Method and system for predicting suitability of fly ash stabilized soil based on physical parameters

Abstract

The application provides a fly ash solidified soil applicability prediction method and system based on physical parameters, and relates to the technical field of civil engineering material performance prediction. The method obtains an original physical parameter set of a fly ash to be measured; generates a plurality of disturbance sequences based on a preset fluctuation range corresponding to each original physical parameter, and superimposes the disturbance sequences to the original physical parameter set to obtain a disturbed parameter set; iteratively trains an initial machine learning model by using the original physical parameter set and the disturbed parameter set, and regards that the positive sample solidified soil performance prediction value approaches a preset target value and the stability characteristic quantity calculated based on the negative sample approaches a minimum value as a training target, so as to obtain a trained machine learning model; obtains a target physical parameter of a fly ash to be evaluated, inputs the trained machine learning model, and outputs a solidified soil applicability prediction result. The application can simultaneously consider the prediction accuracy and the anti-disturbance ability, and provides a scientific basis for the material selection of the fly ash in the soft soil solidification engineering.

Description

Technical Field

[0001] This invention relates to the field of civil engineering material performance prediction technology, and in particular to a method and system for predicting the applicability of fly ash-stabilized soil based on physical property parameters. Background Technology

[0002] In the field of civil engineering, especially in the solidification of soft soil, fly ash is a commonly used industrial waste solidifying agent. Its compatibility with soft soil directly determines the engineering performance of the solidified soil. However, as an industrial by-product, the physicochemical properties of fly ash are affected by raw materials and combustion processes, exhibiting significant batch-to-batch fluctuations. Current technologies for predicting the performance of fly ash-solidified soil typically rely on estimations based on single or static physical properties, neglecting the impact of fluctuations in fly ash parameters on the stability of the final solidified soil performance. This leads to overly idealized predictions that fail to accurately reflect the performance variability caused by fluctuations in fly ash properties in actual engineering applications. Furthermore, traditional methods struggle to quantify the degree of disturbance this fluctuation causes to the solidified soil performance, leaving engineers without reliable data support when selecting fly ash materials and unable to effectively assess their applicability and stability under specific working conditions. Summary of the Invention

[0003] To address the aforementioned problems in existing technologies, the first aspect of this invention proposes a method for predicting the applicability of fly ash-stabilized soil based on physical property parameters, comprising: S1: Obtain the original set of physical property parameters of the fly ash to be tested; S2: Based on the preset fluctuation range corresponding to each original physical property parameter in the original physical property parameter set, multiple disturbance sequences are generated, and the multiple disturbance sequences are superimposed on the original physical property parameter set to obtain the disturbed parameter set; wherein, S2 includes: S21: Obtain the historical extreme values ​​of fluctuations for each original physical property parameter in the original physical property parameter set, and use them as the preset fluctuation range; S22: Generate a set of initial chaotic time series based on the control equations of the chaotic system; S23: Normalize the initial chaotic time series and scale the amplitude of the initial chaotic time series to match the preset fluctuation range to obtain multiple chaotic disturbance value sequences; S24: Add the multiple chaotic perturbation value sequences to the corresponding original physical property parameters in the original physical property parameter set respectively to generate a perturbed parameter set containing multiple perturbed samples; S3: Based on the original set of physical property parameters and the disturbed set of parameters, iteratively train the initial machine learning model until the initial machine learning model outputs the predicted value of the solidified soil performance and the corresponding stability characterization quantity, thus obtaining the trained machine learning model; wherein, S3 includes: S31: Construct an initial machine learning model that includes a feature extraction layer and a prediction output layer; S32: The original set of physical property parameters is used as a positive sample input to the feature extraction layer, and the disturbed set of parameters is used as a negative sample input to the feature extraction layer. After processing by the feature extraction layer, positive sample feature vectors and negative sample feature vectors are generated respectively. S33: Input the positive sample feature vector and the negative sample feature vector into the prediction output layer. After processing by the prediction output layer, the predicted values ​​of the solidified soil performance of the positive sample and the predicted values ​​of the solidified soil performance of the negative sample are generated respectively. S34: Calculate the dispersion of the predicted performance values ​​of the solidified soil in the negative sample as a stability characterization quantity; S35: With the predicted value of solidified soil performance of positive samples approaching the preset target value and the stability characterization value approaching the minimum as the training objective, the parameters of the initial machine learning model are iteratively updated until the preset convergence condition is met, and the trained machine learning model is obtained. S4: Obtain the target physical property parameters of the fly ash to be evaluated, input them into the trained machine learning model, and after processing by the trained machine learning model, output the prediction results of the applicability of the solidified soil corresponding to the fly ash to be evaluated.

