A concrete strength pre-determination method and system based on proportioning parameters

Abstract

This invention relates to the field of building materials testing, and provides a method and system for predicting concrete strength based on mix proportion parameters. The method includes: acquiring design parameters; generating predicted performance values ​​through a prediction model; mapping the predicted performance values ​​to theoretical rebound values ​​and fluctuation ranges; responding to the on-site measured rebound values ​​of the target object; when the on-site measured rebound value exceeds the fluctuation range, converting it to a measured estimated performance value and comparing the deviation with the predicted performance value; triggering a deviation warning only when the deviation exceeds the fluctuation range and the deviation comparison confirms an anomaly; acquiring performance data to back-calculate the estimated true performance value; comparing the true performance value with the predicted performance value; locating at least one weight parameter in the prediction model that causes the deviation, using it as a target weight parameter; and performing self-calibration of the target weight parameter with a micro-step size. This invention achieves real-time, accurate, and self-evolving closed-loop prediction of concrete strength from mix design to on-site testing.

Description

Technical Field

[0001] This application relates to the field of building materials testing, and in particular to a method and system for pre-determining concrete strength based on mix proportion parameters. Background Technology

[0002] Currently, in the field of building materials testing, the quality control of concrete strength mainly follows two relatively independent paths: the laboratory standard curing method based on test blocks and the rebound method based on on-site physical testing. The laboratory standard curing method involves preparing test blocks and curing them under standard conditions for 28 days to conduct compressive strength tests, which serve as the final evaluation basis. The rebound method, on the other hand, uses a rebound hammer to impact the concrete surface, measures the rebound value, and calculates the estimated compressive strength value based on a unified strength measurement curve, thus quickly obtaining a strength reference for the structure in its current state.

[0003] However, in practical use, the aforementioned existing technologies form a separate two-layer architecture for prediction and physical analysis, resulting in significant lag. During routine on-site operations, the strength report from the laboratory and the rebound test results can often only be compared retrospectively, failing to predict whether the 28-day strength will meet the standard at the initial stage of concrete pouring. This lack of effective feedback during routine on-site operations causes construction teams to miss the optimal window for adjusting mix proportions or processes, leading to subsequent quality risks. Furthermore, its static threshold comparison method frequently triggers false alarms: the system directly compares a single measured value with a broad strength range, lacking a refined mapping relationship and quantitative representation of uncertainty. Once the measured value falls to the threshold edge, false alarms are frequently triggered, interfering with normal production judgments and reducing maintenance personnel's trust in early warning signals. Summary of the Invention

[0004] To address the problem that existing technologies cannot achieve real-time, accurate, and self-evolving closed-loop prediction of concrete strength from mix design to on-site testing due to the static and fixed prediction models and the lack of cross-stage data mapping and self-learning mechanisms, this application proposes a concrete strength prediction method and system based on mix parameters. This method uses the theoretical rebound value generated by the mix design as a cross-validation anchor point, utilizes early performance data to trace the root cause of deviations, and uses the measured deviations to drive the model weights to smooth self-calibrate, thereby achieving a complete closed-loop prediction from front-end prediction to back-end correction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a method for predicting concrete strength based on mix proportion parameters includes: acquiring the design parameters of the target object; generating predicted performance values ​​through a prediction model; mapping the predicted performance values ​​to theoretical rebound values ​​and fluctuation ranges; converting the on-site measured rebound values ​​of the target object into measured estimated performance values ​​when the on-site measured rebound values ​​exceed the fluctuation range; comparing the measured estimated performance values ​​with the predicted performance values; triggering a deviation warning when the deviation comparison confirms an anomaly; acquiring the performance data of the target object and retrospectively estimating the true performance values; comparing the true performance values ​​with the predicted performance values, locating at least one weight parameter in the prediction model that causes the deviation, denoted as the target weight parameter, and performing self-calibration of the target weight parameter with a micro-step size.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the prediction model is a strength prediction model based on a backpropagation neural network. The specific construction and application process includes: acquiring historical production data containing an input feature set, mixing parameters, and output labels. The input feature set includes water-cement ratio, unit water consumption, total cementitious material content, sand ratio, mineral admixture dosage, and additive dosage. The output label is the measured compressive strength value. A three-layer backpropagation neural network is constructed. The number of nodes in the input layer corresponds to the dimension of the input feature set, the number of nodes in the hidden layer is determined through trial and error, and the number of nodes in the output layer is used to output the predicted performance value. A particle swarm optimization algorithm is used to globally optimize the initial weight matrix and bias threshold of the backpropagation neural network. The mean squared error of the backpropagation neural network on the validation set is used as the fitness function to iteratively search for the optimal initial parameters. The backpropagation neural network is initialized using the optimal initial parameters, and the prediction model is obtained using historical production data as training samples. The design mix proportion parameters of the target object are input into the prediction model, and after forward propagation calculation, the predicted performance value is output.

[0007] In conjunction with the first aspect mentioned above, one possible implementation involves mapping the predicted performance value to the theoretical rebound value and the fluctuation range. Specifically, this includes: acquiring a historical dataset containing measured compressive strength values ​​and rebound values ​​measured by a rebound hammer. Using the measured compressive strength value as the independent variable and the rebound value measured by the rebound hammer as the dependent variable, a least squares regression fitting is performed to construct a strength-rebound value mapping function. The predicted performance value is used as input and substituted into the strength-rebound value mapping function to obtain the theoretical rebound value. This theoretical rebound value serves as the benchmark anchor for static comparison with the field-measured rebound value. The standard deviation of the predicted residuals generated during the regression fitting process is calculated. Based on the standard deviation of the predicted residuals and the confidence coefficient, a fluctuation range is generated centered on the theoretical rebound value. This fluctuation range serves as the dynamic tolerance boundary for determining whether statistical deviation occurs in the field-measured rebound value.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the process of triggering a deviation warning specifically includes: inputting the predicted performance value into the strength-rebound value mapping function to obtain the theoretical rebound value and its fluctuation range; comparing the fluctuation range of the theoretical rebound value with the measured rebound value of the rebound hammer; if the measured rebound value of the rebound hammer exceeds the fluctuation range, marking a first anomaly; based on the strength measurement curve, converting the measured rebound value of the rebound hammer into a measured estimated strength; based on the maturity equivalence principle, converting the predicted performance value to the expected converted strength of the same age as the measured estimated strength; calculating the relative deviation between the measured estimated strength and the expected converted strength; if the relative deviation exceeds a preset threshold, marking a second anomaly; only when both the first and second anomaly markers are marked is an anomaly confirmed, and a deviation warning is triggered.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves converting the predicted performance value to the expected converted strength at the same age as the measured estimated strength. Specifically, this includes: obtaining the predicted performance value, the baseline maturity, and the equivalent maturity of the target concrete entity from pouring to the test age. The baseline maturity and the equivalent maturity are then input into a preset strength development function to obtain the corresponding theoretical strength development coefficients. Based on the maturity equivalence principle, the expected converted strength is calculated using the theoretical strength development coefficients.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the process of identifying at least one weight parameter in the prediction model that causes the deviation as the target weight parameter specifically includes: In response to a deviation warning, acquiring performance data of the batch of concrete for which the warning was issued, including measured compressive strength values ​​and corresponding temperature history data; inputting the performance data into a strength backtracking model to obtain the backtracking estimated strength; calculating the strength deviation between the backtracking estimated strength and the predicted performance value, and determining the unidirectionality of the strength deviation; in response to the unidirectionality of the strength deviation, performing a correlation analysis on the connection weights with the cementitious material in the prediction model to obtain statistical correlation; and determining the weights with statistical correlations higher than a correlation threshold as the target weight parameters.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves the self-calibration process of the target weight parameters using micro-steps. Specifically, this includes: determining the calibration direction based on the intensity deviation between the backtracked inference intensity and the predicted performance value; calling a preset micro-step size and adjusting the target weight parameters according to the calibration direction to obtain the adjusted weights; and using a validation set to verify and self-calibrate the adjusted weights.

[0012] Secondly, a concrete strength prediction system based on mix proportion parameters is provided, comprising: a communication unit for acquiring design parameters, field-measured rebound values, and performance data of the target object; and a processing unit including a prediction engine module, a dual cross-validation module, and a self-calibration module; the self-calibration module is used to respond to deviation warnings, back-estimate the true performance value based on performance data, locate the target weight parameters causing deviations in the prediction model by comparing the true performance value with the predicted performance value, and self-calibrate the target weight parameters with micro-steps.

[0013] In conjunction with the second aspect mentioned above, in one possible implementation, the prediction engine module specifically includes: a prediction model unit, used to build and run an intensity prediction model based on a backpropagation neural network, and generate predicted performance values ​​based on design parameters; and a mapping model unit, used to build and run an intensity-rebound value mapping function, and map the predicted performance values ​​to theoretical rebound values ​​and their fluctuation ranges.

[0014] In conjunction with the second aspect mentioned above, in one possible implementation, the dual cross-validation module specifically includes: a rebound range verification unit, used to compare the fluctuation range of the theoretical rebound value with the field-measured rebound value, and mark the first anomaly based on the comparison result; and an intensity range verification unit, used to convert the field-measured rebound value into a measured estimated intensity based on the intensity measurement curve, and to convert the predicted performance value into an expected converted intensity based on the maturity equivalence principle, and to mark the second anomaly based on a preset threshold by calculating the relative deviation between the measured estimated intensity and the expected converted intensity. A deviation warning is triggered only when both the first and second anomaly markers are marked.

