AI-based karst region soil-groundwater collaborative pollution assessment system

Abstract

The application relates to the technical field of pollution risk assessment, and discloses a soil-underground water collaborative pollution assessment system based on AI. The system fuses multi-source soil and underground water heavy metal time sequence monitoring data, constructs cross-medium collaborative representation characteristics, introduces an enhanced time sequence modeling method with efficient parameter adaptation capability, and designs a composite dynamic loss function for non-stationary pollution evolution characteristics, so as to realize collaborative prediction and uncertainty description of soil and underground water heavy metal pollution states. Further, the prediction result is combined with spatial gridding analysis to construct a collaborative pollution risk index and a risk zoning result, realize technical transformation from pollution evolution modeling to spatial risk assessment and early warning output, and provide reliable technical support for dynamic early warning, hierarchical prevention and control and environmental management decision of heavy metal pollution in a karst area.

Description

An AI-based soil-groundwater co-contamination assessment system for karst areas Technical Field

[0001] This invention relates to the field of pollution risk assessment technology, and in particular to an AI-based soil-groundwater co-pollution assessment system for karst areas. Background Technology

[0002] In the field of environmental monitoring and pollution risk assessment, existing intelligent technologies have gradually incorporated multi-source sensing monitoring, time-series data analysis, machine learning prediction models, and geographic information systems to identify and assess the heavy metal pollution status in soil or groundwater, and to some extent, have achieved automated analysis of pollution trends. However, most existing intelligent assessment technologies still focus on modeling a single medium or a single time scale, typically predicting or spatially interpolating heavy metal concentrations in soil or groundwater separately. They lack a unified modeling capability for the collaborative migration process under conditions of high soil-groundwater connectivity and strong hydrodynamic response in karst areas, making it difficult to characterize the transport intensity, time-delay characteristics, and multi-metal co-evolution relationships of heavy metals between different media. Meanwhile, existing intelligent models generally rely on static or weak time-series characteristics, which are insufficient in characterizing non-stationary pollution responses caused by external hydrological drivers such as rainfall and water level fluctuations. This results in limited stability and generalization ability of prediction results under complex karst hydrological conditions. In addition, existing technologies are mostly focused on pollution status identification or concentration prediction, lacking a collaborative mechanism to further transform prediction results into spatial zoning risk assessment and hierarchical management decisions. This makes it difficult to meet the actual needs for dynamic early warning and refined control in the prevention and control of heavy metal pollution in karst areas. Summary of the Invention

[0003] To address the shortcomings of existing intelligent pollution assessment technologies in depicting the collaborative migration relationship between soil and groundwater under complex hydrogeological conditions in karst areas, their insufficient response to non-stationary hydrological drivers such as rainfall and water level fluctuations, and the inability of assessment results to directly serve environmental management decisions, this invention proposes an AI-based soil-groundwater collaborative pollution assessment system for karst areas. This system constructs a cross-media collaborative characterization approach from key mechanistic dimensions such as migration intensity, external hydrological driving response, collaborative consistency, and time-delayed transmission. Furthermore, it introduces an enhanced time-series modeling method with efficient parameter adaptation capabilities. During model training, a composite model is further designed to address the non-stationary characteristics of pollution evolution in karst areas. The model employs a dynamic loss function that uses time oscillation weights to co-modulate the classification and continuous concentration prediction errors. This enables the model to adaptively balance classification accuracy and numerical prediction stability at different hydrological stages, thereby achieving co-prediction and uncertainty characterization of heavy metal pollution status in soil and groundwater. Furthermore, the prediction results are integrated with spatial gridded analysis to construct quantifiable co-pollution risk indicators and risk zoning results. Under a unified technical framework, a closed loop is completed from cross-media pollution evolution modeling to spatial risk assessment and early warning output, significantly improving the technical level of heavy metal pollution assessment in karst areas in terms of temporal sensitivity, cross-media consistency, and environmental management applicability.

[0004] This invention proposes an AI-based soil-groundwater co-polluting assessment system for karst areas, which is applied to the monitoring and risk assessment of heavy metal inorganic pollution in karst watersheds. The karst watershed includes overburden soil, karst fissure zone and karst aquifer, and is equipped with soil monitoring sampling points and groundwater monitoring wells. The system includes a field monitoring subsystem, a central data server cluster and an environmental management terminal.

