A method for predicting multiple properties of high ductility concrete

Abstract

The application discloses a high-toughness concrete multi-performance prediction method, and particularly relates to the technical field of building material performance prediction, which comprises the following steps: obtaining high-toughness concrete sample data and preprocessing, establishing a BP neural network multi-performance prediction model with mix proportion parameters, fiber parameters and curing conditions as inputs and compressive strength, bending toughness, ultimate tensile strain and fracture energy as outputs; obtaining compressive strength prediction error and toughness index prediction error through test samples to generate strength bias learning parameters; further obtaining toughness failure process characteristics and constructing toughness instability discrete parameters, and comprehensively analyzing the two to obtain toughness prediction capability scores; then, correcting the BP neural network model according to the scores, and outputting multi-performance prediction results of the high-toughness concrete to be predicted and a toughness prediction reliability grade; the application can reduce the toughness index prediction distortion caused by the model bias of compressive strength fitting, and improve the high-toughness concrete performance prediction reliability.

Description

Technical Field

[0001] This invention relates to the field of building material performance prediction technology, specifically to a method for predicting multiple properties of high-toughness concrete. Background Technology

[0002] High-toughness concrete, due to its superior crack resistance, ductility, and energy dissipation capacity, has been increasingly applied in engineering applications such as bridge expansion joint repair, seismic energy-dissipating components, thin-layer pavement, and prefabricated connection nodes. Existing technologies typically employ data-driven methods such as backpropagation (BP) neural networks, using concrete mix proportions, fiber parameters, and curing conditions as inputs, and outputting performance indicators such as compressive strength, tensile strength, flexural toughness, ultimate tensile strain, and fracture energy to predict multiple properties of high-toughness concrete.

[0003] However, existing BP neural networks tend to fit numerically stable and highly regular indicators such as compressive strength more easily when making joint predictions of multiple properties. They are less accurate in predicting toughness indicators such as bending toughness, ultimate tensile strain, and fracture energy, which are greatly affected by fiber dispersion, crack propagation process, interfacial bonding state, and post-peak softening behavior. At the same time, existing methods usually evaluate model performance based on the overall prediction error, lacking specific identification of model strength fitting bias and toughness instability dispersion. This can lead to materials being judged as having sufficient strength but actually lacking in ductility, energy dissipation capacity, and crack control capabilities, affecting their reliable application in earthquake-resistant, repair, and high-toughness structures.

[0004] Therefore, there is an urgent need for a method to predict multiple properties of high-toughness concrete that can further evaluate and correct the predictive ability of toughness-related indicators based on BP neural network prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting multiple properties of high-toughness concrete to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting multiple properties of high-toughness concrete, comprising: Obtain high-toughness concrete sample data, including mix proportion parameters, fiber parameters, curing conditions, and corresponding measured values ​​of compressive strength, flexural toughness, ultimate tensile strain, and fracture energy; The sample data is preprocessed to obtain standardized sample data, which is then divided into training samples, validation samples, and test samples. Using mix proportion parameters, fiber parameters and curing conditions as inputs, and compressive strength, flexural toughness, ultimate tensile strain and fracture energy as outputs, a BP neural network multivariate performance prediction model was established and trained. The test samples are input into the trained BP neural network multivariate performance prediction model to obtain the predicted values ​​of each performance and compare them with the corresponding measured values ​​to obtain the prediction error of compressive strength and the prediction error of toughness index. Strength bias learning parameters are generated based on the prediction error of compressive strength and the prediction error of toughness index. Obtain the ductile failure process characteristics corresponding to the sample, and construct ductile instability discrete parameters based on the ductile failure process characteristics; The strength bias learning parameters and the toughness instability discrete parameters are combined and analyzed to obtain the toughness prediction capability score; The BP neural network multi-performance prediction model is modified based on the toughness prediction capability score, and the modified model is used to output the multi-performance prediction results and toughness prediction reliability level of the high-toughness concrete to be predicted.

[0007] Preferably, obtaining high-toughness concrete sample data includes: Establish sample numbers according to the same high-toughness concrete sample, and link and store the mix proportion parameters, fiber parameters, curing conditions with the measured values ​​of compressive strength, flexural toughness, ultimate tensile strain and fracture energy. The material usage, fiber size, curing age, and measured performance values ​​from samples from different sources were standardized to the same data unit, and the corresponding data source, specimen size, test method, and loading method were recorded. When the sample uses mixed fibers, the type, length, diameter and doping amount of each fiber are recorded separately, so that the different fiber features in the same sample can independently participate in feature expression in subsequent prediction; The test results of parallel specimens corresponding to each performance measured value and their dispersion are bound to the sample number to form the original sample data.

[0008] Preferably, the sample data is preprocessed, including: Field alignment and missing item screening were performed on the original sample data to remove data records that lacked any measured value of compressive strength, flexural toughness, ultimate tensile strain and fracture energy; Based on the median and robust dispersion of the measured values ​​of compressive strength, flexural toughness, ultimate tensile strain and fracture energy, respectively, the robust deviation values ​​of compressive strength and toughness response are obtained. When the compressive strength robustness deviation value of the same data record is in the central range while the toughness response deviation value is in the discrete range, or when the compressive strength robustness deviation value is in the discrete range while the toughness response deviation value is in the central range, it is determined that the data record has a disconnect between the strength response and the toughness response. Data records whose strength response and toughness response do not correspond are removed to obtain sample data after consistency verification.

[0009] Preferably, standardized sample data is obtained and divided into training samples, validation samples, and test samples, including: The mix proportion parameters, fiber parameters, curing conditions and various measured values ​​after consistency verification were treated as independent parameter categories and scaled to obtain initial standardized data. The fiber volumetric content, ultimate tensile strain and fracture energy were divided into four intervals according to the 25%, 50% and 75% quantiles. After standardization within each interval, the data were merged according to the sample number to obtain standardized sample data. Based on the standardized measured compressive strength and the comprehensive measured toughness obtained from bending toughness, ultimate tensile strain, and fracture energy, the samples are divided into low strength, medium strength, high strength, low toughness, medium toughness, and high toughness. The strength and toughness categories were combined to form nine joint distribution groups, and within each joint distribution group, training samples, validation samples, and test samples were divided in a 7:1:2 ratio.

[0010] Preferably, a BP neural network multivariate performance prediction model is established and trained, including: According to the order of the preprocessed fields, the mix proportion parameters, fiber parameters and curing conditions in the training samples are combined into an input feature sequence, and the compressive strength is used as the strength output label, and the bending toughness, ultimate tensile strain and fracture energy are used as the toughness output label. Each batch of training samples is organized into 9 joint distribution groups, so that each batch of training samples contains data records with different intensity levels and different toughness discrete levels. A multivariate performance prediction model of a BP neural network is established, which includes an input layer, a shared hidden layer, a compressive strength output node, and three toughness output nodes. During joint training, when the output error of compressive strength in the validation samples decreases continuously while the output error of toughness does not decrease synchronously, the connection weights corresponding to the output node of compressive strength are fixed, and only the connection weights from the shared hidden layer to the output node of toughness are updated in reverse until the output error of toughness decreases and then joint training is resumed.

[0011] Preferably, the predicted errors for compressive strength and toughness indicators are obtained, including: The test samples are input into the trained BP neural network multivariate performance prediction model according to the preprocessed field order to obtain the standardized predicted values ​​of compressive strength, flexural toughness, ultimate tensile strain and fracture energy corresponding to the sample number. According to the overall standardization rules or interval standardization rules recorded in the preprocessing stage, the standardized predicted values ​​are restored to the data scale consistent with the corresponding measured values; Based on the sample number, the restored performance prediction values ​​are matched with the corresponding measured values ​​one by one to obtain strength prediction comparison data and toughness prediction comparison data. Based on the strength prediction comparison data, the direction and magnitude of the compressive strength deviation are extracted. Based on the toughness prediction comparison data, the direction and magnitude of the deviation of the bending toughness, ultimate tensile strain and fracture energy are extracted. The prediction errors of the compressive strength and toughness indicators are then summarized according to the joint distribution group.

