence step size adjustment strategy based on information entropy strictly controls the miniaturization step of the algorithm when approaching the global optimum position, which improves the local search accuracy when approaching the optimal solution region, thereby ensuring that the system can ultimately output the optimal mix ratio that takes into account mechanical performance, construction performance, and economic cost.

[0070] Figure 2 This diagram illustrates how the conflict determination threshold changes with the number of iterations. The continuous single curves in the diagram represent the evolution trajectory of the determination threshold during the optimization process.

[0071] The graph shows a clear non-linear decrease in the decision threshold as the number of iterations increases. This demonstrates that the algorithm successfully achieves dynamic tightening of the decision criteria: a higher threshold is used in the early stages of optimization to ensure the fault tolerance of evidence fusion, while a lower threshold is used in the later stages to enhance the capture of subtle conflicts. This corresponds to the control logic described in the specific implementation that uses an adjustment coefficient that decays with the number of iterations to adjust the decision threshold in real time.

[0072] Figure 3 This diagram illustrates how the step size parameter for velocity updates changes with the normalized Pengistic probability entropy. The dashed lines, composed of alternating long and short dashes, represent inertia weights; the dotted lines, composed of closely spaced dots, represent learning factors.

[0073] Observing the images reveals that the trajectories of both the inertia weight and the learning factor decrease synchronously as the entropy value decreases. This demonstrates that the algorithm can sense changes in system uncertainty and adjust the search pace in real time. When the normalized Pengistic probability entropy decreases, the velocity update step size shrinks accordingly, ensuring higher local search accuracy when approaching the global optimum. This phenomenon confirms the effectiveness of the technique in the specific implementation method of adjusting the optimization step size using a monotonically increasing entropy function.

[0074] Figure 4This diagram illustrates the comparison of global joint trust level with iteration count under different schemes. The dotted-dash line (consisting of long dots and dashed lines) represents Scheme 1, which uses standard fusion rules; the dashed line (consisting of short dashed lines of equal length) represents Scheme 2, which introduces a dynamic threshold correction mechanism; and the bold solid line at the top represents the complete scheme of this invention, which integrates chaotic mapping mutation and step-size contraction control.

[0075] Image comparison reveals that the complete scheme represented by the solid line has the steepest curve slope and reaches steady state first. Furthermore, the global joint confidence level achieved by the solid line is significantly higher than that of the schemes corresponding to the dashed and dotted lines. This demonstrates that the present invention, by deeply coupling the evidence fusion results with the improved particle swarm optimization algorithm, significantly improves the system's optimization speed in the multi-dimensional parameter space and can capture the optimal combination scheme that balances various performance indicators with a higher confidence level.

[0076] This invention also discloses an adaptive optimization system for phosphogypsum concrete mix proportions, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an adaptive optimization method for phosphogypsum concrete mix proportions according to the present invention.

[0077] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0078] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for self-adaptive optimization of phosphogypsum concrete mix proportion, characterized in that, include: S1: Construct an identification framework with performance grades of compressive strength, flowability and cost as the hypothesis set respectively, collect the proportion parameters and input them into the proportion prediction model group; use a membership function based on error dispersion and historical hit rate as a weighted factor to map the output of each prediction model to generate the basic quality function; S2: Calculate the distance and probability conflict coefficient between the basic quality functions to obtain the comprehensive conflict measure between pairs of evidence to construct a conflict matrix. Extract the off-diagonal elements of this matrix and calculate the mean to obtain the global scalar of the current comprehensive conflict measure. Extract the eigenvalues ​​of the conflict matrix to construct a consistency index and adjust the hit rate weights. Generate a dynamic threshold based on the historical comprehensive conflict measure. When the global scalar of the current comprehensive conflict measure exceeds the dynamic threshold corrected by the consistency index, calculate the redistribution weights based on the focal element cardinality and evidence credibility to perform non-uniform redistribution of conflict quality and assign the remaining conflict quality to the whole set. S3: Fuse the quality functions and calculate the joint trust degree of the fused quality function in each identification framework. Convert this into a reverse penalty weight to construct a global fitness function using a weighted multi-objective function. Run an optimization algorithm to iteratively search for the fit ratio. Iteratively extract the feature variables of the conflict matrix, dynamically adjust the mutation parameters to perturb the particles, and dynamically shrink the search step size according to the uncertainty index of the probability distribution. After obtaining the optimal fit ratio, update the historical hit rate.

2. The method according to claim 1, wherein, The distance between the basic quality functions is the Jousselme distance, and the probability conflict coefficient is the Pignistic probability conflict coefficient. The calculation of the distance between the basic quality functions and the probability conflict coefficients constructs a conflict matrix, and the calculation of the current comprehensive conflict measure includes: obtaining the Jousselme distance between the i-th and j-th pieces of evidence; Calculate the cosine similarity between the pignistic probability distribution vectors of the i-th and j-th pieces of evidence, and subtract the cosine similarity from 1 to obtain the pignistic probability conflict coefficient; take the square root of the product of the Jousselme distance and the pignistic probability conflict coefficient as the comprehensive conflict measure between each pair of pieces of evidence, fill it into the corresponding position of the symmetric matrix, and construct the conflict matrix; extract the set of off-diagonal elements of the conflict matrix, and calculate its arithmetic mean as the global scalar of this comprehensive conflict measure.