[0005] The on-site monitoring subsystem collects soil-groundwater heavy metal co-monitoring data and heavy metal content analysis results;

[0006] The edge acquisition terminal, deployed near soil monitoring sampling points and observation wells, is equipped with an industrial-grade ARM processor, analog-to-digital converter chip, NB-IoT communication module, and local storage module. The edge acquisition terminal is used to perform timestamp alignment, basic noise reduction, and local caching on soil-groundwater heavy metal collaborative monitoring data to obtain soil-groundwater heavy metal pollution acquisition data. The data and heavy metal content analysis results are then sent to the central data server cluster via the NB-IoT communication module through a wireless network.

[0007] The central data server cluster is deployed in the data center of the environmental monitoring and management center. It is a rack-mounted server equipped with a 64-core CPU, GPU accelerator card and RAID disk array. It is used to run the collaborative pollution assessment software platform, which includes: data preprocessing module, feature construction module, collaborative pollution assessment module and risk zoning module.

[0008] The data preprocessing module calibrates the soil-groundwater heavy metal pollution collection data based on the heavy metal content analysis results to obtain high-precision heavy metal concentration label data. It then performs missing value imputation, outlier detection, unit unification, spatial coordinate transformation, and time-scale resampling on the soil-groundwater heavy metal pollution collection data and the corresponding high-precision heavy metal concentration label data to construct a soil-groundwater heavy metal time series dataset.

[0009] The feature construction module constructs a multi-channel input feature set based on the soil-groundwater heavy metal time series dataset;

[0010] The collaborative pollution assessment module constructs a Chronos-T5 model with enhanced rotational parameters; it introduces a composite dynamic loss function as a constraint mechanism for model training and prediction, and performs temporal encoding and joint modeling on the multi-channel input feature set based on the Chronos-T5 model with enhanced rotational parameters to obtain the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater, and outputs collaborative pollution risk indicators.

[0011] The risk zoning module maps the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater to a unified geospatial grid in the karst region, and generates a continuous spatial distribution field of heavy metal concentration through spatial interpolation algorithms. Based on the spatial distribution field of heavy metal concentration and synergistic pollution risk indicators, it performs threshold determination and exceeds the probability calculation, dividing the study area into safe zones, controllable risk zones, key prevention and control zones, and high-risk zones, and generates synergistic pollution risk zoning results. When the synergistic pollution risk indicator corresponding to any spatial grid unit exceeds the preset threshold, synergistic pollution early warning information is automatically generated and pushed to the environmental management terminal.

[0012] The environmental management terminal includes desktop workstations connected to a central data server cluster. Each workstation is equipped with a monitor, a human-computer interface, and a geographic information system (GIS) client to receive collaborative pollution early warning information. Based on the received collaborative pollution early warning information and collaborative pollution risk zoning results, the environmental management terminal generates corresponding environmental management recommendations. These recommendations include: suggestions for adjusting monitoring frequency, delineating key monitoring areas, prioritizing soil remediation, and restricting groundwater extraction. These recommendations assist in decision-making regarding heavy metal pollution prevention and control and environmental management in karst areas. The environmental management terminal also visualizes the collaborative pollution risk zoning results in the form of a gridded collaborative pollution risk zoning map, pollution evolution curves at typical monitoring points, and statistical analysis reports, thus providing intuitive technical support for decision-making regarding heavy metal pollution prevention and control and environmental management in karst areas.

[0013] Furthermore, the process of constructing a multi-channel input feature set based on the soil-groundwater heavy metal time-series dataset specifically includes the following steps:

[0014] Step B1: Based on the soil-groundwater heavy metal time series dataset, combined with the spatial pairing relationship between soil monitoring points and underlying groundwater monitoring wells, calculate the heavy metal concentration difference between soil and groundwater, the magnitude of concentration gradient change per unit time, the apparent migration rate of heavy metals from soil to groundwater, and the heavy metal enrichment coefficient, and generate migration intensity feature data to characterize the migration intensity of heavy metals between soil and groundwater.

[0015] Step B2: Based on the soil-groundwater heavy metal time series dataset and the corresponding rainfall and groundwater level data, calculate the groundwater level rise rate after the rainfall event, the response amplitude of heavy metal concentration in groundwater, the time lag between rainfall and heavy metal concentration change, and the correlation index between rainfall intensity and heavy metal concentration change, and generate heavy metal response coupling characteristic data to characterize the external hydrological driving effect.

[0016] Step B3: Based on the soil-groundwater heavy metal time series dataset, calculate the time fluctuation coefficient, frequency of abnormal fluctuations, concentration decline rate, and long-term trend drift amplitude of heavy metal concentration in groundwater, and generate pollution stability characteristic data to characterize the response characteristics of the groundwater system to external disturbances.

