Method for predicting strength of cement stabilized chrome-iron slag macadam base based on machine learning

Abstract

The application relates to the technical field of material detection, in particular to a cement stabilized chrome-iron slag macadam base strength prediction method based on machine learning, which comprises the following steps: acquiring the cement mixing amount, glass phase proportion, free calcium oxide proportion, and environment temperature sequence and relative humidity sequence of the material to be detected; if the cement mixing amount is lower than a preset alkalinity threshold value or the cumulative effective accumulated temperature calculated based on the environment temperature sequence is lower than a preset accumulated temperature threshold value, the material is determined to be in an unexcited state, otherwise, the material is determined to be in an excited state; taking the basic prediction strength or the reinforced prediction strength as the initial strength; counting the high-humidity days; if the free calcium oxide proportion and the high-humidity days are respectively greater than preset safety limit values and time threshold values, a discount coefficient is calculated, and the final strength is obtained, otherwise, the initial strength is directly taken as the final strength. The application can convert the material dynamic phase change rule into a model structure constraint, so as to solve the problem of base strength prediction failure under a complex curing environment.

Description

Technical Field

[0001] This application relates to the technical field of materials testing, and in particular to a method for predicting the strength of cement-stabilized ferrochrome slag crushed stone base course based on machine learning. Background Technology

[0002] With the rapid advancement of infrastructure construction, cement-stabilized crushed stone base courses play a crucial role in road engineering. The resource utilization of industrial waste such as ferrochrome slag in these base course materials has become an important development trend in the engineering field in response to green and environmentally friendly requirements. Accurate assessment of base course strength is a core element in ensuring the overall quality, structural safety, and long service life of road engineering projects. Traditional laboratory strength testing typically requires specimens to be cured in a standard environment for a long period before undergoing artificial mechanical destructive testing. This method is not only time-consuming and costly but also significantly deviates from the complex curing conditions of actual construction sites. In recent years, machine learning-based strength prediction technology, with its powerful nonlinear fitting and efficient data processing capabilities, can directly predict material properties without the need for lengthy physical curing cycles. It has gradually become an important technical approach for solving the problem of rapid non-destructive testing and performance evaluation of engineering materials.

[0003] Among existing machine learning-based strength prediction methods, Chinese invention patent authorization announcement number CN120412804B discloses a method and system for joint prediction of cement strength based on a cascade model. This technical solution proposes first obtaining the initial composition information of the material to be predicted, and then inputting it into a pre-trained cascade model to predict the final material strength. Its core mechanism lies in the fact that the cascade model contains two levels: a structure mapping model and a strength prediction model. During the method's operation, the structure mapping model first maps the initial composition information to the final microstructure information. Then, the initial composition information and the mapped structure information are used as input, and the strength prediction model jointly outputs the final predicted strength. This method effectively overcomes the limitations of relying solely on a single type of information for prediction. Under standard curing environments and conventional cement material systems, it can establish relatively stable data correlations, thereby improving the overall prediction accuracy.

[0004] However, the aforementioned existing technologies reveal a serious technical flaw when faced with the complex and ever-changing natural maintenance environment in actual road engineering and with ferrochrome slag aggregates containing special microscopic physicochemical properties. The underlying logic of the model is severely disconnected from the actual dynamic evolution of the material. The core technical problem lies in the fact that existing deep learning prediction architectures essentially solidify the complex microscopic structural evolution of the material into a continuous, static black-box mapping process, completely failing to analyze and quantify the nonlinear phase transitions and hysteresis damage patterns of the material's microscopic physicochemical properties under specific environmental time-series induction. In complex practical application scenarios, the internal mechanisms of ferrochrome slag are highly dependent on threshold triggers from the external environment. For example, in construction scenarios with low temperatures or low alkalinity ratios in autumn and winter, the glassy phase on the slag surface remains permanently inert because it has not broken through the thermodynamic and chemical activation energy barriers. In such cases, existing continuous mapping models, lacking environmental threshold constraints, are prone to inertial derivations based on component ratios and severely overestimate the strength of the base course. Conversely, in rainy and humid environments, trace amounts of free calcium oxide within the material accumulate over time, causing delayed hydration and expansion. This leads to the sudden initiation of microcracks and deterioration of the micro-cementation network. Existing technologies that treat dynamic phase transformation processes as static prediction methods are completely unable to capture this phenomenon of base layer strength collapse triggered by high humidity environments. Ultimately, this can easily lead to unpredictable structural damage and serious safety hazards at the engineering site. Summary of the Invention

[0005] In order to transform the dynamic phase transformation law of materials into model structural constraints and solve the problem of predicting the failure of base course strength under complex curing environment, this application provides a machine learning-based method for predicting the strength of cement-stabilized ferrochrome slag crushed stone base course.

[0006] The machine learning-based method for predicting the strength of cement-stabilized ferrochrome slag crushed stone base courses provided in this application adopts the following technical solution: The machine learning-based method for predicting the strength of cement-stabilized ferrochrome slag crushed stone base courses includes: Obtain the cement content, glass phase ratio, free calcium oxide ratio, ambient temperature series, and relative humidity series of the material to be tested; If the cement content is lower than the preset alkalinity threshold, or the cumulative effective accumulated temperature calculated based on the ambient temperature sequence is lower than the preset accumulated temperature threshold, it is determined to be in an unactivated state; otherwise, it is determined to be in an activated state. When in the unexcited state, the feature weight of the glass phase ratio is set to zero for prediction, and the basic prediction intensity is output; when in the excited state, the feature weights of the glass phase ratio and the cumulative effective accumulated temperature are assigned to be non-zero for prediction, and the enhanced prediction intensity is output; the basic prediction intensity or the enhanced prediction intensity is used as the initial intensity. The number of high-humidity days exceeding the preset warning line in the relative humidity sequence is counted. If the proportion of free calcium oxide and the number of high-humidity days are greater than the preset safety limit and time threshold, respectively, a discount coefficient is calculated based on the excess amount of the proportion of free calcium oxide exceeding the safety limit and the number of high-humidity days. The discount coefficient is multiplied by the initial intensity to obtain the final intensity. Otherwise, the initial intensity is directly used as the final intensity.

[0007] Optionally, obtaining the cement content, glass phase ratio, free calcium oxide ratio, and ambient temperature and relative humidity sequences of the material to be tested includes: Extracting physical proportion characteristics to obtain the cement content provides a comparative basis for determining the unactivated or activated state; Chemical characteristics are extracted to obtain the proportion of the glass phase and the proportion of free calcium oxide, which are used as the basis for outputting the strengthening prediction intensity and calculating the discount factor, respectively. The environmental time-series characteristics of the on-site maintenance environment are collected to obtain the environmental temperature sequence and the relative humidity sequence, which are used as the environmental time-series basis for calculating the cumulative effective accumulated temperature and counting the number of high humidity days, respectively.

[0008] Optionally, the extraction of chemical features to obtain the glass phase ratio and the free calcium oxide ratio includes configuring the following synergistic constraints: The glass phase ratio is configured as a potential active feature and used as a weighted parameter controlled by the unexcited state and the excited state to control the output conditions of the enhancement prediction intensity. The percentage of free calcium oxide is configured as an expansion and deterioration characteristic, serving as a benchmark parameter for comparison with preset safety limits, and is used in conjunction with the number of high humidity days to trigger the calculation condition of the discount coefficient.