[0012] Preferably, strength bias learning parameters are generated based on the prediction errors of compressive strength and toughness-related indices, including: Within the same joint distribution group, the prediction errors of compressive strength and toughness indicators are sorted from low to high to obtain the strength error order and toughness error order for the same sample. Samples whose toughness error ranking is at least 20% higher than the strength error ranking within the same group are extracted as strength-biased candidate samples. The candidate samples with strength bias are screened for directional consistency. Samples with compressive strength deviation not exceeding the median of compressive strength deviation in the group and at least two toughness indicators showing the same direction of deviation are retained as strength bias confirmation samples. Based on the strength bias, the distribution ratio of the confirmed samples, the degree to which the toughness error order exceeds the strength error order, and the same-direction deviation of toughness indicators, strength bias learning parameters are generated.

[0013] Preferably, the ductile failure process characteristics corresponding to the sample are obtained, and ductile instability discrete parameters are constructed based on the ductile failure process characteristics, including: According to the sample number, load change data, deformation change data and crack evolution data are obtained during the bending toughness, ultimate tensile strain and fracture energy test process, and bound with the toughness error record to form a toughness failure process record. The ductile failure process record was divided into the initial cracking stage, the multi-crack propagation stage, and the post-peak attenuation stage, and the characteristics of crack number variation, crack width variation, crack spacing variation, and post-peak bearing capacity attenuation were extracted respectively. Based on the stage failure characteristics, the discrete characteristics of crack distribution, the concentrated development characteristics of deformation, and the post-peak load attenuation characteristics are extracted to form a set of ductile instability characteristics; Based on the degree of synchronous occurrence of the discrete characteristics of crack distribution, concentrated development characteristics of deformation, and post-peak load attenuation characteristics in the same sample, discrete parameters of ductile instability are constructed.

[0014] Preferably, the strength bias learning parameters and the ductility instability discrete parameters are comprehensively analyzed to obtain a ductility prediction capability score, including: Based on the sample number, the single sample strength bias learning parameters and the single sample ductile instability discrete parameters are matched to form a correlation record between model learning bias and material ductile instability. The high-level limits are determined based on the 75th percentile of the intensity bias learning parameter and the discrete parameter of toughness instability, and 0.50, respectively. Samples that reach the corresponding high-level limits at the same time are marked as toughness prediction risk samples. The distribution proportion of resilience prediction risk samples is statistically analyzed according to different strength levels and resilience dispersion levels. A non-compensation fusion method is adopted to fuse the strength bias learning parameter and the toughness instability discrete parameter, so that when either parameter is at a high level, it cannot be completely canceled by the low level of the other parameter. The toughness prediction capability score is obtained based on the fusion result and the distribution ratio of the toughness prediction risk sample.

[0015] Preferably, the BP neural network multivariate performance prediction model is modified based on the resilience prediction capability score, including: The correction method is determined based on the toughness prediction capability score. When the toughness prediction capability score is below 70, the BP neural network multivariate performance prediction model is retrained in a targeted manner. When the toughness prediction capability score is not lower than 70 but lower than 85, the output nodes of bending toughness, ultimate tensile strain and fracture energy are lightly retrained. When the toughness prediction capability score is not lower than 85, the current model parameters are kept unchanged. Samples with a toughness correction sample whose toughness standardization error is higher than that of compressive strength standardization error and whose difference is not less than 0.05 are extracted from the test samples. The BP neural network multivariate performance prediction model was retrained in a targeted manner using toughness-corrected samples. During the retraining process, the connection weights of the compressive strength output node remained unchanged, while the learning ratio of the bending toughness, ultimate tensile strain, and fracture energy output nodes to the toughness-corrected samples was increased, resulting in the corrected BP neural network multivariate performance prediction model.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention does not evaluate the quality of a model solely based on the overall error output of the BP neural network. Instead, under the same sample and strength-toughness distribution conditions, it constructs strength-biased learning parameters to identify whether the model exhibits a learning bias where the compressive strength prediction error is controlled while the toughness-related index prediction error abruptly increases. This invention directly addresses the problem of existing multi-output BP models prioritizing compressive strength and neglecting the learning of bending toughness, ultimate tensile strain, and fracture energy. It separates model defects from hidden comprehensive errors, improving the relevance and reliability of the toughness-related index prediction evaluation.

[0017] This invention also introduces ductile instability discrete parameters and correlates failure process characteristics such as crack distribution discreteness, concentrated deformation development, and post-peak load attenuation with model errors. A toughness prediction score is then formed through non-compensated fusion. This invention differs from existing methods that rely solely on mix proportion data and final strength results. It can identify the toughness prediction distortion caused by the transition from stable multi-crack propagation to concentrated failure of the main crack, and accordingly correct the BP neural network. This reduces the risk of misjudging performance as adequate strength but insufficient ductility and energy dissipation capacity, improving the reliability of multi-performance prediction results for high-toughness concrete in engineering applications. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of a method for predicting multiple properties of high-toughness concrete according to the present invention.

[0020] Figure 2 This is a flowchart of the method for obtaining the toughness prediction capability score of the present invention.

[0021] Figure 3 This is a flowchart of the method for modifying the multinomial performance prediction model of the BP neural network according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0023] Example 1, please refer to Figure 1 As shown in this embodiment, a method for predicting multiple properties of high-toughness concrete includes: Obtain high-toughness concrete sample data, including mix proportion parameters, fiber parameters, curing conditions, and corresponding measured values ​​of compressive strength, flexural toughness, ultimate tensile strain, and fracture energy.

[0024] In one embodiment, obtaining high-toughness concrete sample data specifically includes: for each group of high-toughness concrete test samples, recording the mix proportion parameters, fiber parameters, curing conditions, and corresponding measured performance values ​​of the sample, and storing the above data in association according to the same data format, so that each set of mix proportion parameters, fiber parameters, and curing conditions corresponds to a set of measured values ​​of compressive strength, flexural toughness, ultimate tensile strain, and fracture energy.

[0025] The mix proportion parameters include one or more of the following: cement dosage, fly ash dosage, slag powder dosage, silica fume dosage, fine aggregate dosage, coarse aggregate dosage, water dosage, water-reducing agent dosage, water-cement ratio, sand-cement ratio, and total amount of cementitious materials; preferably, the dosage of each material is uniformly expressed as the mass dosage per cubic meter of concrete, in units of... For high-toughness cement-based composites without coarse aggregate, the amount of coarse aggregate can be recorded as zero or a blank value and uniformly processed in subsequent pretreatment steps. The fiber parameters include fiber type, fiber volume fraction, fiber mass fraction, fiber length, fiber diameter, fiber aspect ratio, fiber elastic modulus, fiber tensile strength, and whether one or more of hybrid fibers are used. The fiber type can be steel fiber, polyvinyl alcohol fiber, polyethylene fiber, polypropylene fiber, basalt fiber, or a combination of the above fibers. When hybrid fibers are used, the type, length, diameter, and fraction of different fibers are recorded separately to allow subsequent model identification of the influence of different fiber systems on toughness properties. The curing conditions include curing age, curing temperature, curing humidity, curing method, and whether steam curing, standard curing, or natural curing is used. The curing age can be recorded as 3 days, 7 days, 28 days, 56 days, or other actual test ages; the curing temperature is recorded in degrees Celsius, and the curing humidity is recorded in relative humidity. If the sample undergoes multiple curing stages, the duration, temperature, and humidity of each stage are recorded separately. The measured compressive strength is obtained by conducting compressive tests on high-toughness concrete cube or prism specimens; the measured flexural toughness is obtained by bending loading tests and can be determined based on load-deflection curves, load-displacement curves, or the area of ​​a curve within a specified deflection range; the measured ultimate tensile strain is obtained by uniaxial tensile tests or equivalent tensile property tests; and the measured fracture energy is obtained by fracture tests or bending fracture tests.