[0015] This application provides a method and system for predicting concrete strength based on mix proportion parameters. It maps predicted performance values ​​to theoretical rebound values ​​and their fluctuation ranges, enabling a direct correlation between predicted values ​​and readily measurable physical quantities on-site. Compared to existing technologies that only estimate the range of unidirectional strength values, the theoretical rebound value generated by this invention is a theoretical anchor point with clear physical meaning and can be directly compared with rebound hammer readings. This allows for high-information cross-validation at the moment of on-site rebound testing, extending the prediction window from obtaining the 28-day specimen strength to any age after concrete hardening, thus solving the core problem of severe prediction lag. Simultaneously, through a dual cross-validation mechanism, it abandons the single and rigid threshold comparison logic, technically achieving the superposition of quantitative uncertainty (fluctuation range) and logically redundant judgment (rebound-strength dual conditions). This mechanism absorbs noise such as random fluctuations in rebound values ​​and operational errors through the fluctuation range, and eliminates conversion model errors through secondary verification of the strength value range. Only systematic deviations that are inherent to the material and exhibit consistency in both dimensions will trigger an effective warning, accurately solving the problem of frequent false anomaly alarms and significantly reducing the workload of ineffective investigation. Attached Figure Description

[0016] Figure 1 A schematic flowchart illustrating a method for pre-determining concrete strength based on mix proportion parameters, provided in an embodiment of this application; Figure 2 A flowchart illustrating the construction and application of a backpropagation neural network prediction model in a concrete strength prediction method based on mix proportion parameters provided in this application embodiment; Figure 3 A schematic diagram illustrating the generation of theoretical rebound value and fluctuation range in a concrete strength pre-determination method based on mix proportion parameters provided in this application embodiment; Figure 4 A flowchart illustrating the double cross-validation-triggered deviation warning process in a concrete strength pre-judgment method based on mix proportion parameters provided in this application embodiment; Figure 5 The flowchart of target weight parameter positioning and micro-step self-calibration in a concrete strength pre-determination method based on mix proportion parameters provided in the embodiments of this application; Figure 6 A system architecture diagram for pre-determining concrete strength based on mix proportion parameters is provided in this application embodiment; Figure 7 This is an example diagram illustrating the application of a concrete strength pre-determination system based on mix proportion parameters, provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0019] Example 1: like Figure 1 As shown, this embodiment provides a method for pre-determining concrete strength based on mix proportion parameters. This method constructs a complete closed-loop logical framework from mix design to on-site testing and model self-evolution, specifically including the following six core steps: Step 100: Obtain the design parameters of the target object and generate predicted performance values ​​through the prediction model.

[0020] The predicted performance value refers to the compressive strength value that concrete is expected to reach under standard curing conditions, calculated by the prediction model based on the current input mix proportion parameters, such as the 28-day standard compressive strength. The prediction model refers to a mathematical model that can establish a nonlinear mapping relationship from multidimensional mix proportion parameters to the final strength. Its specific form is not limited to a particular algorithm; for example, it can be a multiple linear regression model, a support vector machine regression model, or a neural network model, which will be discussed in detail in subsequent embodiments, as long as it can receive design parameters and output a definite predicted strength value.

[0021] Before the actual project commences or during the mix design phase, the system receives design parameters for the target concrete object, such as water-cement ratio, sand ratio, and unit water consumption. These parameters are then used as feature vectors and input into a pre-trained prediction model. The model then performs complex internal mathematical calculations to output a specific predicted performance value. This predicted performance value serves as the source benchmark for the entire closed-loop prediction chain, providing the theoretical starting point for subsequent physical quantity mapping and comparison.

[0022] It should be understood that although this embodiment uses 28-day compressive strength as a preferred example of the predicted performance value, in other embodiments, the predicted performance value may also be a strength index of a longer age, such as 56 days or 90 days, depending on the requirements of project acceptance.

[0023] Step 200: Map the predicted performance values ​​to theoretical rebound values ​​and fluctuation ranges.

[0024] The theoretical rebound value refers to the theoretical physical quantity value corresponding to the reading of the non-destructive testing instrument in the field, obtained by converting the predicted performance value through a specific physical mapping relationship. For example, the theoretical rebound value can serve as a benchmark anchor point for static comparison with the measured data in the field. The fluctuation range refers to the numerical range that fluctuates around the theoretical rebound value. It represents the normal tolerance boundary caused by uncertainties such as the inherent discreteness of the material and the error of the test operation, and is a dynamic statistical confidence range.

[0025] Since the on-site measurements directly measure physical quantities such as rebound value rather than compressive strength, simply comparing the predicted strength with the measured rebound strength across dimensions can easily lead to distorted judgments due to conversion errors. Therefore, during its use, the predicted performance value needs to be converted into a theoretical rebound value through a mapping function, ensuring that subsequent comparisons are conducted within the same physical dimension. Simultaneously, the residual standard deviation is calculated based on the statistical discrete characteristics of historical data, and a fluctuation range is generated by combining it with the confidence coefficient. This provides a dynamic tolerance boundary for quantifying the uncertainty of the on-site measured data, making the comparison judgment no longer a rigid single-point threshold comparison, but a statistically inclusive interval judgment, effectively absorbing normal random fluctuation noise.

[0026] Step 300: In response to the field-measured rebound value of the target object, when the field-measured rebound value exceeds the fluctuation range, convert the field-measured rebound value into a measured estimated performance value.

[0027] Among them, the on-site measured rebound value refers to the average reading obtained after impacting a hardened concrete entity on the construction site using non-destructive testing equipment such as a rebound hammer. The measured estimated performance value refers to the estimated compressive strength value of the concrete at the current age, which is derived by reverse calculation from the on-site measured rebound value based on industry strength curves or special conversion formulas.

[0028] In some implementations, when the on-site testing age is reached, the measured rebound value is obtained, and it can be initially compared with the fluctuation range: If the measured rebound value falls within the fluctuation range, the current intensity development is considered normal, and there is no need to trigger the subsequent in-depth verification process. If the measured rebound value exceeds the fluctuation range, it indicates that the performance of the physical quantity on site has deviated from the statistical tolerance boundary expected by theory. At this time, the system converts the measured rebound value into the measured estimated performance value to prepare data for subsequent cross-dimensional strength value range verification.

[0029] This effectively establishes an operational logic of first filtering physical dimension ranges and then converting the intensity, avoiding the waste of resources that all measured data need to undergo complex conversions. It also ensures that only data showing initial signs of anomalies can enter the high-precision intensity verification stage.

[0030] Step 400: Compare the measured estimated performance value with the predicted performance value to determine the deviation. If the deviation comparison confirms an anomaly, trigger a deviation warning.

[0031] Deviation comparison refers to calculating and determining the numerical difference between the measured estimated performance value and the predicted performance value after converting them to the expected strength at the same age, in order to confirm whether the deviation belongs to a systematic anomaly inherent in the material. Deviation warning refers to a high-level alarm signal issued by the system to external terminals or control centers, indicating that there is a substantial risk to the concrete strength.

[0032] In some implementations, this step establishes a collaborative judgment rule based on dual cross-validation: the first condition is that the measured rebound value exceeds the fluctuation range; the second condition is that the relative deviation between the measured estimated performance value and the expected intensity exceeds a preset threshold. Only when both conditions are met simultaneously can the system confirm the anomaly and trigger a deviation warning.

[0033] Since single-dimensional deviations are very likely to be pseudo-anomalies caused by operational errors or local dispersion, while true systematic material deviations must show consistent deviations in both physical quantity dimensions (springback value) and mechanical quantity dimensions (estimated strength), this kind of dual redundancy judgment can be effectively utilized. By requiring both conditions to be met simultaneously, the system can effectively filter out a large number of false alarms of single-dimensional pseudo-anomalies, thereby improving the accuracy and anti-interference ability of on-site judgment.

[0034] Step 500: In response to the deviation warning, obtain the performance data of the target object and backtrack to estimate the true performance value.

[0035] Among them, performance data refers to the measured mechanical response and environmental history records of the early-age batch of concrete under warning, such as the 3-day or 7-day early compressive strength and the corresponding temperature history curves. The true performance value, on the other hand, refers to the estimated final strength of the batch of concrete at 28 days, derived from the early performance data through a strength backtracking model, reflecting the actual material activity and curing conditions under the current conditions.

[0036] In some implementations, when a deviation warning is triggered, the system no longer stops at comparing surface data but delves into the root cause of the deviation. This involves automatically extracting early-stage test block strength and temperature history data for the batch of concrete and inputting it into a strength backtracking model. The backtracking model then uses the material activity information contained in the early data to reconstruct the strength development trajectory of the batch under real-world conditions, thereby estimating the true performance values. This extends the verification process from simple numerical comparison to a physical backtracking of the material's essential state, providing a true and objective reference benchmark for subsequently locating weight deviations within the model.

[0037] Step 600: Compare the actual performance value with the predicted performance value, locate at least one weight parameter in the prediction model that causes the bias, use it as the target weight parameter, and perform self-calibration on the target weight parameter with a microstep size.

[0038] The target weight parameter refers to the model connection weights within the prediction model that have a high statistical correlation with the root causes of bias and have a significant impact on the output results, such as the weight coefficients characterizing the contribution of cementitious material activity. Microstep self-calibration refers to a smooth evolution process in which the target weight parameter is adjusted in a targeted, minute manner along the bias elimination direction with extremely small and fixed adjustment increments, and the effectiveness of the adjustment is confirmed on the validation set.

[0039] In some implementations, the deviation between the actual performance value and the predicted performance value is calculated, and it is analyzed whether the deviation exhibits continuous unidirectionality. If a unidirectional deviation is confirmed, the system performs correlation analysis on the weights in the prediction model that are connected to input features such as cementitious materials, and accurately locates the target weight parameters that cause the deviation.

[0040] Subsequently, instead of large-scale retraining or random adjustments, the system calls a preset micro-step size (e.g., 0.005) and adjusts the target weight parameters by a very small margin according to the calibration direction for bias elimination. The validation set is then used to verify whether the adjusted model output approaches the true value. This allows the model to slowly and smoothly absorb the systematic drift caused by raw material batch fluctuations without interrupting production services, avoiding model oscillations and production interruptions caused by blind retraining, and evolving a static, dead model into a living model with adaptive and self-evolving capabilities.

[0041] Based on the above technical solution, by constructing a closed-loop system encompassing prediction, mapping, verification, backtracking, and calibration, the theoretical rebound value is used as a precise comparison anchor point to achieve dynamic data mapping across stages. Accuracy is improved by filtering out false anomalies through dual-condition redundancy judgment. Finally, early performance data is used to backtrack and locate model weights, and micro-step smoothing self-calibration is performed, thus achieving real-time, accurate, and self-evolving closed-loop prediction. This solves the problem that existing technologies, due to static and fixed prediction models and the lack of cross-stage data mapping and self-learning mechanisms, cannot achieve real-time, accurate, and self-evolving closed-loop prediction of concrete strength from mix design to on-site testing.

[0042] Example 2: like Figure 2 As shown in Example 1, this example elaborates on the specific algorithm implementation of "generating predicted performance values ​​through a prediction model" in step 100. In this example, the prediction model is an intensity prediction model of a backpropagation neural network, and its specific construction and application process includes the following sub-steps.

[0043] Step 201: Obtain historical production data containing input feature set, mixing parameters and output labels. Input feature set includes water-cement ratio, unit water consumption, total amount of cementitious materials, sand ratio, mineral admixture dosage and admixture dosage. Output label is the measured value of compressive strength.

[0044] Among them, the backpropagation neural network (BPNN) is a multi-layer feedforward neural network trained by the error backpropagation algorithm. Its core mechanism is to use the error between the actual output and the expected output of the output layer to calculate the error gradient of each connection weight layer by layer from the output layer, and to correct the weights and thresholds according to the gradient descent method, so that the actual output of the network continuously approaches the expected output.

