Soil conditioner preparation method based on component detection

Abstract

The invention relates to the technical field of cross-modal data fusion analysis of soil component detection and improvement agent preparation, in particular to a preparation method of a soil improvement agent based on component detection, which comprises the following steps: acquiring soil component data through a near infrared spectrometer, a high performance liquid chromatograph and an inductively coupled plasma mass spectrometer; an incremental principal component analysis algorithm is combined with an oscillation suppression function to update a covariance matrix in real time, a dynamic defect factor priority list is generated, a multi-objective optimization model improved based on NSGA-II is guided to integrate the soil pH value, humidity and organic matter content to construct a dose response curved surface, and a Pareto optimal solution set is solved. A sensor array collects environment feedback data, drives a transfer learning algorithm to calibrate parameters, a dynamic incidence matrix attenuation rate and optimal weight adaptive adjustment, iteratively outputs a modifier synergistic effect solution set through a closed-loop control mechanism, and solves the problems of difficult data fusion, principal component weight offset and component synergistic deviation. And the proportioning accuracy and stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cross-modal data fusion analysis of soil component detection and improver preparation, and in particular to a method for preparing a soil improver based on component detection. Background Art

[0002] Existing methods for preparing soil conditioners based on component testing involve quantitatively characterizing the physical and chemical properties of target soil samples through spectral analysis, ion chromatography, and mass spectrometry, including core indicators such as organic matter content, cation exchange capacity, and heavy metal pollutant concentrations. By establishing a soil component database and combining it with a principal component analysis algorithm, key defect factors that limit soil fertility or ecological function can be identified, allowing the design of appropriate conditioner formulations, such as humic acid complexes for organic matter deficiency or silicate passivation materials for heavy metal contamination. During the preparation process, the raw material ratio and reaction conditions are dynamically adjusted based on the test data, and the effectiveness of the conditioner is verified through in vitro culture experiments simulating soil environments.

[0003] The heterogeneity of multi-source detection data (such as spectra, chromatography, and mass spectrometry) in existing methods makes cross-platform data fusion difficult, leading to feature weight shifts in the principal component analysis algorithm. This is specifically manifested in the inability to directly couple the modeling of organic matter near-infrared spectral absorption peaks and heavy metal X-ray fluorescence intensity values ​​due to dimensional differences, resulting in an increase in the error rate in identifying key defect factors. For example, in the scenario of cadmium-contaminated farmland remediation, there is an order of magnitude mismatch between the cadmium ion concentration detected by inductively coupled plasma mass spectrometry and the humic acid functional group data obtained by Fourier transform infrared spectroscopy. If directly input into the principal component analysis model, it will lead to calculation deviations in the synergistic effect of silicate passivation and humic acid addition, which will manifest as excessive passivation causing soil compaction or insufficient humic acid leading to cadmium ion reactivation. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method for preparing a soil conditioner based on component detection. The present invention solves the technical problem that in the existing soil conditioner preparation process based on multi-source detection data, the heterogeneity of spectral, chromatographic and mass spectrometric data leads to difficulties in cross-platform data fusion, which in turn causes the characteristic weight offset of the principal component analysis model, resulting in errors in the identification of key defect factors and deviations in the calculation of the synergistic effect of the conditioner components.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows: The present invention provides a method for preparing a soil conditioner based on component detection, comprising: Receiving soil composition detection data collected by a spectral device, a chromatographic device, and a mass spectrometric device, wherein the spectral device is a near-infrared spectrometer, the chromatographic device is a high-performance liquid chromatograph, and the mass spectrometric device is an inductively coupled plasma mass spectrometer; Performing cross-modal normalization on the soil composition detection data, including calibrating multi-device dimensional differences through a transfer learning algorithm based on adversarial domain adaptation, and generating dimensionless fused data using a combination of range normalization and Z-score algorithm; Based on the dimensionless fusion data, a time-sensitive feature association model is constructed within a sliding time window by introducing a random forest algorithm with a time decay factor, and dynamic feature weights are output; According to the dynamic feature weights, an incremental principal component analysis algorithm is used to adjust the index weight coefficients of the principal component analysis model in real time to generate a dynamic defect factor ranking; Based on the dynamic defect factor ranking, a dose response surface model of the humic acid and silicate composite material is constructed by using a multi-objective optimization algorithm improved by NSGA-II, and environmental parameters are integrated to generate a Pareto optimal solution set, wherein the environmental parameters include soil pH, soil moisture and soil organic matter content; Applying a preset amendment component in a preset simulated soil environment, and collecting environmental parameter feedback data through a sensor array, wherein the sensor array includes a pH sensor, a humidity sensor, and a conductivity sensor; The calibration parameters of the transfer learning algorithm, the time attenuation factor of the timing-sensitive feature association model, and the objective function weights of the multi-objective optimization algorithm are iteratively updated according to the environmental parameter feedback data to optimize the improver components prepared by the soil conditioner, wherein the iterative update process completes the parameter adaptive adjustment through the gradient descent algorithm.

[0006] Furthermore, in the method for preparing a soil conditioner based on component detection of the present invention, the cross-modal standardization process includes: Using the preset heavy metal standard solution as the reference substance, the equivalent conversion relationship between spectral absorbance and mass spectrometry concentration is established through the standard substance gradient experiment; Based on the equivalent conversion relationship, a transfer learning algorithm based on adversarial domain adaptation is used to perform dimension calibration on the detection data collected by spectral, chromatographic and mass spectrometric equipment to generate calibration data with unified dimensions across devices, wherein the domain adaptation weight of the transfer learning algorithm is dynamically adjusted through the KL divergence loss function; The calibration data are input into a combination algorithm of range normalization and Z-score, and are mapped to a dimensionless space after eliminating dimension differences through linear transformation to generate dimensionless fusion data of spectrum, chromatography and mass spectrometry data.

[0007] Furthermore, in the method for preparing a soil conditioner based on component detection of the present invention, the construction of the time-series sensitive feature association model includes: Based on the dimensionless fusion data, a random forest algorithm with a time decay factor is introduced within a sliding time window to calculate the correlation between the characteristic absorption peaks of humic acid and silicate in the near-infrared spectrum and the heavy metal concentration, and output a dynamic correlation parameter; Inputting the dynamic correlation parameters into a matrix decomposition algorithm, constructing a dynamic correlation matrix based on the data distribution changes in the time window, and extracting the dynamic feature weights corresponding to the spectral absorption peaks, wherein the rank of the dynamic correlation matrix is ​​consistent with the number of heavy metal species; The dynamic correlation matrix is ​​used as input, and the signal response deviation threshold of the spectral and mass spectrometric data is iteratively adjusted through a cross-validation mechanism.

[0008] Furthermore, in the method for preparing a soil conditioner based on component detection of the present invention, the generation of the dynamic defect factor ranking includes: Based on the dynamic feature weights, an incremental principal component analysis algorithm is integrated to calculate a real-time updated value of the covariance matrix; Applying a model oscillation suppression function based on exponential smoothing to the real-time update value to adjust the interference of detection batch differences on the weight distribution coefficient in the covariance matrix, wherein the exponential smoothing coefficient is positively correlated with the rank of the dynamic correlation matrix; The adjusted covariance matrix is ​​input into a principal component contribution ranking algorithm based on the Kaiser criterion, and a priority list of defect factors evolving over time is output. The priority list of defect factors is arranged in descending order of contribution of heavy metals such as cadmium, lead, and arsenic.