[0026] Each performance measured value is preferably determined by the test results of no less than three parallel specimens, and the average value is taken as the performance measured value of the sample group. At the same time, the dispersion of each parallel specimen is recorded so as to analyze the fluctuation characteristics of toughness indicators in subsequent analysis.

[0027] During data acquisition, a unique correspondence is established between the input parameters and output performance values ​​of the same group of high-toughness concrete. For example, the i-th sample group is recorded as: mix proportion parameter set, fiber parameter set, curing condition set, measured compressive strength value, measured flexural toughness value, measured ultimate tensile strain value, and measured fracture energy value. Multiple groups of samples constitute the original sample database. This original sample database can originate from indoor test data, engineering testing data, or publicly available literature data. When data comes from different sources, the data source, specimen size, test method, and loading method are recorded uniformly to facilitate subsequent screening, correction, or standardization of data from different sources.

[0028] The sample data is preprocessed to obtain standardized sample data, which is then divided into training samples, validation samples, and test samples.

[0029] In this embodiment, sample data preprocessing includes field alignment, unit unification, consistency verification, interval standardization, and hierarchical partitioning.

[0030] First, each group of high-toughness concrete samples was assigned a unique sample number. Data records were then established to correspond one-to-one with the sample number for mix proportion parameters, fiber parameters, curing conditions, measured compressive strength, measured flexural toughness, measured ultimate tensile strain, and measured fracture energy. For cases where the same parameter had different names in different data sources, the names of fields such as water-cement ratio, cementitious material dosage, fiber volume fraction, and curing age were standardized. For material dosage, a unified conversion was performed. Fiber length and fiber diameter are uniformly converted to mm; ultimate tensile strain is uniformly converted to percentage; fracture energy is uniformly converted to N / m.

[0031] After the conversion is completed, data records lacking any of the measured values ​​of compressive strength, flexural toughness, ultimate tensile strain, or fracture energy will not be included in the subsequent division of training samples, validation samples, and test samples.

[0032] After completing field alignment and unit unification, consistency verification is performed on the strength response and toughness response of the same sample.

[0033] For the i-th group of high-toughness concrete samples, the measured values ​​of compressive strength fc,i, flexural toughness Tb,i, ultimate tensile strain εu,i, and fracture energy Gf,i were read respectively.

[0034] First, calculate the median of each measured value in all samples, where: the median of the measured compressive strength is denoted as Mc, the median of the measured flexural toughness is denoted as MT, the median of the measured ultimate tensile strain is denoted as Mε, and the median of the measured fracture energy is denoted as MG.

[0035] Next, calculate the median of the absolute values ​​of the differences between the measured values ​​of each performance indicator and the corresponding median, and use this as the robust dispersion measure for each performance indicator, where: the robust dispersion measure for compressive strength is... Bending toughness robust discrete quantity Ultimate tensile strain robust discrete quantity Fracture energy robust discrete quantity .

[0036] Calculate the robustness deviation values ​​for each item in the i-th sample group, where: robustness deviation value for compressive strength Bending toughness robustness deviation value Ultimate tensile strain robustness deviation value Fracture energy robustness deviation value The value 0.000001 is used to avoid a denominator of 0.

[0037] Finally, the arithmetic mean of the bending toughness robustness deviation, the ultimate tensile strain robustness deviation, and the fracture energy robustness deviation is taken to obtain the toughness response deviation value Zt,i = (ZT,i + Zε,i + ZG,i) / 3 for the i-th sample, which is used to characterize the comprehensive deviation of the sample from the three measured toughness values.

[0038] If the absolute value of the robustness deviation of compressive strength in the same data record is less than 1.5 and the absolute value of the toughness response deviation is greater than 3.0; or if the absolute value of the robustness deviation of compressive strength is greater than 3.0 and the absolute value of the toughness response deviation is less than 1.5, then the strength response and toughness response of the data record are determined to be out of correspondence and the record is removed from the sample data.

[0039] The above 1.5 and 3.0 are set according to the central and discrete intervals of the robust deviation distribution, which are used to retain the majority of normal corresponding samples and exclude data records with mismatched strength and toughness labels.

[0040] The data records after consistency verification are standardized by interval. The mixing ratio parameters, fiber parameters, curing conditions and various measured values ​​are treated as independent parameter categories. Within each parameter category, the minimum and maximum values ​​of each field are calculated first. Then, the value of each field is subtracted from the corresponding minimum value, and the result is divided by the sum of the difference between the corresponding maximum and minimum values ​​and 0.000001 to obtain the standardized field value in the range of 0 to 1.

[0041] For fields such as fiber volume content, ultimate tensile strain, and fracture energy, which are concentrated and have a small number of end samples, the field values ​​are first divided into four intervals according to the 25%, 50%, and 75% quantiles. Within each interval, the same minimum and maximum value conversion method is used for standardization. Then, the standardized results of the four intervals are merged according to the original sample number to obtain standardized sample data.

[0042] This treatment ensures that the mix proportion parameters, fiber parameters, curing conditions, and various measured values ​​are standardized within their respective parameter categories, and avoids the numerical range of compressive strength suppressing the numerical expression of toughness measured values.

[0043] After obtaining standardized sample data, the data was stratified according to the joint distribution of measured compressive strength and toughness values. First, the arithmetic mean of the standardized measured values ​​of flexural toughness, ultimate tensile strain, and fracture energy was taken to obtain the comprehensive toughness value. Then, the samples were divided into three categories: low strength, medium strength, and high strength, and three categories: low toughness, medium toughness, and high toughness, based on the 33.33% and 66.67% quantiles of the measured compressive strength and comprehensive toughness values, respectively. Finally, the strength and toughness categories were combined in pairs to form nine joint distribution groups.

[0044] For each joint distribution group, training samples, validation samples, and test samples are drawn in a ratio of 7:1:2. When the number of samples in a certain joint distribution group is less than 10, at least 1 group is drawn as a validation sample, 1 group is drawn as a test sample, and the rest are used as training samples.

[0045] After the above division, the training samples, validation samples, and test samples all contain data records with different strength levels and different toughness discrete levels. This allows the subsequent BP neural network multivariate performance prediction model to access sample distributions with inconsistent strength and toughness responses during the training, validation, and testing phases, thereby improving the reliability of the evaluation of the ability to predict bending toughness, ultimate tensile strain, and fracture energy.

[0046] Using mix proportion parameters, fiber parameters, and curing conditions as inputs, and compressive strength, flexural toughness, ultimate tensile strain, and fracture energy as outputs, a BP neural network multivariate performance prediction model was established and trained.

[0047] In this embodiment, when establishing and training the BP neural network multivariate performance prediction model, the preprocessed field order in the training samples is first read. The mix proportion parameters, fiber parameters, and curing conditions of the i-th training sample are arranged into an input feature sequence Xi, where Xi = [xi1, xi2, ..., xin], and n is the number of input fields. The compressive strength is recorded as the strength output label fci, and the flexural toughness, ultimate tensile strain, and fracture energy are recorded as the toughness output labels Tbi, εui, and Gfi, respectively. During training, each batch of training samples is organized according to the aforementioned nine joint distribution groups, preferably with 36 groups per batch. Four groups are extracted from each joint distribution group. When a joint distribution group has fewer than four groups, samples are extracted cyclically according to the sample number to make up the difference, so that each batch of training samples contains data records of different strength levels and different toughness discrete levels, avoiding continuous participation of samples from a single strength interval in weight updates.

