Machine learning based concrete mix optimization method and system

Abstract

The application discloses a concrete mixing proportion optimization method and system based on machine learning, and particularly relates to the technical field of intelligent proportioning of building materials, and comprises the following steps: a mapping model of target performance indexes is constructed by using raw material performance parameters and historical trial mixing data, cross-station mixing proportion optimization is carried out based on real-time sensing data of multiple mixing stations, a resonance relationship between raw material moisture content and model iteration frequency is monitored in a dynamic construction process, a parameter smoothing and disturbance suppression mechanism is triggered to correct a mixing proportion scheme, and the corrected mixing proportion adjustment scheme is applied to raw material control and feeding of the multiple mixing stations; the application improves training sample quality through unified data processing, realizes cross-mixing station performance consistency optimization by constructing a multi-dimensional mapping model, and introduces resonance detection and disturbance suppression mechanisms to guarantee the stability and anti-interference ability of the mixing proportion adjustment under dynamic construction conditions, so that high-performance, low-cost and strong-robustness concrete intelligent optimization control is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent mix proportioning technology for building materials, and more specifically, to a method and system for optimizing concrete mix proportions based on machine learning. Background Technology

[0002] The design of concrete mix proportions directly determines its mechanical properties, construction adaptability, and durability. In traditional production processes, the determination of mix proportions mainly relies on manual trial mixing and experience regression, which suffers from poor adaptability, long update cycles, and difficulty in dynamic adjustment. To improve the level of intelligent concrete production, some companies have introduced machine learning models to predict performance indicators based on historical trial mixing data and real-time sensor information, thereby achieving automatic adjustment of mix proportions.

[0003] In recent years, with the improvement of sensor networks and edge computing capabilities, collaborative production of multiple mixing plants has gradually become the mainstream. Multiple mixing plants can share the same performance mapping model and generate local mix proportion adjustment schemes based on the real-time status of their respective raw materials. However, in actual deployment, it has been found that when the same model is running in parallel at multiple mixing plants, a specific interference phenomenon is likely to occur: when there are obvious periodic fluctuations in the moisture content of raw materials, if the iteration frequency or mix proportion update frequency of the machine learning model is close to the main frequency of this fluctuation, a "resonance phenomenon" may occur.

[0004] This resonance phenomenon manifests as multiple mixing plants simultaneously making excessive adjustments to key mix proportion parameters such as water-cement ratio and sand ratio within a few hours. This transforms what was originally planned as fine-tuning under performance constraints into large-scale collective disturbances. Such disturbances further amplify the uncertainty of performance output and may even cause structural strength to fail to meet design standards, affecting construction quality and schedule.

[0005] Existing research mainly focuses on static optimization of single-station mix proportions, predictive regression of single performance indicators, or closed-loop control of local parameters. A control mechanism for identifying and suppressing resonance risks in multi-station coordinated optimization has not yet been established. The lack of dynamic intervention methods significantly reduces the robustness and safety of the model when deployed on-site, severely hindering the widespread application of intelligent mix proportioning systems in large-scale engineering projects. Therefore, this paper proposes a machine learning-based concrete mix proportion optimization method and system to address the aforementioned problems. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The machine learning-based method for optimizing concrete mix proportions includes the following steps:

[0008] Collect multi-dimensional raw material performance parameters, including cement strength grade, sand and gravel gradation, and raw material moisture content, as well as construction and curing conditions. Align the collected data with historical concrete mix design data in terms of time reference and format to generate a training dataset for model training.

[0009] Based on the training dataset, a multidimensional nonlinear mapping model between concrete material composition parameters and target performance indicators, including compressive strength, workability, and durability, is constructed using machine learning algorithms to jointly predict the target performance indicators.

[0010] During the concrete production process, real-time sensor data from multiple batching plants are input into the mapping model to perform cross-batching plant mix proportion optimization calculations based on target performance index constraints, and generate mix proportion adjustment schemes.

[0011] During dynamic construction, based on real-time sensor data and the iteration step size of the mapping model, it is detected whether there is a resonance relationship between the periodic fluctuation of the moisture content of raw materials and the iteration step size. When a resonance relationship is detected, the parameter smoothing mechanism and the dynamic disturbance suppression mechanism are triggered to correct the mix proportion adjustment scheme.

[0012] The revised mix proportion adjustment scheme was applied to the concrete production process of multiple batching plants to control the actual execution values ​​of the input of each raw material and the water-cement ratio.

[0013] In a preferred embodiment, the moisture content is corrected through the following steps before the training dataset is generated:

[0014] For the same batch of raw materials, the standard moisture content measured by laboratory drying method and the real-time moisture content measured by on-site electrical sensor are collected simultaneously. The zero-point offset and full-scale proportional coefficient are calculated using multiple sets of corresponding data from the two, and the two are defined as the initial values ​​of the range coefficient.

[0015] The initial value of the range coefficient is applied to the continuous time series moisture content data. The sensor output within each window is smoothed using the sliding window polynomial fitting method. The fitting residual between the smoothed value and the laboratory standard value is calculated at the boundary of each window. When the residual exceeds the preset fitting deviation threshold, the full-range proportional coefficient is corrected according to the preset residual proportional factor, thus forming a dynamic adjustment mechanism for the range coefficient.

[0016] The average temperature and average particle size distribution of the batch of raw materials are calculated. The average temperature is weighted by the first weight and the average particle size distribution by the second weight to obtain the comprehensive temperature-particle index of the raw materials. The comprehensive index is used as the independent variable and the difference between the smoothed sensor output and the laboratory standard value is used as the dependent variable to perform multiple regression modeling to obtain the moisture content correction curve adapted to different temperature and humidity conditions and particle structures.

[0017] The moisture content correction curve is stored in the model parameter library and called when generating the training dataset to standardize and correct all field moisture content data.

[0018] In a preferred embodiment, before the training dataset is generated, outlier samples need to be removed, specifically as follows:

[0019] A multidimensional feature matrix including strength grade, particle size distribution, moisture content, and curing temperature and humidity is constructed. A multivariate consistency test based on Mahalanobis distance is performed on the matrix to detect and remove abnormal samples that deviate from the overall distribution by more than the confidence interval threshold.

[0020] In a preferred embodiment, the steps of constructing a multidimensional nonlinear mapping model based on the training dataset and performing joint prediction include the following process:

[0021] Using material composition parameters (i.e., multidimensional raw material performance parameters), construction conditions, and curing conditions as inputs, and compressive strength, workability, and durability as outputs, a joint prediction model is established and a priori constraints of materials science are introduced. The law that the increase of the ratio of water to cementitious material dosage leads to a decrease in compressive strength after the critical point, and the law that the sand ratio monotonically increases workability within the allowable range, are written into the monotonicity constraint term and the critical point segmentation constraint term.

[0022] At the model output, the expected value and uncertainty of the compressive strength, workability and durability are given simultaneously. The variance of each output is adaptively learned by the objective function based on heteroscedastic Gaussian likelihood, and the loss of multiple outputs is dynamically weighted by the gradient balancing strategy so that different performance indicators achieve balanced convergence during the training process.

[0023] Furthermore, a reverse mapping sub-model is constructed with the target performance index as input and the material composition parameters as output. Together with the forward mapping model, it forms a closed-loop structure, and a cycle consistency penalty term is used to constrain the deviation of the same sample after forward prediction and reverse reconstruction.

[0024] The joint prediction model is trained and screened using sectional cross-validation and independent external validation sets. The uncertainty range of each performance index is calibrated based on the quantile calibration method, and the calibrated model parameters and constraint configurations are used as the final version of the mapping model.

[0025] In a preferred embodiment, the cross-mixing plant mix proportion optimization calculation includes the following steps:

[0026] Real-time sensor data from each mixing plant is input into the mapping model to generate single-plant mix proportion schemes that meet the constraints of target performance indicators such as compressive strength, workability and durability.

[0027] All single-station mix design schemes are summarized and a comprehensive evaluation function is constructed. The comprehensive evaluation function is formed by weighting and summing the total raw material consumption score, production cost score and performance margin score according to preset weights. The total raw material consumption score is calculated based on the difference between the raw material consumption and the target consumption of each scheme. The production cost score is calculated based on the proportion of material cost to total cost of each scheme. The performance margin score is calculated based on the safety margin ratio of the performance index of each scheme to the target value.