[0004] Secondly, this invention proposes a fly ash-stabilized soil applicability prediction system based on physical property parameters. The system employs the aforementioned fly ash-stabilized soil applicability prediction method based on physical property parameters, and includes: The original parameter acquisition module is used to perform step S1: acquire the original physical property parameter set of the fly ash to be tested; The disturbance sequence generation and superposition module is used to execute step S2: based on the preset fluctuation range corresponding to each original physical property parameter in the original physical property parameter set, multiple disturbance sequences are generated, and the multiple disturbance sequences are superimposed on the original physical property parameter set to obtain the disturbed parameter set; wherein, S2 includes: S21: Obtain the historical extreme values ​​of fluctuations for each original physical property parameter in the original physical property parameter set, and use them as the preset fluctuation range; S22: Generate a set of initial chaotic time series based on the control equations of the chaotic system; S23: Normalize the initial chaotic time series and scale the amplitude of the initial chaotic time series to match the preset fluctuation range to obtain multiple chaotic disturbance value sequences; S24: Add the multiple chaotic perturbation value sequences to the corresponding original physical property parameters in the original physical property parameter set respectively to generate a perturbed parameter set containing multiple perturbed samples; The model training module is used to execute step S3: based on the original set of physical property parameters and the disturbed set of parameters, iteratively train the initial machine learning model until the initial machine learning model outputs the predicted value of the solidified soil performance and the corresponding stability characterization quantity, thus obtaining the trained machine learning model; wherein, S3 includes: S31: Construct an initial machine learning model that includes a feature extraction layer and a prediction output layer; S32: The original set of physical property parameters is used as a positive sample input to the feature extraction layer, and the disturbed set of parameters is used as a negative sample input to the feature extraction layer. After processing by the feature extraction layer, positive sample feature vectors and negative sample feature vectors are generated respectively. S33: Input the positive sample feature vector and the negative sample feature vector into the prediction output layer. After processing by the prediction output layer, the predicted values ​​of the solidified soil performance of the positive sample and the predicted values ​​of the solidified soil performance of the negative sample are generated respectively. S34: Calculate the dispersion of the predicted performance values ​​of the solidified soil in the negative sample as a stability characterization quantity; S35: With the predicted value of solidified soil performance of positive samples approaching the preset target value and the stability characterization value approaching the minimum as the training objective, the parameters of the initial machine learning model are iteratively updated until the preset convergence condition is met, and the trained machine learning model is obtained. The prediction output module is used to perform step S4: obtain the target physical property parameters of the fly ash to be evaluated, input them into the trained machine learning model, and after processing by the trained machine learning model, output the prediction result of the applicability of the solidified soil corresponding to the fly ash to be evaluated.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: Obtaining the original set of physical property parameters of the fly ash to be tested provides real basic data for subsequent analysis, ensuring the objectivity and relevance of the prediction process. The historical extreme values ​​of fluctuations for each original physical property parameter are obtained as a preset fluctuation range, providing boundaries for the disturbance amplitude that conform to actual working conditions. Then, an initial chaotic time series is generated based on the governing equations of the chaotic system, and it is normalized and scaled to the preset fluctuation range to obtain multiple chaotic disturbance value sequences. Finally, these disturbance value sequences are superimposed on the original set of physical property parameters to generate a disturbed parameter set containing multiple disturbed samples. The above operations, through the inherent randomness and ergodicity of the chaotic system, simulate the complex fluctuation characteristics of fly ash physical property parameters in a real engineering environment, generating a disturbance sample pool, providing negative samples reflecting parameter variability for subsequent model training.

[0006] Building upon this foundation, a machine learning model was constructed and trained. The original set of physical property parameters was used as positive samples, and the disturbed set of parameters as negative samples were input into the feature extraction layer in parallel. This enabled the model to simultaneously learn the performance output under steady-state conditions and the performance shift under disturbed conditions. Predicted values ​​of solidified soil performance for both positive and negative samples were generated through the output prediction layer. Then, the dispersion of the negative sample predicted values ​​was calculated as a stability characterization, quantifying the degree of variation in solidified soil performance caused by fluctuations in physical property parameters. Iterative optimization was performed with the training objective of ensuring that the positive sample predicted values ​​approach a preset target value and that the stability characterization approaches its minimum. This training objective compels the model to maintain robustness to fluctuations in input parameters while pursuing prediction accuracy; that is, when the input is disturbed, the fluctuation amplitude of the output performance is minimized.

[0007] Finally, by inputting the target physical property parameters of the fly ash to be evaluated into the model, an applicability prediction result that integrates performance level and stability can be output. This result overcomes the limitation of traditional methods that only focus on a single predicted value, providing engineers with a more comprehensive and reliable basis for decision-making, and improving the scientific selection and application safety of fly ash in soft soil consolidation projects. Attached Figure Description

[0008] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0009] Figure 1 The diagram shown is a flowchart illustrating a method for predicting the applicability of fly ash-stabilized soil based on physical property parameters, according to an embodiment of the present invention. Figure 2 The diagram shown is a schematic diagram of a fly ash-stabilized soil applicability prediction system based on physical property parameters provided in an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0011] The specific embodiments of the present invention will be described below.

[0012] Example 1 like Figure 1As shown, the first aspect of this invention proposes a method for predicting the applicability of fly ash-stabilized soil based on physical property parameters, comprising: S1: Obtain the original set of physical property parameters of the fly ash to be tested; S2: Based on the preset fluctuation range corresponding to each original physical property parameter in the original physical property parameter set, multiple disturbance sequences are generated, and the multiple disturbance sequences are superimposed on the original physical property parameter set to obtain the disturbed parameter set; wherein, S2 includes: S21: Obtain the historical extreme values ​​of fluctuations for each original physical property parameter in the original physical property parameter set, and use them as the preset fluctuation range; S22: Generate a set of initial chaotic time series based on the control equations of the chaotic system; S23: Normalize the initial chaotic time series and scale the amplitude of the initial chaotic time series to match the preset fluctuation range to obtain multiple chaotic disturbance value sequences; S24: Add the multiple chaotic perturbation value sequences to the corresponding original physical property parameters in the original physical property parameter set respectively to generate a perturbed parameter set containing multiple perturbed samples; S3: Based on the original set of physical property parameters and the disturbed set of parameters, iteratively train the initial machine learning model until the initial machine learning model outputs the predicted value of the solidified soil performance and the corresponding stability characterization quantity, thus obtaining the trained machine learning model; wherein, S3 includes: S31: Construct an initial machine learning model that includes a feature extraction layer and a prediction output layer; S32: The original set of physical property parameters is used as a positive sample input to the feature extraction layer, and the disturbed set of parameters is used as a negative sample input to the feature extraction layer. After processing by the feature extraction layer, positive sample feature vectors and negative sample feature vectors are generated respectively. S33: Input the positive sample feature vector and the negative sample feature vector into the prediction output layer. After processing by the prediction output layer, the predicted values ​​of the solidified soil performance of the positive sample and the predicted values ​​of the solidified soil performance of the negative sample are generated respectively. S34: Calculate the dispersion of the predicted performance values ​​of the solidified soil in the negative sample as a stability characterization quantity; S35: With the predicted value of solidified soil performance of positive samples approaching the preset target value and the stability characterization value approaching the minimum as the training objective, the parameters of the initial machine learning model are iteratively updated until the preset convergence condition is met, and the trained machine learning model is obtained. S4: Obtain the target physical property parameters of the fly ash to be evaluated, input them into the trained machine learning model, and after processing by the trained machine learning model, output the prediction results of the applicability of the solidified soil corresponding to the fly ash to be evaluated.