[0048] The BP neural network multivariate performance prediction model includes an input layer, a shared hidden layer, a compressive strength output node, and three toughness output nodes.

[0049] When establishing a BP neural network multivariate performance prediction model, the number of input layer nodes is the same as the number of input fields, denoted as n; the number of output labels is 4, representing compressive strength, flexural toughness, ultimate tensile strain, and fracture energy. The expression for calculating the number of shared hidden layer nodes h is as follows: , where ceil means rounding up.

[0050] For the i-th sample group, its input feature sequence is represented as Xi = [xi1, xi2, ..., xin], where xir represents the r-th input field value of the i-th sample group. The input value of the j-th shared hidden layer node is obtained by multiplying each input field value by its corresponding input weight, summing the results, and then adding the hidden layer bias, i.e., netij = Σ(xir × Wrj) + bj; where Wrj represents the connection weight from the r-th input field of the input layer to the j-th shared hidden layer node, and bj represents the hidden layer bias of the j-th shared hidden layer node.

[0051] The activation value of the j-th shared hidden layer node is calculated using the Sigmoid function, i.e., Hij = 1 / [1 + exp(-netij)]; substituting netij, it can be expressed as Hij = 1 / [1 + exp(-(Σ(xir×Wrj) + bj))]. Here, Hij represents the activation value of the i-th sample at the j-th shared hidden layer node, and exp represents the exponential operation with the natural constant as the base.

[0052] The output layer comprises four output nodes, corresponding to compressive strength, flexural toughness, ultimate tensile strain, and fracture energy, respectively. The predicted value of the k-th output node is obtained by multiplying the activation values ​​of each shared hidden layer node by their corresponding output weights, summing the results, and then adding the output layer bias, i.e., Yik = Σ(Hij × Vjk) + ck; where Vjk represents the connection weight from the j-th shared hidden layer node to the k-th output node, ck represents the output layer bias of the k-th output node, and Yik represents the predicted value of the i-th sample for the k-th performance index. When k takes the values ​​1, 2, 3, and 4, Yik represents the predicted values ​​of compressive strength, flexural toughness, ultimate tensile strain, and fracture energy, respectively.

[0053] During the initial joint training, the prediction errors of compressive strength, flexural toughness, ultimate tensile strain, and fracture energy are used together as the basis for weight updates. After the r-th round of training, the output error of compressive strength Ec(r) is calculated on the validation samples. Ec(r) is the average of the absolute values ​​of the differences between the predicted and measured compressive strength values ​​of each validation sample. The output error of toughness Et(r) is calculated. Et(r) is the arithmetic mean of the average absolute errors of flexural toughness, ultimate tensile strain, and fracture energy.

[0054] The initial joint training loss value is L(r) = Ec(r) + Et(r). Backpropagation is used to update Wrj, bj, Vjk and ck. The learning rate is preferably 0.01. After each training round, Ec(r) and Et(r) are recorded in the validation samples.

[0055] When the verification results of three consecutive rounds satisfy that Ec(r) is less than Ec(r-1), and at least two of the three rounds satisfy that Et(r) is not less than Et(r-1), it is determined that the output error of compressive strength continues to decrease while the output error of toughness does not decrease synchronously.

[0056] At this point, the connection weights and biases corresponding to the compressive strength output node are fixed, i.e., Vjc and cc are fixed from the shared hidden layer to the compressive strength output node. Only the connection weights and biases from the shared hidden layer to the flexural toughness output node, the ultimate tensile strain output node, and the fracture energy output node are retained for reverse updating. The loss value in this stage is only Et(r), and the connection weights Wrj from the input layer to the shared hidden layer remain updatable, allowing the shared hidden layer to continue learning feature combinations that are beneficial to the expression of the toughness output label. After each round of toughness-oriented update, Et(r) is recalculated. When Et(r) is less than Et(r-1) of the previous round for two consecutive rounds, the joint update of all weights of the compressive strength output node, the toughness output node, and the shared hidden layer is restored.

[0057] Repeat the joint training and toughness-oriented update process described above until Ec(r) and Et(r) on the validation samples do not decrease within 5 consecutive rounds, or the number of training rounds reaches 500, to obtain the BP neural network multinomial performance prediction model. This training method, without changing the input-output relationship of the main claim, prevents the compressive strength output from dominating the shared hidden layer update in the long term, thereby improving the learning sufficiency of bending toughness, ultimate tensile strain, and fracture energy.

[0058] The test samples are input into the trained BP neural network multivariate performance prediction model to obtain the predicted values ​​of each performance, and compared with the corresponding measured values ​​to obtain the prediction error of compressive strength and the prediction error of toughness index.

[0059] In this embodiment, the test samples are used to evaluate the prediction performance of the trained BP neural network multivariate performance prediction model for unknown samples. First, the test samples are read according to the field order formed in the preprocessing stage. The mix proportion parameters, fiber parameters, and curing conditions of the i-th test sample are arranged into an input feature sequence Xi, which is then input into the trained BP neural network multivariate performance prediction model to obtain four standardized prediction values ​​corresponding to the sample numbers. These are denoted as the standardized predicted value for compressive strength, the standardized predicted value for flexural toughness, the standardized predicted value for ultimate tensile strain, and the standardized predicted value for fracture energy. Each set of predicted values ​​is bound to the sample number of the test sample for storage, avoiding mismatches between predicted and measured values ​​for different samples.

[0060] Since the output of the BP neural network multinomial performance prediction model is on a standardized scale, it needs to be reverse-engineered according to the standardization rules used during preprocessing. For fields standardized using minimum and maximum values, the restoration process is as follows: After obtaining the standardized predicted values ​​from the BP neural network multinomial performance prediction model, it is necessary to restore the standardized predicted values ​​to a data scale consistent with the measured values. For fields that are standardized using the overall minimum and maximum values, let the standardized predicted value of the k-th field be Yk′. The minimum value of this field recorded in the preprocessing stage is Yk,min, and the maximum value is Yk,max. Then, the restored value Yk of this field is calculated as follows: Yk=Yk′×(Yk,max-Yk,min)+Yk,min.

[0061] For fields that have undergone interval standardization, first read the interval number corresponding to the test sample in the preprocessing stage, denoted as q; then, based on this interval number q, read the minimum value Yk,qmin and the maximum value Yk,qmax of the k-th field in the q-th interval; subsequently, use the same reverse restoration method as for the overall standardized fields to calculate the predicted value at the actual scale, i.e., Yk = Yk′ × (Yk,qmax - Yk,qmin) + Yk,qmin. Here, Yk represents the restored actual-scale predicted value, and Yk′ represents the standardized predicted value output by the model.

[0062] Through the above restoration process, the predicted values ​​of compressive strength, flexural toughness, ultimate tensile strain, and fracture energy can be restored to the same data units and numerical ranges as the corresponding measured values.

[0063] After the restoration is completed, the predicted compressive strength value and the measured compressive strength value are in the same unit, and the predicted bending toughness value, the predicted ultimate tensile strain value and the predicted fracture energy value are also in the same unit as the corresponding measured values.

[0064] Subsequently, pairing was performed item by item using the sample number as an index. The predicted and measured compressive strength values ​​of the i-th test sample formed the strength prediction control data; the predicted and measured flexural toughness values, predicted and measured ultimate tensile strain values, and predicted and measured fracture energy values ​​of the i-th test sample formed the toughness prediction control data. During pairing, only data with the same sample number were allowed to be included in the same control data; test samples lacking any measured value or any predicted value were not included in this round of error statistics.