[0028] The combination of mix proportion schemes with the highest score in the comprehensive evaluation function is used as the mix proportion adjustment scheme across mixing plants.

[0029] In a preferred embodiment, the resonance relationship detection step includes the following process:

[0030] First, outlier removal and normalization are performed on the time series of raw material moisture content of each mixing plant. Then, the dominant frequency and corresponding spectral energy of moisture content are extracted using discrete Fourier transform and recorded as the dominant frequency value and the spectral energy of moisture content. At the same time, the iteration frequency is calculated based on the iteration step size of the mapping model and the jitter range of the iteration frequency is estimated within the rolling time window.

[0031] Establish a set of resonance discrimination criteria, and determine whether a resonance relationship exists based on the set of resonance discrimination criteria. Output a resonance flag when a resonance relationship exists.

[0032] In a preferred embodiment, the resonance discrimination criterion set includes the following:

[0033] The frequency proximity criterion is that the relative deviation between the dominant frequency of moisture content and the iterative frequency or its first harmonic should not exceed the allowable deviation coefficient, preferably five percent.

[0034] The phase locking criterion is that the time-averaged absolute change of the phase difference between the water content main frequency component and the iterative frequency reference signal within the window does not exceed the phase change threshold, which is preferably one-eighteenth of radians 3.1415926.

[0035] The energy amplification criterion is that the growth rate of the variance of the mix proportion adjustment range or the variance of the ratio of water to cementitious material dosage is not less than the energy amplification threshold, preferably 30%, within an observation period of not less than twice the iteration step size.

[0036] The determination principle is triggered by any criterion. When any one of the following criteria is met, such as the frequency proximity criterion, the phase locking criterion, or the energy amplification criterion, the resonance relationship is directly determined to exist.

[0037] In a preferred embodiment, correcting the mix proportion adjustment scheme when a resonance relationship is detected includes the following steps:

[0038] The ratio of water to cementitious material, sand ratio and admixture dosage are formed into a time series by batch. Median filtering with an odd number of batches as the window length is used to eliminate pulse anomalies. Then, polynomial fitting of no more than the third order is performed in each segment using the segment boundaries obtained by the change point detection, and piecewise polynomial smoothing is achieved with the continuity of the first derivative of adjacent segments as a constraint.

[0039] For the first-order changes of adjacent batches, upper limits of amplitude and directional constraints are set respectively. The upper limit of amplitude is determined based on the percentile of the historical stable interval and the changes exceeding the limit are clipped. The directional constraints are determined by the trend estimation of the most recent batches to be monotonically non-increasing or monotonically non-decreasing and the changes that violate the trend are projected and corrected. At the same time, based on the main frequency of the moisture content of raw materials in each mixing plant and the phase distribution execution time phase misalignment, the execution start time of each mixing plant is staggered according to the phase division. The upper limit of the adjustment amplitude of each mix proportion parameter is set as a function inversely proportional to the performance margin to suppress synchronous amplification.

[0040] The smoothed and limited mix proportion correction scheme is input into the mapping model to predict compressive strength, workability and durability respectively and compare them with the target performance index constraints. If all are met, it is marked as executable. If any is not met, the mix proportion adjustment range is reduced by a preset scaling factor and the prediction is repeated until the target performance index constraints are met and the baseline mix proportion scheme is marked as executable.

[0041] The mix proportion modification schemes marked as executable will be distributed to each mixing plant and linked to the batch identifier.

[0042] In a preferred embodiment, the machine learning-based concrete mix design optimization system specifically includes:

[0043] The data acquisition module is used to collect multi-dimensional raw material performance parameters, including cement strength grade, sand and gravel gradation, and raw material moisture content, as well as construction and curing conditions. The collected data is aligned with historical concrete mix design data in terms of time reference and format, and a training dataset is generated for model training.

[0044] The model building module is used to construct a multidimensional nonlinear mapping model between concrete material composition parameters and target performance indicators, including compressive strength, workability, and durability, based on the training dataset and using machine learning algorithms, and to jointly predict the target performance indicators.

[0045] The cross-site optimization module is used to input real-time sensor data from multiple batching plants into the mapping model during the concrete production process, perform cross-batch ratio optimization calculations based on target performance index constraints, and generate mix ratio adjustment schemes.

[0046] The resonance detection module is used to detect whether there is a resonance relationship between the periodic fluctuation of the moisture content of raw materials and the iteration step size of the mapping model during dynamic construction. When a resonance relationship is detected, the module triggers the parameter smoothing mechanism and the dynamic disturbance suppression mechanism to correct the mix proportion adjustment scheme.

[0047] The execution control module is used to apply the revised mix proportion adjustment plan to the concrete production process of multiple mixing plants, and to control the actual execution values ​​of the input of each raw material and the water-cement ratio.

[0048] The technical effects and advantages of this invention are as follows:

[0049] This invention collects multi-dimensional raw material performance parameters, including cement strength grade, aggregate gradation, and raw material moisture content, and combines these with construction and curing conditions. The collected data is then aligned with historical concrete mix design data in terms of time and format, forming a high-quality training dataset that can be directly used for machine learning model training. This data processing method effectively eliminates differences in collection time, recording format, and units of measurement among data from different sources, avoiding model input errors caused by heterogeneous data sources or inconsistent time series. Simultaneously, this unified processing ensures the comparability and fusion of data from different batches and under different construction conditions, improving the sample coverage and representativeness of model training, thus providing a solid data foundation for accurate predictions by the subsequent mapping model. This process significantly reduces the need for manual intervention in data preprocessing, improves data utilization efficiency, and ensures that the model has higher generalization ability and robustness when facing actual production data.

[0050] This invention, based on a constructed multidimensional nonlinear mapping model, jointly predicts concrete material composition parameters with multiple target performance indicators, including compressive strength, workability, and durability. During production, real-time sensor data from multiple mixing plants are input into the model to perform cross-mixing plant mix proportion optimization calculations based on target performance indicator constraints, thereby generating a mix proportion adjustment scheme that meets global performance objectives. This optimization mechanism enables unified optimization of mix proportions across different mixing plants, overcoming performance differences that may result from independent adjustments at a single plant. The comprehensive evaluation considers factors such as raw material consumption, production costs, and performance margins, ensuring that the final mix proportion scheme not only meets performance requirements but also achieves optimal raw material utilization and economic efficiency. This method can coordinate raw material input and water-cement ratio control across multiple production units within a region, helping to improve overall production efficiency, reduce resource waste, and ensure consistent concrete performance across different sites.

[0051] This invention, during dynamic construction, analyzes the relationship between the periodic fluctuations of raw material moisture content and the iteration step size based on real-time sensor data and the mapping model, detecting the existence of resonance. When resonance is detected, the system automatically triggers a parameter smoothing mechanism and a dynamic disturbance suppression mechanism to correct the generated mix proportion adjustment scheme. This design effectively avoids the problem of significant fluctuations in performance prediction values ​​caused by the amplification of periodic fluctuations in raw material moisture content during iterative calculations. Through smoothing and disturbance suppression, short-term drastic changes in mix proportion parameters can be limited, reducing execution risks during construction and ensuring a more stable and controllable trend in dosage changes. Simultaneously, this mechanism maintains the stability of concrete performance indicators under the influence of environmental disturbances and material fluctuations at the construction site, improving the anti-interference ability of the production process and the predictability of finished product quality, thereby achieving refined and stable control throughout the entire process. Attached Figure Description

[0052] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0053] Figure 1 This is a schematic diagram of the machine learning-based concrete mix proportion optimization method in this invention.

[0054] Figure 2 This is a schematic diagram of the machine learning-based concrete mix design optimization system of this invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Reference Figure 1 - Figure 2 The following examples were obtained:

[0057] Example 1: Collect multi-dimensional raw material performance parameters, including cement strength grade, sand and gravel gradation, and raw material moisture content, as well as construction and curing conditions. Align the collected data with historical concrete mix design data in terms of time reference and format to generate a training dataset for model training. The significance of this step is to establish a unified data foundation for subsequent learning and optimization. By unifying data from different sources and time scales under the same time reference, it ensures that material composition, environmental and process information can be accurately correlated with the corresponding performance results, thereby reducing systematic biases introduced by time drift and format inconsistencies. It also makes the training dataset traceable, reusable, and scalable, laying the foundation for robust training and generalization of the model.