[0009] Furthermore, in the method for preparing a soil conditioner based on component detection of the present invention, the multi-objective optimization algorithm includes: Based on the defect factor priority list and with the target heavy metal passivation requirement as a constraint, a dose response surface model of humic acid and silicate composite material was constructed; The soil pH value, soil moisture and soil organic matter content are input into the dose-response surface model, and the output results of the principal component analysis model are combined to calculate the interaction relationship between soil porosity and passivation efficiency; By dynamically adjusting the weight coefficient of the interaction relationship, the Pareto optimal solution of the passivation efficiency and soil porosity is solved.

[0010] Furthermore, the method for preparing a soil conditioner based on component detection of the present invention includes applying a preset conditioner component in a preset simulated soil environment and collecting environmental parameter feedback data through a sensor array, which includes: The real-time electrochemical parameters in the soil solution are collected by the pH sensor, humidity sensor and conductivity sensor; Inputting the real-time electrochemical parameters into a heavy metal ion morphological kinetic model to calculate a time-varying curve of the complexation efficiency of the humic acid and silicate composite material on the heavy metal ions of cadmium, lead and arsenic; Based on the preset passivation efficiency threshold deviation between the time-varying curve and the Pareto optimal solution, the objective function weight adjustment amount of the multi-objective optimization algorithm is calculated by a logic controller, and a feedback signal is output to the optimization algorithm parameters.

[0011] Furthermore, in the method for preparing a soil conditioner based on component detection of the present invention, the components of the soil conditioner prepared to optimize the soil conditioner include: The output feedback signal is synchronously input into the transfer learning algorithm to update the dimension calibration parameters of the spectral, chromatographic and mass spectrometric equipment; Inputting the feedback signal into the dynamic correlation matrix, and adjusting the decay rate of the time decay factor to match the real-time data distribution; updating the parameters of the multi-objective optimization algorithm based on the objective function weight adjustment amount in the feedback signal; Recalculating the covariance matrix update period of the incremental principal component analysis algorithm according to the updated dimension calibration parameters, time decay factor and optimization algorithm parameters; Based on the recalculated covariance matrix, the principal component contribution ranking algorithm is used to iteratively output the synergistic effect solution set of the improver components.

[0012] Beneficial effects of the present invention: The present invention uses a transfer learning algorithm to calibrate the dimensional differences of spectral, chromatographic and mass spectrometric equipment, combines range normalization with a Z-score combination algorithm to generate cross-modal consistent dimensionless fusion data, and eliminates the fusion difficulties caused by heterogeneity of multi-source detection data. The time-sensitive feature association model introduces a time decay factor and a random forest algorithm to dynamically capture the correlation pattern between spectral absorption peaks and heavy metal concentrations, and combines the non-negative matrix decomposition algorithm to extract dynamic feature weights, thereby reducing feature offsets caused by differences in detection principles. The incremental principal component analysis algorithm updates the covariance matrix in real time and suppresses batch noise interference, generates a dynamic defect factor priority list, guides the multi-objective optimization model to construct a dose-response surface, and integrates environmental parameters to solve the Pareto optimal solution set. The closed-loop feedback mechanism collects real-time electrochemical parameters through a sensor array, drives the transfer learning algorithm, the dynamic correlation matrix and the optimization algorithm parameters to adaptively adjust, realizes iterative optimization of the synergistic solution set of the improver components, improves the accuracy and stability of the soil improver ratio, and overcomes the technical defects of the principal component weight offset and component synergistic calculation deviation in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0014] Figure 1 A flow chart of a method for preparing a soil conditioner based on component detection provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0016] See also Figure 1 The present invention provides a method for preparing a soil conditioner based on component detection, comprising: Step 1: receiving soil composition detection data collected by a spectral device, a chromatographic device, and a mass spectrometric device, wherein the spectral device is a near-infrared spectrometer, the chromatographic device is a high-performance liquid chromatograph, and the mass spectrometric device is an inductively coupled plasma mass spectrometer; Step 2: performing cross-modal normalization on the soil composition detection data, including calibrating multi-device dimensional differences through a transfer learning algorithm based on adversarial domain adaptation, and generating dimensionless fused data using a combination of range normalization and Z-score algorithm; Step 3: Based on the dimensionless fusion data, a time-sensitive feature association model is constructed within a sliding time window by using a random forest algorithm with a time decay factor, and dynamic feature weights are output; Step 4: Based on the dynamic feature weights, an incremental principal component analysis algorithm is used to adjust the index weight coefficients of the principal component analysis model in real time to generate a dynamic defect factor ranking; Step 5, based on the dynamic defect factor ranking, constructing a dose response surface model of the humic acid and silicate composite material by using a multi-objective optimization algorithm improved by NSGA-II, and integrating environmental parameters to generate a Pareto optimal solution set, wherein the environmental parameters include soil pH, soil moisture, and soil organic matter content; Step 6: applying a preset modifier component in a preset simulated soil environment, and collecting environmental parameter feedback data through a sensor array, wherein the sensor array includes a pH sensor, a humidity sensor, and a conductivity sensor; Step 7, iteratively updating the calibration parameters of the transfer learning algorithm, the time attenuation factor of the time-sensitive feature association model, and the objective function weight of the multi-objective optimization algorithm based on the environmental parameter feedback data to optimize the improver components prepared by the soil conditioner, wherein the iterative update process completes the parameter adaptive adjustment through the gradient descent algorithm.

[0017] In the method for preparing a soil conditioner based on component detection provided by the present invention, soil component detection data are first collected separately by a spectral device, a chromatographic device and a mass spectrometric device, wherein a near-infrared spectrometer is used to obtain characteristic absorption peaks of organic matter, a high-performance liquid chromatograph is used to separate and detect components related to cation exchange capacity, and an inductively coupled plasma mass spectrometer is used to quantitatively analyze the concentration of heavy metal pollutants. The data output by different detection devices are heterogeneous due to differences in dimensions and detection principles, and need to be standardized through cross-modal processing to unify the data space. In the standardization process, a preset heavy metal standard solution is used as a reference substance, and an equivalent conversion relationship between spectral absorbance and mass spectrum concentration is established through a gradient experiment. Based on this relationship, the domain adaptation layer of the transfer learning algorithm is used to perform dimension calibration on the multi-source data to generate calibration data with unified dimensions across devices. The calibrated data is further linearly transformed by a range standardization algorithm to eliminate the difference in numerical span, and then normalized to a standard normal distribution by a Z-score algorithm to generate dimensionless fusion data of spectral, chromatographic and mass spectrometric data, providing an input basis for subsequent time series modeling.

[0018] Based on dimensionless fused data, a time decay factor is introduced within the sliding time window to adjust the data weight decay rate, and a time-sensitive feature association model is constructed in combination with the random forest algorithm. The sliding time window dynamically adjusts the window length according to the data acquisition frequency, and the time decay factor calculates the weight decay rate based on the data timestamp interval, giving more recent data a higher weight in the association analysis. The random forest algorithm uses multiple decision trees to perform regression analysis on the spectral absorption peak wavelength, peak area, and heavy metal concentration, outputting a set of parameters that characterize the dynamic correlation strength. The dynamic correlation parameters are input into the non-negative matrix factorization algorithm and decomposed into the product of a basis matrix and a coefficient matrix. The basis matrix represents the basic correlation pattern between the spectral absorption peak and the heavy metal concentration, and the coefficient matrix reflects the changes in correlation strength under different time windows. The dynamic feature weights corresponding to the spectral absorption peak are extracted.