[0065] For the strength prediction comparison data, calculate the compressive strength difference. The compressive strength difference is equal to the predicted compressive strength value minus the measured compressive strength value. When the compressive strength difference is positive, the direction of the compressive strength deviation is recorded as overestimation. When the compressive strength difference is negative, the direction of the compressive strength deviation is recorded as underestimation. When the compressive strength difference is 0, the direction of the compressive strength deviation is recorded as no deviation. The magnitude of the compressive strength deviation is taken as the absolute value of the compressive strength difference.

[0066] For the toughness prediction comparison data, the bending toughness difference, ultimate tensile strain difference, and fracture energy difference are calculated respectively. Each difference is obtained by subtracting the corresponding measured value from the corresponding predicted value. When the difference is positive, the corresponding deviation direction is recorded as overestimation; when the difference is negative, the corresponding deviation direction is recorded as underestimation; and when the difference is 0, the corresponding deviation direction is recorded as no deviation. The magnitude of the corresponding deviation is taken as the absolute value of the corresponding difference. This forms the strength error record and toughness error record for the i-th test sample.

[0067] In the grouping and summarization stage, the strength and toughness dispersion levels established in the preprocessing stage are used to assign test samples to corresponding joint distribution groups. Within each joint distribution group, the compressive strength prediction error is the arithmetic mean of all compressive strength deviations in that group; the flexural toughness prediction error is the arithmetic mean of all flexural toughness deviations in that group; the ultimate tensile strain prediction error is the arithmetic mean of all ultimate tensile strain deviations in that group; and the fracture energy prediction error is the arithmetic mean of all fracture energy deviations in that group. The prediction errors for toughness-related indicators are obtained by calculating the arithmetic mean of the flexural toughness prediction error, the ultimate tensile strain prediction error, and the fracture energy prediction error.

[0068] The summary results retain the sample number, strength level, toughness dispersion level, compressive strength deviation direction, flexural toughness deviation direction, ultimate tensile strain deviation direction, and fracture energy deviation direction. This ensures that the compressive strength prediction error and the toughness index prediction error maintain a corresponding relationship under the same sample and the same joint distribution group, providing a data foundation for the subsequent generation of strength bias learning parameters.

[0069] In this embodiment, strength bias learning parameters are generated based on the prediction errors of compressive strength and toughness indicators.

[0070] In this embodiment, the strength bias learning parameter is used to characterize the extent to which the BP neural network multivariate performance prediction model has sufficiently learned about compressive strength in the test samples, but insufficiently learned about toughness-related indicators. Based on the strength error record and toughness error record obtained in the previous step, the strength level and toughness discrete level of each test sample are read according to the sample number, and a joint distribution group is formed by the combination of the strength level and the toughness discrete level, denoted as the g-th group. The number of test samples in the g-th group is denoted as Ng.

[0071] Within group g, the compressive strength prediction errors are sorted from low to high. The ranking number of the compressive strength prediction error of the i-th test sample within group g is denoted as Rci, with later ranking numbers indicating higher prediction errors within the same group. To eliminate the influence of different sample sizes in different joint distribution groups, the strength error ranking is denoted as Pci, where Pci equals Rci divided by Ng. Similarly, the toughness index prediction errors are sorted from low to high. The ranking number of the toughness index prediction error of the i-th test sample is denoted as Rti, and the toughness error ranking is denoted as Pti, where Pti equals Rti / Ng. Therefore, the same sample simultaneously possesses both the strength error ranking Pci and the toughness error ranking Pti, and both fall within the range of 0 to 1.

[0072] After completing the intra-group ranking and calibration, the differences in the strength error ranking and toughness error ranking of the same sample are identified. The ranking difference of the i-th test sample is denoted as Dri, where Dri = Pti - Pci. When Dri is not less than 0.20, it indicates that the sample's intra-group ranking of the toughness index prediction error is at least 20% higher than that of the compressive strength prediction error ranking, and the sample is extracted as a strength-biased candidate sample; when Dri is less than 0.20, the sample is not included in the strength-biased candidate sample set.

[0073] It should be noted that 0.20 is the preferred position difference control value, and its setting rule is as follows: when there are at least 5 test samples in the g group, 20% of the positions in the group can distinguish the error difference caused by random sorting fluctuations and the error difference caused by model learning bias; when Ng is less than 5, the g group is not used alone for strength bias candidate sample extraction, but is merged into the adjacent toughness discrete level group with the same strength level before sorting and calibration.

[0074] After obtaining the candidate samples with strength bias, further screening for directional consistency is performed. First, the median Mgc of the compressive strength deviation amplitude within group g is calculated; when the compressive strength deviation amplitude of the candidate samples with strength bias does not exceed Mgc, the compressive strength deviation amplitude is determined to be controlled. Then, the flexural toughness deviation direction, ultimate tensile strain deviation direction, and fracture energy deviation direction in the candidate sample are read, with overestimation recorded as 1, underestimation as -1, and no deviation as 0; when at least two of the three toughness indicators have the same value and are all non-zero, the toughness indicators are determined to show the same directional deviation.

[0075] Only candidate samples that simultaneously meet the criteria of controlled compressive strength deviation and at least two toughness indicators showing a unidirectional deviation are retained as strength bias confirmation samples. This screening process can eliminate error differences caused by accidental fluctuations in a single toughness indicator, ensuring that the retained samples reflect the consistency deviation of the BP neural network multi-performance prediction model in toughness output.

[0076] Subsequently, strength bias learning parameters are generated based on the strength bias confirmation samples. First, the distribution proportion A is calculated, which equals the number of strength bias confirmation samples divided by the total number of test samples. Then, the degree of rank excess B is calculated, which equals the average Dri value of all strength bias confirmation samples. Next, the degree of toughness homogeneity deviation C is calculated: for each strength bias confirmation sample, if all three toughness indicators show homogeneity deviation, the homogeneity deviation value of that sample is 1; if only two toughness indicators show homogeneity deviation, the homogeneity deviation value of that sample is 2 divided by 3; C equals the average homogeneity deviation value of all strength bias confirmation samples. Finally, the strength bias learning parameter Ks is obtained by weighted summation of A, B, and C, where the weight of A is 0.40, the weight of B is 0.35, and the weight of C is 0.25, that is, Ks equals the sum of 0.40 times A, 0.35 times B, and 0.25 times C.

[0077] When there are no samples confirming the strength bias, A, B, C, and Ks are all set to 0. Ks is in the range of 0 to 1. The closer the value is to 1, the more concentrated the BP neural network multivariate performance prediction model is in the test samples, exhibiting controlled compressive strength prediction error, increased ranking of toughness index prediction errors, and consistent toughness deviation direction. This parameter transforms a simple comparison of error magnitude into an evaluation of learning bias under the same sample, the same joint distribution group, and the same deviation direction, providing quantifiable input for subsequent comprehensive analysis with toughness instability discrete parameters.

[0078] Obtain the toughness failure process characteristics corresponding to the sample, and construct the toughness instability discrete parameters based on the toughness failure process characteristics.

[0079] In this embodiment, the bending toughness, ultimate tensile strain, and fracture energy test data corresponding to the test samples are read according to the sample number. For each test process, the loading time, load value, deformation value, number of cracks, width of each crack, and spacing between adjacent cracks are recorded. For the i-th sample, the load value at the q-th acquisition time is denoted as Piq, the deformation value as diq, the number of cracks as Niq, the set of crack widths as Wiq, and the set of crack spacings as Siq.

[0080] The above data is bound to the toughness error record of the i-th sample according to the sample number to form a toughness failure process record, so that the failure characteristics calculated subsequently can maintain the same sample correspondence with the prediction errors of bending toughness, ultimate tensile strain and fracture energy.