[0058] Based on the training dataset, a multidimensional nonlinear mapping model is constructed using machine learning algorithms to connect concrete material composition parameters with target performance indicators, including compressive strength, workability, and durability, to jointly predict the target performance indicators. The significance of this step lies in using a data-driven approach to characterize the joint influence of material composition, construction conditions, and curing conditions on the target performance indicators. This enables the model to simultaneously consider the mutual constraints and coupling effects between various properties, achieve synergistic prediction of target performance indicators, improve the fitting ability to complex nonlinear laws, and obtain higher prediction accuracy and more stable decision-making basis under limited experimental costs.

[0059] In the concrete production process, real-time sensor data from multiple mixing plants is input into a mapping model to perform cross-mixing plant mix proportion optimization calculations based on target performance index constraints, generating mix proportion adjustment schemes. The significance of this step lies in transforming static training results into dynamic production guidance, mapping the constantly changing raw material states and on-site environment of each mixing plant to the performance space in real time, and forming an executable mix proportion adjustment scheme under the premise of meeting the target performance index constraints. This enables consistent performance output and comprehensive optimization of resource allocation in parallel production at multiple sites, reducing trial and error and rework, and improving overall production efficiency.

[0060] During dynamic construction, based on real-time sensor data and the iteration step size of the mapping model, it detects whether there is a resonance relationship between the periodic fluctuations of the raw material moisture content and the iteration step size. When a resonance relationship is detected, a parameter smoothing mechanism and a dynamic disturbance suppression mechanism are triggered to correct the mix proportion adjustment scheme. The significance of this step is to proactively protect against the amplification risk of time-varying disturbances on site. By comparing the periodic fluctuation characteristics of the moisture content with the update rhythm of the model, it can promptly identify resonance relationships that may cause synchronous over-amplitude adjustments, and break unfavorable couplings with smoothing and disturbance suppression methods to avoid batch performance mutations caused by excessive short-term parameter fluctuations, thereby ensuring the stability and safety of the optimization process.

[0061] The revised mix design was applied to the concrete production processes of multiple batching plants to control the actual values ​​of raw material inputs and water-cement ratios. The significance of this step lies in implementing upstream data processing, model prediction, and optimization decisions into downstream production control. By transforming the mix design into executable input and water-cement ratio instructions, it ensures consistent execution across batching plants under the same objective. Furthermore, it establishes a continuous verification and feedback channel for the results within the production closed loop, providing a reliable basis for subsequent data accumulation, model updates, and process improvements.

[0062] Before the training dataset is generated, the moisture content is corrected using the following steps:

[0063] For the same batch of raw materials, both laboratory drying and on-site electrical sensors were used to measure the moisture content. The laboratory drying method involved drying the raw material sample in a 105°C constant-temperature oven until its mass became constant. The standard moisture content was then calculated by dividing the mass difference before and after drying by the mass before drying. The formula was: Standard Moisture Content (%) = (Wet Mass - Dry Mass) ÷ Wet Mass × 100. The real-time moisture content measured by the on-site electrical sensors was an instantaneous value derived from the known functional relationship between the material's conductivity or dielectric constant and the moisture content. By pairing multiple sets of laboratory standard moisture contents with the sensor's real-time moisture content for the same batch of raw materials, the least squares fitting method was used to calculate the zero-point offset (representing the sensor's output deviation at zero moisture content) and the full-scale proportionality coefficient (representing the scaling factor between the sensor's output value and the true value). The initial calculation results of the zero-point offset and the full-scale proportionality coefficient were defined as the initial values ​​of the range coefficient. For example, in a batch of 10 tons of sand and gravel raw materials testing, 15 sets of laboratory and sensor paired data were collected. After fitting, the zero-point offset was found to be 0.4% moisture content and the full-scale proportionality coefficient was 1.08.

[0064] The initial value of the range coefficient is applied to the moisture content data of the continuous time series. First, zero-point offset correction and proportional coefficient scaling are performed on the sensor output. Then, a sliding window polynomial fitting method is used to smooth the corrected output within each window. The length of the sliding window is determined based on the field detection frequency and industry recommendations, preferably 5 to 9 sampling points. The order of the fitting polynomial within the window is no higher than the third order to ensure curve smoothness while preserving trend change characteristics. At the boundary of each window, the fitting residual between the fitted value at the end of the window and the laboratory standard value at the corresponding time point is calculated. The residual calculation formula is: Residual = Fitted value - Standard value. When the absolute value of the residual exceeds the preset fitting deviation threshold, the proportional coefficient adjustment is triggered. The preset fitting deviation threshold is preferably ±0.5 percentage points. The proportional coefficient adjustment value = Current proportional coefficient + Residual × Residual proportional factor. The residual proportional factor can be obtained statistically based on the historical error distribution, preferably ranging from 0.05 to 0.15, thus forming a dynamic adjustment mechanism for the range coefficient. For example, in a single test, if the window boundary residual is -0.6% and the scaling factor is 0.1, then the scaling coefficient is adjusted from 1.08 to 1.074.

[0065] The average temperature and average particle size distribution of the batch of raw materials are calculated. The average temperature is obtained by measuring multiple points on different parts of the raw material surface using an infrared thermometer and taking the arithmetic mean. The average particle size distribution is obtained by sieving or laser particle size analysis. The average temperature is weighted according to a first weight, and the average particle size distribution is weighted according to a second weight to form the raw material temperature-particle composite index. The weight values ​​are determined based on materials science research and historical regression analysis, with the preferred first weight being 0.6 and the second weight being 0.4. The composite index is calculated as follows: Temperature-Particle Composite Index = 0.6 × Average Temperature + 0.4 × Average Particle Size Distribution. In a practical embodiment, if the average temperature is 25 degrees Celsius and the average particle size distribution is 3.5 mm, then the temperature-particle composite index = 0.6 × 25 + 0.4 × 3.5 = 15 + 1.4 = 16.4. Using the comprehensive index as the independent variable and the difference between the smoothed sensor output and the laboratory standard value as the dependent variable, a multivariate regression modeling method is used to establish a functional relationship. The regression model can be a multinomial regression, support vector regression, or a gradient boosting-based regression tree model to fit the moisture content difference under different temperature and humidity environments and particle characteristics, and obtain a moisture content correction curve adapted to different temperature and humidity conditions and particle structures.

[0066] The moisture content calibration curve is stored in the model parameter library, which can be implemented using a relational database or a key-value database. The stored content includes curve fitting parameters, applicable temperature-grain composite index ranges, calibration accuracy evaluation indicators, and generation date. When generating the training dataset, this moisture content calibration curve is used to standardize and correct all field moisture content data. The correction formula is: Standardized Moisture Content = Original Moisture Content × Proportional Coefficient Correction Value + Zero-Point Correction Value + Regression Correction Based on the Composite Index. This ensures that the data input to the model has a consistent measurement benchmark and comparability under different batches and environmental conditions, and significantly reduces the impact of moisture content measurement errors on the accuracy of mix proportion optimization.

[0067] Before generating the training dataset, outlier samples need to be removed. Specifically, a multi-dimensional feature matrix is ​​constructed, including strength grade, particle size distribution, moisture content, and curing temperature and humidity. The strength grade is the 28-day compressive strength measured according to GB / T17671 "Test Methods for Physical and Mechanical Properties of Cement Part 3: Strength of Cement Mortar". The particle size distribution is the cumulative percentage vector of different particle sizes retained on the sieve after sieving tests according to JGJ52 "Standard for Quality and Testing Methods of Sand and Stone for Ordinary Concrete". The moisture content is the mass percentage obtained by correcting the laboratory drying method. The curing temperature and humidity are the curing environment temperature (degrees Celsius) and relative humidity (percentage) recorded according to GB / T50081 "Standard for Test Methods for Physical and Mechanical Properties of Concrete". These parameters are organized into an N-row, M-column matrix according to the sample number. Each row of the matrix represents a multi-dimensional feature vector X_i=[x_1,x_2,…,x_M] for a sample, where M is the number of features.