[0013] In specific engineering applications, fly ash, as a commonly used mineral admixture, exhibits physical properties such as fineness, loss on ignition, and water requirement ratio that fluctuate depending on the origin, batch, and production process. These fluctuations directly impact the engineering performance of soft soil after solidification treatment, including strength and stability. To accurately predict the suitability of fly ash for solidification of specific soft soils, this invention proposes a prediction method based on physical property parameters. First, the original set of physical property parameters for the fly ash to be tested is obtained. This parameter set serves as the foundation for all subsequent analyses and predictions, including key indicators characterizing the chemical activity and physical state of fly ash, such as fineness, loss on ignition, and activity index. Since these parameters may fluctuate randomly within a certain range during the production and transportation of fly ash, predictions based solely on a single original parameter value may not fully reflect the performance under actual working conditions. Therefore, based on a preset fluctuation range corresponding to each original physical property parameter, multiple perturbation sequences are generated, and these perturbation sequences are superimposed one by one onto the original physical property parameter set, thereby obtaining a perturbed parameter set containing a large number of perturbed samples. This process essentially simulates various possible variations of fly ash parameters within a reasonable range, constructing a sample set that can cover the actual fluctuation space of the parameters. The preset fluctuation range is determined based on the extreme values ​​of fluctuations in historical statistical data to ensure that the generated disturbance values ​​do not deviate from physical reality.

[0014] The original set of physical property parameters is used as positive samples, and the disturbed set of parameters is used as negative samples. Both are input into the initial machine learning model for iterative training. During training, the model outputs the predicted values ​​of the solidified soil performance corresponding to the negative samples, and calculates the dispersion of these predicted values ​​as a stability characterization. The model parameters are adjusted by an optimization algorithm so that the predicted values ​​of the positive samples continuously approach the preset target values, while the stability characterization continuously decreases, that is, the fluctuation of the predicted values ​​of the negative samples becomes smaller and smaller. The final trained machine learning model has both high-precision prediction ability and robustness to fluctuations in input parameters.

[0015] The preset fluctuation range is determined by obtaining the historical extreme values ​​of each original physical property parameter. Historical extreme values ​​refer to the maximum and minimum values ​​that each physical property parameter has ever experienced, statistically analyzed from long-term quality testing data of the fly ash source or similar materials. This extreme value range constitutes the physical boundary of the parameter's potential variation under natural conditions, ensuring that subsequent introduced disturbances do not deviate from reality. Additionally, a disturbance sequence conforming to the inherent laws of material fluctuation needs to be generated, introducing the governing equations of the chaotic system, such as using classical chaotic models like the Lorentz equation or logistic mapping. By setting the initial state values ​​and control parameter values ​​of the chaotic system, the system is brought into a chaotic state, and iterative solutions are applied to generate a set of temporally correlated, inherently deterministic initial chaotic time series. The characteristic of chaotic time series is that while they appear random on the surface, they are actually generated by nonlinear deterministic equations, capable of simulating the non-periodic fluctuations of engineering material parameters caused by complex factors.

[0016] After generating the initial sequence, it needs to be normalized to map the values ​​in the sequence to a standardized interval near zero, thus eliminating the dimensional influence of the original chaotic values. Then, based on the predefined fluctuation range of each original physical property parameter, the amplitude of the normalized chaotic sequence is linearly scaled so that its numerical range precisely matches the actual fluctuation range of the parameter, resulting in multiple sequences of chaotic perturbation values ​​corresponding to the physical meaning of the original parameters. Finally, the chaotic perturbation value sequences are algebraically added to the corresponding original parameter values ​​in the original physical property parameter set, i.e., the original parameter values ​​are added to the perturbation values, generating a new sample. Since chaotic sequences typically contain values ​​at multiple time points, multiple different perturbation values ​​can be generated, resulting in multiple different perturbed samples. All these samples together form the perturbed parameter set. The perturbations generated by the chaotic system can deterministically traverse a large number of possible states in the parameter space, making the negative samples encountered in subsequent model training more diverse and representative, thereby effectively improving the model's adaptability to parameter fluctuations in the real world.

[0017] The initial machine learning model comprises two main parts: a feature extraction layer and a prediction output layer. The feature extraction layer, composed of a multi-layer neural network, automatically extracts deep-level features from the input raw physical property parameters that decisively influence the performance of the solidified soil, mapping the low-dimensional raw parameters to a high-dimensional feature space. The prediction output layer, based on the extracted features, performs linear or non-linear transformations to ultimately output predicted values ​​for the solidified soil performance. During the training phase, the raw physical property parameter set is input as positive samples to the feature extraction layer, while the perturbed parameter set is input as negative samples. Positive samples represent inputs under undisturbed or ideal conditions, while negative samples represent actual, fluctuating inputs. After processing by the feature extraction layer, positive samples are converted into positive sample feature vectors, and negative samples are converted into negative sample feature vectors. These two feature vectors carry key information under stable and perturbed states, respectively. Then, these two types of feature vectors are input to the prediction output layer, which, after processing, generates predicted values ​​for the solidified soil performance of both positive and negative samples. To quantify the impact of parameter fluctuations on prediction results, it is necessary to calculate the dispersion of the predicted values ​​of the solidified soil properties of negative samples, such as calculating their standard deviation or variance, and define this dispersion as a stability characterization quantity. The smaller the value of this characterization quantity, the more concentrated and stable the model's prediction results are, even if the input parameters fluctuate within a certain range; that is, the less sensitive the model is to input disturbances.