[0045] The input feature set refers to the set of parameters used to describe the full-dimensional information of concrete mix design. The seven types of parameters listed in this embodiment, such as water-cement ratio and unit water consumption, are the core factors affecting strength. It should be understood that in actual engineering, the input feature set can be expanded to include more dimensions of features such as concrete type (e.g., pumped, large volume) and admixture type, depending on the specific material system, as long as it can fully characterize the material mix characteristics.

[0046] Output labels refer to the supervisory signals used to calculate errors during network training; in this case, they are the measured compressive strength values ​​after 28 days of standard curing. Specifically, the system extracts a large number of verified mix proportion parameters and corresponding 28-day test block strength data pairs from historical production databases to construct a high-quality training sample set, providing a data foundation for subsequent supervised learning of the neural network.

[0047] Step 202: Construct a three-layer backpropagation neural network. The number of nodes in the input layer corresponds to the dimension of the input feature set. The number of nodes in the hidden layer is determined by trial and error. The number of nodes in the output layer is used to output the prediction performance value.

[0048] Among these, the trial-and-error method refers to an empirical optimization method that determines the optimal number of nodes by continuously trying different numbers of hidden layer nodes and observing the network's error performance on the validation set. The initial weight matrix is ​​the numerical matrix assigned to the connection weights between neurons in each layer during network initialization; its initial values ​​directly affect the convergence trajectory and final result of network training. Furthermore, the bias threshold is a constant term appended to the neuron input, used to adjust the neuron's activation conditions and improve the model's fitting flexibility.

[0049] In some implementations, a three-layer network structure is first constructed, consisting of an input layer, a single hidden layer, and an output layer. The number of nodes in the input layer strictly corresponds to the dimension of the input feature set, ensuring that all ratio parameters can be received by the network. The number of nodes in the hidden layer is determined through a trial-and-error method, with multiple rounds of training and validation conducted within a reasonable range (e.g., starting with half the number of nodes in the input layer and gradually increasing). The number of nodes that minimizes the validation error is selected as the final configuration. Finally, the number of nodes in the output layer is set to 1, specifically for outputting the predicted 28-day compressive strength value.

[0050] This three-layer structure design ensures that the network has sufficient nonlinear mapping capability to fit complex ratio and intensity relationships, while avoiding overfitting and training difficulties caused by excessively deep network layers.

[0051] Step 203: Using the particle swarm optimization algorithm, the initial weight matrix and bias threshold of the backpropagation neural network are globally optimized. The mean square error of the backpropagation neural network on the validation set is used as the fitness function to iteratively search for the optimal initial parameters.

[0052] Particle Swarm Optimization (PSO) is a swarm intelligence optimization algorithm that simulates the foraging behavior of bird flocks. It achieves global search by dynamically tracking individual and group optimum positions and velocities through particle flight and cooperation in the solution space. Mean squared error (MSE) is the average of the squared differences between the network's predicted and actual values, a key indicator of the network's prediction accuracy. Furthermore, the fitness function is a standard function for evaluating the quality of particle positions. In the scenario of PSO and BPNN fusion, the MSE of the BPNN on the validation set is defined as the fitness function, meaning that PSO seeks the initial weight and threshold combination that minimizes the BPNN's prediction error.

[0053] Since BPNN uses gradient descent for training, if the initial weight matrix and bias threshold are not randomly assigned, the network is very likely to get stuck in a local minimum on the error surface. That is, the training process seems to converge, but the actual error is still large. This is called the local optimum problem.

[0054] This step introduces the PSO algorithm, encoding all initial weights and thresholds of the BPNN as particle position vectors, and performing parallel searches within the global solution space. Each particle represents a combination of initial parameters, which is substituted into the BPNN for brief training, and the mean squared error is calculated on the validation set as fitness. PSO guides the particle swarm towards regions with smaller errors based on the fitness value. After multiple iterations, PSO can escape the trap of local minima, ultimately locating the valley region of the globally optimal error surface and outputting the optimal initial parameters. This allows the global macro-search capability of PSO to complement the local fine-grained gradient descent capability of the BPNN, avoiding the model's tendency to get trapped in local optima during initial training and improving the accuracy and stability of the prediction benchmark.

[0055] Step 204: Initialize the backpropagation neural network using the optimal initial parameters, and use historical production data as training samples to obtain the prediction model.

[0056] In some implementations, the optimal initial weight matrix and bias threshold obtained from PSO search are assigned to the corresponding parameters of the BPNN, completing the precise initialization of the network. Subsequently, historical production data is used as training samples and input into the initialized BPNN. At this point, the network undergoes conventional error backpropagation iterative training based on the gradient descent algorithm. Since the starting point is already in a globally optimal region, the network can quickly and stably converge to a high-precision fitting state, ultimately yielding a mature prediction model. This two-stage strategy of global optimization followed by local fine-tuning ensures that the model not only has high training efficiency but also that the final output prediction benchmark does not suffer from systematic inaccuracies due to initial randomness.

[0057] Step 205: Input the design mix ratio parameters of the target object into the prediction model, and output the predicted performance value after forward propagation calculation.

[0058] Forward propagation computation refers to the forward information flow process in which data flows from the network input layer through weight calculation and activation function processing, and finally reaches the output layer to generate prediction results.

[0059] In some implementations, the system receives the design mix proportion parameters of the target concrete object, arranges them into an input vector according to feature dimensions, and feeds it into a pre-trained prediction model. At this point, the data undergoes weighted summation and nonlinear activation operations in the hidden and output layers of the network, ultimately yielding a specific numerical value at the output layer node. This value represents the predicted 28-day compressive strength performance of the mix proportion under standard curing conditions. This data serves as the source data for subsequent mapping and comparison links, and its high accuracy directly determines the reliability of the entire closed-loop prediction system.

[0060] Based on the above technical solution, a particle swarm optimization algorithm is introduced to globally optimize the initial weight matrix and bias threshold. Using mean squared error as the fitness function, the optimal initial parameters are iteratively searched, thus avoiding the risk of the network getting trapped in local optima from the outset. This ensures the global superiority of the prediction model's starting point, significantly improving the accuracy of the prediction benchmark and providing a solid and reliable theoretical benchmark anchor for subsequent cross-validation and self-calibration. This solves the problem that traditional backpropagation neural networks, which use random initialization strategies, are prone to convergence to local optima rather than global optima in complex high-dimensional nonlinear mapping spaces with varying proportions and intensities. This can easily lead to the initial weights and thresholds falling near local minima on the error surface, resulting in inaccurate prediction performance benchmarks and subsequent systemic biases in the entire closed-loop prediction chain.

[0061] Example 3: like Figure 3 As shown, based on Example 1, this example elaborates on the specific mathematical regression fitting implementation of step 200, "mapping the predicted performance value to the theoretical rebound value and fluctuation range". In this example, the mapping process specifically includes the following sub-steps.

[0062] Step 301: Obtain a historical dataset containing measured values ​​of concrete compressive strength and rebound values ​​measured by a rebound hammer.

[0063] The historical dataset refers to a set of paired data of the measured compressive strength of past batches of concrete under the same project or material system, obtained by destructive testing with a pressure testing machine at a standard curing period of 28 days, and the measured rebound value obtained by non-destructive testing with a rebound hammer on the corresponding entity.

[0064] In some implementations, a large number of real strength and rebound data pairs can be extracted from historical testing databases to directly reflect the real mapping relationship between the intrinsic mechanical properties and the extrinsic physical response of a specific material system.

[0065] It should be understood that although this embodiment preferably uses data pairs with an age of 28 days as the historical dataset, in other embodiments, if the project acceptance standard points to a longer age of 56 days or 90 days, the historical dataset can also be adjusted to the measured paired data of the target age, as long as it can represent the mapping relationship under the final acceptance status.

[0066] Step 302: Using the measured compressive strength as the independent variable and the measured rebound value from the rebound hammer as the dependent variable, the least squares method is used for regression fitting to construct a mapping function between strength and rebound value.

[0067] The strength-rebound mapping function refers to the analytical expression that describes the mathematical dependence between the intrinsic compressive strength of concrete and the surface rebound value. For example, it can be a linear function, a polynomial function, or a power function.

[0068] In some implementations, in the physical mechanism of concrete, compressive strength is the intrinsic property of the material, while rebound value is the external physical response of intrinsic strength through surface hardness. Therefore, mapping from the inside to the outside with strength as the independent variable and rebound value as the dependent variable conforms to the physical causal logic.

[0069] The system substitutes the data obtained in step 201 into a preset function form and uses the least squares method for iterative calculation to find the function parameters that minimize the sum of squared residuals, thereby constructing a strength-rebound value mapping function specifically for the current material system. This operation of fitting a dedicated function based on real data avoids the problem of poor applicability caused by a unified strength measurement curve, making the mapping relationship highly consistent with the actual material properties of the current project.

[0070] Step 303: Substitute the predicted performance value into the strength-rebound value mapping function to obtain the theoretical rebound value. The theoretical rebound value is used as the benchmark anchor point for static comparison with the field measured rebound value.

[0071] The theoretical rebound value refers to the theoretically calculated rebound hammer reading that the concrete surface should exhibit at the expected strength, obtained by transforming the predicted performance value output by the prediction model through a mapping function. The benchmark anchor point, on the other hand, is a key value used as a fixed reference standard during data comparison; it does not fluctuate with random errors in on-site operations.

[0072] In some implementations, the predicted performance value is input as an independent variable into the mapping function. In this case, the function outputs a specific theoretical rebound value through forward calculation. This theoretical rebound value is derived entirely from the design mix proportion parameters and has not been affected by any random factors on-site, thus possessing theoretical determinism and self-consistency. It can then be used as a benchmark anchor point for subsequent field measurement data to determine deviations, thereby achieving a direct mapping path from abstract mix proportion parameters to specific on-site physical quantities.

[0073] Step 304: Calculate the standard deviation of the predicted residuals generated during the regression fitting process. Based on the standard deviation of the predicted residuals and the confidence coefficient, generate a fluctuation range centered on the theoretical rebound value. The fluctuation range serves as the dynamic tolerance boundary for determining whether the measured rebound value on site has statistical deviation.

[0074] The standard deviation of the predicted residual refers to the statistical dispersion measure of the deviation between the measured rebound values ​​and the predicted rebound values ​​of the mapping function during the regression fitting process. It quantifies the magnitude of random fluctuations that the mapping model itself cannot explain. The confidence coefficient is a multiplier factor selected according to statistical principles to determine the width of the fluctuation range. For example, the coefficient corresponding to a 95% confidence level is usually 1.96, and that corresponding to a 99% confidence level is 2.58. Finally, the dynamic tolerance boundary refers to the numerical range limit around the benchmark anchor point that can accommodate normal random dispersion without triggering false alarms.