[0068] A multivariate consistency test based on Mahalanobis distance was performed on the multidimensional feature matrix. Mahalanobis distance measures the distance between sample points and the population center, and can simultaneously consider the covariance relationship between each dimension. The formula is as follows: , where μ is the feature mean vector of all samples, and Σ is the covariance matrix of the features. This is the inverse of the covariance matrix. In the calculation, μ and Σ are first calculated based on all samples. Then, for each sample X_i, the squared Mahalanobis distance D²_i is calculated, and the corresponding chi-square distribution critical value χ²_(α,M) is obtained, where α is the significance level and M is the number of feature dimensions. For example, when the feature dimension M=5 and the significance level α=0.05, the critical value of χ²_(0.05,5) is approximately 11.07 (refer to the standard table in Mathematical Statistics or the chi-square distribution table in the appendix of GB / T4883 "Statistical Vocabulary").

[0069] Anomalies exceeding the confidence interval threshold are detected and removed. Specifically, if a sample's D²_i > χ²_(α,M), it is considered an outlier and removed from the training dataset. In a specific example, consider 1000 historical concrete mix design samples, each containing six features (M=6): strength grade, sand ratio, maximum aggregate size, moisture content, curing temperature, and curing humidity. Assuming a chi-square critical value of approximately 12.59 for α=0.05, 18 samples have D² values ​​exceeding 12.59 and are removed. This multivariate consistency test based on Mahalanobis distance effectively removes anomalies that may not be extreme in a single dimension but exhibit unreasonable deviations in multivariate combinations. This ensures the feature distribution of the training dataset matches the statistical characteristics of real production conditions, reducing parameter shifts and prediction instability caused by outlier data during model training.

[0070] Multi-dimensional raw material performance parameters, including cement strength grade, aggregate gradation, and raw material moisture content, as well as construction and curing conditions, were collected. The collected data was aligned with historical concrete mix design data for time reference and format standardization to generate a training dataset for model training. Cement strength grade was determined by conducting 3-day and 28-day compressive and flexural strength tests on cement mortar specimens according to the standards of GB175 "General Portland Cement" or GB / T17671 "Test Methods for Physical and Mechanical Properties of Cement Part 3: Cement Mortar Strength". The 28-day compressive strength was used as the basis for cement grade classification, such as 42.5 grade, 52.5 grade, etc. Cement samples were taken from the same batch of delivered cement, cured in a standard curing chamber at a temperature of 20±1℃ and relative humidity ≥90% until the required age, and then subjected to a pressure testing machine until specimen failure to obtain the strength value.

[0071] The gradation of sand and gravel was determined by sieving tests according to the "Standard for Quality and Testing Methods of Sand and Stone for Ordinary Concrete" JGJ52. Sand samples were sieved stepwise using square-hole sieves (sieve aperture sizes such as 4.75mm, 2.36mm, 1.18mm, 0.6mm, 0.3mm, and 0.15mm) and weighed. The particle size distribution curve was plotted by the cumulative percentage of sieve residue and compared with the standard range. Stone was sieved using large-aperture sieves (31.5mm, 19mm, 9.5mm, etc.) and the gradation ratio of each particle size was calculated to obtain the sand ratio and particle size distribution parameters.

[0072] Construction conditions include parameters such as ambient temperature, relative humidity, wind speed, pouring method, pumping pressure, transportation time, and concrete placement temperature. The measurement of ambient temperature, humidity, and wind speed is based on the requirements for meteorological monitoring in the "Construction Safety Inspection Standard" JGJ59, and is obtained at the construction site using digital thermometers and anemometers. Pumping pressure is recorded by the pressure sensor built into the concrete pump. Transportation time is recorded by the vehicle's GPS. The concrete placement temperature is measured by an infrared thermometer before the concrete is placed in the formwork.

[0073] Curing conditions include parameters such as curing method (e.g., standard curing, natural curing, steam curing), curing temperature, curing humidity, and curing time; standard curing refers to the conditions of 20±2℃ and relative humidity ≥95% specified in GB / T50081 "Standard for Test Methods of Physical and Mechanical Properties of Concrete"; steam curing is based on the process parameters such as heating rate, constant temperature time and cooling rate determined according to JG / T169 "Cure System for Precast Concrete Components".

[0074] Time reference alignment refers to synchronizing newly collected data with historical concrete mix design data using a unified timestamp system, with timestamp accuracy no less than the second level. On-site data collection uses GPS time synchronization to ensure time consistency at each collection point, while historical data is extrapolated to a unified timeline by recording the mix design date, mix design period, and sample number, and interpolation methods are used to fill in missing time period data when necessary.

[0075] Standardized formatting refers to converting data from different sources into standardized data fields and units of measurement. For example, cement strength is standardized to megapascals (MPa), sand and gravel gradation particle size is standardized to millimeters (mm), and moisture content is standardized to mass percentage (%). The data is then stored in tables or database fields with the same structure. For historical data with inconsistent units or symbols, they are converted to a standardized format according to the national measurement unit standard GB3100-3102.

[0076] The steps for constructing a multidimensional nonlinear mapping model and performing joint prediction based on the training dataset include the following process: Using material composition parameters (i.e., multidimensional raw material performance parameters), construction conditions, and curing conditions as inputs, and compressive strength, workability, and durability as outputs, a joint prediction model is established, and material science prior constraints are introduced into the model structure. Multidimensional raw material performance parameters include cement strength grade, sand and gravel gradation, raw material moisture content, and admixture type and dosage; construction conditions include ambient temperature, relative humidity, wind speed, transportation time, pumping pressure, and placement temperature; curing conditions include curing method, curing temperature, curing humidity, and curing duration. Material science prior constraints include two aspects: First, the principle that increasing the ratio of water to cementitious materials decreases compressive strength after the critical point is incorporated into the monotonicity constraint term and the critical point segmentation constraint term. The critical point is referenced from the recommended values ​​regarding the relationship between water-cement ratio and strength in the "Specification for Mix Proportion Design of Ordinary Concrete" (JGJ55), preferably 0.5. When the water-cement ratio is less than or equal to 0.5, the compressive strength increases monotonically with decreasing water-cement ratio; when the water-cement ratio is greater than 0.5, the compressive strength decreases monotonically with increasing water-cement ratio. Secondly, the monotonically increasing relationship between sand ratio and workability within an allowable range is incorporated into the monotonicity constraint term. This allowable range references the recommended sand ratio values ​​in JGJ55, preferably 30% to 46%. Within this range, a higher sand ratio results in better workability, but exceeding the upper limit may lead to bleeding and decreased strength. These constraints are implemented in the model's loss function through a penalty term, imposing weighted penalties on predictions that violate monotonicity or piecewise relationships, thereby guiding the model to learn a mapping relationship consistent with engineering practice.

[0077] At the model output, expected values ​​and uncertainties are simultaneously provided for compressive strength, workability, and durability, respectively. The uncertainty represents the width of the confidence interval for the prediction result. An objective function based on heteroscedastic Gaussian likelihood is used to adaptively learn the variance of each output. This objective function introduces a variance parameter into the loss term, enabling the model to dynamically adjust the prediction confidence based on sample characteristics. For example, for predicting concrete strength under different environmental temperature and humidity conditions, the model can output a wider uncertainty interval for conditions with large data fluctuations. To prevent any single performance index from dominating the training process in multi-task learning, a gradient balancing strategy is used to dynamically weight the losses of multiple outputs. Specifically, in each iteration, the norm difference of the gradients for each task is calculated, and the loss weights are adjusted to ensure that the rate of loss descent for different performance indices remains consistent, thereby achieving balanced convergence during training. Furthermore, a reverse mapping sub-model is constructed, with the target performance index as input and material composition parameters as output, forming a closed-loop structure together with the forward mapping model. The forward mapping model is used to predict performance from input parameters, while the backward mapping sub-model is used to infer material composition parameters from target performance. Both are constrained by a cycle consistency penalty term: the input is reconstructed through backward mapping after forward prediction of the same sample, and the sum of squared differences between the reconstructed and backward predictions is used as the cycle consistency loss in the training objective. A small loss indicates that the forward and backward mappings are consistent under the same data distribution. This design can be used for backward mix design and to enhance the robustness of the model.