[0018] Through backpropagation and an optimizer, the internal parameters of the initial machine learning model are iteratively updated, continuously optimizing the model in terms of both accuracy and stability until a preset convergence condition is met, such as the loss function value no longer decreasing or reaching a preset number of iterations. The resulting trained machine learning model then possesses the ability to provide both accurate predictions and stability assessments in practical applications.

[0019] In some implementations, S22 includes: S221: Select the Lorentz system or logistic mapping as the governing equation for the chaotic system; S222: Set the initial state value and control parameter value of the control equation, iteratively solve the control equation, and generate an initial chaotic time series containing values ​​at multiple time points.

[0020] When generating initial chaotic time series using chaotic systems, several specific mathematical models are available, among which the Lorentz system and the logistic mapping are two typical examples. The Lorentz system is a set of ordinary differential equations describing atmospheric convection. Its solutions exhibit chaotic trajectories with strange attractors in phase space. The system is highly sensitive to initial conditions; small initial differences lead to significant deviations in subsequent trajectories. This characteristic makes the generated sequences highly complex and unpredictable, making it suitable for simulating the fluctuations in fly ash parameters affected by multiple coupled factors. The logistic mapping is a discrete-time nonlinear equation. When the control parameters take values ​​within a specific range, the system enters a chaotic state, and the generated sequences appear random while simultaneously following deterministic patterns.

[0021] In practice, a chaotic system is first selected as the governing equation, and then the initial state value and control parameter values ​​are set. The initial state value is the starting point for iterative calculations, and the control parameter values ​​determine whether the system is in a chaotic region and the complexity of the chaos. The governing equation is iteratively solved repeatedly, and each iteration yields a value at a specific time point. Continuous iterations generate a set of numerical sequences correlated along the time axis, i.e., the initial chaotic time series. Generating perturbation values ​​based on the initial chaotic time series ensures that the perturbation modes are not only diverse but also cover a large region of the parameter space. The perturbations generated by the initial chaotic time series can better simulate various complex fluctuations in material parameters that may occur in reality.

[0022] In some implementations, the feature extraction layer includes a noise injection sublayer, and S32 includes: S321: Input the sample noise into the sub-layer of the original physical property parameter set. After adding random perturbation through the noise injection sub-layer, generate positive sample intermediate features. Input the positive sample intermediate features into the rest of the feature extraction layer to generate positive sample feature vectors. S322: The samples in the disturbed parameter set are directly input into the rest of the feature extraction layer, bypassing the noise injection sub-layer, to generate negative sample feature vectors.

[0023] Specifically, when samples from the original set of physical property parameters are input as positive samples, they first pass through a noise injection sub-layer. In this sub-layer, the system adds small random perturbations to the original samples. These perturbations can be white noise following a Gaussian distribution. After noise injection, the original clean samples become positive sample intermediate features with slight noise. The purpose of this operation is to enable the model to not only remember the surface form of the input data when learning feature representations, but also to mine deep-level features that remain stable under small perturbations, thereby improving the model's generalization ability. Then, the positive sample intermediate features are fed into the rest of the feature extraction layer, and after layers of transformation in the subsequent network, the positive sample feature vector is finally generated. For samples from the perturbed parameter set, since the perturbed parameter set itself has already been superimposed with a large amount of perturbation by the chaotic system, passing them through the noise injection sub-layer might introduce too much human intervention. Therefore, samples from the perturbed parameter set are directly input into the rest of the feature extraction layer, bypassing the noise injection sub-layer, thereby directly generating negative sample feature vectors.

[0024] Through the aforementioned differentiated processing methods, the positive sample path enhances the robustness of features by injecting noise, while the negative sample path retains the original perturbation information for evaluating model stability. During training, the model continuously adjusts its internal weights by comparing the feature representations of positive and negative samples with the predicted output. Ultimately, while maintaining a high-precision fit to the original data, it effectively resists performance degradation caused by fluctuations in input parameters, thus exhibiting excellent prediction accuracy and reliability in practical applications.

[0025] In some implementations, S34 includes: S341: Obtain the predicted values ​​of the negative sample solidified soil properties of multiple disturbed samples corresponding to the same original physical property parameter; S342: Calculate the standard deviation of the predicted performance values ​​of solidified soil from multiple negative samples, and use the standard deviation as a measure of stability.

[0026] In practice, for the same set of original physical property parameters, multiple perturbed samples are generated by superimposing different chaotic perturbation sequences. These perturbed samples are then sequentially input into the currently trained machine learning model, which outputs a corresponding negative sample predicted value for the solidified soil performance for each perturbed sample. Because the input parameters of each perturbed sample have slight differences, the output predicted values ​​are not entirely identical, but rather form a distribution. This distribution reflects the fluctuation in the model's predicted performance when fluctuations occur around the original parameter values. To quantitatively describe the concentration or dispersion of this distribution, the standard deviation of the predicted performance values ​​for the multiple negative samples needs to be calculated. A larger standard deviation indicates greater differences between these predicted values, a more dispersed distribution, meaning that parameter fluctuations have a more significant impact on performance, and the model's stability is worse. Conversely, a smaller standard deviation indicates that the predicted values ​​are tightly clustered around a central value, a more concentrated distribution, meaning that parameter fluctuations have almost no impact on the model output, and the model's stability is stronger.

[0027] In some implementations, S4 includes: S41: Obtain the target physical property parameters of the fly ash to be evaluated, input them into the trained machine learning model, and after processing by the trained machine learning model, output the predicted values ​​of solidified soil performance in multiple dimensions. S42: The predicted values ​​of solidified soil performance in multiple dimensions are used as coordinate values ​​and mapped to a preset multidimensional coordinate system to generate a multidimensional map characterizing the applicability of solidified soil.