[0019] Dynamic feature weights are fed into an incremental principal component analysis algorithm to adjust the index weight coefficients of the principal component analysis model in real time. The incremental algorithm uses a sliding window mechanism to process dynamic data streams, updating the covariance matrix batch by batch. This algorithm, combined with an exponential smoothing-based model oscillation suppression function, adjusts for interference from batch differences in the weight distribution coefficients. The oscillation suppression function sets the smoothing strength based on the statistical variance of the covariance matrix within the sliding window to suppress sudden changes in weights caused by equipment calibration errors or environmental fluctuations. The adjusted covariance matrix is ​​fed into a principal component contribution ranking algorithm based on the Kaiser criterion. The variance explanation rate is extracted through eigenvalue decomposition, and combined with the dynamic feature weights, a priority list of defect factors that evolves over time is generated to quantify the defect level of the components corresponding to each spectral absorption peak.

[0020] Based on dynamic defect factor sorting, a dose-response surface model of humic acid and silicate composites was constructed. With heavy metal passivation requirements as constraints, soil pH, moisture, and organic matter content were integrated as environmental parameters into a multi-physics coupling model. This multi-physics coupling model simulated the diffusion, adsorption, and ion exchange processes of the composite in soil solution, calculated the effect of porosity on the diffusion efficiency of the passivator, and established an interaction matrix between porosity and passivation efficiency. A constraint relaxation strategy was used to dynamically adjust the weight coefficients of the interaction relationships. A non-dominated sorting genetic algorithm was employed to determine the Pareto optimal solution set for passivation efficiency and porosity, resulting in a formulation that balances chemical remediation with physical structural protection.

[0021] After applying the amendment components in a simulated soil environment, an array of pH, humidity, and conductivity sensors collected real-time electrochemical parameters from the soil solution. These electrochemical parameters were input into a cadmium ion speciation kinetic model. Based on the chemical thermodynamic equilibrium equation and the kinetic rate equation, the complexation efficiency of cadmium ions between the humic acid carboxyl functional groups and the silicate interlayer domains was calculated, generating a time-varying curve of the complexation efficiency. The deviation between the time-varying curve and a preset passivation efficiency threshold was input into a fuzzy logic controller, activating the corresponding membership function and rule base, and outputting the weight adjustment for the objective function of the multi-objective optimization algorithm.

[0022] The feedback signal is synchronously input into the transfer learning algorithm's dimensional calibration model. Backpropagation is used to update the calibration parameters of the spectral, chromatographic, and mass spectrometric instruments, eliminating data drift caused by equipment aging or environmental interference. Simultaneously, the decay rate of the time decay factor in the dynamic correlation matrix is ​​adjusted to match the real-time data distribution characteristics. The updated dimensional calibration parameters, time decay factor, and optimization algorithm parameters are used to recalculate the covariance matrix update cycle of the incremental principal component analysis algorithm. The principal component contribution ranking algorithm is then used to iteratively output the synergistic solution set of the modifier components, forming a closed-loop control chain of detection, optimization, and feedback.

[0023] Specifically, in the method for preparing a soil conditioner based on component detection according to the present invention, the cross-modal standardization process includes: Using the preset heavy metal standard solution as the reference substance, the equivalent conversion relationship between spectral absorbance and mass spectrometry concentration is established through the standard substance gradient experiment; Based on the equivalent conversion relationship, a transfer learning algorithm based on adversarial domain adaptation is used to perform dimension calibration on the detection data collected by spectral, chromatographic and mass spectrometric equipment to generate calibration data with unified dimensions across devices, wherein the domain adaptation weight of the transfer learning algorithm is dynamically adjusted through the KL divergence loss function; The calibration data are input into a combination algorithm of range normalization and Z-score, and are mapped to a dimensionless space after eliminating dimension differences through linear transformation to generate dimensionless fusion data of spectrum, chromatography and mass spectrometry data.

[0024] In the cross-modal standardization processing step described in the present invention, a preset heavy metal standard solution is first used as a reference substance, and an equivalent conversion relationship between spectral absorbance and mass spectrum concentration is established through a standard substance gradient experiment. The standard substance gradient experiment uses heavy metal contaminated soil samples containing different concentration gradients, and synchronously collects near-infrared spectral absorption curves and inductively coupled plasma mass spectrometry ion current intensity data under the same environmental conditions. The linear regression equation of the spectral characteristic peak area and mass spectrum concentration is fitted by the least squares method to determine the equivalent conversion coefficient between the two. This step establishes the physical quantity mapping relationship between spectral absorbance and mass spectrum concentration through experimental calibration, providing a benchmark reference for subsequent cross-device data calibration.

[0025] Based on the equivalent conversion relationship, a transfer learning algorithm based on adversarial domain adaptation is used to perform dimension calibration on the detection data collected by spectral, chromatographic and mass spectrometric equipment. The pre-training model of the transfer learning algorithm is constructed based on a historical cross-device data set. The equivalent conversion coefficient is embedded in the model parameters through the fine-tuning layer, and the spectral absorbance, chromatographic retention time and mass spectrometric fragment ion intensity collected in real time are mapped across devices. The domain adaptation weight is dynamically adjusted through the KL divergence loss function to calculate the distribution difference between the source domain (historical data) and the target domain (real-time data), optimize the domain adaptability of the transfer learning model, and generate calibration data with unified dimensions across devices. This step uses the feature alignment mechanism of transfer learning to eliminate the magnitude offset caused by sensitivity differences between different detection devices, making heterogeneous data comparable.

[0026] The calibrated data is input into the range normalization and Z-score combination algorithm for dimensionless processing. The range normalization algorithm performs a linear transformation on the calibration data, compressing the numerical range of each modal data to the interval [0,1], eliminating the impact of the numerical span caused by differences in the dynamic range of the detection equipment. The Z-score algorithm further calculates the mean and standard deviation of the standardized data, and through normalization operations, makes the data of different modalities obey the standard normal distribution. The combined application of the two algorithms takes into account the preservation of data distribution morphology and cross-modal consistency adjustment, and finally generates dimensionless fusion data of spectral, chromatographic and mass spectrometric data. The fused data serves as the input for subsequent time series modeling, providing a cross-modal consistent data foundation for the calculation of dynamic feature weights.

[0027] The above steps, through a progressive process of experimental calibration, transfer learning calibration, and combined algorithm normalization, form a closed-loop technology loop from physical quantity correlation mapping to data space unification. Reference material gradient experiments provide a physical conversion benchmark for cross-modal data, transfer learning algorithms enable dynamic alignment of dimensions between devices, and a combined range normalization and Z-score algorithm further eliminates differences in numerical distributions. The resulting dimensionless fused data provides standardized input for subsequent time-sensitive feature modeling, addressing the difficulties associated with fusing multi-source detection data due to heterogeneity.