[0078] Partial cross-validation and independent external validation sets were used to train and select the joint prediction model. Partial cross-validation involves randomly dividing the training data into several equal parts, selecting one part as the validation set each time, and using the rest as the training set. This process is repeated multiple times to calculate the average performance, thereby evaluating the model's generalization ability on different data subsets. The independent external validation set is an independent dataset that was not involved in the model training process, used to test the model's prediction accuracy on real new data.

[0079] The uncertainty intervals of each performance index are calibrated using a quantile calibration method. This involves matching the quantiles of the predicted distribution with the actual observed distribution, ensuring that the proportion of observed values ​​within the 95% confidence interval of the prediction is close to 95%, thus guaranteeing the statistical significance of the model output uncertainty. Finally, the calibrated model parameters and constraint configurations are used as the final version of the mapping model and stored in the model parameter library for subsequent optimization calculations of concrete mix proportions and real-time on-site predictions. This ensures that the prediction results conform to both data patterns and materials science principles, while also possessing operability and stability in engineering practice.

[0080] A gradient balancing strategy is employed to dynamically weight the multi-output loss. Specifically, during each iteration of model parameter updates, for the three prediction tasks (compression strength, workability, and durability), the gradient magnitude of the current loss function across all model parameters is calculated. The square root of the sum of the squares of these magnitudes is used as the gradient norm for that task. After obtaining the gradient norms for each of the three tasks, their average is calculated. The average is then compared to the gradient norm for each task to obtain the gradient adjustment coefficient. This gradient adjustment coefficient is used to multiply and adjust the loss weights for the current task, decreasing the weights of tasks with larger gradient norms and increasing the weights of tasks with smaller gradient norms. After adjustment, the weights for the three tasks are normalized so that the sum of the weights is one. This dynamic adjustment ensures that the loss decreases at a more consistent rate during training, preventing any single task from dominating the training results.

[0081] The forward mapping model predicts performance from input parameters, while the backward mapping sub-model infers material composition parameters from target performance. The two are constrained by a cycle consistency penalty mechanism. Specifically, for any training sample, the input raw material performance parameters, construction conditions, and curing conditions are first fed into the forward mapping model to obtain the predicted performance result for that sample. Then, this predicted performance result is input into the backward mapping sub-model to obtain the reconstructed result of the original input parameters. Finally, the degree of difference between the original input parameters and the reconstructed result is calculated, and the average of the squares of these differences is used as the cycle consistency penalty value. This penalty value is multiplied by a cycle consistency coefficient in the total model loss and added to the overall optimization objective, thereby ensuring consistency between forward prediction and backward reconstruction during training. The value of the cycle consistency coefficient can be selected through multiple rounds of cross-validation on historical experimental data and validation sets, generally ranging from 0.1 to 0.5.

[0082] For monotonicity and piecewise constraints, these are implemented as penalty terms in the model's loss function. The monotonicity constraint states that within a certain critical value range where the ratio of water to cementitious material is less than or equal to that ratio, the predicted compressive strength should continuously increase as that ratio decreases. If a sample with a smaller ratio has a lower predicted strength than a sample with a larger ratio, the square of the strength difference between them is calculated as the monotonicity penalty and accumulated. The piecewise constraint states that when the ratio of water to cementitious material is greater than the critical value, the predicted compressive strength should continuously decrease as that ratio increases. If a sample with a larger ratio has a higher predicted strength than a sample with a smaller ratio, the square of the strength difference between them is calculated as the piecewise penalty and accumulated. The monotonicity penalty and the piecewise penalty are multiplied by their respective penalty coefficients and then summed to form the total constraint penalty value. The critical value can be referenced from the recommended value of the relationship between water-cement ratio and strength in the "Specification for Mix Proportion Design of Ordinary Concrete", preferably 0.5. The penalty coefficient can be determined by testing the stability of the model prediction and the acceptable error in engineering, and is generally between 0.05 and 0.2.

[0083] The model's total loss consists of the following components: the sum of the predicted losses of the three performance indicators multiplied by the weights dynamically adjusted by the gradient balancing strategy; the product of the cycle consistency penalty and the cycle consistency coefficient; and the weighted sum of the monotonicity and piecewise constraint penalties. By simultaneously optimizing these three loss components, the model can satisfy material science principles while ensuring prediction accuracy, and possess both positive and negative consistency. This allows for a balance of accuracy, stability, and engineering interpretability in practical concrete mix design optimization.

[0084] The cross-mixing plant mix design optimization calculation includes the following steps: Real-time sensor data from each mixing plant is input into the mapping model to generate single-plant mix design schemes that meet the constraints of target performance indicators such as compressive strength, workability, and durability. Real-time sensor data includes raw material moisture content, sand and gravel gradation, admixture dosage, cement strength grade, pour temperature, mixing time, discharge slump, and construction site environmental parameters. This data is collected by online monitoring equipment at each mixing plant and transmitted to the data processing module after being timestamped to the second level. Based on the nonlinear relationship between raw material composition and performance indicators learned from the training dataset, the mapping model performs inference calculations on the input data to obtain the performance prediction results for each mixing plant under the current state. Based on this, the mix design is selected through constraints, namely, compressive strength not lower than the design strength, workability not lower than the design slump requirement, and durability indicators such as the chloride ion diffusion coefficient not higher than a set limit, thereby generating single-plant mix design schemes that can directly guide production. For example, if a mixing plant inputs data such as the current raw material moisture content of 3%, sand content of 35%, and water-cement ratio of 0.45, the mapping model will output a mix proportion scheme of 390 kg / m³ cement, 760 kg / m³ sand, 1040 kg / m³ stone, and 6 kg / m³ admixture.

[0085] All individual mix design schemes are summarized and a comprehensive evaluation function is constructed. This function is formed by weighting and summing the total raw material consumption score, production cost score, and performance margin score according to preset weights. The total raw material consumption score is obtained by calculating the difference between the total raw material consumption in each mix design and the target consumption; the smaller the difference, the higher the score. The target consumption can be determined based on the recommended values ​​in the "Specification for Mix Design of Ordinary Concrete" or the owner's economic requirements. The production cost score is calculated by comparing the sum of the costs of all raw materials in each mix design with the total cost; the lower the cost percentage, the higher the score. The unit price of raw materials can be updated in real time based on current market prices or long-term procurement contract prices. The performance margin score is obtained by calculating the safety margin ratio between the predicted compressive strength, workability, and durability of each mix design and their respective target values. For example, if the predicted compressive strength is 10% higher than the target value, the safety margin ratio for compressive strength is 10%. The safety margin ratios of the three performance indicators can be weighted and synthesized into the total performance margin score. The allocation of preset weights can be determined based on project priority. For example, for bridge projects with extremely high structural safety requirements, the performance margin weight can be set at 50%, the cost weight at 30%, and the total raw material consumption weight at 20%. For cost-sensitive commercial concrete supply, the cost weight can be set at 50%, the performance margin weight at 30%, and the total raw material consumption weight at 20%. The calculation process of the comprehensive evaluation function is as follows: first, calculate the three score values ​​separately, then multiply them by their corresponding weights, and finally add the three weighted scores to obtain the comprehensive score.

[0086] The mix proportion scheme with the highest score in the comprehensive evaluation function is used as the cross-mixing plant mix proportion adjustment scheme. The generation of the combined scheme not only considers the optimality of each individual plant but also aims for overall optimization across all mixing plants. For example, one mixing plant may have slightly higher performance margin but slightly higher cost, while another mixing plant may have a significant cost advantage but slightly lower performance margin. Through weight balancing of the comprehensive evaluation function, the mix proportion scheme with the highest comprehensive score after combination may ultimately be selected, even if neither is perfectly optimal. In practical applications, all individual plant mix proportion schemes can be sorted according to their comprehensive scores, and the combination with the highest score can be selected. The adjustment command after combination is then sent to the production control system of each mixing plant to adjust the input of each raw material and the water-cement ratio in real time. This achieves optimal production cost and stable performance globally while meeting the constraints of the target performance indicators.