[0028] In practice, the target physical property parameters of the fly ash to be evaluated are obtained. These parameters include key indicators characterizing the physicochemical properties of fly ash, such as loss on ignition, water demand ratio, activity index, and fineness modulus. After inputting these parameters into a trained machine learning model, the model performs calculations based on its internally learned mapping relationships, ultimately outputting multi-dimensional predicted values ​​for the solidified soil performance. Multi-dimensional predicted values ​​refer to the model simultaneously outputting different types of engineering performance indicators for the same input sample. For example, it can simultaneously output unconfined compressive strength, flexural strength, permeability coefficient, or deformation modulus. These different dimensional predicted values ​​reflect the comprehensive performance of the solidified soil under specific working conditions from different perspectives. After obtaining these multi-dimensional predicted values, they are mapped to a preset multi-dimensional coordinate system. The number of coordinate axes in the multi-dimensional coordinate system corresponds to the number of dimensions of the predicted values, with each dimension of the performance prediction value corresponding to one coordinate axis. For example, if the model outputs performance predictions in three dimensions, a three-dimensional coordinate system is constructed, and the three predictions are used as the coordinates of the x-axis, y-axis, and z-axis, respectively, thereby determining a unique point in space.

[0029] In this embodiment, the abstract performance prediction values ​​originally presented in numerical form are transformed into visual points in a multi-dimensional space. The comparison of the applicability of different fly ash samples becomes a comparison of the distribution of points in a multi-dimensional space. Engineers can intuitively judge the differences in the curing effect of different fly ash materials by observing the spatial position of these points. For example, they can observe whether a point of a fly ash sample falls within a certain ideal area in multi-dimensional space, or compare the Euclidean distance between different sample points to measure their overall performance similarity.

[0030] In some implementations, the preset multidimensional coordinate system includes a soft soil moisture content axis, a soft soil organic matter content axis, and a strength guarantee rate axis. S42 includes: S421: Use the target physical property parameters of the fly ash to be evaluated as fixed values; S422: Obtain multiple different soft soil moisture content values ​​and soft soil organic matter content values ​​as variable input values; S423: Input the fixed value and each set of variable input values ​​into the trained machine learning model. After processing by the trained machine learning model, output the predicted value of the solidified soil performance and the corresponding stability characterization quantity for each set of inputs. S424: Convert the stability characterization quantity corresponding to each set of inputs into the strength guarantee rate, and generate a multidimensional map with the soft soil moisture content as the first axis coordinate, the soft soil organic matter content as the second axis coordinate, and the strength guarantee rate as the third axis coordinate.

[0031] In constructing a multidimensional graph, the specific dimensions of the coordinate system can be set according to the specific needs and application scenarios of the actual project, so that the graph can directly serve specific engineering design goals. For example, in the application scenario of soft soil solidification treatment, the properties of the soft soil itself have a decisive influence on the solidification effect, among which the moisture content and organic matter content of the soft soil are two of the most critical variables. The moisture content of soft soil refers to the percentage of the mass of water contained in the soft soil to the mass of dry soil, which determines the sufficiency of the hydration reaction during the solidification process and the dense structure of the solidified soil in the end. The organic matter content of soft soil refers to the percentage of the mass of animal and plant residues and their decomposition products in the soft soil. The presence of organic matter will adsorb calcium ions produced by the hydration of cement or fly ash, hindering the volcanic ash reaction, thus significantly affecting the development of solidification strength. At the same time, the strength guarantee rate of the solidified soil is a core engineering indicator for measuring the stability of solidification, reflecting the reliability of the solidified soil strength in meeting the design requirements under the premise of considering the fluctuation of material parameters. Therefore, a three-dimensional coordinate system including the soft soil moisture content axis, the soft soil organic matter content axis, and the strength guarantee rate axis is preset so that the graph can intuitively show the relationship between these three key variables.

[0032] In practice, the target physical properties of the fly ash to be evaluated are first set as fixed values. This means that in subsequent analyses, the inherent properties of this batch of fly ash, such as loss on ignition, water demand ratio, activity index, and fineness modulus, are assumed to remain constant, and only the influence of changes in soft soil conditions on the solidification effect is considered. Then, multiple different soft soil moisture content and organic matter content values ​​are obtained through sampling tests in actual engineering projects or by consulting relevant geological data, serving as variable input values. The fixed values ​​and each set of variable input values ​​are input into the trained machine learning model. The model processes each input combination independently, outputting the corresponding predicted value of the solidified soil performance and the corresponding stability characterization quantity. Here, the output predicted value of the solidified soil performance refers to the unconfined compressive strength value, and the stability characterization quantity is a statistical index obtained by calculating the dispersion of multiple predicted values ​​under the disturbed parameter set, reflecting the degree of variation in the solidified soil performance corresponding to this set of input conditions under fluctuating fly ash physical properties.

[0033] Each set of inputs is converted into a stability characteristic quantity, which transforms the originally abstract discreteness index into an intuitive probability value, making it easier for engineers to understand and apply. Finally, points are plotted in three-dimensional space with the soft soil moisture content as the first axis, the soft soil organic matter content as the second axis, and the strength guarantee rate as the third axis. All points together form a three-dimensional surface or scatter plot. This plot visually displays the distribution of solidification strength guarantee rates achievable under fixed fly ash conditions for different soft soil conditions. Engineers can directly read the strength guarantee rate corresponding to a specific combination of soft soil moisture content and organic matter content from the plot, providing a clear visual reference for selecting appropriate fly ash under specific soft soil conditions, thus improving the practicality and interpretability of the prediction results.

[0034] In some implementations, the stability characteristic corresponding to each set of inputs is converted into a strength guarantee rate, including: Obtain the preset engineering strength threshold; The probability value that the predicted value of solidified soil performance is not lower than the engineering strength threshold under the disturbance conditions of the disturbed parameter set is statistically analyzed, and the probability value is used as the strength guarantee rate.