[0075] In some implementations, the standard deviation of the fitted residuals is calculated, and a confidence coefficient is selected based on the engineering requirements for the stringency of the judgment. The two are then multiplied and added or subtracted from the theoretical rebound value on both sides to generate a fluctuation range. This range is not a rigid, fixed threshold, but a statistical boundary that dynamically adjusts with the magnitude of the predicted performance value and the material's discrete characteristics. When the measured rebound value falls within this range, the system considers the deviation to be within normal random dispersion; only when it exceeds this range is it considered to have a statistically significant deviation. This allows for the quantification of uncertainties such as the inherent discreteness of the material and operational errors, providing a scientific tolerance buffer for on-site judgment and effectively avoiding the problem of frequent false alarms under a single static threshold comparison.

[0076] Based on the above technical solution, by constructing a mapping function between strength and rebound value, and establishing a direct mapping path from mix proportion parameters to theoretical rebound value, and then generating a fluctuation range by calculating the standard deviation of the predicted residual and the confidence coefficient, a dynamic tolerance boundary for quantifying uncertainty in field measured data is provided. This allows for direct statistical comparison between theoretical predicted values ​​and readily measurable physical quantities in the field, fundamentally solving the problems of prediction lag and judgment distortion. This addresses the issue that existing technologies only perform unidirectional numerical calculations from mix proportion parameters to strength prediction, failing to establish a dynamic mapping path with readily measurable physical quantities such as rebound value in the field, resulting in a lag in the comparison between predicted values ​​and actual physical testing, which are essentially two independent systems.

[0077] Example 4: like Figure 4 As shown, based on Example 1, this example elaborates on the specific redundant judgment logic and time base conversion of "deviation comparison confirmation anomaly" in step 400. In this example, the process of triggering the deviation warning specifically includes the following sub-steps.

[0078] Step 401: Input the predicted performance value into the strength and rebound mapping model to obtain the theoretical rebound value and fluctuation range.

[0079] Among them, the strength and rebound mapping model refers to the strength and rebound value mapping function and its statistical characteristic model constructed as in Example 3. The fluctuation range refers to the numerical range that fluctuates around the theoretical rebound value, calculated based on the standard deviation of the prediction residual and the confidence coefficient. It represents the reasonable statistical boundary that the field-measured rebound value should fall into under a given probability.

[0080] In some implementations, the predicted performance value is input again into the strength-rebound mapping model. In this case, the predicted performance value not only outputs the theoretical rebound value as a benchmark anchor, but also its corresponding fluctuation range, such as the range at a 95% confidence level. This provides a statistically significant dynamic comparison benchmark for subsequent first-level rebound value range verification, making the judgment no longer a rigid single-point threshold comparison.

[0081] It should be understood that although this embodiment preferably uses a 95% confidence level, in engineering scenarios where there are special requirements for the rigor of the judgment, the confidence level can also be adjusted to 90% or 99%, as long as it can scientifically reflect the discrete characteristics of the material.

[0082] Step 402: Compare the fluctuation range of the theoretical rebound value with the actual rebound value measured by the rebound hammer. If the actual rebound value measured by the rebound hammer exceeds the fluctuation range, mark the first abnormality.

[0083] The first anomaly indicator refers to the preliminary anomaly status marker signal generated by the system in the physical dimension of the rebound value, which characterizes that the on-site nondestructive testing data has deviated from the theoretically expected statistical tolerance boundary. It can be a Boolean variable or a specific status code. The actual rebound value measured by the rebound hammer refers to the actual physical reading obtained from on-site testing.

[0084] In some implementations, the measured rebound value is compared with the fluctuation range: If the measured rebound value falls within the fluctuation range, it means that the performance of the physical quantity dimension conforms to the theoretically expected statistical fluctuation range, and no anomaly is marked. If the measured rebound value exceeds the upper or lower limit of the fluctuation range, it indicates that the surface hardness response of the actual object has deviated in a statistical sense, and the system will mark the first abnormality.

[0085] By introducing statistical tolerance of the fluctuation range, normal fluctuation noise caused by operational errors or local material dispersion can be effectively absorbed, avoiding the problem of frequent false alarms due to small deviations under a single static threshold comparison.

[0086] Step 403: Based on the strength measurement curve, convert the rebound value measured by the rebound hammer into the estimated strength measured by the rebound hammer.

[0087] Among them, the strength curve refers to a mathematical conversion curve or function relationship promulgated by national or industry standards, or customized for specific projects, used to convert rebound hammer readings into estimated values ​​of concrete compressive strength. The measured estimated strength refers to the estimated compressive strength value representing the current actual mechanical properties of the concrete structure, which is derived by using the strength curve to reverse the rebound value measured at the current age on site.

[0088] In some implementations, in order to cross from the physical quantity dimension to the mechanical quantity dimension for in-depth verification, the system calls a preset strength measurement curve, takes the rebound value measured by the rebound hammer that exceeds the fluctuation range as input, and calculates the estimated strength of the entity at the test age.

[0089] It should be understood that the selection of the strength test curve should be compatible with the current engineering material system. For example, a dedicated pumping strength test curve should be used for pumped concrete to ensure the accuracy of the conversion.

[0090] Step 404: Based on the maturity equivalence principle, the predicted performance value is converted to the expected converted intensity at the same age as the measured estimated intensity.

[0091] The maturity equivalence principle refers to the physical principle that the development of concrete strength is closely related to the cumulative effect of temperature and time (i.e., maturity). By equivalently converting the maturity under different curing processes, the strength at different ages or under different temperature conditions can be placed under the same comparable benchmark. The expected converted strength refers to the theoretical expected strength value at the same age and under the same curing conditions as the estimated strength measured on site, based on the maturity equivalence principle, after converting the predicted performance value of 28-day standard curing output by the prediction model.

[0092] Since the estimated strength obtained from on-site measurements is acquired at an early age (e.g., 7 days) under natural on-site curing conditions, while the predicted performance value is a theoretical value under 28-day standard curing conditions, a direct comparison between the two involves significant differences in time reference and curing environment, inevitably leading to serious comparison errors. Therefore, some implementation methods introduce the maturity equivalence principle, down-calculating the predicted performance value under 28-day standard curing conditions to the expected converted strength under equivalent on-site 7-day conditions. This ensures that subsequent comparisons are conducted under the same time and maturity scale, fundamentally eliminating systematic comparison errors caused by age differences.

[0093] Step 4041: Obtain the predicted performance value, baseline maturity, and equivalent maturity of the target concrete entity from the time of pouring to the test age. The baseline maturity refers to the accumulated maturity value of concrete under standard curing conditions (e.g., 20℃) from pouring to 28 days of age, which is usually an integral function of time and temperature. The equivalent maturity of the concrete entity, on the other hand, refers to the accumulated maturity value of the target concrete entity under the actual temperature history on site from pouring to the current testing age, which is calculated by integrating the temperature history curve recorded by on-site temperature sensors.

[0094] In some implementations, the predicted performance value for 28 days is first obtained from the prediction model, and the baseline maturity after 28 days of standard curing is calculated. At the same time, the temperature history data of this batch of concrete from pouring to the current testing age is extracted from the on-site monitoring system, and the equivalent maturity of the solid is calculated through integral calculation.

[0095] It should be understood that the maturity calculation formula can use the classic Nurse-Saul function or the Arrhenius function, and the choice can be made according to the sensitivity of the specific material system.

[0096] Step 4042: Input the baseline maturity and the entity equivalent maturity into the preset intensity development function to obtain the corresponding theoretical intensity development coefficients.

[0097] The strength development function is a mathematical empirical formula or fitted curve describing the development of concrete compressive strength with increasing maturity. Its input is the maturity value, and its output is the strength development coefficient, which represents the proportion of the current maturity strength to the final 28-day strength. The theoretical strength development coefficient, on the other hand, is a dimensionless proportionality factor that characterizes the degree of strength development, calculated based on the benchmark maturity or the equivalent maturity of the concrete.

[0098] In some implementations, the baseline maturity is input into the strength development function to obtain the theoretical strength development coefficient under 28 days of standard curing (usually close to 1.0). Then, the equivalent maturity of the entity is input into the same strength development function to obtain the theoretical strength development coefficient for the current age in the field (e.g., 0.6 or 0.7). Thus, through a unified strength development function, the abstract time and temperature history can be transformed into a specific quantitative indicator of strength proportion.

[0099] Step 4043: Based on the maturity equivalence principle, calculate the expected converted intensity using the theoretical intensity development coefficient.

[0100] In some implementations, the 28-day predicted performance value is scaled and converted proportionally using the proportional relationship between the two theoretical intensity development coefficients obtained in step 4042.

[0101] For example, if the theoretical strength development coefficient over 7 days is 0.65, then the expected converted strength is equal to the predicted performance value multiplied by 0.65. This conversion method effectively ensures that the expected converted strength and the measured estimated strength are compared under the same maturity benchmark, completely eliminating the interference of age and curing temperature differences.

[0102] Step 405: Calculate the relative deviation between the measured estimated strength and the expected converted strength. If the relative deviation exceeds the preset threshold, mark the second anomaly.

[0103] The relative deviation refers to the percentage of the absolute value of the difference between the measured estimated strength and the expected converted strength relative to the expected converted strength, quantifying the degree of deviation in the mechanical dimension. The preset threshold, on the other hand, refers to the percentage limit set according to engineering quality control standards for determining whether a strength deviation constitutes a substantial risk, such as 10% or 15%. Furthermore, the second anomaly indicator refers to a deep anomaly status marker signal generated by the system in the strength mechanical dimension, characterizing that the development of the on-site entity's strength has deviated from the theoretical expected converted value.

[0104] In some implementations, the relative deviation between the measured estimated strength and the expected converted strength is calculated and compared with a preset threshold. If the relative deviation is within the threshold, it indicates that the mechanical properties are still normal; if the relative deviation exceeds the threshold, it indicates that not only is the surface physical response abnormal, but the internal mechanical properties have also deviated substantially, and a second anomaly marker is then indicated.

[0105] It should be understood that the preset threshold is not fixed. For high-strength concrete or critical structural parts, the threshold can be set more strictly (e.g., 8%), while for ordinary concrete it can be appropriately relaxed (e.g., 12%) to accommodate different risk tolerance levels.

[0106] Step 406: Only when both the first abnormality flag and the second abnormality flag are marked is an abnormality confirmed and a deviation warning is triggered.

[0107] Among them, deviation warning refers to a high-level alarm signal sent by the system to external terminals, indicating that there is a substantial high risk to the concrete strength.