[0087] The resonance relationship detection process includes the following steps: First, outlier removal and normalization are performed on the raw material moisture content time series for each mixing plant. The raw material moisture content time series is a continuous sequence of moisture content measurements collected by a moisture content sensor at a fixed sampling frequency during concrete production. Each data point has a timestamp accurate to the second. Outlier removal can be achieved using a sliding window statistical method. This involves calculating the average and standard deviation of the moisture content within a window containing several consecutive sampling points. Measurement points deviating from the average by more than three times the standard deviation are identified as outliers and removed. This threshold is based on the statistical requirements for outlier identification in the "Standard for Quality Inspection and Evaluation of Building Engineering". Normalization converts the moisture content data after outlier removal into a standard normal distribution with a mean of zero and a standard deviation of one. This eliminates differences in measurement range and dimensions between different mixing plants and batches, ensuring comparability in subsequent spectral analysis.

[0088] The dominant frequency and corresponding spectral energy of the water content are then extracted using Discrete Fourier Transform (DFT), and denoted as the dominant frequency value and spectral energy of the water content. The DFT calculation process involves decomposing the normalized water content time series into sine and cosine wave components of different frequencies, obtaining the amplitude and phase of each frequency component. The spectral energy is the square of the amplitude of each frequency component, reflecting the energy magnitude on the spectrum. The dominant frequency value of the water content is the frequency value of the frequency component with the largest spectral energy, representing the main period of water content change within the observation period. For example, within an observation window of 1200 seconds, a calculated dominant frequency value of 0.005 Hz means that the main period of water content change is 200 seconds, and the corresponding spectral energy accounts for 65% of the total spectral energy.

[0089] Simultaneously, the iteration frequency is calculated based on the iteration step size of the mapping model, and the jitter range of the iteration frequency is estimated within a rolling time window. The iteration step size is the time interval required for each update of the mix proportion adjustment scheme by the mapping model, in seconds. The iteration frequency is the reciprocal of the iteration step size, representing the number of times the model updates per unit time. The jitter range is the difference between the maximum and minimum values ​​of the iteration frequency within the rolling time window, used to reflect the stability of the iteration frequency. For example, if the average iteration step size of the model in a certain stage is 300 seconds, then the iteration frequency is 0.0033 Hz. If the iteration frequency fluctuates between 0.0030 and 0.0036 Hz within a one-hour rolling window, then the jitter range is 0.0006 Hz.

[0090] A set of resonance discrimination criteria is established, and the existence of resonance relationships is determined based on this set. A resonance flag is output when a resonance relationship exists. The resonance flag is a data record containing information such as trigger time, satisfied discrimination criterion type, dominant frequency value, spectral energy, iteration frequency, and its jitter range. This flag is passed to the parameter smoothing mechanism and dynamic disturbance suppression mechanism as a trigger condition. In practical applications, for example, if the detected water content dominant frequency is 0.005 Hz, the iteration frequency is 0.0033 Hz, and its first harmonic is 0.0066 Hz, with a relative deviation of 3% from the dominant frequency value, which is less than the 5% threshold, then the frequency proximity criterion is satisfied. At this point, the resonance flag is immediately output, and the suppression process begins to prevent sudden changes in concrete performance caused by synchronization parameter adjustments.

[0091] The resonance discrimination criterion set includes the following: Frequency proximity criterion, which refers to the relative deviation between the dominant moisture content value obtained after spectral analysis and the iterative frequency or the first harmonic of the iterative frequency not exceeding the allowable deviation coefficient. The relative deviation is calculated by dividing the absolute value of the difference between the two by the reference frequency value, and then multiplying by 100% to obtain a percentage deviation value. The allowable deviation coefficient is preferably 5%, and this preferred value can be referenced from the consistency limit for determining the relative frequency data in the "Standard for Test Methods of Building Materials". In practical applications, for example, if the dominant moisture content frequency is 0.005 Hz and the iterative frequency is 0.0048 Hz, the relative deviation between the two is 4%, which is less than the 5% threshold, then the frequency proximity criterion is satisfied.

[0092] The phase-locking criterion refers to the fact that the time-averaged absolute change of the phase difference between the water content dominant frequency component and the iterative frequency reference signal within a set time window does not exceed a phase change threshold. The phase difference is calculated by taking the absolute value of the phase difference between the water content dominant frequency component and the iterative frequency reference signal at each sampling point and averaging it over the entire observation window. The preferred phase change threshold is 1 / 18th of 3.1415926 radians. This threshold is derived from a fine-grained phase stability judgment standard commonly used in phase synchronization detection, ensuring that the phases of the two signals remain relatively stable and do not drift significantly under resonance conditions. For example, if the average phase change is 0.017 radians within a one-hour observation window, which is below this threshold, the phase-locking criterion is satisfied.

[0093] The energy amplification criterion refers to the requirement that, within an observation period of at least twice the iteration step size, the growth rate of the variance of the mix proportion adjustment range or the variance of the water-to-cement ratio is not lower than the energy amplification threshold. The growth rate is calculated by subtracting the variance at the beginning of the observation from the variance at the end of the observation, then dividing by the initial variance, and multiplying by 100% to obtain a percentage. The preferred energy amplification threshold is 30%, which is referenced from the sensitivity assessment standards for the variation range of key parameters in the "Code for Construction of Concrete Structures". For example, if, within four iteration cycles, the variance of the water-to-cement ratio increases from 0.004 to 0.0056, the growth rate is 40%, which is higher than the 30% threshold, thus satisfying the energy amplification criterion.

[0094] The system employs a trigger-based judgment principle: when any one of the following criteria—frequency proximity, phase locking, or energy amplification—is met, a resonance relationship can be directly determined to exist in the current state, without the need for multiple criteria to be met simultaneously. In actual operation, for example, if the frequency proximity criterion is detected to be met within a certain period, a resonance flag can be immediately output and the mix proportion adjustment suppression mechanism can be activated. Even if the phase locking and energy amplification criteria are not met simultaneously, it does not affect the confirmation of the resonance relationship. This ensures that the system intervenes when early signs of resonance are detected, preventing mix proportion fluctuations from causing instability in concrete performance.

[0095] When resonance is detected, the specific process for correcting the mix proportion adjustment scheme is as follows: First, the ratio of water to cementitious materials, sand ratio, and admixture dosage are generated as a time series according to the order of concrete production batches. Here, the ratio of water to cementitious materials refers to the ratio of water usage to the total amount of cementitious materials (including cement, mineral admixtures, etc.) in each batch of concrete; the sand ratio refers to the percentage of sand usage to the total amount of fine and coarse aggregates; and the admixture dosage refers to the percentage of chemical admixture usage to the mass of cementitious materials. To eliminate occasional pulse-type outliers, a median filtering method with an odd number of batches as the window length is first used. The median filtering is calculated by arranging the batch data in numerical order within the window range and taking the middle value as the smoothed value of that batch. For example, when the window length is five batches, the value of the middle batch is calculated by medianing the data of the two batches before and after it and its own data, thereby effectively removing single-point anomalous jumps.

[0096] After median filtering, a change point detection method is used to perform structural analysis on the time series, identifying locations where the numerical trend changes significantly as segmentation boundaries. Change point detection can be based on cumulative sum statistics, sliding window variance comparison, or Bayesian segmentation models, with the aim of dividing the overall series into several stable segments. Within each stable segment, a polynomial fitting of no higher than order three is performed to smooth the numerical curves. To ensure the smoothness of the fitted curves at the junctions, a first-order derivative continuity constraint is introduced between segments, requiring that the rate of change of adjacent segments at the boundaries be equal, thereby avoiding abrupt jumps or discontinuities in the adjustment curves during batch switching.

[0097] Next, upper limits and directional constraints are set for the changes in mix proportion parameters of adjacent batches. The upper limit refers to the maximum allowable change in the same mix proportion parameter between two adjacent batches. This value is determined based on the percentile of the stable operating range in historical production data. For example, the change at the 95th percentile of the historical stable range is taken as the upper limit. This threshold, referencing the empirical range in the "Concrete Quality Control Standard," can avoid performance fluctuations caused by large adjustments. The directional constraint determines whether the current parameter should remain monotonically unchanged or monotonically unchanged by estimating the data trend of recent batches. If the new adjustment direction is opposite to the trend direction, it is corrected to the allowed trend direction through projection correction to avoid frequent reversals in the short term.