[0009] Optionally, the calibration logic for the preset alkalinity threshold and the preset accumulated temperature threshold is as follows: Chromium slag test specimens with different cement admixtures were prepared and cured. By comparing the strength changes of the chromium slag test specimens, the minimum cement admixture that triggers the activation of the glass phase was extracted as the preset alkalinity threshold. The minimum cumulative effective accumulated temperature required to trigger the activation of the glass phase is extracted as the preset accumulated temperature threshold. By using the preset alkalinity threshold and the preset accumulated temperature threshold as prerequisites, the determination of the activated state is controlled by the common conditions of the cement dosage and the accumulated effective temperature.

[0010] Optionally, the process of determining whether a state is unexcited or excited includes: If the cement content is lower than the preset alkalinity threshold, it is determined to be inert stage, and a first state identifier is output to characterize the unactivated state. If the cement content is not lower than the preset alkalinity threshold and the cumulative effective accumulated temperature is lower than the preset accumulated temperature threshold, it is determined to be in a dormant stage, and the first state identifier is output to characterize the unactivated state. If the cement content is not lower than the preset alkalinity threshold and the cumulative effective accumulated temperature is not lower than the preset accumulated temperature threshold, it is determined to be in the activation stage, and a second state identifier is output to characterize the activated state.

[0011] Optionally, the process of outputting the base prediction strength or the enhanced prediction strength includes: When the first state identifier is received, the feature weight of the glass phase ratio is set to zero, and prediction is performed using only other acquired features besides the glass phase ratio, and the basic prediction intensity is output. When the second state identifier is received, the glass phase ratio and the cumulative effective accumulated temperature are assigned non-zero feature weights, and these are used together with the other acquired features to make a prediction, and the enhanced prediction intensity is output.

[0012] Optionally, the process of outputting the base prediction strength or the enhanced prediction strength is performed based on a pre-segmented trained dual-branch prediction network, which is pre-trained in the following manner: The preset alkalinity threshold and the preset accumulated temperature threshold are extracted as the dividing boundary to divide the historical engineering sample dataset into an unexcited dataset and an excited dataset. Using the unexcited dataset and the excited dataset, the first branch and the second branch of the dual-branch prediction network are trained independently, so that the first branch and the second branch respectively perform the prediction process in the unexcited state and the excited state.

[0013] Optionally, the conditions for triggering the calculation of the discount coefficient based on the number of high-humidity days include: The number of days with high humidity is defined as the number of consecutive days exceeding the preset warning line; When the proportion of free calcium oxide configured as the expansion and deterioration characteristic and the number of high humidity days are greater than the preset safety limit and the time threshold, respectively, the physical expansion condition is determined to be met. In response to the physical inflation condition being met, the calculated discount coefficient is non-linearly reduced as the proportion of free calcium oxide exceeds the safety limit and the number of high humidity days increases, and a non-zero lower limit threshold is set for the discount coefficient.

[0014] Optionally, the process of calculating the discount factor includes: A preset penalty weight coefficient is introduced, and a nonlinear time factor is generated from the number of high humidity days; The excess amount of free calcium oxide exceeding the safety limit, the penalty weighting coefficient, and the nonlinear time factor are multiplied together to calculate and output the comprehensive reduction amount. Calculate the difference between the preset benchmark value and the comprehensive reduction amount, and limit it to the initial discount coefficient; The maximum value between the initial discount coefficient and the non-zero lower limit threshold is selected and used as the final output discount coefficient.

[0015] Optional, also includes: When the proportion of free calcium oxide configured as having the expansion and deterioration characteristics is not greater than the preset safety limit, or the number of high humidity days is not greater than the time threshold, the physical expansion condition is determined to be invalid. In response to the failure of the physical inflation condition, the calculation of the discount factor is blocked; The initial intensity is directly used as the final intensity output.

[0016] In summary, this application includes the following beneficial technical effects: 1. This application addresses the core technical problem of existing technologies that solidify the evolution of material microstructure into a static black box mapping and cannot analyze the threshold-type phase transformation and hysteresis damage law of ferrochrome slag under complex curing environment. By introducing a dual judgment mechanism of alkalinity and cumulative effective temperature, the dynamic phase transformation law of the material is transformed into a model structural constraint, fundamentally solving the engineering problems of overestimation of strength in low temperature or low alkalinity scenarios and failure prediction of strength collapse in rainy and humid scenarios.

[0017] 2. This application achieves accurate modeling of the glass phase activation process of ferrochrome slag by configuring the glass phase ratio as a weight parameter controlled by the unexcited and excited states, and training a dual-branch prediction network based on alkalinity threshold and accumulated temperature threshold. This significantly improves the accuracy and physical interpretability of the base layer strength prediction under different phase transformation stages.

[0018] 3. This application constructs a discount coefficient calculation mechanism based on superscalar and nonlinear time factors by configuring the proportion of free calcium oxide as an expansion and deterioration characteristic and introducing a dual judgment logic of high humidity days and safety limits. This effectively captures the strength deterioration law caused by the hysteretic hydration expansion of free calcium oxide in high humidity environment, filling the gap in the existing technology for predicting long-term service performance degradation. Attached Figure Description

[0019] Figure 1 This is the overall flowchart of the prediction method of this study; Figure 2 This is a diagram of the dual-branch prediction network architecture of this application; Figure 3 This is a comparison chart of the long-term intensity evolution of this application. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1-3 This application will be described in further detail.

[0021] Existing strength prediction technologies for cement-based materials mostly employ a static black-box mapping architecture, failing to analyze the unique microscopic physicochemical dynamic evolution laws of ferrochrome slag aggregates. This makes them unsuitable for adapting to the material phase transformation characteristics under complex on-site curing environments, easily leading to problems such as overestimation of strength in low-temperature, low-alkali scenarios and failure to predict collapse in rainy, high-humidity scenarios, posing a threat to the safety of road base structures. To address these technical issues, such as... Figure 1 As shown in the embodiments of this application, a method for predicting the strength of cement-stabilized ferrochrome slag crushed stone base course based on machine learning is disclosed. The following steps are described in detail: S1 Acquisition of the characteristics of the test material and the environmental time series characteristics This step comprehensively acquires fundamental data directly related to the entire process of strength generation, evolution, and deterioration of cement-stabilized ferrochrome slag aggregate base course from three dimensions: material physical proportions, microscopic physicochemical properties, and on-site curing environment. This forms a standardized feature dataset with mutual matching and complete dimensions, providing a traceable and highly correlated input foundation for the entire strength prediction system. The acquisition process for all features is designed around the unique hydration reaction mechanism of ferrochrome slag aggregate, precisely matching the inherent laws of material activity activation and long-term deterioration.

[0022] S11 Physical Proportion Feature Extraction and Acquisition This step extracts the core physical proportion characteristics of the cement-stabilized ferrochrome slag-aggregate mixture to be tested, with the cement content being the key target. Cement content is defined as the percentage of cement mass in the total dry aggregate mass of the cement-stabilized ferrochrome slag-aggregate mixture. It is a core parameter determining the alkalinity environment of the cement hydration system and directly affects the activation process of the ferrochrome slag glass phase.