[0108] This step establishes a collaborative judgment rule for dual cross-validation: only when both the rebound value range verification (first anomaly indicator) and the strength value range verification (second anomaly indicator) confirm an anomaly can the system ultimately determine it as a systemic anomaly in the material's nature and trigger a deviation warning.

[0109] If only the first anomaly marker is marked while the second anomaly marker is not marked, it indicates that the physical quantity may be a one-dimensional pseudo-anomaly caused by surface carbonization or operational error; the reverse is also true.

[0110] The mechanism of this dual redundancy judgment lies in the fact that a true systematic material deviation must exhibit consistent deviations in both the physical and mechanical dimensions, while a single-dimensional deviation is highly likely to be random noise. Therefore, by requiring both conditions to be met simultaneously, the system can effectively filter out a large number of false alarms from single-dimensional anomalies, improving the accuracy and anti-interference capability of on-site judgment.

[0111] Based on the above technical solution, by introducing a dual cross-validation collaborative judgment rule of rebound and intensity value ranges, the uncertainty of measured values ​​can be absorbed through the fluctuation range, and the conversion model error can be eliminated through secondary verification of the intensity value range. Anomalies are only confirmed when both conditions are met simultaneously, effectively filtering out false anomalies. Simultaneously, based on the maturity equivalence principle and intensity development function, the 28-day predicted performance value is accurately converted to the expected converted intensity of the same age as the measured estimated intensity, eliminating comparison errors caused by age differences and ensuring that the dual cross-validation is conducted under the same time benchmark, thus fundamentally improving the accuracy of deviation judgment. This solves the technical problem of existing technologies directly comparing a single measured value with a broad intensity range, lacking refined mapping and uncertainty quantification, leading to frequent false alarms at threshold edges. It also solves the problem of directly comparing the measured estimated intensity of early ages with the 28-day predicted intensity across ages, ignoring the significant differences in maintenance time and temperature history, resulting in serious systematic comparison errors.

[0112] Example 5: like Figure 5As shown in Example 1, this example elaborates on the specific correlation analysis and targeted adjustment of "backtracking to estimate the true performance value" in step 500 and "locating the target weight parameter" and "micro-step self-calibration" in step 600. In this example, the process of locating the target weight parameter and self-calibration specifically includes the following sub-steps.

[0113] Step 501: In response to the deviation warning, obtain the early performance data of the batch of concrete under warning. The early performance data includes the measured value of early-age compressive strength and the corresponding temperature history data.

[0114] Early performance data refers to the objective record set of the mechanical response and environmental history exhibited by concrete in the early stages after pouring (e.g., 1 day, 3 days, or 7 days). Early-age compressive strength measurements refer to short-term strength values ​​obtained by conducting accelerated or conventional decomposition tests on early-age specimens from the same batch. Furthermore, temperature history data refers to the sequence of temperature changes over time continuously recorded from the time of pouring using temperature sensors embedded in or attached to the concrete structure.

[0115] In some implementations, when the dual cross-validation mechanism triggers a deviation warning, the system no longer merely compares the predicted performance values ​​over 28 days with the estimated values ​​on-site. Instead, it automatically initiates query requests to the laboratory data terminal and the on-site IoT monitoring nodes to extract the measured 1-day and 3-day strength values ​​of the early-stage concrete test blocks for the batch under warning, as well as the temperature history curve of the corresponding physical structure. This extends the verification process from simple numerical comparison to a physical retrospective of the material's essential state, providing the most original and objective underlying data support for subsequently reconstructing the true strength trajectory.

[0116] It should be understood that although this embodiment preferably lists 1 day and 3 days as early age periods, in other embodiments, depending on the specific detection frequency of the project and the design input requirements of the backtracking model, the early age period can also be extended to include 7 days or only select a single key node such as 3 days, as long as it can provide sufficient information on the early activity characterization of the material.

[0117] Step 502: Input the early performance data into the strength backtracking model to obtain the backtracking inferred strength.

[0118] The strength backtracking model refers to a pre-trained mathematical model that can map and reconstruct the final strength at a longer age (e.g., 28 days) from a mixed feature vector of early short-term mechanical response and environmental temperature history. It can be a convolutional neural network, a recurrent neural network, or another time-series regression model. The backtracking estimated strength refers to the 28-day strength estimate of the current batch, calculated by the strength backtracking model based on real early data from the current batch, reflecting the batch's strength under real material activity and on-site curing conditions.

[0119] In some implementations, the early compressive strength sequence and temperature history sequence are concatenated and preprocessed to form a feature vector that conforms to the input specifications of the backtracking model. This vector is then fed into a well-trained strength backtracking model for forward inference. The backtracking model then extracts deep features of the early strength development rate and the cumulative effect of temperature to inversely deduce the hydration process of the material, ultimately outputting a backtracked estimated strength.

[0120] This allows us to leverage the inherent physicochemical bonding between the early and final properties of concrete, revealing the final strength potential of a batch of concrete under current real-world conditions without waiting for the long 28-day curing period. This provides a real and objective benchmark for evaluating the deviation of the prediction model.

[0121] Step 503: Calculate the strength deviation between the retrospective estimated strength and the predicted performance value, and determine the one-wayness of the strength deviation.

[0122] Intensity bias refers to the absolute or relative difference between the retrospective inferred intensity and the predicted performance value output by the prediction model, quantifying the degree to which the model prediction deviates from the actual physical state. The unidirectionality of intensity bias refers to the consistent trend of the bias direction (positive or negative) across multiple consecutive batches or tests; for example, five consecutive batches may exhibit a negative unidirectional bias where the retrospective inferred intensity is lower than the predicted performance value.

[0123] In some implementations, the difference between the retrospective inference strength and the predicted performance value of the current batch is calculated, and the deviation records of recent historical batches are retrieved for trend analysis. If the deviation shows random alternation between positive and negative, it indicates that it is a normal material discrete fluctuation and there is no need to trigger the underlying calibration of the model. If the deviation exhibits a clear one-way tendency, such as multiple consecutive batches of retrospective inferences showing a one-way deviation of 2 to 3 MPa lower, then the system determines that the one-way tendency is valid.

[0124] The mechanism behind this judgment logic lies in the fact that random discreteness is an inherent property of materials, while one-way deviation clearly indicates a systematic overestimation or underestimation of the contribution of a certain type of material within the model, usually stemming from essential factors such as batch-to-batch activity fluctuations in cementitious materials. Therefore, by rigorously screening one-way deviation conditions, the risk of erroneously triggering model self-calibration due to occasional discrete noise can be effectively avoided.

[0125] Step 504: In response to the unidirectionality of strength deviation, perform correlation analysis on the connection weights with cementitious materials in the prediction model to obtain statistical correlation.

[0126] In this context, connection weights refer to the connection coefficients between neurons in the input layer representing cementitious material parameters and neurons in the hidden layer of a prediction model (such as a backpropagation neural network). Their magnitude and sign directly determine the contribution of cementitious material activity to the final predicted intensity. Statistical correlation, on the other hand, refers to a quantitative indicator calculated using mathematical statistical methods (such as Pearson correlation coefficient or sensitivity analysis) that characterizes the degree of linear or nonlinear correlation between model output deviation and changes in specific connection weights.

[0127] In some implementations, once the deviation is confirmed to be unidirectional, the system locks the search range of the root cause of the deviation and no longer performs an indiscriminate scan of all model weights. Instead, it specifically extracts the cementitious material characteristic parameters such as "total amount of cementitious materials" and "mineral admixture content" from the input layer of the prediction model to the connection weight matrix of the hidden layer.

[0128] By injecting small perturbations into these weights and observing the direction and magnitude of changes in the predicted output performance values, the statistical correlation between each weight and the strength deviation can be calculated. This allows for precise narrowing of the error root cause location to the subset of weights most sensitive to material activity, providing a precise point of application for subsequent targeted calibration and avoiding inaccurate positioning and wasted computational resources caused by blindly searching the entire high-dimensional parameter space.

[0129] Step 505: Determine the weights of statistically correlated parameters that are higher than the correlation threshold as target weight parameters.

[0130] The relevant threshold refers to the critical numerical limit preset based on model structure experience or engineering control accuracy requirements, used to screen highly sensitive weights. For example, a threshold is set where the absolute value of the Pearson correlation coefficient is greater than 0.7. The target weight parameter refers to the few key connection weights that are ultimately selected by the system, directly applied to subsequent microstep self-calibration, and have a strong statistical binding relationship with the intensity deviation.

[0131] In some implementations, the statistical correlation values ​​of each connection weight calculated in step 504 are compared one by one with a preset correlation threshold. Low-sensitivity weights below the threshold are filtered out, and only those weights with significantly excessive correlation values ​​are retained. This set is then determined as the target weight parameter. Thus, through the dynamic screening design of the correlation threshold, it is ensured that each calibration accurately targets the core factors causing the current systematic bias, achieving targeted intervention on the model's internal active characterization mechanisms.

[0132] It should be understood that the number of target weight parameters is not fixed. It depends on the actual activation path inside the model under the current bias scenario. It may contain only 1 to 2 core active weights, or it may contain a small cluster of associated weight recombinations.

[0133] The process of self-calibrating the target weight parameters using micro-step sizes specifically includes: Step 601: Determine the calibration direction based on the strength deviation between the retrospective estimated strength and the predicted performance value.

[0134] The calibration direction refers to the mathematical trend of increase or decrease that the target weight parameters should follow when adjusting their values. For example, when the retrospective inference strength is lower than the predicted performance value in one direction, the calibration direction should be to decrease the value of the relevant weights, and vice versa.

[0135] In some implementations, the calibration direction is determined based on the sign of the strength deviation confirmed in step 5003: If the deviation is negative (the actual value is lower than the predicted value), it means that the model overestimates the contribution of the cementitious material to its activity, and the target weight parameter needs to be adjusted to decrease in order to reduce the predicted output. If the deviation is positive, then the opposite applies. This provides a clear physical guide for subsequent microstep adjustments, ensuring that each parameter fine-tuning proceeds on the correct path to eliminate systematic deviations and avoiding model oscillations caused by undirected random perturbations.

[0136] Step 602: Call the preset microstep size and adjust the target weight parameters according to the calibration direction to obtain the adjusted weight.

[0137] The preset microstep size refers to a very small and fixed numerical increment, such as 0.005 or 0.001, which is much smaller than the learning rate used in conventional neural network training. It is specifically used for slow and smooth incremental fine-tuning of the model. The adjusted weights refer to the new values ​​obtained by adding or subtracting the preset microstep size from the original values ​​of the target weight parameters.