[0098] While implementing directional constraints, time phase misalignment must also be performed based on the dominant frequency and phase distribution of the raw material moisture content at each mixing plant. The calculation method for time phase misalignment involves staggering the start times of mix proportion adjustments at all mixing plants according to equal phase divisions. For example, the phase misalignment angles for four mixing plants within an adjustment cycle are 0 degrees, 90 degrees, 180 degrees, and 270 degrees, respectively. This reduces the overlap of adjustment actions in time and prevents resonance effects caused by synchronous execution. Furthermore, the upper limit of the adjustment amplitude for each mix proportion parameter is set as a function inversely proportional to the performance margin; that is, the smaller the performance margin, the smaller the allowed adjustment amplitude. This function can be taken as the upper limit of the adjustment amplitude equal to the base amplitude multiplied by the reciprocal of the performance margin, or its piecewise attenuation form, thereby suppressing the synchronous amplification effect when performance approaches the lower limit.

[0099] The modified mix proportion scheme, after the aforementioned smoothing and limiting processes, is input into the mapping model to predict compressive strength, workability, and durability, and then compared item by item with preset target performance constraints. If all performance indicators meet the requirements, the scheme is marked as an executable baseline mix proportion scheme; if any performance indicator does not meet the requirements, the mix proportion adjustment range is reduced according to a preset scaling factor, and the prediction and comparison steps are repeated until all performance indicators meet the requirements, ultimately generating an executable baseline mix proportion scheme. Finally, the executable modified mix proportion scheme is distributed to each mixing plant and bound to the production batch identifier to ensure that each batch is executed according to the modified mix proportion during production, thereby suppressing the resonance effect and maintaining the stability of concrete performance.

[0100] The basis for setting the scaling factor is as follows: Performance safety margin assessment: Performance safety margin refers to the relative difference between the target performance index and the current predicted performance value. For example, if the target value of compressive strength is 45 MPa and the current predicted value is 42 MPa, then the performance safety margin is 6.67%. According to the requirements for the lower limit of performance in the "Code for Acceptance of Construction Quality of Concrete Structures", if the margin is less than 10%, the adjustment range should be significantly reduced, and the scaling factor is usually taken as 0.5 to 0.7; if the margin is between 10% and 20%, the scaling factor is taken as 0.7 to 0.9; if the margin is greater than 20%, it can be taken as 0.95 to 1.

[0101] Historical Adjustment Sensitivity Analysis: Sensitivity refers to the rate of change in performance indicators resulting from adjustments to a certain parameter in the formula. It can be obtained by fitting historical batch data, for example, the decrease in compressive strength for every 0.1% increase in the water-cement ratio. If the sensitivity is high (i.e., performance is very sensitive to parameter changes), the scaling factor should be set smaller to prevent excessive performance fluctuations caused by small adjustments.

[0102] Raw material volatility level: The volatility level of raw materials can be measured by the ratio of the standard deviation to the mean of the performance parameters (such as moisture content, fineness, and gradation) of similar raw materials over a recent period. When the volatility level is high, the proportion factor decreases. For example, when the fineness volatility of cement is greater than 5%, the proportion factor is between 0.6 and 0.8.

[0103] Industry experience and safety margins: Referring to industry experience, the "Code for Quality Management of Ready-Mixed Concrete" suggests that the adjustment range of parameters for a single batch should not exceed 25% of the historical stable value. Based on this experience, the range of the scaling factor can be deduced. In special structural engineering projects (such as long-span bridges and nuclear power plants), the scaling factor should be preset more conservatively, usually not exceeding 0.75.

[0104] Alternatively, a segmented weighted calculation formula can be established by combining the above three types of indicators (margin, sensitivity, and volatility level): First, determine the basic proportional factor based on the safety margin, then multiply the sensitivity and volatility levels by correction coefficients respectively to obtain the final proportional factor. For example, a safety margin of 12% corresponds to a basic proportional factor of 0.85, a sensitivity coefficient of 0.9, and a volatility level coefficient of 0.95. The final proportional factor is 0.85 multiplied by 0.9 multiplied by 0.95, which equals 0.726, meaning the adjustment range is reduced by approximately 27%.

[0105] Implementation Method 1: Applicable to large-scale, cross-regional, centralized concrete enterprises. Within four mixing plants in different regions, outliers are removed using a median filter with a window length of five batches. Change point detection employs the sliding variance comparison method, with the upper limit of the amplitude taken as the 95th percentile of the historical stable interval's variation. Directional constraints utilize the trend determination method based on the most recent three batches, with time phase misalignment divided into 90-degree segments. Performance margin is strictly inversely proportional to the upper limit of the adjustment amplitude. Under this method, the mix proportion correction scheme, after smoothing and amplitude limiting, can converge to an executable scheme that meets performance indicators within two iteration cycles.

[0106] Implementation Method 2: Applicable to scenarios with a single large-scale mixing plant but multiple production lines. The median filtering window length is set to three batches. Change point detection uses the cumulative sum statistical method. The upper limit of amplitude is taken as the change amplitude at 90% of the positions. Directional constraints use the trend determination method of the most recent five batches. Time phase misalignment is divided equally according to the number of production lines. The performance margin and the upper limit of the adjustment amplitude have a piecewise attenuation relationship; that is, when the performance margin is less than 10%, the upper limit of the adjustment amplitude is executed at one-third of the original baseline amplitude. Under this method, the adjustment scheme can achieve rapid stabilization within one iteration cycle, suitable for engineering projects with high performance margin requirements.

[0107] Example 2: A machine learning-based concrete mix design optimization system, specifically including:

[0108] The data acquisition module is used to collect multi-dimensional raw material performance parameters, including cement strength grade, sand and gravel gradation, and raw material moisture content, as well as construction and curing conditions. The collected data is aligned with historical concrete mix design data in terms of time reference and format, and a training dataset is generated for model training.

[0109] The model building module is used to construct a multidimensional nonlinear mapping model between concrete material composition parameters and target performance indicators, including compressive strength, workability, and durability, based on the training dataset and using machine learning algorithms, and to jointly predict the target performance indicators.

[0110] The cross-site optimization module is used to input real-time sensor data from multiple batching plants into the mapping model during the concrete production process, perform cross-batch ratio optimization calculations based on target performance index constraints, and generate mix ratio adjustment schemes.

[0111] The resonance detection module is used to detect whether there is a resonance relationship between the periodic fluctuation of the moisture content of raw materials and the iteration step size of the mapping model during dynamic construction. When a resonance relationship is detected, the module triggers the parameter smoothing mechanism and the dynamic disturbance suppression mechanism to correct the mix proportion adjustment scheme.

[0112] The execution control module is used to apply the revised mix proportion adjustment plan to the concrete production process of multiple mixing plants, and to control the actual execution values ​​of the input of each raw material and the water-cement ratio.