[0081] The ductile failure process record was divided into stages. The initial cracking stage was determined by the acquisition time when the number of cracks first changed from 0 to 1; if the crack number record was missing, it was determined by the acquisition time when the slope of the load-deformation curve first decreased to 15% of the initial slope, and the initial slope was the average of the ratio of load increment to deformation increment of the first 5 acquisition times.

[0082] The multi-crack propagation stage is the data acquisition interval from the end of the initial cracking stage to before the load reaches its peak value. The post-peak attenuation stage is the data acquisition interval from after the load peak value to the end of the test. Through the above division, the continuous process of crack formation, crack propagation, and load attenuation can be obtained simultaneously in the same time sequence.

[0083] Stage-specific failure characteristics are extracted within each stage. The crack number variation characteristic is calculated by dividing the difference between the number of cracks at the end of the multi-crack propagation stage and the number of cracks at the end of the initial cracking stage by the deformation increment of the multi-crack propagation stage. The crack width variation characteristic is calculated by dividing the difference between the maximum crack width at peak load and the maximum crack width at initial cracking by the difference between the deformation corresponding to peak load and the deformation at initial cracking. The crack spacing variation characteristic is represented by the coefficient of variation of the spacing between adjacent cracks; the calculation process is as follows: When extracting the crack spacing variation characteristics, first read the set of adjacent crack spacings Siq of the i-th group of samples at the q-th acquisition time, where the m-th adjacent crack spacing is denoted as siqm, and the number of adjacent crack spacings is denoted as M. First calculate the average value of all adjacent crack spacings at the same acquisition time, denoted as Siq,avg, calculated as Siq,avg = (siq1 + siq2 + ... + siqM) / M.

[0084] Subsequently, the squared difference between the spacing of each adjacent crack and this average value is calculated, and the average of all squared differences is taken. Then, the square root of this average value is taken to obtain the standard deviation of the spacing between adjacent cracks, denoted as Siq,std. The calculation method is as follows: .

[0085] Next, divide the standard deviation of the spacing between adjacent cracks by the sum of the average spacing between adjacent cracks and 0.000001 to obtain the crack spacing variation characteristic Ciq, which is calculated as Ciq = Siq,std / (Siq,avg + 0.000001). Here, 0.000001 is used to avoid division by zero when the average spacing between adjacent cracks is 0.

[0086] When extracting the post-peak bearing capacity attenuation characteristics, the peak load Pi,peak, test end load Pi,end, peak load corresponding deformation di,peak, and test end deformation di,end of the i-th sample are read. First, the post-peak deformation increment Δdi is calculated as Δdi = di,end - di,peak; then, the relative attenuation of the post-peak load is calculated as (Pi,peak - Pi,end) / (Pi,peak + 0.000001); finally, the relative attenuation of the post-peak load is divided by the post-peak deformation increment to obtain the post-peak bearing capacity attenuation characteristic Fi, calculated as Fi = [(Pi,peak - Pi,end) / (Pi,peak + 0.000001)] / (Δdi + 0.000001). Here, Fi characterizes the rate at which the bearing capacity of the sample decreases with increasing deformation after the peak load. A higher value indicates a more concentrated post-peak bearing capacity attenuation and a weaker ability to maintain toughness.

[0087] The process of crack propagation transitioning from dispersed expansion to concentrated main crack propagation is identified based on staged failure characteristics. The discrete crack distribution characteristic Ci is the average of the crack spacing variation characteristics at peak load and the proportion of the maximum crack width to the total crack width at peak load. The concentrated deformation development characteristic Di is the proportion of the maximum crack width increase to the total crack width increase during the post-peak attenuation stage. The post-peak bearing capacity attenuation characteristic Fi is the characteristic of bearing capacity attenuation after the peak. Ci, Di, and Fi are all limited to a range of 0 to 1; a value less than 0 is taken as 0, and a value greater than 1 is taken as 1. The ductile instability characteristic set of the i-th sample group is composed of Ci, Di, and Fi.

[0088] The toughness instability discrete parameters are constructed based on the degree of synchronization of Ci, Di, and Fi in the same sample. First, the synchronization value Qi is calculated, which is equal to the minimum value among Ci, Di, and Fi. Then, the equilibrium value Bi is calculated, which is equal to 1 minus the standard deviations of Ci, Di, and Fi; when Bi is less than 0, it is set to 0. Finally, the toughness instability discrete parameter Kdi for the i-th sample group is calculated, which is equal to the sum of Bi multiplied by 0.40 times Ci, 0.30 times Di, 0.20 times Fi, and 0.10 times Qi. If the overall toughness instability discrete parameters of the test sample are required, the arithmetic mean of all Kdi values ​​is taken.

[0089] This calculation method shows that Kdi will only achieve a higher value when the discrete distribution of cracks, concentrated development of deformation, and post-peak load attenuation all occur in the same sample. This reflects the degree of dispersion of high-toughness concrete as it transforms from a stable multi-crack response to a concentrated failure response, and provides a quantitative basis for subsequent comprehensive analysis with strength bias learning parameters.

[0090] Please see Figure 2 As shown, the strength bias learning parameters and the toughness instability discrete parameters are comprehensively analyzed to obtain the toughness prediction capability score.

[0091] In this embodiment, the sample number is used for data correspondence in the comprehensive analysis stage. For the i-th test sample, the single-sample strength bias learning parameter Ksi and ductile instability discrete parameter Kdi retained in the previous step are read. Both are expressed in the range of 0 to 1 and are stored together with the strength level and ductile discrete level to which the sample belongs, forming a model learning bias-material ductile instability correlation record.

[0092] If the aforementioned strength bias learning parameters are represented by an overall value Ks, then the single-sample contribution value of each strength bias confirmation sample is retained as Ksi before forming the overall value Ks; for test samples that are not confirmed as strength bias confirmation samples, Ksi is set to 0. Kdi directly adopts the ductile instability discrete parameters of the corresponding sample.

[0093] After completing the matching, high-level identification is performed on Ksi and Kdi. First, the 75th percentile of Ksi and the 75th percentile of Kdi for all test samples are calculated, and each is compared with 0.50. The higher of these two values ​​is taken as the high-level threshold Ls for strength bias and the high-level threshold Ld for ductile instability. This setting rule ensures that high-level samples reach at least 0.50, while also being adjustable according to changes in the distribution of test samples. If the i-th group of samples simultaneously satisfies Ksi not less than Ls and Kdi not less than Ld, the sample is recorded as a ductile prediction risk sample; otherwise, it is not recorded as a ductile prediction risk sample. This yields a set of ductile prediction risk samples, and the strength level and ductile dispersion level of each risk sample are retained.

[0094] The risk samples for toughness prediction are statistically distributed according to different strength levels and toughness dispersion levels. The number of test samples in the g-th joint distribution group is denoted as Ng, and the number of risk samples is denoted as Mg. The risk proportion of this group, Ag = Mg / Ng, is calculated. When Ng is 0, this group is not included in the calculation. The total number of joint distribution groups involved in the calculation is denoted as G, and the overall risk proportion A is equal to the arithmetic mean of each Ag. This processing ensures that even when risk samples are concentrated at a certain strength level or toughness dispersion level, they are still retained in the grouping results, avoiding being masked by the overall average error of the test samples.

[0095] The non-compensated fusion method employs a two-level calculation: single-sample fusion value and group fusion value. The single-sample fusion value Zi of the i-th group is determined as follows: first, the higher of Ksi and Kdi is taken as the dominant value Hi; then, the product of Ksi and Kdi is taken as the coupling value Ji. Zi equals the sum of 0.60 times Hi and 0.40 times Ji. Since Hi is taken as the higher of the two, a high value for either parameter will directly increase Zi, while a low value for the other parameter cannot offset it. Simultaneously, Ji is used to express the superimposed effect when model learning bias and material ductile instability occur simultaneously. The group fusion value Zg of the g-th joint distribution group is equal to the arithmetic mean of all Zi in that group; the overall fusion value Z is equal to the arithmetic mean of all Zg. The above calculation process ensures that the strength bias learning parameter and the ductile instability discrete parameter have independent influences, while also reflecting the weakening effect on the reliability of ductile prediction when they occur simultaneously.