[0023] Two traceable implementation paths were used to obtain the cement content, adapted to different engineering stages. For the unpaved mixture, the nominal cement content was directly extracted from the approved mix design documents. For the paved base course, representative samples were obtained through on-site core drilling, and the actual cement content was determined using the cement stone burning method. The burning method was implemented as follows: the drilled core samples were crushed to a particle size not exceeding 5mm, reduced to a sample size of not less than 200g using the quartering method, and dried to constant weight in an oven at 105℃±5℃. The total mass of the dried sample was recorded. The sample was then placed in a muffle furnace and ignited at 950℃-1000℃ for 60 minutes. After removal, it was placed in a desiccator to cool to room temperature, and the mass of the sample after ignition was recorded. Simultaneously conduct blank calibration tests, taking cement samples from the same batch and determining the loss on ignition rate of the cement under the same ignition conditions. Collective aggregate samples from the same origin and with the same gradation were taken and their ignition loss rate was determined under the same ignition conditions. The actual cement content is finally calculated using a formula: First, calculate the actual absolute mass of the cement. , Then calculate the actual cement dosage using the external admixture method. Cement dosage = .

[0024] This step simultaneously acquires three auxiliary proportioning characteristics: ferrochrome slag volume replacement rate, aggregate gradation parameters, and initial moisture content. The ferrochrome slag volume replacement rate, defined as the proportion of ferrochrome slag crushed stone to the total volume of coarse aggregate in the mixture, is directly extracted from the mix design document and is a key parameter affecting the total amount of active components in the system. Aggregate gradation parameters are obtained through mixture sieving tests, covering the individual and cumulative sieve residue data for each grade of aggregate on standard sieves for road base courses. Standardized and quantified using the cumulative passing rate of each standard sieve aperture, these are core parameters determining the load-bearing capacity of the base course's physical framework. Initial moisture content is obtained through the drying method, corresponding to the measured moisture content at the time of mixture paving and molding, directly affecting the early hydration reaction process of cement.

[0025] S12 Chemical Characterization and Cooperative Constraint Configuration This step uses standardized physicochemical testing methods to obtain two core microscopic characteristics of the ferrochrome slag under test: the proportion of glassy phase and the proportion of free calcium oxide. Simultaneously, based on the differentiated effects of these two characteristics on the evolution of base strength, corresponding synergistic constraint rules are configured to ensure a precise match between the characteristic attributes and the physicochemical properties of the ferrochrome slag itself.

[0026] The glass phase ratio is defined as the percentage of the mass of the amorphous glass phase in the ferrochrome slag aggregate that can participate in the hydration reaction to the total mass of the ferrochrome slag. This step configures the glass phase ratio as a potential active characteristic, whose contribution to strength directly depends on the alkalinity environment and hydration thermodynamic conditions of the system, matching the physicochemical properties of the ferrochrome slag glass phase, which only undergoes depolymerization under specific conditions. The determination of the glass phase ratio was performed using powder X-ray diffraction combined with the Rietveld full-spectrum fitting refinement method. The specific implementation procedure was as follows: a representative sample of ferrochrome slag used in the project was taken, crushed by a jaw crusher, and then ground by a vibratory mill until all particles passed through a 75 μm square-hole sieve to ensure that the particle fineness of the sample met the requirements of X-ray diffraction testing; the prepared powder sample was uniformly filled into the sample cell and flattened into a test piece using a glass slide; the test was performed using an X-ray diffractometer with a scanning range of 5° to 70°2θ, a scanning step size of 0.02°, and a dwell time of 0.5 s per step; the acquired diffraction patterns were subjected to background subtraction and smoothing, and a phase search was performed based on a database of common crystalline phases of ferrochrome slag, covering typical crystalline phases such as olivine, pyroxene, spinel, and iron oxides. The Rietveld full-spectrum fitting method was used to quantitatively separate and fit the diffraction peaks of the crystalline phases and the diffuse scattering peaks of the amorphous phases to obtain the mass fraction of the amorphous glass phase, which is the quantitative result of the glass phase ratio of the ferrochrome slag.

[0027] The percentage of free calcium oxide is defined as the percentage of the mass of free calcium oxide in ferrochrome slag relative to the total mass of the slag. This step configures the percentage of free calcium oxide as an expansion degradation characteristic, a core parameter affecting the long-term volume stability and strength retention of the substrate, and matches the degradation pattern of microcrack initiation within the material caused by the hydration expansion of free calcium oxide. The determination of the proportion of free calcium oxide was performed using the ethylene glycol-anhydrous ethanol titration method. The specific procedure was as follows: A representative sample of chromium slag was crushed, ground with agate, and passed through a 75 μm square-hole sieve. The powder passing through the sieve was dried to constant weight at 105℃±5℃ and placed in a desiccator for cooling. 0.5 g of the prepared powder sample was accurately weighed and placed in a dry conical flask. 20 mL of anhydrous ethanol-ethylene glycol mixed solution and a few drops of phenolphthalein indicator were added, and a reflux condenser was installed. The conical flask was placed on a constant-temperature magnetic stirrer and heated to a state of slight boiling. Reflux was maintained for 10 min, and titration was performed using a 0.1 mol / L benzoic acid anhydrous ethanol standard solution. The titration endpoint was reached when the red color of the solution completely disappeared. The amount of standard solution consumed was recorded. A blank test was performed simultaneously to eliminate reagent errors. Finally, the quantitative result of the proportion of free calcium oxide was calculated based on the actual amount of standard solution consumed.

[0028] S13 Maintenance Environment Temporal Characteristics Collection and Preprocessing This step involves collecting and standardizing environmental time-series data throughout the entire base course curing process to obtain continuous and complete environmental temperature and relative humidity sequences. Both sequences start on the day the base course mixture is laid and compacted, and end at the predicted target age, covering the entire predicted time range to ensure the time synchronization between the time-series data and the base course hydration curing process.

[0029] The ambient temperature series, composed of daily average ambient temperature data arranged chronologically within the predicted time range, is the core time-series data reflecting the thermodynamic conditions for cement hydration and the activation of ferrochrome slag. The relative humidity series, composed of daily average relative humidity data arranged chronologically within the predicted time range, is the core environmental data reflecting the progress of free calcium oxide hydration.

[0030] The collection of the two time-series data employed two implementation paths adapted to different site conditions. For projects with on-site monitoring equipment, hourly temperature and relative humidity monitoring data were extracted from fixed environmental monitoring stations deployed at the construction site, and daily average data were calculated using an arithmetic mean. For projects without on-site monitoring equipment, daily average temperature and daily average relative humidity data for the corresponding time range were extracted from daily surface meteorological observation data released by the meteorological authorities in the project location.

[0031] For missing data in the collected temperature and relative humidity sequences, this step employs a tiered completion strategy. If consecutive missing data in a single sequence does not exceed 3 days, it is completed using the moving average of the valid data from the adjacent 3 days. If consecutive missing data exceeds 3 days, it is replaced with daily observation data released by the national benchmark meteorological station in the project location during the same period. The completion process does not alter the temporal characteristics and numerical distribution of the original valid data, ensuring the continuity and integrity of the time series data.