[0113] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0114] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0115] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0116] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0117] The above are merely specific embodiments of this application, but the scope of protection of this application 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 this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for optimizing a concrete mix based on machine learning, characterized in that, The method comprises the following steps: Collecting multi-dimensional raw material performance parameters and construction conditions and maintenance conditions as material composition parameters, aligning the collected data with historical concrete test data in time base and unifying the format, and generating a training data set; Based on the training data set, a multi-dimensional nonlinear mapping model between the material composition parameters and the target performance indicators of compressive strength, workability and durability is constructed by using a machine learning algorithm, and the target performance indicators are jointly predicted; In the concrete production process, real-time sensing data from multiple mixing stations is input into the mapping model, cross-station mix proportion optimization calculation based on target performance indicator constraints is performed, and a mix proportion adjustment scheme is generated; In the dynamic construction process, based on the iteration step of real-time sensing data and the mapping model, it is detected whether there is a resonance relationship between the periodic fluctuation of raw material moisture content and the iteration step, including: First, the raw material moisture content time series of each mixing station is processed by outlier removal and normalization, then the discrete Fourier transform is used to extract the moisture content main frequency and corresponding spectral energy, and at the same time, the iteration frequency is calculated according to the iteration step of the mapping model and the jitter range of the iteration frequency is estimated within the rolling time window; A set of resonance discrimination criteria is established, and whether the resonance relationship exists is determined according to the set of resonance discrimination criteria, and a resonance mark is output when the resonance relationship exists; When the existence of the resonance relationship is detected, the parameter smoothing mechanism and the dynamic disturbance suppression mechanism are triggered to modify the mix proportion adjustment scheme, including: The ratio of water to cementitious material, sand ratio and admixture dosage are formed into time series by batch, and the median filter with odd batch number as window length is used to eliminate pulse anomalies, then the segmented boundary obtained by change point detection is used to perform polynomial fitting in each segment not higher than three orders, and the first derivative continuity of adjacent segments is used as a constraint to realize segmented polynomial smoothing; The first order change amount of adjacent batches is respectively set to an amplitude upper limit and a direction constraint, the amplitude upper limit is determined according to the percentile of the historical stable interval and the out-of-limit change is clipped, the direction constraint is determined by trend estimation of the recent batches to realize monotonic non-increasing or monotonic non-decreasing and the change violating the trend is projected and corrected, at the same time, the time phase dislocation is performed according to the raw material moisture content main frequency and phase distribution of each mixing station, so that the execution start time of each mixing station is staggered according to the phase equalization, and the adjustment amplitude upper limit of each mix proportion parameter is set as a function inversely proportional to the performance margin to suppress synchronous amplification; The smoothed and amplitude-limited mix proportion modification scheme is input into the mapping model, the compressive strength, workability and durability are respectively predicted and compared with the target performance indicator constraints, if all meet the target performance indicator constraints, it is marked as executable, if any does not meet the target performance indicator constraints, the mix proportion adjustment amplitude is reduced by a preset proportion factor and the prediction is repeated until the baseline mix proportion scheme that meets the target performance indicator constraints is obtained and marked as executable; The mix proportion modification scheme marked as executable is issued to each mixing station and bound with the batch identification; The modified mix proportion adjustment scheme is applied to the concrete production link of multiple mixing stations to control the actual execution values of each raw material feeding amount and water-cement ratio. 2.The machine learning-based concrete mix optimization method of claim 1, wherein, Before the training data set is generated, the moisture content is corrected by the following steps: The standard moisture content measured by the laboratory drying method and the real-time moisture content measured by the on-site electrical sensor are collected simultaneously for the same batch of raw materials. The zero-point offset and the full-scale proportionality coefficient are calculated using multiple sets of corresponding data from the two methods, and they are defined as the initial range coefficient; The initial range coefficient is applied to the continuous time series of moisture content data. The sliding window polynomial fitting method is used to smooth the sensor output within each window. The fitting residual between the smoothed value and the laboratory standard value is calculated at the boundary of each window. When the residual exceeds the pre-set fitting deviation threshold, the full-scale proportionality coefficient is corrected according to the pre-set residual proportion factor, thereby forming a dynamic adjustment mechanism for the range coefficient; The temperature mean and particle size distribution mean of the batch of raw materials are calculated. The temperature mean is weighted by the first weight, and the particle size distribution mean is weighted by the second weight to obtain the raw material temperature-particle comprehensive index. The comprehensive index is used as the independent variable, and the difference between the smoothed sensor output and the laboratory standard value is used as the dependent variable to perform multivariate regression modeling, obtaining a moisture content correction curve that adapts to different temperature and humidity conditions and particle structures; The moisture content correction curve is stored in the model parameter library, and when generating the training data set, the curve is called to standardize and correct all on-site moisture content data. 3.The machine learning based concrete mix optimization method of claim 1, wherein, Before generating the training data set, abnormal samples need to be removed. Specifically: A multi-dimensional feature matrix containing strength grade, particle gradation, moisture content, and curing temperature and humidity is constructed. The multivariate consistency test based on Mahalanobis distance is performed to detect abnormal samples that deviate from the overall distribution beyond the confidence interval threshold and remove them. 4.The method of claim 1, wherein, The steps of building a multi-dimensional nonlinear mapping model based on the training data set and performing joint prediction include the following process: With material composition parameters, i.e., multi-dimensional raw material performance parameters, construction conditions, and curing conditions as inputs, and compressive strength, workability, and durability as outputs, a joint prediction model is established, and material science prior constraints are introduced. The law that the increase in the ratio of water to cementitious material usage shows a decreasing relationship with compressive strength after the critical point, and the law that the sand ratio within the allowed range shows a monotonically increasing relationship with workability are written into the monotonicity constraint term and the critical point piecewise constraint term; The expected value and uncertainty of compressive strength, workability, and durability are simultaneously given at the output end of the model. The objective function based on heteroscedastic Gaussian likelihood is used to adaptively learn the variance of each output, and the gradient balancing strategy is used to dynamically weight the multi-output loss, allowing different performance indicators to converge evenly during the training process; And a reverse mapping sub-model is constructed with target performance indicators as input and material composition parameters as output, which together with the forward mapping model forms a closed-loop structure, and the deviation between the same sample after forward prediction and reverse reconstruction is constrained by a cyclic consistency penalty term; The joint prediction model is trained and selected using split-fold cross-validation and independent external validation sets. The uncertainty intervals of each performance indicator are calibrated based on the quantile calibration method, and the calibrated model parameters and constraints are used as the final version of the mapping model. 5.The method of claim 4, wherein, The cross-station mix proportion optimization calculation includes the following steps: Input real-time sensing data from each mixing station into the mapping model to generate a single-station mix proportion scheme meeting the constraints of target performance indicators such as compressive strength, workability and durability, respectively; Collect all single-station mix proportion schemes and construct a comprehensive evaluation function, which is formed by weighting and summing the total raw material consumption score, production cost score and performance margin score according to preset weights, wherein the total raw material consumption score is calculated according to the difference between the raw material usage of each scheme and the target usage, the production cost score is calculated according to the proportion of material cost in total cost of each scheme, and the performance margin score is calculated according to the safety margin proportion of performance indicators and target values of each scheme; The mix proportion scheme with the highest score of the comprehensive evaluation function is combined as the cross-mixing station mix proportion adjustment scheme. 6.The method of claim 1, wherein, The resonance discrimination criterion set includes the following: Frequency proximity criterion: the relative deviation of the main frequency value of the water content from the iteration frequency or the first harmonic thereof does not exceed the allowable deviation coefficient; Phase locking criterion: the time-averaged absolute change of the phase difference between the main frequency component of the water content and the iteration frequency reference signal within the window does not exceed the phase change threshold; Energy amplification criterion: the growth rate of the mix proportion adjustment amplitude variance or the water to cementitious material usage ratio variance within an observation duration of not less than twice the iteration step length is not less than the energy amplification threshold; Any criterion triggers the determination principle, and when any of the frequency proximity criterion or the phase locking criterion or the energy amplification criterion is met, it is directly determined that the resonance relationship exists.

7. A machine learning based concrete mix optimization system based on the machine learning based concrete mix optimization method of any one of claims 1-6, characterized in that, Specifically, it includes: A data acquisition module for acquiring multi-dimensional raw material performance parameters including cement strength grade, sand and stone gradation, and raw material moisture content, as well as construction conditions and curing conditions, aligning the collected data with historical concrete trial data in time reference and unifying the format, and generating a training data set for model training; A model construction module for constructing a multi-dimensional nonlinear mapping model between concrete material composition parameters and target performance indicators including compressive strength, workability and durability based on the training data set using machine learning algorithms, and jointly predicting the target performance indicators; A cross-station optimization module for inputting real-time sensing data from multiple mixing stations into the mapping model during concrete production to perform cross-mixing station mix proportion optimization calculation based on target performance indicator constraints to generate a mix proportion adjustment scheme; A resonance detection module for detecting whether there is a resonance relationship between the periodic fluctuation of raw material moisture content and the iteration step length based on real-time sensing data and the iteration step length of the mapping model during dynamic construction, and triggering parameter smoothing and dynamic disturbance suppression mechanisms to modify the mix proportion adjustment scheme when the resonance relationship exists; An execution control module for applying the modified mix proportion adjustment scheme to the concrete production link of multiple mixing stations to control the actual execution values of the raw material dosage and water-cement ratio.

Images