[0096] A resilience prediction capability score S is generated based on the non-compensated fusion results. The score is based on a 100-point scale and is calculated as follows: S = 100 - 100Z - 20A; when the calculated result is less than 0, S is 0; when the calculated result is greater than 100, S is 100.

[0097] A represents the average proportion of risk samples in each joint distribution group for toughness prediction. Therefore, as the proportion of risk samples increases, S decreases in a fixed direction. S is used to characterize the reliability of the BP neural network multivariate performance prediction model for bending toughness, ultimate tensile strain, and fracture energy. The closer S is to 100, the higher the reliability of the toughness index prediction; the closer S is to 0, the less suitable the model is for judging toughness indexes under the current sample distribution.

[0098] Please refer to Figure 3. The BP neural network multi-performance prediction model is modified according to the toughness prediction capability score, and the modified model is used to output the multi-performance prediction results and toughness prediction reliability level of the high-toughness concrete to be predicted.

[0099] In this embodiment, the correction method of the BP neural network multivariate performance prediction model is determined based on the toughness prediction capability score S. S is expressed as a score from 0 to 100. When S is not lower than 85, the current model parameters remain unchanged; when S is not lower than 70 but lower than 85, the output nodes of bending toughness, ultimate tensile strain, and fracture energy are lightly retrained; when S is lower than 70, the reliability of the toughness index prediction is deemed insufficient, and the BP neural network multivariate performance prediction model is retrained in a targeted manner. 70 and 85 are used as the scoring cutoff values, and the rules for setting them are as follows: when S is lower than 70, the non-compensated fusion results and the proportion of risky samples in toughness prediction have jointly caused a score loss of more than 30 points, indicating that the prediction error of toughness indexes cannot be ignored; when S is not lower than 85, the score loss does not exceed 15 points, indicating that the model maintains stability in predicting toughness indexes.

[0100] When extracting toughness correction samples, the prediction errors of compressive strength and toughness-related indices corresponding to the test samples are read. To ensure comparability of different performance dimensions, the i-th group of test samples is first subjected to error standardization. The standardization error of compressive strength Eci is equal to the compressive strength deviation amplitude divided by the sum of the measured compressive strength value and 0.000001. The standardization errors of bending toughness, ultimate tensile strain, and fracture energy are obtained in the same way. The arithmetic mean of the three standardization errors is then taken as the standardization error of toughness Eti.

[0101] When Eti is greater than Eci and the result of Eti minus Eci is not less than 0.05, the i-th group of test samples is used as the toughness correction sample. 0.05 is used as the error difference boundary value to exclude small differences caused by numerical rounding, and is preferably taken from the 5% relative error allowable range in engineering tests.

[0102] When performing targeted retraining based on toughness-corrected samples, both the original training samples and the toughness-corrected samples are input into the model. In each round of retraining, the toughness-corrected samples participate in training twice, while the original training samples participate once, ensuring that the output nodes for bending toughness, ultimate tensile strain, and fracture energy receive a higher learning ratio from the toughness-corrected samples. The retraining loss value L is calculated as follows: L equals the sum of the compressive strength stability term, bending toughness error term, ultimate tensile strain error term, and fracture energy error term; where the compressive strength stability term equals the average of the squared differences between the predicted compressive strength values ​​before and after retraining, and the bending toughness error term, ultimate tensile strain error term, and fracture energy error term equals the average of the squared differences between the corresponding predicted and measured values, respectively. The weights of the three toughness error terms are all set to 1.5, and the weight of the compressive strength stability term is set to 1.0.

[0103] During retraining, the connection weights of the compressive strength output nodes remain unchanged. Only the connection weights from the input layer to the shared hidden layer, the bias of the shared hidden layer, and the connection weights and biases from the shared hidden layer to the three resilience output nodes are updated. When the resilience standardization error does not decrease in five consecutive training rounds, or when the number of retraining rounds reaches 200, retraining is stopped, and the corrected BP neural network multivariate performance prediction model is obtained.

[0104] When predicting high-toughness concrete, its mix proportion parameters, fiber parameters, and curing conditions are processed according to the same field order and standardization rules as the training samples to form a sequence of input features to be predicted. This sequence is then input into the modified BP neural network multivariate performance prediction model to obtain standardized output values. The standardized output values ​​are then restored to the predicted compressive strength, flexural toughness, ultimate tensile strain, and fracture energy according to the reverse restoration rules recorded in the preprocessing stage.

[0105] The reliability level of toughness prediction is determined jointly by the toughness prediction capability score S and the output consistency C. The calculation process for output consistency C is as follows: First, the bending toughness prediction result, ultimate tensile strain prediction result, and fracture energy prediction result are converted to a range of 0 to 1 according to the minimum and maximum values ​​of the corresponding fields in the training samples, respectively; then, the average of the three converted values ​​is calculated; subsequently, the standard deviation of the three converted values ​​relative to this average is calculated; C equals 1 minus this standard deviation, and if C is less than 0, it is taken as 0. Finally, when S is not less than 85 and C is not less than 0.90, the toughness prediction reliability level is recorded as Level 1; when S is not less than 70 and C is not less than 0.80, it is recorded as Level 2; otherwise, it is recorded as Level 3. Finally, the compressive strength prediction result, bending toughness prediction result, ultimate tensile strain prediction result, fracture energy prediction result, and toughness prediction reliability level are associated and output according to the same sample number to be predicted.

[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.

Claims

1. A method for predicting multiple properties of high-toughness concrete, characterized in that, include: Obtain high-toughness concrete sample data, including mix proportion parameters, fiber parameters, curing conditions, and corresponding measured values ​​of compressive strength, flexural toughness, ultimate tensile strain, and fracture energy; The sample data is preprocessed to obtain standardized sample data, which is then divided into training samples, validation samples, and test samples. Using mix proportion parameters, fiber parameters and curing conditions as inputs, and compressive strength, flexural toughness, ultimate tensile strain and fracture energy as outputs, a BP neural network multivariate performance prediction model was established and trained. The test samples are input into the trained BP neural network multivariate performance prediction model to obtain the predicted values ​​of each performance and compare them with the corresponding measured values ​​to obtain the prediction error of compressive strength and the prediction error of toughness index. Strength bias learning parameters are generated based on the prediction error of compressive strength and the prediction error of toughness index. Obtain the ductile failure process characteristics corresponding to the sample, and construct ductile instability discrete parameters based on the ductile failure process characteristics; The strength bias learning parameters and the toughness instability discrete parameters are combined and analyzed to obtain the toughness prediction capability score; The BP neural network multi-performance prediction model is modified based on the toughness prediction capability score, and the modified model is used to output the multi-performance prediction results and toughness prediction reliability level of the high-toughness concrete to be predicted.

2. The method for predicting multiple properties of high-toughness concrete according to claim 1, characterized in that, Obtain high-toughness concrete sample data, including: Establish sample numbers according to the same high-toughness concrete sample, and link and store the mix proportion parameters, fiber parameters, curing conditions with the measured values ​​of compressive strength, flexural toughness, ultimate tensile strain and fracture energy. The material usage, fiber size, curing age, and measured performance values ​​from samples from different sources were standardized to the same data unit, and the corresponding data source, specimen size, test method, and loading method were recorded. When the sample uses mixed fibers, the type, length, diameter and doping amount of each fiber are recorded separately, so that the different fiber features in the same sample can independently participate in feature expression in subsequent prediction; The test results of parallel specimens corresponding to each performance measured value and their dispersion are bound to the sample number to form the original sample data.