[0032] S2 Material Active Phase Transition State Determination This step directly uses the cement content obtained in S11 and the ambient temperature sequence obtained in S13 as core inputs. Based on the hydration reaction kinetics mechanism of glass phase activation in ferrochrome slag, it constructs a nested judgment logic with dual constraints of alkalinity and thermodynamics to accurately classify the active phase transition stages of the material and output standardized state identifiers. This step overcomes the limitations of existing technologies that map material performance evolution as a continuous static black box, transforming the threshold-type activation law unique to ferrochrome slag into a rigid judgment threshold. This fundamentally avoids the interference of invalid features on the subsequent prediction process, ensuring that the prediction logic completely matches the actual reaction process of the material.

[0033] S21 State Determination Threshold Calibration This step standardizes and calibrates the preset alkalinity threshold and preset accumulated temperature threshold, clarifying the quantitative values ​​and calibration basis of the two thresholds, providing a reproducible and highly adaptable comparison benchmark for subsequent state determination. Both thresholds are specifically calibrated for the ferrochrome slag raw material and cement type used in the project, ensuring that the determination logic is highly consistent with the actual reaction characteristics of the materials.

[0034] The preset alkalinity threshold is defined as the minimum cement dosage required to trigger the activation of the glass phase in ferrochrome slag. The depolymerization of the glass phase is a typical chemical reaction, and its reaction process is controlled by the alkalinity environment of the system. Only when the alkalinity released by cement hydration breaks through the chemical barrier of the reaction can large-scale depolymerization and activation of the glass phase be achieved. If the alkalinity is insufficient, the glass phase will be in an inert state and will not produce a significant cementitious strengthening effect.

[0035] The standardized calibration process for the preset alkalinity threshold is as follows: First, prepare standard specimens with gradient cement admixtures. The cement admixture gradients are set to 2.0%, 3.0%, 4.0%, 5.0%, and 6.0%. Prepare at least six parallel specimens for each admixture gradient. The specimens use the standard 150mm×150mm×150mm cubic size commonly used in road engineering. The chromium slag volume substitution rate, aggregate gradation, and initial moisture content of the mixture all adopt fixed parameters from the project design. The specimen compaction degree is controlled to 98% of the requirements for road base construction. Second, place all prepared specimens in a standard curing environment. The standard curing environment parameters are: temperature 20℃±2℃, relative humidity ≥95%. Third, use a universal testing machine to determine the unconfined compressive strength of specimens with each admixture at 7d, 28d, 60d, and 90d. Plot the strength-age growth curves under different cement admixtures. Simultaneously test the strength of the control group specimens with the same cement admixture and no chromium slag substitution. The fourth step is to determine the inflection point of strength increase. The criterion is: at the same age, the unconfined compressive strength of the chromium-iron slag specimen increases by more than 15% compared to the pure cement-stabilized crushed stone control group, and the later-stage strength growth rate from 60 to 90 days of age increases by more than 30% compared to the control group. The later-stage strength growth rate is calculated as: Growth rate = (90-day strength - 60-day strength) / 60-day strength × 100%. The minimum cement content that meets this criterion is determined as the preset alkalinity threshold. In this embodiment, after the above calibration process, the preset alkalinity threshold is determined to be 4.0%.

[0036] The preset accumulated temperature threshold is defined as the minimum cumulative effective accumulated temperature required to trigger the large-scale activation reaction of the glass phase on the surface of ferrochrome slag, provided that the cement admixture meets the alkalinity threshold. Glass phase depolymerization is a typical chemical reaction, and the reaction process is directly determined by the ambient temperature. Only when the cumulative effective reaction energy of the system breaks through the activation energy barrier can large-scale glass phase depolymerization and activation be achieved. If the cumulative effective energy is insufficient, even if the alkalinity meets the requirements, the glass phase will remain in a dormant state and will not produce a significant cementitious strengthening effect.

[0037] The standardized calibration process for the preset accumulated temperature threshold is as follows: First, prepare standard specimens containing chromium-iron slag with a cement content not lower than the preset alkalinity threshold calibrated above. Simultaneously prepare control specimens of pure cement-stabilized crushed stone with the same cement content and no chromium-iron slag substitution. Prepare no fewer than 15 parallel specimens for each group. The specimen specifications, molding compaction degree, and mixture ratio parameters are consistent with the aforementioned alkalinity threshold calibration steps. Second, divide the chromium-iron slag specimens and the pure cement-stabilized crushed stone control specimens into multiple test groups and place them in different constant temperature curing environments. The curing temperature gradients are set at 5℃, 10℃, 15℃, 20℃, and 25℃. At least 3 parallel specimens are set for each temperature gradient. The relative humidity of the curing environment for each group of specimens is uniformly controlled at ≥95%. Third, for specimens cured at different temperatures, measure the unconfined compressive strength of the chromium-iron slag specimens in the same group and the control specimens at the corresponding temperature every 3 days. Record the curing time at which the chromium-iron slag specimens in each temperature group show a significant early strength increase. The strength leap judgment criterion here is: at the current test age, the unconfined compressive strength of the chromium slag-containing specimen increases by more than 15% compared to the control group specimens of pure cement-stabilized crushed stone cured at the same temperature. The fourth step clarifies the calculation rules for effective accumulated temperature. The daily effective accumulated temperature is the difference between the daily curing temperature and the reference temperature, which is set at 5℃. This value corresponds to the critical activity temperature of the hydration reaction of cement-based materials. When the daily curing temperature is below 5℃, the hydration reaction basically stops, and the daily effective accumulated temperature is calculated as 0. The cumulative effective accumulated temperature is the sum of all daily effective accumulated temperatures within the curing period. The fifth step calculates the cumulative effective accumulated temperature at which the early strength leap occurs in the chromium slag-containing specimens of each temperature group. The minimum value among the calculated results for each group is taken as the preset accumulated temperature threshold. In this embodiment, after the above calibration process, the preset accumulated temperature threshold is determined to be 600℃·d.

[0038] S22 Nested State Determination This step, based on the basic data obtained in S1 and the two thresholds calibrated in S21, performs a two-level nested logical judgment to accurately classify the active phase transition stage of the material. The judgment logic is fully matched with the dual constraint characteristics of the active activation of ferrochrome slag, avoiding misjudgment of the state caused by a single condition judgment.

[0039] The first stage is the pre-judgment of alkalinity conditions. The cement dosage obtained in S11 is compared with the preset alkalinity threshold calibrated in S21. If the cement dosage is lower than the preset alkalinity threshold, the system is directly judged to be in the inert stage. This stage corresponds to a permanently insufficient alkalinity in the system, and the chromium ferrochrome slag glass phase cannot be activated in an unactivated state. If the cement dosage is not lower than the preset alkalinity threshold, it means that the alkalinity conditions of the system meet the basic requirements for activation, and the system proceeds to the second stage of accumulated temperature condition judgment.