[0035] A higher strength guarantee rate indicates that even if parameters such as fly ash loss on ignition and water demand ratio vary randomly within a certain range, the strength of the solidified soil can still stably meet or exceed the design requirements, and the material's applicability and engineering reliability are also better. This embodiment transforms the abstract stability characteristic into a strength guarantee rate with clear engineering significance, making the prediction results closer to the decision-making needs in engineering practice. Designers and construction personnel can directly judge the applicability risk of specific fly ash under specific soft soil conditions based on the strength guarantee rate, thereby making more scientific and reasonable material selection decisions.

[0036] In some implementations, the method also includes: Identify all coordinate points in the multidimensional map whose strength guarantee rate is greater than the preset guarantee rate threshold, and mark the area covered by all coordinate points as the compatibility recommendation area; Obtain the measured moisture content and measured organic matter content of the soft soil to be evaluated; Determine whether the coordinate points formed by the measured moisture content and measured organic matter content fall within the compatibility recommendation area in the multidimensional map, and generate a compatibility evaluation result.

[0037] In the multidimensional map, coordinate points with a strength guarantee rate greater than a preset guarantee rate threshold are identified through image processing or numerical search methods. The combination of soft soil moisture content and soft soil organic matter content corresponding to these points represents the range of soft soil conditions under which the engineering requirements can be met with high reliability under this fly ash condition. The area covered by all coordinate points that meet the conditions is marked or colored on the map. This marked area is the compatibility recommendation zone, which visually delineates the range of soft soil conditions suitable for this fly ash. Soft soil within this range has good compatibility with this fly ash. Soft soil outside this range requires caution or other technical measures.

[0038] In practical engineering applications, for a specific soft soil to be evaluated, it is necessary to take samples on-site and determine the measured moisture content and organic matter content in the laboratory. These two measured values ​​constitute a coordinate point. This coordinate point is then placed in a multidimensional graph, and the coordinate comparison is used to determine whether its location falls within the previously marked compatibility recommendation zone. If the location falls within the recommendation zone, it indicates good compatibility between the soft soil and the fly ash. Under normal fluctuations in the physical properties of the fly ash, the strength of the solidified soil is likely to meet engineering requirements, and it can be used with confidence. If the location falls outside the recommendation zone, it indicates poor compatibility. Using this fly ash to solidify the soft soil may result in insufficient strength. In this case, it is necessary to change the fly ash source, adjust the mix design, or adopt other reinforcement measures.

[0039] By using the above-mentioned landing point discrimination method, compatibility evaluation results for specific soft soil and specific fly ash combinations can be quickly generated, providing intuitive and quantitative decision support for engineering material selection.

[0040] In some implementations, the original set of physical property parameters includes parameters such as loss on ignition of fly ash, water demand ratio, activity index, and fineness modulus. The predicted value of solidified soil performance includes the unconfined compressive strength value. The output prediction result of the suitability of solidified soil corresponding to the fly ash to be evaluated includes the output unconfined compressive strength value.

[0041] The initial set of physical property parameters can specifically include the loss on ignition (LOI) parameter, water requirement ratio parameter, activity index parameter, and fineness modulus parameter of fly ash. These four parameters are the core indicators characterizing the quality and engineering applicability of fly ash. Among them, the LOI parameter refers to the percentage of mass lost by fly ash under high-temperature burning conditions, mainly reflecting the content of unburned carbon in the fly ash. Unburned carbon is a porous, lightweight, inert substance that not only lacks cementitious activity but also adsorbs a large amount of water and admixtures. Therefore, an excessively high LOI parameter will significantly affect the strength development and workability of the solidified soil. The water requirement ratio parameter refers to the ratio of the water requirement for fly ash paste to reach standard consistency to the water requirement for silicate cement to reach the same consistency. It measures the degree of water demand of fly ash. An excessively high water requirement ratio means that under the same water content, the fluidity of the mixture is poor, requiring additional water to meet construction requirements. This leads to increased porosity and decreased density within the solidified soil, thus affecting the final strength. The activity index parameter refers to the ratio of the strength of mortar after mixing fly ash and cement to the strength of pure cement mortar. It reflects the pozzolanic reactivity of fly ash and is a key factor determining the later strength of solidified soil. A higher activity index indicates a higher content of active silica and alumina in the fly ash, and a stronger ability to react with hydration products in soft soil to form cementitious substances. The fineness modulus parameter is usually expressed as the percentage of fly ash residue on a specific sieve. It characterizes the coarseness of fly ash particles. The finer the particle size, the larger the specific surface area of ​​the fly ash, and the more complete the pozzolanic reaction. However, the water demand may also increase accordingly, so a comprehensive balance needs to be made.

[0042] Regarding the predicted performance values ​​of solidified soil, the unconfined compressive strength value reflects the maximum ability of solidified soil to resist axial pressure under unconfined lateral conditions. When outputting the predicted applicability of solidified soil corresponding to the fly ash to be evaluated, the unconfined compressive strength value is output, allowing engineers to intuitively understand the strength level that the soft soil after solidification of the fly ash can achieve. This allows for comparison with the strength value required by the engineering design to determine whether it meets the specifications.