3. The method for predicting multiple properties of high-toughness concrete according to claim 2, characterized in that, Preprocessing of sample data includes: Field alignment and missing item screening were performed on the original sample data to remove data records that lacked any measured value of compressive strength, flexural toughness, ultimate tensile strain and fracture energy; Based on the median and robust dispersion of the measured values ​​of compressive strength, flexural toughness, ultimate tensile strain and fracture energy, respectively, the robust deviation values ​​of compressive strength and toughness response are obtained. When the compressive strength robustness deviation value of the same data record is in the central range while the toughness response deviation value is in the discrete range, or when the compressive strength robustness deviation value is in the discrete range while the toughness response deviation value is in the central range, it is determined that the data record has a disconnect between strength response and toughness response. Data records whose strength response and toughness response do not correspond are removed to obtain sample data after consistency verification.

4. The method for predicting multiple properties of high-toughness concrete according to claim 3, characterized in that, Obtain standardized sample data and divide it into training samples, validation samples, and test samples, including: The mix proportion parameters, fiber parameters, curing conditions and various measured values ​​after consistency verification were treated as independent parameter categories and scaled to obtain initial standardized data. The fiber volumetric content, ultimate tensile strain and fracture energy were divided into four intervals according to the 25%, 50% and 75% quantiles. After standardization within each interval, the data were merged according to the sample number to obtain standardized sample data. Based on the standardized measured compressive strength and the comprehensive measured toughness obtained from bending toughness, ultimate tensile strain, and fracture energy, the samples are divided into low strength, medium strength, high strength, low toughness, medium toughness, and high toughness. The strength and toughness categories were combined to form nine joint distribution groups, and within each joint distribution group, training samples, validation samples, and test samples were divided in a 7:1:2 ratio.

5. The method for predicting multiple properties of high-toughness concrete according to claim 1, characterized in that, Establish and train a BP neural network multivariate performance prediction model, including: According to the order of the preprocessed fields, the mix proportion parameters, fiber parameters and curing conditions in the training samples are combined into an input feature sequence, and the compressive strength is used as the strength output label, and the bending toughness, ultimate tensile strain and fracture energy are used as the toughness output label. Each batch of training samples is organized into 9 joint distribution groups, so that each batch of training samples contains data records with different intensity levels and different toughness discrete levels. A multivariate performance prediction model of a BP neural network is established, which includes an input layer, a shared hidden layer, a compressive strength output node, and three toughness output nodes. During joint training, when the output error of compressive strength in the validation samples decreases continuously while the output error of toughness does not decrease synchronously, the connection weights corresponding to the output node of compressive strength are fixed, and only the connection weights from the shared hidden layer to the output node of toughness are updated in reverse until the output error of toughness decreases and then joint training is resumed.

6. The method for predicting multiple properties of high-toughness concrete according to claim 5, characterized in that, The prediction errors for compressive strength and toughness-related indices are obtained, including: The test samples are input into the trained BP neural network multivariate performance prediction model according to the preprocessed field order to obtain the standardized predicted values ​​of compressive strength, flexural toughness, ultimate tensile strain and fracture energy corresponding to the sample number. According to the overall standardization rules or interval standardization rules recorded in the preprocessing stage, the standardized predicted values ​​are restored to the data scale consistent with the corresponding measured values; Based on the sample number, the restored performance prediction values ​​are matched with the corresponding measured values ​​one by one to obtain strength prediction comparison data and toughness prediction comparison data. Based on the strength prediction comparison data, the direction and magnitude of the compressive strength deviation are extracted. Based on the toughness prediction comparison data, the direction and magnitude of the deviation of the bending toughness, ultimate tensile strain and fracture energy are extracted. The prediction errors of the compressive strength and toughness indicators are then summarized according to the joint distribution group.

7. The method for predicting multiple properties of high-toughness concrete according to claim 6, characterized in that, Strength-biased learning parameters are generated based on the prediction errors of compressive strength and toughness indices, including: Within the same joint distribution group, the prediction errors of compressive strength and toughness indicators are sorted from low to high to obtain the strength error order and toughness error order for the same sample. Samples whose toughness error ranking is at least 20% higher than the strength error ranking within the same group are extracted as strength-biased candidate samples. The candidate samples with strength bias are screened for directional consistency. Samples with compressive strength deviation not exceeding the median of compressive strength deviation in the group and at least two toughness indicators showing the same direction of deviation are retained as strength bias confirmation samples. Based on the strength bias, the distribution ratio of the confirmed samples, the degree to which the toughness error order exceeds the strength error order, and the same-direction deviation of toughness indicators, strength bias learning parameters are generated.

8. The method for predicting multiple properties of high-toughness concrete according to claim 7, characterized in that, Obtain the ductile failure process characteristics corresponding to the sample, and construct ductile instability discrete parameters based on the ductile failure process characteristics, including: According to the sample number, load change data, deformation change data and crack evolution data are obtained during the bending toughness, ultimate tensile strain and fracture energy test process, and bound with the toughness error record to form a toughness failure process record. The ductile failure process record was divided into the initial cracking stage, the multi-crack propagation stage, and the post-peak attenuation stage, and the characteristics of crack number variation, crack width variation, crack spacing variation, and post-peak bearing capacity attenuation were extracted respectively. Based on the stage failure characteristics, the discrete characteristics of crack distribution, the concentrated development characteristics of deformation, and the post-peak load attenuation characteristics are extracted to form a set of ductile instability characteristics; Based on the degree of synchronous occurrence of the discrete characteristics of crack distribution, concentrated development characteristics of deformation, and post-peak load attenuation characteristics in the same sample, discrete parameters of ductile instability are constructed.

9. The method for predicting multiple properties of high-toughness concrete according to claim 8, characterized in that, By comprehensively analyzing the strength-biased learning parameters and the ductile instability discrete parameters, a ductile prediction capability score is obtained, including: Based on the sample number, the single sample strength bias learning parameters and the single sample ductile instability discrete parameters are matched to form a correlation record between model learning bias and material ductile instability. The high-level limits are determined based on the 75th percentile of the intensity bias learning parameter and the discrete parameter of toughness instability, and 0.50, respectively. Samples that reach the corresponding high-level limits at the same time are marked as toughness prediction risk samples. The distribution proportion of resilience prediction risk samples is statistically analyzed according to different strength levels and resilience dispersion levels. A non-compensation fusion method is adopted to fuse the strength bias learning parameter and the toughness instability discrete parameter, so that when either parameter is at a high level, it cannot be completely canceled by the low level of the other parameter. The toughness prediction capability score is obtained based on the fusion result and the distribution ratio of the toughness prediction risk sample.

10. The method for predicting multiple properties of high-toughness concrete according to claim 9, characterized in that, The BP neural network multivariate performance prediction model was modified based on the resilience prediction capability score, including: The correction method is determined based on the toughness prediction capability score. When the toughness prediction capability score is below 70, the BP neural network multivariate performance prediction model is retrained in a targeted manner. When the toughness prediction capability score is not lower than 70 but lower than 85, the output nodes of bending toughness, ultimate tensile strain and fracture energy are lightly retrained. When the toughness prediction capability score is not lower than 85, the current model parameters are kept unchanged. Samples with a toughness correction sample whose toughness standardization error is higher than that of compressive strength standardization error and whose difference is not less than 0.05 are extracted from the test samples. The BP neural network multivariate performance prediction model was retrained in a targeted manner using toughness-corrected samples. During the retraining process, the connection weights of the compressive strength output node remained unchanged, while the learning ratio of the bending toughness, ultimate tensile strain, and fracture energy output nodes to the toughness-corrected samples was increased, resulting in the corrected BP neural network multivariate performance prediction model.

Images