[0138] In some implementations, a preset microstep size constant is retrieved from the configuration file and multiplied by the calibration direction sign determined in step 6001 to obtain a small directional adjustment. This small adjustment is then applied to each target weight parameter located in step 505, causing a very small shift in its value, thus obtaining the adjusted weight.

[0139] This micro-step adjustment mechanism avoids the traditional approach of large-scale retraining or gradient descent with a large learning rate when encountering model biases. Instead, it employs a smooth evolution strategy with extremely small step sizes. The underlying mechanism is that the activity drift caused by batch fluctuations in raw materials is usually slow and minute. Micro-step adjustment is not only sufficient to absorb these minute drifts but also ensures that the overall topology of the model does not undergo drastic changes. This allows the model to slowly and adaptively adapt to the properties of new materials without interrupting production services.

[0140] Step 603: Use the validation set to validate and self-calibrate the adjusted weights.

[0141] The validation set refers to a set of standard historical data, independent of the training samples and the current warning batch, used to objectively evaluate the generalization performance of the model. Validation and self-calibration involve substituting the adjusted weights into the prediction model, running forward inference on the validation set, and calculating the error to confirm whether the fine-tuning has effectively improved the model's accuracy without causing side effects. If effective, the adjusted weights are officially updated to the model's current weights, completing the self-calibration loop.

[0142] In some implementations, the adjusted weights obtained in step 602 temporarily replace the original target weight parameters in the prediction model, while keeping other non-target weights unchanged. Then, validation set data is input into this temporary model for inference testing. The mean squared error of the temporary model on the validation set is then compared with the mean squared error of the original model. If the error decreases significantly or remains stable and the predicted value approaches the true value, it indicates that the microstep calibration direction and magnitude are correct. Therefore, the adjusted weights are formally solidified into the prediction model to complete this round of self-calibration. If the error increases or fluctuates abnormally, it indicates that the step size may be too large or the direction judgment is incorrect. The system will then cancel the adjustment and record the failure log, waiting for the next batch of deviation signals to trigger a more careful calibration.

[0143] This closed-loop verification mechanism provides a final safeguard for microstep self-calibration, ensuring that each model evolution is rigorously tested against an objective validation set. This eliminates the risk of new systematic biases or oscillations that the calibration operation itself may introduce, thus guaranteeing the stability of the model during long-term self-evolution.

[0144] Based on the above technical solution, by utilizing early performance data backtracking to estimate intensity and determine the unidirectionality of deviation, correlation analysis is performed on the connection weights related to the activity of cementitious materials in the model to accurately locate the target weight parameters. Then, with a very small preset micro-step size, directional smoothing adjustments and closed-loop validation on the validation set are performed along the calibration direction. This precisely pinpoints the root cause of error to the weights representing the activity of cementitious materials in the model, achieving slow and smooth self-calibration of the prediction model. This allows the model to gradually adapt to new fluctuations in the activity of cementitious materials without interrupting production, evolving a static, dead model into a living model with adaptive and self-evolving capabilities. This maintains high accuracy over the long term and avoids the continuous accumulation of systematic deviations and the oscillation risks caused by blind retraining. This solves the technical problem caused by the parameter solidification of existing prediction models once training is complete.

[0145] Example 6: like Figure 6As shown, this embodiment provides a concrete strength pre-determination system based on mix proportion parameters. This system is applied to a concrete strength pre-determination method based on mix proportion parameters as described in any of the preceding embodiments 1 to 5. The system constructs a complete closed-loop architecture from the perspectives of hardware and logic modules, specifically including two core components: a communication unit and a processing unit.

[0146] The communication unit is used to acquire the design parameters of the target object, the measured rebound value on site, and early performance data.

[0147] The communication unit refers to the hardware interface module or logical component inside the system that is responsible for exchanging information with external data sources. Its physical form can be a network communication adapter, a wireless radio frequency module, an IoT gateway data receiving port, or an API call interface that directly reads the local database.

[0148] The system continuously listens for and responds to data requests during operation: In the early design stage of the project, it retrieves the design parameters of the target concrete object, such as water-cement ratio, sand ratio, and total amount of cementitious materials, from the mix design terminal or production control system, and transmits them to the processing unit. During the on-site testing phase, it receives measured rebound value data uploaded by the on-site rebound hammer equipment; During the retrospective phase following the triggering of the deviation warning, it initiates queries to the laboratory data server and on-site temperature sensor nodes to extract the measured values ​​of the early-age compressive strength of the batch of concrete under warning and the corresponding temperature history data.

[0149] It should be understood that although the communication unit is preferably described as a unified data aggregation gateway in this embodiment, in other embodiments, depending on the field network topology and device heterogeneity, the communication unit can also be deployed in a distributed manner as multiple independent sub-interfaces, each specifically responsible for receiving design parameters, receiving field bounce data, and receiving early IoT data, as long as it can complete the function of fully aggregating the three types of heterogeneous source data and transmitting them to the processing unit.

[0150] This allows the communication unit to serve as a unified front-end portal for the system's interaction with the physical world, shielding it from the diversity of underlying device protocols and ensuring that the processing unit can acquire the complete, real-time, multi-source heterogeneous data streams required for the entire link, thus providing a solid data supply foundation for closed-loop prediction.

[0151] The processing unit includes a prediction engine module, a dual cross-validation module, and a self-calibration module, which are used to run the prediction model to generate prediction performance values ​​and map the prediction performance values ​​to theoretical rebound values ​​and their fluctuation ranges.

[0152] When the measured rebound value exceeds the fluctuation range, the measured rebound value is converted into a measured estimated performance value, and the measured rebound value is converted into a measured estimated strength. Based on the maturity equivalence principle, the predicted performance value is converted into an expected converted strength of the same age as the measured estimated strength. The relative deviation between the measured estimated strength and the expected converted strength is calculated, and whether it is abnormal is determined according to a preset threshold. A deviation warning is triggered only when the measured rebound value exceeds the fluctuation range and the relative deviation exceeds the preset threshold.

[0153] In response to deviation warnings, the true performance value is estimated by backtracking based on early performance data. The target weight parameters in the prediction model are located by comparing the true performance value with the predicted performance value, and the target weight parameters are self-calibrated with micro-steps.

[0154] The processing unit refers to the central control core within the system, capable of data processing, logical judgment, and model execution. Its physical carrier can be the main processor of a high-performance industrial control computer, a virtual computing instance of a cloud server, or a specially deployed edge computing AI chip module. The prediction engine module, the dual cross-validation module, and the self-calibration module are three core functional components with decoupled logic within the processing unit. They respectively encapsulate the algorithm logic and control instructions for forward prediction, redundancy verification, and feedback correction.

[0155] Specifically, the three main modules within the processing unit work closely together with the control signal flow via an internal data bus. During operation, the prediction engine module first receives the design parameters from the communication unit, runs the trained prediction model and mapping function within it, performs forward calculations to generate predicted performance values, and further maps them to theoretical rebound values ​​and their fluctuation ranges. These forward prediction results are then sent to the dual cross-validation module.

[0156] When the communication unit delivers the measured rebound value from the field, the dual cross-validation module immediately initiates the dual-dimensional judgment logic, compares the measured rebound value with the fluctuation range, and converts it into the measured estimated performance value when it exceeds the range. At the same time, it converts the predicted performance value into the expected converted intensity for deviation comparison. Only when both conditions are met and an anomaly is confirmed, does the dual cross-validation module send a deviation warning control signal to the self-calibration module.

[0157] Finally, in response to the warning signal, the self-calibration module obtains early performance data through the communication unit, runs the intensity backtracking model internally to estimate the true performance value, compares and locates the target weight parameters, and performs directional adjustment and verification with extremely small microsteps. After completing the self-calibration, the updated weight parameters are written back to the prediction model inside the prediction engine module, thus completing a complete closed-loop iteration.

[0158] It should be understood that although this embodiment describes the three modules as a logical division within the processing unit, in large-scale distributed computing scenarios, these three modules can also be deployed on different physical computing nodes and remotely invoked and coordinated through a microservice architecture, as long as their control connection relationship and data flow logic remain closed loop. This system architecture design opens up the entire link of data flow and control connection from forward prediction of design parameters, dual cross-validation of on-site data to early data backtracking and self-calibration, making the system no longer a one-way open-loop alarm, but evolving into a closed-loop intelligent entity that is interconnected and self-healing. At the system level, it achieves complete integrated management from source prediction to on-site verification and model self-evolution.

[0159] Example 7: like Figure 6 As shown, based on Example 6, this example elaborates on the specific sub-module structure binding of the prediction engine module and the dual cross-validation module within the processing unit: First, the prediction engine module specifically includes a prediction model unit. This prediction model unit refers to an independent hardware logic circuit or software microserver within the processing unit that is specifically responsible for performing nonlinear mapping operations from multidimensional ratio parameters to long-term compressive strength. It encapsulates the weight matrix and bias threshold parameters of the PSO-BPNN model, which has completed particle swarm optimization and gradient training as detailed in Example 2, and is used to build and run an intensity prediction model based on a backpropagation neural network, and generate prediction performance values ​​based on design parameters.

[0160] In some implementations, after the communication unit transmits the design parameters of the target object to the processing unit, the prediction model unit is activated as the first stage entry point for forward computation. It receives feature vectors such as water-cement ratio and sand ratio, and performs forward propagation calculations of weighted summation and nonlinear activation of a multi-layer neural network internally. Finally, it generates a predicted performance value representing the expected compressive strength after 28 days at the output node.

[0161] It should be understood that although this embodiment preferably describes the prediction model unit as a dedicated unit that runs a backpropagation neural network, in other system architecture deployments, the prediction model unit can also be configured as an execution engine that runs other types of prediction algorithms such as support vector machine regression or random forest, as long as it can stably output a determined prediction performance value based on the design parameters.

[0162] The advantage of this design, which encapsulates the prediction model as a dedicated unit, is that it physically or logically decouples the deep learning inference process, which consumes the most computational resources and has the highest algorithm update frequency, from other mapping logic. When only the neural network structure needs to be optimized or the training samples need to be changed, the system can perform hot updates on only the prediction model unit without downtime and reconstructing the entire prediction engine module, which greatly improves the modularity and computational maintenance efficiency of the system's forward prediction architecture.

[0163] Secondly, the prediction engine module also includes a mapping model unit. The mapping model unit refers to an independent logic operation module within the processing unit that is specifically responsible for converting abstract mechanical strength indicators into on-site measurable physical quantity indicators and their statistical tolerance boundaries. It internally contains parameters of a specific engineering strength and rebound value mapping function fitted based on the least squares method as shown in Example 3, as well as the corresponding prediction residual standard deviation and confidence coefficient. This is used to construct and run the strength and rebound value mapping function and map the predicted performance value to the theoretical rebound value and its fluctuation range.