[0028] Specifically, in the method for preparing a soil conditioner based on component detection of the present invention, the construction of the time-series sensitive feature association model includes: Based on the dimensionless fusion data, a random forest algorithm with a time decay factor is introduced within a sliding time window to calculate the correlation between the characteristic absorption peaks of humic acid and silicate in the near-infrared spectrum and the heavy metal concentration, and output a dynamic correlation parameter; Inputting the dynamic correlation parameters into a matrix decomposition algorithm, constructing a dynamic correlation matrix based on the data distribution changes in the time window, and extracting the dynamic feature weights corresponding to the spectral absorption peaks, wherein the rank of the dynamic correlation matrix is ​​consistent with the number of heavy metal species; The dynamic correlation matrix is ​​used as input, and the signal response deviation threshold of the spectral and mass spectrometric data is iteratively adjusted through a cross-validation mechanism.

[0029] In the construction steps of the time-sensitive feature association model described in the present invention, first, based on the dimensionless fusion data generated by cross-modal normalization processing, the correlation between the characteristic absorption peaks of humic acid and silicate in the near-infrared spectrum and the heavy metal concentration is calculated by introducing a random forest algorithm with a time decay factor within a sliding time window. The length of the sliding time window is dynamically adjusted according to the data acquisition frequency. For example, when data is collected every five minutes, the window length is set to thirty minutes, covering six groups of continuous data points. The time decay factor calculates the weight decay rate based on the data timestamp interval, so that recent data has a higher weight in the association analysis, thereby reflecting the real-time change trend of soil composition. The random forest algorithm performs regression analysis on the spectral absorption peak wavelength, peak area and heavy metal concentration through multiple decision trees, and outputs a parameter set that characterizes the dynamic correlation strength, providing input for subsequent matrix decomposition.

[0030] The dynamic correlation parameters are input into the non-negative matrix factorization algorithm, and a dynamic correlation matrix is ​​constructed based on the changes in data distribution within the time window. Non-negative matrix factorization decomposes the parameter set into the product of a basis matrix and a coefficient matrix. The basis matrix represents the basic correlation pattern between humic acid absorption peaks and heavy metal concentrations, and the coefficient matrix reflects the changes in correlation strength under different time windows. The rank of the dynamic correlation matrix is ​​set to be consistent with the number of heavy metal types. For example, when detecting cadmium, lead, and arsenic, the matrix rank is three. This ensures that the dimensions of the decomposed basis matrix match the types of pollutants and ensures the physical meaning of feature extraction is clear. The dynamic feature weights corresponding to the spectral absorption peaks are extracted through the decomposed basis matrix, and the contribution of each spectral feature to the prediction of heavy metal concentration is quantified, providing data support for the ranking of defect factors.

[0031] Using the dynamic correlation matrix as input, a cross-validation mechanism iteratively adjusts the signal response deviation threshold for spectral and mass spectrometric data. Cross-validation employs a leave-one-out strategy, dividing the dataset into a training set and a validation set. By alternately masking portions of the spectral or mass spectrometric data, the stability of the model's feature weights in the absence of a signal is evaluated. Based on the validation results, the deviation matching threshold between the spectral absorption peak response intensity and the mass spectrometric ion current signal is dynamically adjusted. For example, the initial threshold is set to ±5%, which is optimized to within ±3% after validation to reduce the impact of detection principle deviations on the dynamic feature weights due to spectral scattering effects or differences in mass spectrometric ionization efficiency. The adjusted threshold is re-input into the matrix decomposition algorithm to generate updated dynamic feature weights, forming an iterative optimization closed loop for modeling and validation.

[0032] The above steps establish a dynamic mapping relationship between spectral absorption peaks and heavy metal concentrations through the synergistic effect of time-series weight decay, matrix factorization correlation modeling, and cross-validation bias calibration. The sliding time window captures temporal variations in data distribution, the time decay factor enhances the analytical weight of recent data, non-negative matrix factorization extracts interpretable feature correlation patterns, and the cross-validation mechanism suppresses noise introduced by differences in detection principles. This technical solution forms a complete logical chain from dynamic data processing to model robustness optimization, providing reliable dynamic feature input for defect factor identification and multi-objective optimization.

[0033] Specifically, in the method for preparing a soil conditioner based on component detection of the present invention, the generation of the dynamic defect factor ranking includes: Based on the dynamic feature weights, an incremental principal component analysis algorithm is integrated to calculate a real-time updated value of the covariance matrix; Applying a model oscillation suppression function based on exponential smoothing to the real-time update value to adjust the interference of detection batch differences on the weight distribution coefficient in the covariance matrix, wherein the exponential smoothing coefficient is positively correlated with the rank of the dynamic correlation matrix; The adjusted covariance matrix is ​​input into a principal component contribution ranking algorithm based on the Kaiser criterion, and a priority list of defect factors evolving over time is output. The priority list of defect factors is arranged in descending order of contribution of heavy metals such as cadmium, lead, and arsenic.

[0034] In the dynamic defect factor ranking generation step described in the present invention, an incremental principal component analysis algorithm is first integrated to calculate the real-time updated value of the covariance matrix based on the dynamic feature weights output by the time-sensitive feature association model. The incremental principal component analysis algorithm uses a sliding window mechanism to process dynamic data streams, updating the covariance matrix batch by batch based on real-time soil composition detection data, capturing the temporal variation characteristics of the data distribution. The dynamic feature weights are used as input variables, and the correlation strength between spectral absorption peaks and heavy metal concentrations is mapped to the weight space of the principal component analysis model through weighted covariance calculation, providing a dynamic data foundation for defect factor identification.

[0035] A model oscillation suppression function based on exponential smoothing is applied to the real-time updated covariance matrix to adjust the interference of test batch differences on the weight allocation coefficient. The model oscillation suppression function sets the smoothing coefficient based on the statistical variance value of the covariance matrix within the sliding window. For example, when the variance exceeds the preset threshold, the smoothing coefficient is adjusted to 0.8 to enhance the smoothing strength. The exponential smoothing coefficient is positively correlated with the rank of the dynamic correlation matrix, which is equal to the number of target heavy metal types. For example, when testing cadmium, lead, and arsenic, the matrix rank is three, and the corresponding smoothing coefficient is scaled proportionally to suppress weight mutations caused by equipment calibration errors or fluctuations in environmental parameters, while retaining effective signals of dynamic changes in soil composition.

[0036] The adjusted covariance matrix is ​​input into the principal component contribution ranking algorithm based on the Kaiser criterion, and the principal component contributions of the covariance matrix are extracted through eigenvalue decomposition. The principal component contributions are calculated based on the product of the variance explanation rate and the dynamic feature weight, quantifying the defect level of the components corresponding to each spectral absorption peak in the soil improvement needs. The defect factor priority list is arranged in descending order of the contribution of heavy metals such as cadmium, lead, and arsenic. For example, when cadmium ion concentration contributes most to soil pollution, the weighted principal component contribution of its corresponding spectral absorption peak is the largest and ranks first in the list. As the time window slides, the principal component contributions are dynamically adjusted based on the real-time updated covariance matrix to generate a priority list that reflects the current soil environmental defect characteristics.

[0037] The above steps, through a progressive process of incremental covariance updates, oscillation suppression smoothing, and dynamic principal component sorting, form a closed-loop technology from dynamic data capture to quantitative output of defect factors. The incremental algorithm ensures the real-time performance of the covariance matrix, the oscillation suppression function reduces batch noise interference, and the Kaiser criterion sorting algorithm transforms statistical features into executable remediation priority decisions, providing dynamic constraints for subsequent multi-objective optimization algorithms. This overcomes the technical limitation of traditional static principal component analysis models, which cannot adapt to the temporal changes in soil composition.