[0040] The second stage is the thermodynamic condition determination. First, based on the ambient temperature sequence obtained in S13, the cumulative effective accumulated temperature is calculated from the day the base layer is laid to the predicted age, according to the effective accumulated temperature calculation rules specified in S21. Then, the calculated cumulative effective accumulated temperature is compared with the preset accumulated temperature threshold calibrated in S21. If the cumulative effective accumulated temperature is lower than the preset accumulated temperature threshold, the system is determined to be in a dormant stage. This stage corresponds to the alkalinity condition being met, but the cumulative thermodynamic energy is insufficient, and the glass phase is in an unactivated state. If the cumulative effective accumulated temperature is not lower than the preset accumulated temperature threshold, the system is determined to be in an activated stage. This stage corresponds to the alkalinity and thermodynamic conditions being met, and the chromium slag glass phase has undergone large-scale depolymerization and activation, reaching an activated state.

[0041] S23 Status Identifier Output This step, based on the state determination result of S22, outputs standardized state identifiers to achieve unified and digital transmission of the material's active phase transition state. For the two scenarios determined to be inert or dormant stages, a unified first state identifier is output; both scenarios represent unactivated states where the ferrochrome slag glass phase has no active cementing contribution. For the scenario determined to be in the activated stage, a second state identifier is output, corresponding to the activated state, where the ferrochrome slag glass phase has exhibited an active cementing enhancement effect.

[0042] S3 base initial strength prediction This step uses the state identifier output by S2 as the core routing basis, synchronously calls the full feature data obtained by S1, dynamically adjusts the feature participation rules based on the material's active phase transition state, and completes the strength prediction for the corresponding state through a pre-constructed dual-branch prediction network, outputting the initial strength of the base layer. This step breaks through the limitations of the black-box mapping of static input of full features in existing technologies, directly transforming the determination result of the material's active state into the model's feature control rules, allowing the prediction process to fully conform to the strength contribution mechanism of ferrochrome slag in different states, and fundamentally avoiding the interference of invalid features on the prediction results.

[0043] Combination Figure 2 As shown, Figure 2 This is a diagram of the mechanism-driven dual-branch prediction network architecture of an embodiment of this application. Figure 2The comparison demonstrates the differences in the internal network structure constraints between the first and second branches, particularly showcasing the forced blocking state in the unexcited first branch where the feature weights of the glass phase proportion are set to zero. This directly proves how this application breaks through the limitations of continuous static black-box mapping, transforming the complex phase transition physics mechanism into structural constraints at the model's underlying level.

[0044] S31 dual-branch prediction network pre-built This step pre-constructs the dataset for the dual-branch prediction network, performs mechanism-driven data segmentation, and trains the branches independently, forming two independent prediction branches adapted to different material activity states: the first branch corresponding to the unexcited state and the second branch corresponding to the excited state. The construction of both branches relies entirely on the material mechanism and threshold parameters defined in the previous steps, ensuring a deep fit between the model architecture and the hydration reaction law of ferrochrome slag, rather than simply data fitting.

[0045] First, a historical engineering sample dataset is constructed to provide comprehensive and reliable basic data support for model training. Each sample in the dataset contains complete input feature labels and measured strength labels. The input feature labels completely correspond to the feature dimensions specified in step S1, specifically including cement content, glass phase ratio, cumulative effective accumulated temperature, free calcium oxide ratio, curing age, chromium slag volume replacement rate, aggregate gradation parameters, and initial moisture content. Among them, the cumulative effective accumulated temperature is calculated based on the daily ambient temperature data within the corresponding curing period of the sample, according to the effective accumulated temperature calculation rules specified in step S21; the aggregate gradation parameters are standardized and quantified using the cumulative passing rate values ​​of each standard sieve aperture. The measured strength label is the measured value of the unconfined compressive strength of the corresponding sample at the corresponding age. The data sources include accurate test data from indoor standard curing tests and core sampling data from actual engineering sites, ensuring the data's adaptability to the actual engineering scenario. The effective sample size of the sample dataset is no less than 500 groups, which fully covers actual engineering scenarios with different cement admixtures, different ferrochrome slag raw materials, different curing environments, and different ages, ensuring the scenario coverage and feature distribution balance of the dataset, and providing a data foundation for the generalization ability of the model.

[0046] Subsequently, a rigid partitioning of the dataset was performed based on the material mechanism. The partitioning boundaries were entirely based on the preset alkalinity threshold and preset accumulated temperature threshold calibrated in S21, and the partitioning rules were completely consistent with the nested state determination rules in S22. If the cement content corresponding to a sample is lower than the preset alkalinity threshold, or the cumulative effective accumulated temperature is lower than the preset accumulated temperature threshold, the sample is assigned to the unexcited dataset; if the cement content corresponding to a sample is not lower than the preset alkalinity threshold, and the cumulative effective accumulated temperature is not lower than the preset accumulated temperature threshold, the sample is assigned to the excited dataset. This data partitioning method based on the material mechanism, unlike the random partitioning mode of conventional machine learning models, ensures that the training data of the two branches completely correspond to their respective suitable material phase transition states, guaranteeing the prediction accuracy of the branch models from the data source, and also ensuring that the model fitting process completely fits the actual reaction law of the material.

[0047] Finally, the first and second branches were trained independently using both unexcited and excited datasets. This embodiment employs the XGBoost regression model as the foundation for both branches. This model, in prediction scenarios involving tabular engineering data, exhibits excellent nonlinear fitting ability, anti-overfitting ability, and interpretability, making it suitable for engineering needs such as predicting the strength of cement-stabilized materials. Both branches use the same core hyperparameter settings: maximum tree depth of 6, learning rate of 0.1, number of iterations of 200, and mean squared error as the loss function. This ensures that the basic fitting ability of the two branches remains consistent, achieving targeted adaptation to different states only through differences between input features and training data.

[0048] The training process for the first branch is as follows: The unexcited dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set is used for model parameter fitting, the validation set is used for overfitting monitoring during training, and the test set is used for final model accuracy verification. The glass phase ratio is excluded from the model's input features, and only all other input features are retained. This solidifies the constraint that the weights of the glass phase ratio feature in the unexcited state should be zero during the model training phase, avoiding interference from invalid features during the prediction phase. During training, an early stopping strategy is implemented based on the validation set loss. The number of early stopping rounds is set to 20. When the validation set loss does not decrease for 20 consecutive iterations, training is terminated early to avoid model overfitting. After training is completed, the model weights and fixed input feature configuration of the first branch are saved for subsequent prediction calls.

[0049] The training process for the second branch is as follows: the excited dataset is divided into training, validation, and test sets in an 8:1:1 ratio, with the dataset division rules consistent with those of the first branch; the model's input features include all input features, ensuring that the proportion of the glass phase and the cumulative effective accumulated temperature can fully participate in model fitting, and the corresponding non-zero feature weights are automatically learned during training to accurately match the gelation enhancement effect produced by the depolymerization of the chromium slag glass phase in the excited state; independent training is completed using the same training strategy and hyperparameter settings as the first branch, and the model weights and full input feature configuration of the second branch are saved after training for subsequent prediction calls.

[0050] Calculation of predicted base strength under unexcited state of S32 Upon receiving the first state flag output from S2, the prediction process in this step is triggered, directly calling the pre-trained first branch model to perform strength calculation. This step filters out the valid input features determined during the pre-training of the first branch from the full set of features obtained from S1, completely excluding the glass phase proportion feature. The filtered feature data is then input into the first branch model, and the model's output prediction result is the basic predicted strength. This process, through fixed feature selection rules, achieves the core requirement of zeroing the weight of the glass phase proportion feature, mapping only the strength contribution of cement bulk hydration and aggregate physical skeleton. This perfectly matches the inert material properties of the glass phase in ferrochrome slag in the unexcited state, fundamentally avoiding the strength overestimation problem caused by indiscriminate feature input in existing technologies.