3. The adaptive optimization method for phosphogypsum concrete mix proportion according to claim 1, characterized in that, The step of extracting eigenvalues ​​from the conflict matrix to construct a consistency index and adjust the hit rate weight includes: performing eigenvalue decomposition on the conflict matrix, extracting the first eigenvalue with the largest absolute value and the second eigenvalue with the second largest absolute value; using the ratio of the absolute value of the second eigenvalue to the absolute value of the first eigenvalue as a conflict dispersion index, and subtracting this index from 1 to obtain the consistency index; taking the absolute value of the eigenvector corresponding to the first eigenvalue, normalizing it using the L1 norm as the conflict participation degree of each piece of evidence, and performing inverse normalization on the conflict participation degree to obtain the credibility weight, thereby adjusting the hit rate.

4. The adaptive optimization method for phosphogypsum concrete mix proportion according to claim 1, characterized in that, The method of calculating redistribution weights based on focal element cardinality and evidence credibility to perform non-uniform redistribution of conflict quality and assign the remaining conflict quality to the entire set includes: when the intersection of the focal elements that cause conflict is an empty set, extracting the focal elements involved in the conflict, the union of the focal elements involved in the conflict, and their internal non-empty subsets as redistribution objects; taking the inverse of the focal element cardinality of each redistribution object, multiplying it by the mean of the evidence credibility that causes the conflict, and normalizing it to obtain the redistribution weight; splitting the total conflict quality according to a set retention coefficient: multiplying the total conflict quality by 1 and subtracting the retention coefficient, then multiplying the difference by the redistribution weight, and adding it to the basic probability assignment of each non-empty subset; assigning the remaining part obtained by multiplying the total conflict quality by the retention coefficient to the entire set of focal elements of the identification frame to ensure that the total probability sum of the global quality function is equal to 1.

5. The adaptive optimization method for phosphogypsum concrete mix proportion according to claim 1, characterized in that, The process of obtaining the dynamic threshold corrected by the consistency index and the state determination process include: constructing a fixed-time window cache queue, and sequentially storing the comprehensive conflict measure and the corresponding consistency index calculated in multiple rounds; calculating the arithmetic mean of the comprehensive conflict measure in the queue, correcting the arithmetic mean using the consistency index and an adjustment coefficient that decays with the number of iterations, and outputting the dynamic threshold; when the current comprehensive conflict measure is greater than the dynamic threshold, triggering the non-uniform redistribution; when the current comprehensive conflict measure is less than or equal to the dynamic threshold, determining it to be a low-conflict state, skipping redistribution and directly entering the Dempster rule fusion step.

6. The adaptive optimization method for phosphogypsum concrete mix proportion according to claim 1, characterized in that, The method of generating a basic quality function by employing a membership function based on a dual-factor weighting of error dispersion and historical hit rate includes: calculating the standard deviation of the prediction error of each prediction model on the historical validation set, normalizing and inverting it to obtain a dimensionless error stability index; linearly weighting it with the model's historical hit rate, and normalizing it using the L1 norm to obtain a comprehensive confidence weight; dynamically scaling the Gaussian kernel width of the Gaussian membership function based on the comprehensive confidence weight, thereby calculating the nonlinear probability mapping value of each performance level and generating the basic quality function.

7. The adaptive optimization method for phosphogypsum concrete mix proportion according to claim 6, characterized in that, The Gaussian kernel width satisfies the following relationship: In the formula, Characterizes the Gaussian kernel width; Characterizing the first The confidence weight of individual prediction models; A preset lower bound characterizing the width of the Gaussian kernel; A preset upper limit characterizing the width of the Gaussian kernel.

8. The adaptive optimization method for phosphogypsum concrete mix proportion according to claim 1, characterized in that, The process of constructing a global fitness function using a weighted multi-objective function and running an optimization algorithm to iteratively search for the mix ratio includes: normalizing the joint confidence level corresponding to the target levels of compressive strength, flowability, and cost, and using 1 minus the normalized value or its reciprocal as the inverse penalty weight of the corresponding sub-objective function; performing directional processing on each sub-objective function, converting compressive strength into a minimization term, and flowability into an interval deviation penalty term, and directionalizing it with the cost target; summing the product of the inverse penalty weight and the corresponding directional sub-objective function to obtain the global fitness function for single-objective optimization; initializing a multi-dimensional position vector representing the mix ratio parameters, and running a particle swarm optimization algorithm within the set mix ratio constraint boundary, with the goal of minimizing the global fitness function, and iteratively updating the individual and global optimal positions of the particles.

9. The adaptive optimization method for phosphogypsum concrete mix proportion according to claim 1, characterized in that, The feature variable is the spectral radius of the contemporary conflict matrix, and the optimization algorithm is the particle swarm optimization algorithm. The process of extracting the feature variables of the conflict matrix, dynamically adjusting the mutation parameters to perturb the particles, and dynamically shrinking the search step size according to the uncertainty index of the probability distribution includes: linearly mapping the spectral radius to a specific numerical range [1.5,2] of the Tent chaotic map to generate the mutation parameters; When a particle gets stuck in a local optimum, its position variable is substituted into the Tent mapping equation containing the mutation parameter to perform perturbation mutation; the Shannon entropy of the fused output Pignistic probability distribution is calculated as the uncertainty index, and after the Shannon entropy is normalized, the learning factor and inertia weight in the particle swarm velocity update equation are synchronously adjusted by its monotonically increasing function to achieve adaptive shrinkage of the search step size.

10. An adaptive optimization system for phosphogypsum concrete mix proportions, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement an adaptive optimization method for the mix proportion of phosphogypsum concrete according to any one of claims 1-9.

Images