[0038] Specifically, the method for preparing a soil conditioner based on component detection of the present invention, the multi-objective optimization algorithm includes: Based on the defect factor priority list and with the target heavy metal passivation requirement as a constraint, a dose response surface model of humic acid and silicate composite material was constructed; The soil pH value, soil moisture and soil organic matter content are input into the dose-response surface model, and the output results of the principal component analysis model are combined to calculate the interaction relationship between soil porosity and passivation efficiency; By dynamically adjusting the weight coefficient of the interaction relationship, the Pareto optimal solution of the passivation efficiency and soil porosity is solved.

[0039] In the multi-objective optimization algorithm steps described in the present invention, first, based on the priority list generated by dynamic defect factor sorting, a dose response surface model of a humic acid and silicate composite material is constructed with the target heavy metal passivation requirement as a constraint condition. The model obtains the binding efficiency data of humic acid functional groups and silicate lattice structures under different dosage ratios through calibration experiments, sets the constraint boundary in combination with the heavy metal ion concentration threshold, for example, the cadmium ion concentration threshold ≤0.1mg / kg is used as the constraint upper limit, and establishes a nonlinear response surface in which the passivation efficiency changes with the dose. The model input parameters include the amount of humic acid added, the silicate particle size and the reaction time, and the output parameter is the correlation curve between the heavy metal passivation efficiency and the soil porosity, providing a dose-effect mapping relationship for multi-objective optimization.

[0040] When soil pH, moisture, and organic matter content were input as environmental parameters into the dose-response surface model, the dynamic defect factor weight coefficients output by the principal component analysis model were integrated to calculate the interaction relationship between soil porosity and passivation efficiency. The multi-physics coupling model simulated the diffusion, adsorption, and ion exchange processes of the composite material in the soil solution. The passivation agent migration rate and reaction kinetic constant were calculated in combination with the porosity parameter to quantify the inhibitory effect of porosity on the passivation agent diffusion efficiency. The interaction relationship matrix established a negative correlation between porosity and passivation efficiency through regression analysis. For example, a 10% decrease in porosity resulted in a 5% decrease in passivation efficiency, forming a quantitative association between physical structure and chemical remediation efficiency.

[0041] By dynamically adjusting the weight coefficients of the interaction relationships, a non-dominated sorting genetic algorithm is used to find the Pareto optimal solution for passivation efficiency and porosity. Weight coefficients are dynamically assigned based on real-time soil environmental parameters. For example, when soil moisture exceeds a critical value, the passivation efficiency weight coefficient is increased to 0.7, while the porosity weight coefficient is reduced to 0.3. A constraint relaxation strategy is used to gradually relax the limits of non-critical constraints. During the algorithm iteration process, the objective function space is redefined based on the real-time updated weight coefficients. A set of amendment ratios that meet the passivation efficiency threshold and minimize porosity loss is output, forming a multi-objective optimization decision that balances chemical remediation and soil structure protection.

[0042] The above steps establish a logical closed loop from defect factor prioritization to formulation generation through the synergistic effect of dose-response modeling, multi-physics coupling analysis, and dynamic weight optimization. The dose-response surface model transforms dynamic defect factors into dose constraint boundaries, the multi-physics coupling model quantifies the impact of environmental parameters on remediation effectiveness, and the dynamic weight adjustment mechanism balances conflicting optimization objectives. The resulting Pareto-optimal solution set provides an executable basis for formulation decision-making in soil amendment preparation.

[0043] Specifically, the method for preparing a soil conditioner based on component detection of the present invention includes applying a preset conditioner component in a preset simulated soil environment and collecting environmental parameter feedback data through a sensor array, including: The real-time electrochemical parameters in the soil solution are collected by the pH sensor, humidity sensor and conductivity sensor; Inputting the real-time electrochemical parameters into a heavy metal ion morphological kinetic model to calculate a time-varying curve of the complexation efficiency of the humic acid and silicate composite material on the heavy metal ions of cadmium, lead and arsenic; Based on the preset passivation efficiency threshold deviation between the time-varying curve and the Pareto optimal solution, the objective function weight adjustment amount of the multi-objective optimization algorithm is calculated by a logic controller, and a feedback signal is output to the optimization algorithm parameters.

[0044] In the simulated soil environment feedback data collection and optimized parameter adjustment steps described in the present invention, real-time electrochemical parameters in the soil solution are first collected using an array of pH sensors, humidity sensors, and conductivity sensors. The pH sensor monitors changes in hydrogen ion concentration during the reaction between humic acid and silicate, the humidity sensor detects the moisture distribution gradient along the modifier diffusion path, and the conductivity sensor quantifies fluctuations in solution conductivity caused by heavy metal ion migration. The sensor array utilizes a multi-channel synchronous acquisition module to acquire spatially distributed data with a time resolution of seconds. Signal conditioning circuitry eliminates environmental noise interference, generating a dataset of electrochemical parameters reflecting the ion migration state of the soil solution.

[0045] Real-time electrochemical parameters are input into a heavy metal ion speciation kinetic model, and the complexation adsorption efficiency of cadmium, lead, and arsenic ions by the humic acid / silicate composite material is calculated based on chemical thermodynamic equilibrium equations and kinetic rate equations. The model input parameters include real-time pH, ionic strength, and redox potential. By solving the distribution coefficients of heavy metal ions in different complexation phases, a time-varying curve is output showing the proportion of bound cadmium, lead, and arsenic ions as a function of time. For example, the dissociation degree of the humic acid carboxyl functional group increases at pH > 6, enhancing its complexation capacity for cadmium ions. The adsorption efficiency of the silicate interlayer domain is affected by the humidity gradient, with the adsorption rate accelerating under high humidity conditions. The model generates a dynamic efficiency curve by coupling these parameters.

[0046] Based on the preset passivation efficiency threshold deviation between the time-varying curve and the Pareto optimal solution, the objective function weight adjustment of the multi-objective optimization algorithm is calculated through the logic controller. The preset threshold is set according to the soil remediation standard. For example, the cadmium ion passivation efficiency target is ≥85%. The logic controller presets a three-level membership function of "negative deviation", "critical deviation" and "positive deviation", and activates the corresponding rule base based on the difference and change rate between the real-time passivation efficiency and the target value. For example, when the cadmium ion passivation efficiency deviation exceeds -10%, the "negative deviation" rule is activated and an adjustment signal is output to increase the passivation efficiency weight; when the deviation is within the range of ±5%, the "critical deviation" rule is activated to maintain the current weight distribution. The adjustment signal is transmitted to the multi-objective optimization algorithm parameters through the digital interface, driving the algorithm to recalculate the Pareto frontier solution set and realize dynamic iterative optimization of the amendment ratio parameters.

[0047] The above steps, through the synergy of multi-dimensional data acquisition, dynamic modeling, and logical feedback control, form a closed-loop technology from environmental monitoring to dynamic adjustment of optimization parameters. The sensor array provides real-time environmental state input, the morphological dynamics model quantifies the effect of the amendment, and the logic controller converts effect deviations into algorithm parameter correction signals. Ultimately, this drives the multi-objective optimization process to generate a ratio solution that is tailored to current soil conditions, overcoming the technical limitation of traditional open-loop control methods, which cannot track remediation results in real time.