[0051] Calculation of enhanced prediction intensity in the S33 excited state Upon receiving the second state flag output from S2, the prediction process in this step is triggered, directly calling the pre-trained second branch model to perform intensity calculation. This step inputs all feature data acquired in S1 into the second branch model. The proportion of the glass phase and the cumulative effective accumulated temperature have corresponding non-zero feature weights obtained through the pre-training process. The model can fully fit the gelation enhancement effect brought about by secondary hydration after glass phase depolymerization, and the output prediction result is the enhanced predicted intensity. This process fully covers the nonlinear growth law of the intensity after ferrochrome slag activation. Unlike the static continuous mapping of existing technologies, it can accurately capture the intensity leap effect brought about by activation, ensuring the prediction accuracy and physical rationality in the activated state.

[0052] S34 initial strength determination This step determines the basic predicted strength output by S32 or the enhanced predicted strength output by S33 as the initial strength of the base layer, which serves as the benchmark data for subsequent strength calibration.

[0053] Final strength determination of S4 base layer This step uses the initial strength output by S3 as the correction benchmark, and simultaneously calls the core data of free calcium oxide percentage and relative humidity sequence obtained by S1. Targeting the delayed hydration expansion and deterioration effect of free calcium oxide in ferrochrome slag, it constructs a mechanism-driven condition judgment and strength correction logic to accurately calibrate the initial strength and output the final predicted strength of the base layer. This step overcomes the core deficiency of existing technologies that only focus on the strength growth process and cannot capture the delayed strength decay during long-term service. It directly transforms the intrinsic mechanism of material degradation and environmental triggering conditions into quantifiable correction rules, allowing the prediction results to fully cover the entire cycle of base layer strength evolution from hydration and formation to long-term service, effectively avoiding the engineering risk of strength collapse prediction failure in high-humidity scenarios.

[0054] Combination Figure 3 As shown, Figure 3 This is a comparison chart of the long-term strength evolution of embodiments of this application. Figure 3 The paper presents a comparison curve of the long-term evolution of measured values, existing static mapping prediction results, and the predicted intensity of this application. This demonstrates the scientific basis for the introduced mechanism and intuitively verifies the significant technological advancements of the method in capturing hysteretic hydration expansion and solving the problem of predicting failure due to strength collapse under high humidity conditions.

[0055] S41 Degradation Correction Threshold Calibration This step standardizes and calibrates the preset relative humidity warning line, preset free calcium oxide safety limit, preset time threshold, non-zero lower limit threshold, and penalty weight coefficient, clarifying the quantitative values ​​and calibration basis of each threshold. This provides a reproducible and highly adaptable comparison benchmark for subsequent condition judgment and intensity correction. All thresholds are specifically calibrated for the ferrochrome slag raw materials and engineering mix proportions used in the project, ensuring a high degree of match between the correction logic and the actual deterioration characteristics of the materials.

[0056] The preset relative humidity warning line is defined as the critical relative humidity value that triggers a significant hydration reaction of free calcium oxide. The hydration reaction of free calcium oxide requires a sufficient moisture environment. Only when the relative humidity reaches the critical level can the hydration reaction continue and produce a cumulative expansion effect. Below this critical value, the hydration reaction rate is extremely slow, the reaction process essentially stops, and no expansion deformation leading to strength degradation occurs. The standardized calibration process for the preset relative humidity warning line is as follows: First, prepare cement-stabilized ferrochrome slag crushed stone specimens with an engineering mix proportion suitable for the project. The specimens adopt the standard 150mm×150mm×150mm cubic size commonly used in road engineering, with a compaction degree controlled at 98%. The cement content, ferrochrome slag replacement rate, and other parameters of the mixture are completely consistent with the project design parameters. Second, divide the same batch of specimens into multiple test groups and place them in constant temperature curing environments with different relative humidity levels. The relative humidity gradients are set to 60%, 70%, 80%, 85%, 90%, and 95%, and the curing temperature is uniformly set to 20℃. The first step involves preparing at least three parallel test specimens for each group. The second step involves measuring the volumetric expansion rate and unconfined compressive strength of each group of specimens at 28d, 60d, 90d, and 180d, and plotting performance change curves under different humidity conditions. The third step involves determining the critical humidity inflection point. The criterion is: when the relative humidity reaches a certain value, if the volumetric expansion rate of the specimen at the corresponding age increases by more than 80% compared to the volumetric expansion rate of the specimen at the same age under the next lower relative humidity gradient, and the unconfined compressive strength attenuation rate shows a significant increase during the same period, this critical relative humidity value is determined as the preset relative humidity warning line. In this embodiment, after the above calibration process, the preset relative humidity warning line is determined to be 85%.

[0057] The preset safety limit for free calcium oxide is defined as the maximum percentage of free calcium oxide that will not cause significant expansion and deterioration of the base material. When the percentage of free calcium oxide is below this limit, the slight expansion and deformation caused by hydration can be accommodated by the internal pore structure of the cement paste, without causing microcrack damage and strength reduction in the cementing system. When the limit is exceeded, the expansion stress caused by hydration will exceed the ultimate tensile strength of the cement paste, initiating continuous microcracks inside the material, which will continue to expand as the hydration process progresses, ultimately leading to a continuous decrease in strength. The standardized calibration procedure for the preset safety limit of free calcium oxide is as follows: First, prepare standard specimens with gradient free calcium oxide percentages of 0.5%, 1.0%, 1.5%, 2.0%, 2.5%, and 3.0%. Prepare at least six parallel specimens for each gradient, ensuring the specimen specifications, compaction degree, and basic mix proportions are consistent with the engineering design. Second, place all specimens in a standard curing environment with 95% relative humidity and 20°C for long-term curing up to 180 days. Third, periodically measure the unconfined compressive strength and volumetric expansion rate of specimens at each gradient, and fully record the performance changes at different curing ages. Fourth, determine the safety threshold. The criterion is: when the free calcium oxide percentage is below a certain value, the strength decay rate of the specimen at 180 days is less than 5%, the volumetric expansion rate is less than 0.05%, and there are no significant deterioration characteristics; when this value is exceeded, both the strength decay rate and expansion rate show a significant jump. This threshold value is then determined as the preset safety limit for free calcium oxide. In this embodiment, the preset safety limit for free calcium oxide is determined to be 1.0% after the above calibration process.