[0043] Example 2 like Figure 2 As shown, in a second aspect, the present invention proposes a fly ash-stabilized soil applicability prediction system based on physical property parameters. The system adopts the fly ash-stabilized soil applicability prediction method based on physical property parameters proposed in any of the above embodiments. The system includes: The original parameter acquisition module is used to perform step S1: acquire the original physical property parameter set of the fly ash to be tested; The disturbance sequence generation and superposition module is used to execute step S2: based on the preset fluctuation range corresponding to each original physical property parameter in the original physical property parameter set, multiple disturbance sequences are generated, and the multiple disturbance sequences are superimposed on the original physical property parameter set to obtain the disturbed parameter set; wherein, S2 includes: S21: Obtain the historical extreme values ​​of fluctuations for each original physical property parameter in the original physical property parameter set, and use them as the preset fluctuation range; S22: Generate a set of initial chaotic time series based on the control equations of the chaotic system; S23: Normalize the initial chaotic time series and scale the amplitude of the initial chaotic time series to match the preset fluctuation range to obtain multiple chaotic disturbance value sequences; S24: Add the multiple chaotic perturbation value sequences to the corresponding original physical property parameters in the original physical property parameter set respectively to generate a perturbed parameter set containing multiple perturbed samples; The model training module is used to execute step S3: based on the original set of physical property parameters and the disturbed set of parameters, iteratively train the initial machine learning model until the initial machine learning model outputs the predicted value of the solidified soil performance and the corresponding stability characterization quantity, thus obtaining the trained machine learning model; wherein, S3 includes: S31: Construct an initial machine learning model that includes a feature extraction layer and a prediction output layer; S32: The original set of physical property parameters is used as a positive sample input to the feature extraction layer, and the disturbed set of parameters is used as a negative sample input to the feature extraction layer. After processing by the feature extraction layer, positive sample feature vectors and negative sample feature vectors are generated respectively. S33: Input the positive sample feature vector and the negative sample feature vector into the prediction output layer. After processing by the prediction output layer, the predicted values ​​of the solidified soil performance of the positive sample and the predicted values ​​of the solidified soil performance of the negative sample are generated respectively. S34: Calculate the dispersion of the predicted performance values ​​of the solidified soil in the negative sample as a stability characterization quantity; S35: With the predicted value of solidified soil performance of positive samples approaching the preset target value and the stability characterization value approaching the minimum as the training objective, the parameters of the initial machine learning model are iteratively updated until the preset convergence condition is met, and the trained machine learning model is obtained. The prediction output module is used to perform step S4: obtain the target physical property parameters of the fly ash to be evaluated, input them into the trained machine learning model, and after processing by the trained machine learning model, output the prediction result of the applicability of the solidified soil corresponding to the fly ash to be evaluated.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the applicability of fly ash-stabilized soil based on physical property parameters, characterized in that, include: S1: Obtain the original set of physical property parameters of the fly ash to be tested; S2: Based on the preset fluctuation range corresponding to each original physical property parameter in the original physical property parameter set, multiple disturbance sequences are generated, and the multiple disturbance sequences are superimposed on the original physical property parameter set to obtain the disturbed parameter set; wherein, S2 includes: S21: Obtain the historical extreme values ​​of fluctuations for each original physical property parameter in the original physical property parameter set, and use them as the preset fluctuation range; S22: Generate a set of initial chaotic time series based on the control equations of the chaotic system; S23: Normalize the initial chaotic time series and scale the amplitude of the initial chaotic time series to match the preset fluctuation range to obtain multiple chaotic disturbance value sequences; S24: Add the multiple chaotic perturbation value sequences to the corresponding original physical property parameters in the original physical property parameter set respectively to generate a perturbed parameter set containing multiple perturbed samples; S3: Based on the original set of physical property parameters and the disturbed set of parameters, iteratively train the initial machine learning model until the initial machine learning model outputs the predicted value of the solidified soil performance and the corresponding stability characterization quantity, thus obtaining the trained machine learning model; wherein, S3 includes: S31: Construct an initial machine learning model that includes a feature extraction layer and a prediction output layer; S32: The original set of physical property parameters is used as a positive sample input to the feature extraction layer, and the disturbed set of parameters is used as a negative sample input to the feature extraction layer. After processing by the feature extraction layer, positive sample feature vectors and negative sample feature vectors are generated respectively. S33: Input the positive sample feature vector and the negative sample feature vector into the prediction output layer. After processing by the prediction output layer, the predicted values ​​of the solidified soil performance of the positive sample and the predicted values ​​of the solidified soil performance of the negative sample are generated respectively. S34: Calculate the dispersion of the predicted performance values ​​of the solidified soil in the negative sample as a stability characterization quantity; S35: With the predicted value of solidified soil performance of positive samples approaching the preset target value and the stability characterization value approaching the minimum as the training objective, the parameters of the initial machine learning model are iteratively updated until the preset convergence condition is met, and the trained machine learning model is obtained. S4: Obtain the target physical property parameters of the fly ash to be evaluated, input them into the trained machine learning model, and after processing by the trained machine learning model, output the prediction results of the applicability of the solidified soil corresponding to the fly ash to be evaluated. The feature extraction layer includes a noise injection sublayer, and S32 includes: S321: Input the sample noise into the sub-layer of the original physical property parameter set. After adding random perturbation through the noise injection sub-layer, generate positive sample intermediate features. Input the positive sample intermediate features into the rest of the feature extraction layer to generate positive sample feature vectors. S322: The samples in the disturbed parameter set are directly input into the rest of the feature extraction layer, bypassing the noise injection sub-layer, to generate negative sample feature vectors.

2. The method for predicting the applicability of fly ash-stabilized soil based on physical property parameters according to claim 1, characterized in that, S22 includes: S221: Select the Lorentz system or logistic mapping as the governing equation for the chaotic system; S222: Set the initial state value and control parameter value of the control equation, iteratively solve the control equation, and generate an initial chaotic time series containing values ​​at multiple time points.

3. The method for predicting the applicability of fly ash-stabilized soil based on physical property parameters according to claim 1, characterized in that, S34 includes: S341: Obtain the predicted values ​​of the negative sample solidified soil properties of multiple disturbed samples corresponding to the same original physical property parameter; S342: Calculate the standard deviation of the predicted performance values ​​of solidified soil from multiple negative samples, and use the standard deviation as a measure of stability.