[0048] Specifically, the soil conditioner preparation method based on component detection of the present invention, the improver components prepared to optimize the soil conditioner include: The output feedback signal is synchronously input into the transfer learning algorithm to update the dimension calibration parameters of the spectral, chromatographic and mass spectrometric equipment; Inputting the feedback signal into the dynamic correlation matrix, and adjusting the decay rate of the time decay factor to match the real-time data distribution; updating the parameters of the multi-objective optimization algorithm based on the objective function weight adjustment amount in the feedback signal; Recalculating the covariance matrix update period of the incremental principal component analysis algorithm according to the updated dimension calibration parameters, time decay factor and optimization algorithm parameters; Based on the recalculated covariance matrix, the principal component contribution ranking algorithm is used to iteratively output the synergistic effect solution set of the improver components.

[0049] In the modifier component optimization step described in the present invention, the sensor feedback signal is first synchronously input into the transfer learning algorithm, and the dimensional calibration parameters of the spectral, chromatographic, and mass spectrometric equipment are updated through the back-propagation mechanism. The pre-training model of the transfer learning algorithm is constructed based on a historical cross-device data set. By comparing the distribution differences between the real-time feedback signal and the historical data, the KL divergence loss function is used to dynamically adjust the domain adaptation weights to eliminate the detection data offset caused by equipment aging or environmental interference. The updated calibration parameters realign the dimensional mapping relationship between spectral absorbance, chromatographic retention time, and mass spectrometric ion current intensity, providing standardized input for the next round of data fusion.

[0050] After the feedback signal is input into the dynamic correlation matrix, the decay rate of the time decay factor is adjusted based on the statistical variance of the real-time data distribution. The dynamic correlation matrix calculates the decay rate adjustment step size based on the covariance differences of the data within the sliding window. For example, when the data variance within the window exceeds a preset threshold, the decay rate is increased to 0.15 / minute, increasing the focus on recent data. The adjusted time decay factor matches the real-time soil composition trends, optimizes the adaptability of the time-sensitive feature correlation model to dynamic data, and improves the accuracy of the correlation analysis between spectral absorption peaks and heavy metal concentrations.

[0051] Based on the objective function weight adjustments from the feedback signal, a fuzzy logic control rule base is used to update the parameters of the multi-objective optimization algorithm. The fuzzy logic controller maps the weight deviations between passivation efficiency and porosity to membership function activation levels. For example, when the passivation efficiency deviation exceeds -10%, the "high priority" rule is activated, increasing the passivation efficiency weight coefficient from 0.5 to 0.7. These updated parameters are fed into a non-dominated sorting genetic algorithm, which redefines the objective function space and drives the Pareto optimal solution set to iteratively generate an amendment ratio suitable for the current environmental conditions.

[0052] The covariance matrix update period for the incremental principal component analysis algorithm is recalculated based on the updated dimension calibration parameters, time decay factor, and optimized algorithm parameters. The covariance matrix update period is dynamically linked to the sliding window length. For example, when the window length is shortened to 20 minutes, the update period is adjusted to every 5 minutes. The incremental algorithm processes the real-time data stream batch by batch and uses an oscillation suppression function to smooth covariance matrix fluctuations. This captures the temporal variations in soil component distribution and provides an updated data foundation for defect factor ranking.

[0053] Based on the recalculated covariance matrix, a principal component contribution ranking algorithm is used to iteratively output a solution set for the synergistic effects of the amendment components. This ranking algorithm generates a prioritized list of components that reflects the current soil environmental needs, based on a weighted combination of variance explanation and dynamic feature weights. For example, when the weighted contribution of cadmium ions is highest, the corresponding humic acid-silicate ratio is prioritized in the solution set. This solution set is fed back to the dose-response surface model through a closed-loop control chain, forming a complete technical closed loop from parameter calibration to dynamic modeling to optimization decision-making, enabling adaptive iterative optimization of soil amendment ratios.

[0054] The above steps address the problem of ratio deviation caused by data drift in traditional open-loop control through multi-dimensional parameter adjustment of feedback signals and coordinated model updates. Transfer learning calibration ensures the stability of data fusion, time decay factor optimization enhances the sensitivity of time series modeling, dynamic weight allocation balances multi-objective conflicts, and covariance update cycle adjustment improves the real-time identification of defect factors. Finally, principal component sorting generates an executable synergistic solution set, enabling precise control of the soil amendment preparation process.

[0055] The technical terms of the present invention are explained as follows: The transfer learning algorithm uses an adversarial domain adaptation mechanism to process heterogeneous data from spectral, chromatographic, and mass spectrometric devices. Based on gradient experiments with preset heavy metal standard solutions, it establishes an equivalent conversion relationship between spectral absorbance and mass spectrometric concentration. It then uses the KL divergence loss function to dynamically adjust the distribution differences between the source domain (historical calibration data) and the target domain (real-time detection data). This optimizes domain adaptation weights, eliminates magnitude shifts caused by device sensitivity differences, and generates calibration data with a unified dimension across all devices.

[0056] The time-sensitive feature association model is constructed collaboratively using the random forest algorithm and the non-negative matrix factorization algorithm. The random forest algorithm introduces a time decay factor within the sliding time window, dynamically adjusting the weight decay rate based on the data timestamp interval, strengthening the analytical weight of recent data, and outputting a set of dynamic correlation parameters between spectral absorption peaks and heavy metal concentrations. The non-negative matrix factorization algorithm decomposes the correlation parameters into a basis matrix and a coefficient matrix. The basis matrix represents the basic correlation pattern between spectral absorption peaks and heavy metal concentrations, while the coefficient matrix reflects the changes in correlation strength under different time windows. Its rank is consistent with the number of heavy metal species detected, for example, cadmium, lead, and arsenic have a rank of three, ensuring that the feature weight dimensions match actual needs.

[0057] The incremental principal component analysis model uses a sliding window mechanism to update the covariance matrix batch by batch, incorporating an exponential smoothing function to mitigate sudden weight changes caused by equipment calibration errors or environmental fluctuations. The oscillation suppression function dynamically adjusts the smoothing coefficient based on the covariance matrix variance, for example, increasing the smoothing coefficient to 0.8 when the variance exceeds a threshold. The adjusted covariance matrix is ​​input into a principal component contribution ranking algorithm based on the Kaiser criterion. The algorithm generates a prioritized list of defect factors by multiplying the variance explained by the dynamic feature weights, quantifying the contribution of each component to soil contamination.

[0058] A multi-objective optimization model, based on an improved NSGA-II algorithm, integrates environmental parameters such as soil pH, moisture, and organic matter content to construct a dose-response surface model for the humic acid and silicate composite material. Multi-physics coupling is used to simulate the inhibitory effect of porosity on the diffusion rate of the passivation agent. A matrix for the interaction between passivation efficiency and porosity is established, and weight coefficients are dynamically adjusted to solve for the Pareto optimal solution set. A logic controller activates a membership function rule base based on sensor feedback on passivation efficiency deviations. For example, if the cadmium ion efficiency deviation exceeds -10%, the passivation weight is increased to 0.7. This drives the iterative update of optimization parameters, forming a closed-loop control chain.

[0059] The heavy metal ion morphological kinetic model is based on the chemical thermodynamic equilibrium equation and the kinetic rate equation. It inputs real-time electrochemical parameters such as pH value, ionic strength and redox potential, and calculates the time-varying curve of the complexation adsorption efficiency of cadmium, lead and arsenic ions between the carboxyl group of humic acid and the interlayer domain of silicate, providing a quantitative feedback basis for the logic controller.