[0058] The preset time threshold is defined as the minimum number of consecutive high-humidity days required to trigger significant degradation of material strength. The hydration and expansion of free calcium oxide is a cumulative process over time. Short-term high-humidity environments only produce trace hydration reactions, and the expansion and deformation are insufficient to trigger structural degradation within the material. Only when the number of consecutive high-humidity days reaches a critical value will the cumulative effect of hydration and expansion exceed the material's tolerance limit, leading to microcrack propagation and strength reduction. The standardized calibration procedure for the preset time threshold is as follows: First, using standard specimens with a free calcium oxide content exceeding the aforementioned safety limit, prepare no fewer than 15 parallel specimens as test groups, with specimen parameters consistent with the aforementioned calibration steps; simultaneously prepare the same number and mix proportion of specimens as a baseline control group; Second, place the test group specimens in a constant high-humidity curing environment with relative humidity ≥95% and temperature 20℃, setting different continuous high-humidity curing durations, with duration gradients of 3d, 7d, 14d, 21d, and 28d, and setting no fewer than 3 parallel specimens for each duration gradient; the baseline... The control group specimens were simultaneously cured in a baseline curing environment at 20°C and relative humidity below a preset relative humidity warning line (e.g., 60%). In the third step, after curing for the corresponding duration, the unconfined compressive strength of each experimental group specimen was measured and compared with that of the baseline control group specimens of the same age to calculate the corresponding strength attenuation rate. In the fourth step, a time threshold was determined. The criterion was: when the number of consecutive high-humidity days was below a certain value, the strength attenuation rate of the specimen was less than 3%, showing no significant deterioration characteristics; when this value was reached and exceeded, the strength attenuation rate increased significantly, and this critical number of days was determined as the preset time threshold. In this embodiment, after the above calibration process, the preset time threshold was determined to be 7 days.

[0059] The non-zero lower limit threshold is defined as the minimum value of the discount factor, used to avoid extremely low values ​​of strength reduction that are inconsistent with engineering realities. Even under the most severe high-humidity environments and the highest levels of free calcium oxide exceeding the standard, the aggregate interlocking skeleton of the cement-stabilized chromium slag crushed stone base course will still retain the basic bearing capacity, and the overall strength of the material will not be completely lost. In this embodiment, based on the long-term service performance test data of cement-stabilized base courses for high-grade highways in China, and the minimum bearing capacity requirements of semi-rigid base course structures, the non-zero lower limit threshold is set to 0.6.

[0060] The penalty weighting coefficient is a fixed parameter in the degradation correction process, used to quantify the strength degradation magnitude under the combined effects of excessive free calcium oxide and high humidity days. The standardized calibration process for the penalty weighting coefficient is as follows: First, prepare multiple sets of specimens with different excessive free calcium oxide levels, covering a gradient from 0.2% to 2.0%. Second, place each set of specimens in a constant high humidity environment, setting different continuous high humidity curing durations. Compare these specimens with similar specimens under a low humidity baseline environment to determine the measured strength decay rate under different combinations. Third, perform univariate linear fitting analysis using multiple sets of experimental data, using the product of excessive free calcium oxide and the natural logarithm of the continuous high humidity days as the single independent variable, and the measured strength decay rate as the dependent variable. The resulting regression coefficient is the penalty weighting coefficient for the matching multiplication operation, ensuring that the deviation between the strength reduction magnitude calculated using this coefficient and the measured strength decay rate is controlled within 10%. In this embodiment, after the above calibration and fitting process, the penalty weighting coefficient is set to 0.08.

[0061] S42 High Humidity Days Statistics This step, based on the relative humidity sequence obtained in S1, completes the standardized statistics of high-humidity days. The statistical time range is completely consistent with the prediction period, covering the entire natural day period from the day the base layer is laid to the predicted age. The statistical rules strictly follow the continuous day requirement defined in the claims, traversing the entire relative humidity sequence, identifying each natural day whose daily average relative humidity exceeds the preset relative humidity warning line calibrated in S41, counting the number of consecutive days exceeding the preset warning line, and extracting the longest consecutive high-humidity days among all consecutive periods of exceeding the standard as the core parameter for subsequent condition determination. The statistical rules for continuous days perfectly match the cumulative effect characteristics of free calcium oxide hydration and expansion, avoiding misjudgments under intermittent high-humidity environments, and ensuring the accuracy and rationality of degradation determination.

[0062] S43 Determination of Physical Induced Inflation Conditions This step, based on the free calcium oxide percentage obtained in S1, the various thresholds calibrated in S41, and the number of consecutive high-humidity days counted in S42, performs a simultaneous determination of dual conditions to clarify the trigger boundary for strength correction. The determination logic perfectly matches the mechanism of free calcium oxide expansion and degradation; expansion and degradation, as well as strength attenuation, will only occur when both internal material factors and external environmental factors are present simultaneously. Specifically, the determination rule is as follows: when the free calcium oxide percentage exceeds the preset free calcium oxide safety limit, and the number of consecutive high-humidity days exceeds the preset time threshold, the physical expansion condition is deemed valid; if either condition is not met, the physical expansion condition is deemed invalid. This dual determination logic effectively avoids over-correction caused by a single trigger condition, ensuring the physical rationality of the strength prediction results.

[0063] S44 Final Strength Calculation and Output This step, based on the judgment result of S43, executes the corresponding strength processing flow and outputs the final predicted strength of the base layer.

[0064] When the physical expansion condition is not met, the calculation process of the discount coefficient is directly blocked, and the initial strength determined by S34 is directly used as the final strength output.

[0065] When the physical inflation condition is determined to be met, the discount factor is first calculated according to a standardized procedure, and then the initial strength is corrected using the discount factor to obtain the final strength. The calculation process of the discount factor strictly follows the nonlinear law of the degradation effect and consists of five consecutive steps: The first step is to calculate the excess amount of free calcium oxide. The excess amount equals the measured percentage of free calcium oxide minus the preset safety limit for free calcium oxide. During the calculation, the percentage value before the percentile sign of the excess amount of free calcium oxide is directly extracted into the calculation (for example, if the excess amount is 1.0%, then 1.0 is used in the calculation). Since the physical expansion condition has been determined to be met, the excess amount calculated here is a positive value.

[0066] The second step is to determine the nonlinear time factor. The nonlinear time factor is calculated from the number of consecutive days of high humidity using a natural logarithmic function. The formula is: Nonlinear time factor = ln(number of consecutive days of high humidity). The core reason for using the natural logarithmic function is that the hydration rate of free calcium oxide shows a trend of initially being fast and then slowing down with curing time. The impact of expansion degradation on strength also gradually slows down over time. The logarithmic function can accurately fit this nonlinear degradation pattern, avoiding excessive reduction in the later stages caused by linear calculations, and ensuring that the correction result highly matches the actual degradation process of the material.

[0067] The third step is to calculate the overall reduction. The overall reduction is equal to the product of the excess amount of free calcium oxide, the penalty weighting coefficient, and the nonlinear time factor. The calculated overall reduction is a dimensionless positive value, which characterizes the overall extent of intensity degradation.

[0068] The fourth step is to calculate the initial discount factor. The initial discount factor is equal to the preset baseline value minus the total reduction amount, where the preset baseline value is 1.0, representing the baseline state with no intensity reduction.

[0069] The fifth step is to determine the final discount factor. The maximum value between the initial discount factor and the non-zero lower limit threshold is taken. That is, when the initial discount factor is greater than or equal to 0.6, the initial discount factor is used as the final discount factor; when the initial discount factor is less than 0.6, 0.6 is directly used as the final discount factor. This limits the lower limit of strength reduction, ensuring that the predicted results conform to the actual bearing characteristics of the semi-rigid base course.

[0070] The calculated final discount factor is multiplied by the initial strength determined in S34, and the result is the final strength, thus completing the entire base layer strength prediction process.