4. The method for predicting the applicability of fly ash-stabilized soil based on physical property parameters according to claim 1, characterized in that, S4 include: S41: Obtain the target physical property parameters of the fly ash to be evaluated, input them into the trained machine learning model, and after processing by the trained machine learning model, output the predicted values ​​of solidified soil performance in multiple dimensions. S42: The predicted values ​​of solidified soil performance in multiple dimensions are used as coordinate values ​​and mapped to a preset multidimensional coordinate system to generate a multidimensional map characterizing the applicability of solidified soil.

5. The method for predicting the applicability of fly ash-stabilized soil based on physical property parameters according to claim 4, characterized in that, The preset multidimensional coordinate system includes the soft soil moisture content axis, the soft soil organic matter content axis, and the strength guarantee rate axis. S42 includes: S421: Use the target physical property parameters of the fly ash to be evaluated as fixed values; S422: Obtain multiple different soft soil moisture content values ​​and soft soil organic matter content values ​​as variable input values; S423: Input the fixed value and each set of variable input values ​​into the trained machine learning model. After processing by the trained machine learning model, output the predicted value of the solidified soil performance and the corresponding stability characterization quantity for each set of inputs. S424: Convert the stability characterization quantity corresponding to each set of inputs into the strength guarantee rate, and generate a multidimensional map with the soft soil moisture content as the first axis coordinate, the soft soil organic matter content as the second axis coordinate, and the strength guarantee rate as the third axis coordinate.

6. The method for predicting the applicability of fly ash-stabilized soil based on physical property parameters according to claim 5, characterized in that, The stability characteristics corresponding to each set of inputs are converted into strength guarantee rates, including: Obtain the preset engineering strength threshold; The probability value that the predicted value of solidified soil performance is not lower than the engineering strength threshold under the disturbance conditions of the disturbed parameter set is statistically analyzed, and the probability value is used as the strength guarantee rate.

7. The method for predicting the applicability of fly ash-stabilized soil based on physical property parameters according to claim 5, characterized in that, The method also includes: Identify all coordinate points in the multidimensional map whose strength guarantee rate is greater than the preset guarantee rate threshold, and mark the area covered by all coordinate points as the compatibility recommendation area; Obtain the measured moisture content and measured organic matter content of the soft soil to be evaluated; Determine whether the coordinate points formed by the measured moisture content and measured organic matter content fall within the compatibility recommendation area in the multidimensional map, and generate a compatibility evaluation result.

8. The method for predicting the applicability of fly ash-stabilized soil based on physical property parameters according to claim 1, characterized in that, The original set of physical property parameters includes parameters such as loss on ignition, water requirement ratio, activity index, and fineness modulus of fly ash. The predicted value of solidified soil performance includes the unconfined compressive strength value. The output prediction result of the applicability of solidified soil corresponding to the fly ash to be evaluated includes the output unconfined compressive strength value.

9. A fly ash-stabilized soil applicability prediction system based on physical property parameters, characterized in that, The system employs the fly ash-stabilized soil applicability prediction method based on physical property parameters as described in any one of claims 1 to 8, and the system comprises: The original parameter acquisition module is used to perform step S1: acquire the original physical property parameter set of the fly ash to be tested; The disturbance sequence generation and superposition module is used to execute step S2: based on the preset fluctuation range corresponding to each original physical property parameter in the original physical property parameter set, multiple disturbance sequences are generated, and the multiple disturbance sequences are superimposed on the original physical property parameter set to obtain the disturbed parameter set; wherein, S2 includes: S21: Obtain the historical extreme values ​​of fluctuations for each original physical property parameter in the original physical property parameter set, and use them as the preset fluctuation range; S22: Generate a set of initial chaotic time series based on the control equations of the chaotic system; S23: Normalize the initial chaotic time series and scale the amplitude of the initial chaotic time series to match the preset fluctuation range to obtain multiple chaotic disturbance value sequences; S24: Add the multiple chaotic perturbation value sequences to the corresponding original physical property parameters in the original physical property parameter set respectively to generate a perturbed parameter set containing multiple perturbed samples; The model training module is used to execute step S3: based on the original set of physical property parameters and the disturbed set of parameters, iteratively train the initial machine learning model until the initial machine learning model outputs the predicted value of the solidified soil performance and the corresponding stability characterization quantity, thus obtaining the trained machine learning model; wherein, S3 includes: S31: Construct an initial machine learning model that includes a feature extraction layer and a prediction output layer; S32: The original set of physical property parameters is used as a positive sample input to the feature extraction layer, and the disturbed set of parameters is used as a negative sample input to the feature extraction layer. After processing by the feature extraction layer, positive sample feature vectors and negative sample feature vectors are generated respectively. S33: Input the positive sample feature vector and the negative sample feature vector into the prediction output layer. After processing by the prediction output layer, the predicted values ​​of the solidified soil performance of the positive sample and the predicted values ​​of the solidified soil performance of the negative sample are generated respectively. S34: Calculate the dispersion of the predicted performance values ​​of the solidified soil in the negative sample as a stability characterization quantity; S35: With the predicted value of solidified soil performance of positive samples approaching the preset target value and the stability characterization value approaching the minimum as the training objective, the parameters of the initial machine learning model are iteratively updated until the preset convergence condition is met, and the trained machine learning model is obtained. The prediction output module is used to perform step S4: obtain the target physical property parameters of the fly ash to be evaluated, input them into the trained machine learning model, and after processing by the trained machine learning model, output the prediction result of the applicability of the solidified soil corresponding to the fly ash to be evaluated. The feature extraction layer includes a noise injection sublayer, and S32 includes: S321: Input the sample noise into the sub-layer of the original physical property parameter set. After adding random perturbation through the noise injection sub-layer, generate positive sample intermediate features. Input the positive sample intermediate features into the rest of the feature extraction layer to generate positive sample feature vectors. S322: The samples in the disturbed parameter set are directly input into the rest of the feature extraction layer, bypassing the noise injection sub-layer, to generate negative sample feature vectors.

Images