[0060] The transfer learning algorithm calibrates the dimensionality differences of spectral, chromatographic, and mass spectrometric devices through an adversarial domain adaptation mechanism. Specifically, this algorithm establishes an equivalent conversion relationship between spectral absorbance and mass spectrometric concentration based on gradient experiments with preset heavy metal standard solutions. It then uses the KL divergence loss function to dynamically adjust the distribution differences between the source domain (historical data) and the target domain (real-time data). This optimizes domain adaptation weights, eliminates magnitude shifts caused by device sensitivity differences, and generates calibration data with uniform dimensions across all devices.

[0061] A combined range normalization and Z-score algorithm processes the calibrated data in steps. Range normalization linearly compresses each modal data to the [0, 1] interval, eliminating differences in device dynamic range. The Z-score algorithm further normalizes the data to a standard normal distribution, generating dimensionless fused data that is consistent across modalities and avoiding weight allocation bias caused by direct input of heterogeneous data into the model. A random forest algorithm, combined with a time decay factor, constructs a time-sensitive feature association model. The time decay factor dynamically adjusts the weight decay rate based on the data timestamp interval, emphasizing the contribution of recent data in the regression analysis and outputting a set of dynamic correlation parameters between spectral absorption peaks and heavy metal concentrations.

[0062] The non-negative matrix factorization algorithm decomposes the dynamic correlation parameters into a basis matrix and a coefficient matrix. The basis matrix represents the basic correlation pattern between spectral absorption peaks and heavy metal concentrations, and the coefficient matrix reflects the changes in correlation strength under different time windows. The dimension of the extracted dynamic feature weights is consistent with the number of heavy metal types. The incremental principal component analysis algorithm uses a sliding window mechanism to update the covariance matrix batch by batch, combined with an exponential smoothing function to suppress batch noise interference and dynamically adjust the principal component weights. The principal component contribution ranking algorithm based on the Kaiser criterion is based on the product of the variance explanation rate and the dynamic feature weight, and outputs a priority list of defect factors that evolves over time. The multi-objective optimization algorithm improved based on NSGA-II integrates environmental parameters and solves the Pareto optimal solution set of passivation efficiency and porosity through dynamic weight adjustment. The logic controller activates the membership function based on real-time feedback data to adjust the objective function weight, forming a closed-loop iterative optimization mechanism.

[0063] In a specific embodiment of the present invention, the spectral absorption peak, cation exchange capacity and heavy metal concentration data of the soil sample are first collected respectively by a near-infrared spectrometer, a high performance liquid chromatograph and an inductively coupled plasma mass spectrometer. In view of the dimensional difference between spectral absorbance and mass spectrum concentration, a gradient experiment is carried out using a preset heavy metal standard solution to establish a linear regression equation between the near-infrared spectral characteristic peak area and the mass spectrum ion current intensity, and determine the equivalent conversion coefficient. Based on this coefficient, the multi-source data is dynamically calibrated by the domain adaptation layer of the transfer learning algorithm, wherein the domain adaptation weight is dynamically adjusted according to the KL divergence of the real-time data and the historical data to eliminate the magnitude offset caused by the difference in equipment sensitivity. The calibrated data is compressed to the [0,1] interval by the range normalization algorithm, and then normalized to the standard normal distribution by the Z-score algorithm to generate dimensionless fusion data that is consistent across modalities, thereby avoiding the weight distribution deviation caused by the direct input of heterogeneous data into the model.

[0064] When constructing a time-sensitive feature association model based on dimensionless fused data, the sliding time window dynamically adjusts its length based on the data acquisition frequency. For example, if data is collected every five minutes, the window is set to thirty minutes. A time decay factor calculates the weight decay rate based on the data timestamp interval, giving more recent data a higher weight in the random forest algorithm. The random forest algorithm uses multiple decision trees to perform regression analysis on the wavelength and peak area of ​​humic acid's characteristic absorption peaks and heavy metal concentrations, outputting a set of dynamic association parameters. This set of parameters is then fed into a non-negative matrix factorization algorithm, where it is decomposed into a basis matrix and a coefficient matrix. The basis matrix represents the underlying correlation pattern between spectral absorption peaks and heavy metal concentrations, while the coefficient matrix reflects the changes in correlation strength across different time windows. The dimension of the extracted dynamic feature weights corresponds to the number of heavy metal species. For example, a three-dimensional weight vector is generated when detecting cadmium, lead, and arsenic. The spectral and mass spectrometry signal response deviation thresholds are iteratively adjusted using leave-one-out cross-validation, optimizing the initial ±5% threshold to within ±3%, thereby minimizing the impact of differences in detection principles on weight stability.

[0065] Dynamic feature weights are fed into an incremental principal component analysis algorithm. A sliding window is used to update the covariance matrix batch by batch, incorporating an exponential smoothing function to mitigate sudden weight changes caused by equipment calibration errors. The smoothing coefficient is dynamically set based on the covariance matrix variance and is positively correlated with the rank of the dynamic correlation matrix. For example, when testing three heavy metals, the smoothing coefficient is scaled to 0.6. The adjusted covariance matrix is ​​then sorted using the Kaiser criterion to extract principal component contributions, generating a list of defect factors ranked in descending order by contribution of cadmium, lead, and arsenic. Based on this list, a dose-response surface model for a humic acid and silicate composite material is constructed. Soil pH, moisture, and organic matter content are input. Multiphysics coupled simulations are used to calculate the inhibitory effect of porosity on the diffusion rate of the passivation agent, establishing a negative correlation matrix between passivation efficiency and porosity. A non-dominated sorting genetic algorithm is used to dynamically adjust the weight coefficients. For example, when moisture exceeds a critical value, the passivation efficiency weight is increased to 0.7, resulting in a Pareto-optimal solution set. Soil solution parameters are collected in real time through pH sensors, humidity sensors and conductivity sensors, and the heavy metal ion morphological kinetic model is input to generate a time-varying curve of complexation efficiency. The logic controller is activated based on the deviation value to adjust the objective function weight, drive the transfer learning algorithm to update the dimensional calibration parameters, dynamic correlation matrix decay rate and covariance update period, forming a closed-loop control chain to achieve adaptive iterative optimization of the amendment ratio.

[0066] The present invention solves the fusion difficulty problem caused by the heterogeneity of multi-source detection data through cross-modal data standardization. A transfer learning algorithm is used to perform dimension calibration on the heterogeneous data collected by spectral, chromatographic and mass spectrometric equipment. The equivalent conversion relationship between spectral absorbance and mass spectrometric concentration is established through a gradient experiment of a preset heavy metal standard solution to eliminate the magnitude differences between different detection equipment. The calibration data is linearly transformed and normalized through a combination of range standardization and Z-score algorithms to generate unified dimensionless fusion data across devices, providing standardized input for subsequent modeling and avoiding feature weight offsets caused by dimensional incomparability.

[0067] A time-sensitive feature association model is constructed based on dimensionless fused data. A random forest algorithm with a time decay factor introduced within a sliding time window dynamically captures the correlation between spectral absorption peaks and heavy metal concentrations. The time decay factor adjusts the weight decay rate based on the data collection interval, strengthening the analytical weight of recent data. A dynamic correlation matrix, whose rank is consistent with the number of heavy metal species, is extracted using a non-negative matrix factorization algorithm to quantify the dynamic feature weights corresponding to spectral absorption peaks. A cross-validation mechanism iteratively adjusts the signal response deviation threshold of spectral and mass spectrometric data to reduce the impact of differences in detection principles on the association analysis and ensure that the dynamic feature weights reflect the true temporal changes in soil composition.