[0071] This embodiment, through a series of nested technical steps, deeply integrates the microscopic hydration mechanism, environment-induced phase transition law, and strength prediction process of ferrochrome slag materials. It breaks through the limitations of existing static black-box mapping technologies, achieving accurate prediction of strength under different active states and effectively capturing hysteretic strength degradation in high-humidity environments. The prediction results highly match the actual service performance of the material, providing reliable technical support for quality control and safety assessment of cement-stabilized ferrochrome slag crushed stone base courses in road engineering. The above prediction method can be executed by a computer device equipped with a processor and memory. The memory stores a computer program that implements the logic of the above steps, and the processor executes the computer program to complete the entire prediction process.

[0072] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for predicting the strength of cement-stabilized ferrochrome slag crushed stone base course based on machine learning, characterized in that, include: Obtain the cement content, glass phase ratio, free calcium oxide ratio, ambient temperature series, and relative humidity series of the material to be tested; If the cement content is lower than the preset alkalinity threshold, or the cumulative effective accumulated temperature calculated based on the ambient temperature sequence is lower than the preset accumulated temperature threshold, it is determined to be in an unactivated state; otherwise, it is determined to be in an activated state. When in the unexcited state, the feature weight of the glass phase ratio is set to zero for prediction, and the basic prediction intensity is output; when in the excited state, the feature weights of the glass phase ratio and the cumulative effective accumulated temperature are assigned to be non-zero for prediction, and the enhanced prediction intensity is output. Use the basic predicted strength or the enhanced predicted strength as the initial strength; The number of high-humidity days exceeding the preset warning line in the relative humidity sequence is counted. If the proportion of free calcium oxide and the number of high-humidity days are greater than the preset safety limit and time threshold, respectively, a discount coefficient is calculated based on the excess amount of the proportion of free calcium oxide exceeding the safety limit and the number of high-humidity days. The discount coefficient is multiplied by the initial intensity to obtain the final intensity; otherwise, the initial intensity is directly used as the final intensity.

2. The method according to claim 1, characterized in that, The acquisition of the cement content, glass phase ratio, free calcium oxide ratio, and ambient temperature and relative humidity sequences of the material to be tested includes: Extracting physical proportion characteristics to obtain the cement content provides a comparative basis for determining the unactivated or activated state; Chemical characteristics are extracted to obtain the proportion of the glass phase and the proportion of free calcium oxide, which are used as the basis for outputting the predicted strengthening intensity and calculating the discount factor, respectively. The environmental time-series characteristics of the on-site maintenance environment are collected to obtain the environmental temperature sequence and the relative humidity sequence, which are used as the environmental time-series basis for calculating the cumulative effective accumulated temperature and counting the number of high humidity days, respectively.

3. The method according to claim 2, characterized in that, The extraction of chemical features to obtain the glass phase ratio and the free calcium oxide ratio includes configuring the following synergistic constraints: The glass phase ratio is configured as a potential active feature and used as a weighted parameter controlled by the unexcited state and the excited state to control the output conditions of the enhancement prediction intensity. The percentage of free calcium oxide is configured as an expansion and deterioration characteristic, serving as a benchmark parameter for comparison with preset safety limits, and is used in conjunction with the number of high humidity days to trigger the calculation condition of the discount coefficient.

4. The method according to claim 1, characterized in that, The calibration logic for the preset alkalinity threshold and the preset accumulated temperature threshold is as follows: Chromium slag test specimens with different cement admixtures were prepared and cured. By comparing the strength changes of the chromium slag test specimens, the minimum cement admixture that triggers the activation of the glass phase was extracted as the preset alkalinity threshold. The minimum cumulative effective accumulated temperature required to trigger the activation of the glass phase is extracted as the preset accumulated temperature threshold. By using the preset alkalinity threshold and the preset accumulated temperature threshold as prerequisites, the determination of the activated state is controlled by the common conditions of the cement dosage and the accumulated effective temperature.

5. The method according to claim 4, characterized in that, The process of determining whether a state is unexcited or excited includes: If the cement content is lower than the preset alkalinity threshold, it is determined to be inert stage, and a first state identifier is output to characterize the unactivated state. If the cement content is not lower than the preset alkalinity threshold and the cumulative effective accumulated temperature is lower than the preset accumulated temperature threshold, it is determined to be in a dormant stage, and the first state identifier is output to characterize the unactivated state. If the cement content is not lower than the preset alkalinity threshold and the cumulative effective accumulated temperature is not lower than the preset accumulated temperature threshold, it is determined to be in the activation stage, and a second state identifier is output to characterize the activated state.

6. The method according to claim 5, characterized in that, The process of outputting the base prediction strength or the enhanced prediction strength includes: When the first state identifier is received, the feature weight of the glass phase ratio is set to zero, and prediction is performed using only other acquired features besides the glass phase ratio, and the basic prediction intensity is output. When the second state identifier is received, the glass phase ratio and the cumulative effective accumulated temperature are assigned non-zero feature weights, and these are used together with the other acquired features to make a prediction, and the enhanced prediction intensity is output.

7. The method according to claim 6, characterized in that, The process of outputting the base prediction strength or the enhanced prediction strength is performed based on a pre-trained dual-branch prediction network: the dual-branch prediction network is pre-trained in the following manner: The preset alkalinity threshold and the preset accumulated temperature threshold are extracted as the dividing boundary to divide the historical engineering sample dataset into an unexcited dataset and an excited dataset. Using the unexcited dataset and the excited dataset, the first branch and the second branch of the dual-branch prediction network are trained independently, so that the first branch and the second branch respectively perform the prediction process in the unexcited state and the excited state.

8. The method according to claim 3, characterized in that, The conditions for triggering the calculation of the discount coefficient in conjunction with the number of high humidity days include: The number of days with high humidity is defined as the number of consecutive days exceeding the preset warning line; When the proportion of free calcium oxide configured as the expansion and deterioration characteristic and the number of high humidity days are greater than the preset safety limit and the time threshold, respectively, the physical expansion condition is determined to be met. In response to the physical inflation condition being met, the calculated discount coefficient is non-linearly reduced as the proportion of free calcium oxide exceeds the safety limit and the number of high humidity days increases, and a non-zero lower limit threshold is set for the discount coefficient.

9. The method according to claim 8, characterized in that, The process of calculating the discount factor includes: A preset penalty weight coefficient is introduced, and a nonlinear time factor is generated from the number of high humidity days; The excess amount of free calcium oxide exceeding the safety limit, the penalty weighting coefficient, and the nonlinear time factor are multiplied together to calculate and output the comprehensive reduction amount. Calculate the difference between the preset benchmark value and the comprehensive reduction amount, and limit it to the initial discount coefficient; The maximum value between the initial discount coefficient and the non-zero lower limit threshold is selected and used as the final output discount coefficient.

10. The method according to claim 8, characterized in that, Also includes: When the proportion of free calcium oxide configured as having the expansion and deterioration characteristics is not greater than the preset safety limit, or the number of high humidity days is not greater than the time threshold, the physical expansion condition is determined to be invalid. In response to the failure of the physical inflation condition, the calculation of the discount factor is blocked; The initial intensity is directly used as the final intensity output.

Images