[0164] In some implementations, the predicted performance value generated by the prediction model unit is not directly output to the verification module. Instead, it is first sent to the mapping model unit, which is also within the prediction engine module. The mapping model unit receives the predicted performance value as an independent variable, substitutes it into its internally stored mapping function for forward calculation, and simultaneously calls the internally fixed residual standard deviation and confidence coefficient to calculate and generate the fluctuation range centered on the theoretical rebound value.

[0165] This allows for the separation of complex deep learning nonlinear mappings from linear or polynomial mappings based on physical statistical regression, enabling the two stages to employ different computational precision and parallel strategies. For example, the prediction model unit can be deployed on a cloud GPU cluster for high-load inference, while the mapping model unit can be deployed on an edge computing gateway for lightweight and fast conversion. This achieves refined allocation of computing resources and efficient collaboration of decoupled two-stage processing at the overall system architecture level.

[0166] Furthermore, the dual cross-validation module specifically includes a rebound value range verification unit. This rebound value range verification unit refers to an independent judgment logic component within the dual cross-validation module that is specifically responsible for performing preliminary anomaly screening in the physical quantity dimension. It only receives and processes data streams directly related to the rebound meter readings, and is used to compare the fluctuation range of the theoretical rebound value with the actual rebound value measured on site, and mark the first anomaly based on the comparison result.

[0167] In some implementations, after the communication unit transmits the measured rebound value to the processing unit, the data stream is first routed to the rebound value range verification unit. This unit obtains the theoretical rebound value and its fluctuation range from the mapping model unit of the prediction engine module, and compares the measured rebound value with the upper and lower limits of the fluctuation range: if the measured rebound value falls within the fluctuation range, the rebound value range verification unit determines that the physical quantity dimension is normal and does not generate any abnormal flag signal; if the measured rebound value exceeds the upper or lower limit of the fluctuation range, it indicates that the surface hardness response of the physical entity has deviated from the theoretically expected statistical tolerance boundary, and the rebound value range verification unit immediately marks the first abnormal flag in the internal register or output port, for example, by setting a specific Boolean variable to True or outputting a high-level signal.

[0168] It should be understood that the specific physical or logical form of the first anomaly flag can be flexibly set according to the underlying bus protocol of the system, and its core function is only to characterize the deviation state of the rebound value range dimension. The advantage of setting up a dedicated rebound value range verification unit in the system architecture is that it separates the preliminary screening logic of the physical quantity dimension from the complex intensity conversion and comparison logic, enabling the system to quickly filter out a large amount of measured data that is obviously within the normal statistical fluctuation range at the data entry point with extremely low computational latency, preventing these normal data from entering the subsequent high-load intensity conversion process, thereby significantly improving the overall data throughput efficiency and real-time response speed of the system.

[0169] Meanwhile, the dual cross-validation module also includes an intensity range verification unit. This unit is a dedicated, high-precision computational component within the dual cross-validation module responsible for performing in-depth anomaly confirmation in the mechanical quantity dimension. It not only embeds industry-specific or custom-defined strength measurement curve conversion formulas but also integrates maturity equivalence conversion algorithms and intensity development functions, as described in Example 4. Based on the strength measurement curve, it converts the field-measured rebound value into a measured estimated intensity, and based on the maturity equivalence principle, converts the predicted performance value into an expected converted intensity. By calculating the relative deviation between the measured estimated intensity and the expected converted intensity, a second anomaly flag is marked according to a preset threshold.

[0170] In some implementations, the activation of the strength range verification unit typically depends on the preliminary screening results of the rebound range verification unit. Once the first anomaly flag is marked, the system control flow sends the field-measured rebound value exceeding the fluctuation range, along with the corresponding predicted performance value, to the strength range verification unit. This unit first calls the internal strength measurement curve to inversely convert the field-measured rebound value into the measured estimated strength. Subsequently, based on the maturity equivalence principle, the 28-day predicted performance value is converted to the expected converted strength at the same age as the measured estimated strength. Finally, the relative deviation between the two is calculated and compared with an internally preset threshold: if the relative deviation does not exceed the threshold, it indicates that the mechanical quantity dimension is still within normal dispersion, and the strength range verification unit does not mark an anomaly; if the relative deviation exceeds the threshold, it indicates that the internal mechanical performance has also deviated substantially, and the strength range verification unit marks the second anomaly flag.

[0171] Ultimately, a deviation warning is triggered only when both the first and second anomaly flags are marked. The dual cross-validation module internally incorporates a collaborative decision logic gate or software state machine that continuously monitors the first anomaly flag output by the rebound value range verification unit and the second anomaly flag output by the intensity value range verification unit. Only when both flag signals arrive simultaneously and are in a confirmed anomaly state will the collaborative decision logic send a deviation warning control signal to the self-calibration module and external terminals. If only the first anomaly flag is marked while the second anomaly flag is not, the system determines it to be a one-dimensional pseudo-anomaly of physical quantities caused by surface carbonization or operational error, and the collaborative decision logic blocks the warning output; conversely, if both flags are marked, the warning output is blocked.

[0172] This embodiment, by setting up a rebound value range verification unit and an intensity value range verification unit in the system to perform redundant cross-judgment in two dimensions, can fundamentally block the path of single-dimensional pseudo-abnormal signals to the downstream self-calibration module at the architecture level, eliminate the serious systemic risk of blind self-calibration of the model due to single-dimensional misjudgment, and improve the robustness and anti-interference capability of the prediction system at the system architecture level.

[0173] Example 8: like Figure 7 As shown, in order to more intuitively demonstrate the operational effect and closed-loop self-evolution capability of the technical solution of the present invention in a real engineering environment, this embodiment takes the on-site pouring and testing of C50 pumped concrete in a bridge project as a specific application scenario, and comprehensively demonstrates the application of the methods and systems described in the aforementioned embodiments 1 to 7.

[0174] It should be understood that the specific mix proportions, rebound readings, estimated strength values, and microstep constants listed in this embodiment are merely illustrative examples and not restrictive. In actual engineering, these parameters will be dynamically adjusted according to the specific material system and design specifications.

[0175] In this application scenario, the target material is C50 pumped concrete, and its specific design parameters include: a water-cement ratio of 0.35, a unit water consumption of 175 kg / m³, a total cementitious material content of 500 kg / m³ (including 350 kg / m³ of cement, 100 kg / m³ of fly ash, and 50 kg / m³ of slag powder), a sand ratio of 42%, and an admixture dosage of 1.2%. The system communication unit first obtains the above design parameters and transmits them to the prediction engine module of the processing unit.

[0176] The prediction model unit within the prediction engine module receives the feature vector of the design parameters, runs the trained PSO-BPNN prediction model for forward propagation calculation, and outputs a predicted performance value, namely, a predicted 28-day standard compressive strength of 52.4 MPa. This predicted performance value is then sent to the mapping model unit, which also belongs to the prediction engine module. At this point, the mapping model unit uses 52.4 MPa as the independent variable, substitutes it into the internally stored strength-rebound value mapping function, and calculates a theoretical rebound value of 38.5. Simultaneously, the mapping model unit calls upon the internally stored prediction residual standard deviation (e.g., σ=1.2) and 95% confidence coefficient (1.96) to generate a fluctuation range centered on the theoretical rebound value of 38.5: [38.5-1.96×1.2, 38.5+1.96×1.2], i.e., [36.2, 40.8]. This range serves as the dynamic tolerance boundary for determining whether the measured rebound value in the field exhibits statistical deviation.

[0177] When the concrete structure was poured and reached its 7-day curing period, on-site testing personnel used a rebound hammer to conduct an impact test on the structure, obtaining a measured rebound value of 34.8. The communication unit then transmitted this measured data to the dual cross-validation module of the processing unit. The rebound value range verification unit first initiated the first layer of verification, comparing the measured rebound value of 34.8 with the fluctuation range [36.2, 40.8]. Since 34.8 is lower than the lower limit of the fluctuation range of 36.2, the rebound value range verification unit determined that the dimension of the on-site physical quantity had deviated from the theoretically expected statistical tolerance boundary, and subsequently marked the first anomaly.

[0178] In response to the first anomaly indicator, the strength range verification unit initiated a second layer of in-depth verification: based on the industry-specific strength measurement curve, the field-measured rebound value of 34.8 was converted into a measured estimated strength at 7 days of age, resulting in a measured estimated strength of 35.6 MPa. Simultaneously, to eliminate comparison errors caused by age differences, the strength range verification unit, based on the maturity equivalence principle, converted the predicted performance value of 52.4 MPa to the expected converted strength at the same 7-day age as the measured estimated strength.

[0179] The specific conversion process is as follows: The system obtains the baseline maturity after 28 days of standard curing and the equivalent maturity of the solid at the site from pouring to 7 days of testing age (calculated by integrating the temperature history curve recorded by the on-site temperature sensor). Both are then input into a preset strength development function to obtain the theoretical strength development coefficient for 28 days (close to 1.0) and the theoretical strength development coefficient for 7 days (e.g., 0.65). The expected converted strength for 7 days is then calculated using the coefficient ratio: 52.4 × 0.65 = 34.1 MPa. The strength range verification unit calculates the relative deviation between the measured estimated strength of 35.6 MPa and the expected converted strength of 34.1 MPa as (35.6 - 34.1) / 34.1 = 4.4%. This deviation is lower than the preset threshold (e.g., 10%), indicating that the mechanical quantity dimension is still within normal dispersion, and the strength range verification unit does not mark a second anomaly.

[0180] Since only the first anomaly flag is marked while the second anomaly flag is not marked, the collaborative judgment logic inside the dual cross-validation module determines that this deviation is only a single-dimensional pseudo-anomaly in the dimension of surface physical quantities, which may be caused by local operational errors or surface carbonization. This blocks the output of deviation warnings and avoids the trouble caused to on-site engineers by invalid warnings.

[0181] As the project progressed, similar phenomena were observed in the rebound tests of three consecutive batches of C50 concrete at 7 days of age: the measured rebound values ​​consistently showed a unidirectionally low trend, and the converted estimated strength was consistently lower than the preset threshold of over 10% below the expected strength. For example, the fourth batch had a measured rebound value of 33.5 MPa, a converted estimated strength of 33.2 MPa, while the expected strength was 34.3 MPa, resulting in a relative deviation of 13.1%. At this point, the rebound value range verification unit in the dual cross-validation module marked the first anomaly, and the strength value range verification unit also marked the second anomaly. Only when both the first and second anomaly markers were marked did the collaborative judgment logic confirm a systematic deviation in the material's inherent nature, formally triggering a deviation warning.