[0068] An incremental principal component analysis algorithm updates the covariance matrix in real time, integrating an oscillation suppression function to smooth batch noise and generate a dynamic defect factor priority list. A multi-objective optimization model is constructed based on defect factor ranking, integrating environmental parameters to calculate the interaction between soil porosity and passivation efficiency. Dynamic weight adjustment is used to find the Pareto optimal solution. Feedback data collected by the sensor array is fed into a closed-loop control mechanism, which iteratively updates dimensional calibration parameters, time decay factors, and optimization algorithm weights. This drives the adjustment of the covariance matrix update cycle and the ranking of principal component contributions. The resulting solution sets for the synergistic effects of the amendments are tailored to the real-time soil conditions, addressing issues such as feature weight shift and component synergy deviation.

Claims

1. A method for preparing a soil conditioner based on component detection, characterized in that: include: Receiving soil composition detection data collected by a spectral device, a chromatographic device, and a mass spectrometric device, wherein the spectral device is a near-infrared spectrometer, the chromatographic device is a high-performance liquid chromatograph, and the mass spectrometric device is an inductively coupled plasma mass spectrometer; Performing cross-modal normalization on the soil composition detection data, including calibrating multi-device dimensional differences through a transfer learning algorithm based on adversarial domain adaptation, and generating dimensionless fused data using a combination of range normalization and Z-score algorithm; Based on the dimensionless fusion data, a time-sensitive feature association model is constructed within a sliding time window by introducing a random forest algorithm with a time decay factor, and dynamic feature weights are output; According to the dynamic feature weights, an incremental principal component analysis algorithm is used to adjust the index weight coefficients of the principal component analysis model in real time to generate a dynamic defect factor ranking; Based on the dynamic defect factor ranking, a dose response surface model of the humic acid and silicate composite material is constructed by using a multi-objective optimization algorithm improved by NSGA-II, and environmental parameters are integrated to generate a Pareto optimal solution set, wherein the environmental parameters include soil pH, soil moisture and soil organic matter content; Applying a preset amendment component in a preset simulated soil environment, and collecting environmental parameter feedback data through a sensor array, wherein the sensor array includes a pH sensor, a humidity sensor, and a conductivity sensor; The calibration parameters of the transfer learning algorithm, the time attenuation factor of the timing-sensitive feature association model, and the objective function weights of the multi-objective optimization algorithm are iteratively updated according to the environmental parameter feedback data to optimize the improver components prepared by the soil conditioner, wherein the iterative update process completes the parameter adaptive adjustment through the gradient descent algorithm.

2. The method for preparing a soil conditioner based on component detection according to claim 1, characterized in that: The cross-modal normalization process includes: Using the preset heavy metal standard solution as the reference substance, the equivalent conversion relationship between spectral absorbance and mass spectrometry concentration is established through the standard substance gradient experiment; Based on the equivalent conversion relationship, a transfer learning algorithm based on adversarial domain adaptation is used to perform dimension calibration on the detection data collected by spectral, chromatographic and mass spectrometric equipment to generate calibration data with unified dimensions across devices, wherein the domain adaptation weight of the transfer learning algorithm is dynamically adjusted through the KL divergence loss function; The calibration data are input into a combination algorithm of range normalization and Z-score, and are mapped to a dimensionless space after eliminating dimension differences through linear transformation to generate dimensionless fusion data of spectrum, chromatography and mass spectrometry data.

3. The method for preparing a soil conditioner based on component detection according to claim 2, characterized in that: The construction of the time-sensitive feature association model includes: Based on the dimensionless fusion data, a random forest algorithm with a time decay factor is introduced within a sliding time window to calculate the correlation between the characteristic absorption peaks of humic acid and silicate in the near-infrared spectrum and the heavy metal concentration, and output a dynamic correlation parameter; Inputting the dynamic correlation parameters into a matrix decomposition algorithm, constructing a dynamic correlation matrix based on the data distribution changes in the time window, and extracting the dynamic feature weights corresponding to the spectral absorption peaks, wherein the rank of the dynamic correlation matrix is ​​consistent with the number of heavy metal species; The dynamic correlation matrix is ​​used as input, and the signal response deviation threshold of the spectral and mass spectrometric data is iteratively adjusted through a cross-validation mechanism.

4. The method for preparing a soil conditioner based on component detection according to claim 3, characterized in that: The generation of the dynamic defect factor ranking includes: Based on the dynamic feature weights, an incremental principal component analysis algorithm is integrated to calculate a real-time updated value of the covariance matrix; Applying a model oscillation suppression function based on exponential smoothing to the real-time update value to adjust the interference of detection batch differences on the weight distribution coefficient in the covariance matrix, wherein the exponential smoothing coefficient is positively correlated with the rank of the dynamic correlation matrix; The adjusted covariance matrix is ​​input into a principal component contribution ranking algorithm based on the Kaiser criterion, and a priority list of defect factors evolving over time is output. The priority list of defect factors is arranged in descending order of contribution of heavy metals such as cadmium, lead, and arsenic.

5. The method for preparing a soil conditioner based on component detection according to claim 4, characterized in that: The multi-objective optimization algorithm includes: Based on the defect factor priority list and with the target heavy metal passivation requirement as a constraint, a dose response surface model of humic acid and silicate composite material was constructed; The soil pH value, soil moisture and soil organic matter content are input into the dose-response surface model, and the output results of the principal component analysis model are combined to calculate the interaction relationship between soil porosity and passivation efficiency; By dynamically adjusting the weight coefficient of the interaction relationship, the Pareto optimal solution of the passivation efficiency and soil porosity is solved.

6. The method for preparing a soil conditioner based on component detection according to claim 5, characterized in that: The step of applying a preset modifier component in a preset simulated soil environment and collecting environmental parameter feedback data through a sensor array includes: The real-time electrochemical parameters in the soil solution are collected by the pH sensor, humidity sensor and conductivity sensor; Inputting the real-time electrochemical parameters into a heavy metal ion morphological kinetic model to calculate a time-varying curve of the complexation efficiency of the humic acid and silicate composite material on the heavy metal ions of cadmium, lead and arsenic; Based on the preset passivation efficiency threshold deviation between the time-varying curve and the Pareto optimal solution, the objective function weight adjustment amount of the multi-objective optimization algorithm is calculated by a logic controller, and a feedback signal is output to the optimization algorithm parameters.

7. The method for preparing a soil conditioner based on component detection according to claim 6, characterized in that: The improver components prepared by optimizing the soil improver include: The output feedback signal is synchronously input into the transfer learning algorithm to update the dimension calibration parameters of the spectral, chromatographic and mass spectrometric equipment; Inputting the feedback signal into the dynamic correlation matrix, and adjusting the decay rate of the time decay factor to match the real-time data distribution; updating the parameters of the multi-objective optimization algorithm based on the objective function weight adjustment amount in the feedback signal; Recalculating the covariance matrix update period of the incremental principal component analysis algorithm according to the updated dimension calibration parameters, time decay factor and optimization algorithm parameters; Based on the recalculated covariance matrix, the principal component contribution ranking algorithm is used to iteratively output the synergistic effect solution set of the improver components.

Images