[0182] Upon triggering the deviation warning, the self-calibration module within the processing unit responds immediately. The communication unit queries the laboratory data terminal and the on-site IoT node to obtain early performance data for the batch of concrete under warning, including the measured value of the 3-day early-age compressive strength (e.g., 18.5 MPa, significantly lower than expected) and the corresponding temperature history data. The self-calibration module inputs the aforementioned early performance data into the internally integrated strength backtracking model. By extracting the early strength development rate and temperature cumulative effect characteristics, the model inversely extrapolates the material's hydration process and outputs a backtracked estimated strength of 49.8 MPa.

[0183] The self-calibration module calculated the strength deviation between the retrospectively estimated strength of 49.8 MPa and the predicted performance value of 52.4 MPa to be -2.6 MPa, and retrieved recent historical batch records to determine the unidirectionality of this deviation. Since multiple consecutive batches showed a unidirectional negative deviation trend where the retrospectively estimated strength was lower than the predicted performance value, the system confirmed that the unidirectionality of the strength deviation was valid.

[0184] In response to this unidirectionality, the self-calibration module performs correlation analysis on the connection weights in the prediction model related to cementitious materials (especially fly ash content characteristics). The calculation shows that a certain hidden layer connection weight, characterizing the contribution of fly ash activity, has the highest statistical correlation, far exceeding the correlation threshold. Therefore, the system determines this weight as the target weight parameter. This precise positioning reveals that the root cause of the deviation lies in the recent fluctuations in fly ash batch activity, leading to an overestimation of its contribution by the model.

[0185] Based on the negative unidirectionality of the intensity deviation, the self-calibration module determines the calibration direction as reducing the value of the target weight parameter. Subsequently, the system calls a preset micro-step size of 0.005 to adjust the target weight parameter in the direction of reduction, subtracting 0.005 from its original value to obtain the adjusted weight. The self-calibration module uses an independent validation set to perform forward inference verification on the adjusted weight. The results show that the overall mean square error of the model decreases, and the predicted output moves closer to the true value, indicating successful verification. The system then formally writes the adjusted weight back to update the PSO-BPNN prediction model in the prediction engine module, completing this round of micro-step self-calibration closed loop. This smooth calibration with extremely small steps allows the model to slowly and accurately absorb the systematic drift caused by batch fluctuations in fly ash without interrupting production services, evolving the static dead model into a living model with adaptive and self-evolving capabilities, maintaining high-precision predictions over the long term.

[0186] This application example fully demonstrates the closed-loop operation effect of the entire chain, from forward prediction of design parameters to generate theoretical rebound values ​​and fluctuation ranges, to dual cross-validation of field measured data to filter false anomalies and accurately identify systematic deviations, and then to backtracking the fly ash activity weight based on early performance data and performing smooth self-calibration with a micro-step size of 0.005. This scenario intuitively confirms that the present invention effectively solves the core pain points of prediction lag, frequent false anomaly false alarms, and the accumulation of systematic deviations caused by model solidification by constructing a complete closed loop of prediction, mapping, verification, backtracking, and calibration. It achieves real-time, accurate, and self-evolving closed-loop prediction of concrete strength from mix design to field testing.

[0187] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention, such as replacing the backpropagation neural network with other machine learning models with nonlinear mapping capabilities, replacing the particle swarm optimization algorithm with other global optimization algorithms, adjusting the microstep size constant to other minimal values ​​that adapt to the drift characteristics of the current material system, or adjusting the early age to other time nodes that can characterize the material activity, should all be covered within the scope of protection of the present invention.

Claims

1. A method for pre-determining concrete strength based on mix proportion parameters, characterized in that, include: Obtain the design parameters of the target object and generate predicted performance values ​​through a prediction model; The predicted performance values ​​are mapped to theoretical rebound values ​​and fluctuation ranges; In response to the field-measured rebound value of the target object, when the field-measured rebound value exceeds the fluctuation range, the field-measured rebound value is converted into a measured estimated performance value; The measured estimated performance value is compared with the predicted performance value. When the deviation comparison is confirmed to be abnormal, a deviation warning is triggered. In response to the deviation warning, the performance data of the target object is obtained, and the true performance value is back-estimated. By comparing the actual performance value with the predicted performance value, at least one weight parameter in the prediction model that causes the deviation is located and denoted as the target weight parameter. The target weight parameter is then self-calibrated using a microstep size.

2. The method for pre-determining concrete strength based on mix proportion parameters according to claim 1, characterized in that, The prediction model is an intensity prediction model of a backpropagation neural network, and its specific construction and application process includes: Acquire historical production data containing an input feature set, mixing parameters, and an output label. The input feature set includes water-cement ratio, unit water consumption, total amount of cementitious materials, sand ratio, mineral admixture dosage, and admixture dosage. The output label is the measured value of compressive strength. A three-layer backpropagation neural network is constructed. The number of nodes in the input layer corresponds to the dimension of the input feature set. The number of nodes in the hidden layer is determined by trial and error. The number of nodes in the output layer is used to output the prediction performance value. The particle swarm optimization algorithm is used to globally optimize the initial weight matrix and bias threshold of the backpropagation neural network. The mean square error of the backpropagation neural network on the validation set is used as the fitness function to iteratively search for the optimal initial parameters. The backpropagation neural network is initialized using the optimal initial parameters, and the historical production data is used as training samples to obtain the prediction model; The design mix ratio parameters of the target object are input into the prediction model, and after forward propagation calculation, the predicted performance value is output.

3. The method for pre-determining concrete strength based on mix proportion parameters according to claim 1, characterized in that, The process of mapping the predicted performance value to the theoretical rebound value and fluctuation range specifically includes: Obtain a historical dataset containing measured compressive strength values ​​and rebound values ​​measured by a rebound hammer. Using the measured compressive strength as the independent variable and the measured rebound value from the rebound hammer as the dependent variable, the least squares method is used for regression fitting to construct a mapping function between strength and rebound value. The predicted performance value is used as input and substituted into the strength-rebound value mapping function to obtain the theoretical rebound value. The theoretical rebound value is used as a benchmark anchor point for static comparison with the field measured rebound value. The standard deviation of the predicted residuals generated during the regression fitting process is calculated. Based on the standard deviation of the predicted residuals and the confidence coefficient, a fluctuation range is generated with the theoretical rebound value as the center. The fluctuation range serves as the dynamic tolerance boundary for determining whether the measured rebound value in the field has a statistical deviation.

4. The method for pre-determining concrete strength based on mix proportion parameters according to claim 1, characterized in that, The process of triggering the deviation warning specifically includes: The predicted performance value is input into the strength-rebound value mapping function to obtain the theoretical rebound value and fluctuation range; The fluctuation range of the theoretical rebound value is compared with the actual rebound value measured by the rebound hammer. If the actual rebound value measured by the rebound hammer exceeds the fluctuation range, a first abnormality mark is made. Based on the strength measurement curve, the rebound value measured by the rebound hammer is converted into the measured estimated strength. Based on the maturity equivalence principle, the predicted performance value is converted to the expected converted intensity of the same age as the measured estimated intensity; Calculate the relative deviation between the measured estimated strength and the expected converted strength. If the relative deviation exceeds a preset threshold, mark a second anomaly. An anomaly is confirmed and the deviation warning is triggered only when both the first anomaly flag and the second anomaly flag are marked.

5. The method for pre-determining concrete strength based on mix proportion parameters according to claim 4, characterized in that, The process of converting the predicted performance value to the expected converted intensity at the same age as the measured estimated intensity specifically includes: Obtain the predicted performance value, the baseline maturity, and the equivalent maturity of the target concrete entity from the time of pouring to the test age; The baseline maturity and the entity equivalent maturity are respectively input into a preset intensity development function to obtain the corresponding theoretical intensity development coefficient; Based on the maturity equivalence principle, the expected converted intensity is calculated using the theoretical intensity development coefficient.

6. The method for pre-determining concrete strength based on mix proportion parameters according to claim 1, characterized in that, The process of locating at least one weight parameter in the prediction model that causes the bias, and using it as a target weight parameter, specifically includes: In response to the deviation warning, the performance data of the batch of concrete under warning is obtained, including the measured compressive strength value and the corresponding temperature history data. The performance data is input into the strength backtracking model to obtain the backtracking estimated strength; Calculate the strength deviation between the retrospectively estimated strength and the predicted performance value, and determine the one-wayness of the strength deviation; In response to the unidirectionality of the strength deviation, a correlation analysis is performed on the connection weights with the cementitious material in the prediction model to obtain statistical correlation. The weights with statistical relevance higher than the relevance threshold are determined as the target weight parameters.

7. The method for pre-determining concrete strength based on mix proportion parameters according to claim 6, characterized in that, The process of self-calibrating the target weight parameters with a micro-step size specifically includes: The calibration direction is determined based on the strength deviation between the retrospectively estimated strength and the predicted performance value; The preset microstep size is called, and the target weight parameter is adjusted according to the calibration direction to obtain the adjusted weight; The adjusted weights are validated and self-calibrated using a validation set.

8. A concrete strength pre-determination system based on mix proportion parameters, characterized in that, The concrete strength pre-determination system, applied to any one of claims 1-7, specifically includes: The communication unit is used to acquire the design parameters, on-site measured springback values, and performance data of the target object. The processing unit includes a prediction engine module, a dual cross-validation module, and a self-calibration module. The self-calibration module is used to respond to the deviation warning, backtrack and estimate the true performance value based on the performance data, locate the target weight parameter in the prediction model that causes the deviation by comparing the true performance value with the predicted performance value, and self-calibrate the target weight parameter with a microstep.

9. A concrete strength pre-determination system based on mix proportion parameters according to claim 8, characterized in that, The prediction engine module specifically includes: The prediction model unit is used to build and run an intensity prediction model based on a backpropagation neural network, and generate prediction performance values ​​based on design parameters. The mapping model unit is used to build and run the strength-rebound value mapping function, and map the predicted performance value to the theoretical rebound value and fluctuation range.

10. A concrete strength pre-determination system based on mix proportion parameters according to claim 8, characterized in that, The dual cross-validation module specifically includes: The rebound value range verification unit is used to compare the fluctuation range of the theoretical rebound value with the actual rebound value measured on site, and mark the first anomaly based on the comparison result; The strength value range verification unit is used to convert the field measured rebound value into the measured estimated strength based on the strength measurement curve, and to convert the predicted performance value into the expected converted strength based on the maturity equivalence principle. By calculating the relative deviation between the measured estimated strength and the expected converted strength, a second anomaly flag is marked according to a preset threshold. A deviation warning is triggered only when both the first anomaly flag and the second anomaly flag are marked.

Images