[0017] Step B4: Based on the soil-groundwater heavy metal time series dataset, calculate the synchronicity of concentration changes, consistency of migration trends, and order of peak responses of different heavy metal elements in soil and groundwater, and generate cooperative migration consistency feature data to characterize the cooperative migration behavior of multiple heavy metals.

[0018] Step B5: Based on the soil-groundwater heavy metal time series dataset, calculate the time difference of the occurrence of the peak concentration of heavy metals in soil and groundwater, the peak response decay rate, and the time interval between consecutive peak events to generate time-delay transport characteristic data to characterize the time delay characteristics of heavy metal transport from soil to groundwater.

[0019] Step B6: Organize, align, and integrate the migration intensity feature data, heavy metal response coupling feature data, pollution stability feature data, co-migration consistency feature data, and time-delayed transport feature data at a unified time scale to form a multi-channel input feature set for model input.

[0020] Furthermore, the composite dynamic loss function is constructed as follows: A cross-entropy loss function is constructed, and a time-oscillation weighting factor based on a triangular wave is introduced to dynamically modulate the class gradient contribution and loss landscape curvature characteristics of the cross-entropy loss function, resulting in a dynamic cross-entropy loss function; a mean square error loss function is constructed, and a time-oscillation weighting factor based on a square wave is introduced to time-varyingly modulate the target side of the mean square error loss function, achieving dynamic shaping of the class gradient distribution and second-order geometry (curvature / valley width), resulting in a dynamic mean square error loss function; a composite dynamic loss function is constructed by combining the dynamic cross-entropy loss function and the dynamic mean square error loss function; this composite dynamic loss function is used to optimize the temporal feature modeling capability and prediction accuracy of the rotation parameter-enhanced Chronos-T5 model.

[0021] Furthermore, the Chronos-T5 model, enhanced by rotational parameters, performs temporal encoding and joint modeling on the multi-channel input feature set to obtain the distribution of heavy metal content in soil and the concentration distribution of heavy metals in groundwater, and outputs a synergistic pollution risk indicator. This process specifically includes the following steps:

[0022] Step S1: Normalize and quantize the multi-channel input feature set to form a token sequence;

[0023] Step S2: Build a Chronos-T5 model. Input the token sequence into the encoder part of the Chronos-T5 model. Use the multi-head self-attention mechanism and residual connection structure to perform contextual modeling on historical time segments and extract cross-time period latent feature representations.

[0024] Step S3: Based on the cross-time period latent feature representation, perform efficient parameter adaptation processing on the Chronos-T5 model based on rotational degrees of freedom. Specifically, this includes: introducing an efficient parameter adaptation mechanism based on cross-layer joint decomposition and rotational degrees of freedom, adjusting the axial strength and optimizing the linear mapping weight parameters of the encoder and decoder in the Chronos-T5 model with sparse rotational updates, and constructing a Chronos-T5 model with enhanced rotational parameters.

[0025] Step S4: Input the cross-time period latent feature representation into the decoder part of the Chronos-T5 model with rotational parameter enhancement, and use an autoregressive sampling mechanism to predict the output step by step; constrain the probability distribution of the predicted output according to the composite dynamic loss function, generate a predicted token sequence, and obtain a set of predicted trajectories to characterize the uncertainty of changes in heavy metal content in soil and heavy metal concentration in groundwater in multiple future time steps by sampling the predicted output multiple times.

[0026] Step S5: Perform dequantization and descaling operations on the predicted token sequence to restore it to a continuous real-valued predicted sequence, thus obtaining the dequantized predicted sequence; and calculate statistical features based on the predicted trajectory set.

[0027] Step S6: Summarize the inverse quantification prediction sequence and statistical characteristics, determine the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater in multiple future time steps as the prediction results, and calculate the synergistic pollution risk index to characterize the degree of soil-groundwater synergistic pollution based on the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater.

[0028] Furthermore, step S3 specifically includes the following steps:

[0029] Step S31: Based on the cross-time period latent feature representation, extract the linear mapping weight parameters corresponding to the encoder and decoder in the Chronos-T5 model; perform cross-layer weight projection and structure alignment processing on each linear mapping weight parameter to map the weight parameters of different layers to a unified latent representation space, so as to eliminate the scale difference of parameters of different layers and align their structural features, and obtain the aligned cross-layer weight projection result.

[0030] Step S32: Based on the aligned cross-layer weight projection results, perform cross-layer joint decomposition on the linear mapping weight parameters of multiple layers in the Chronos-T5 model; during the joint decomposition process, extract the latent orthogonal basis that is applicable to multiple layers to form a cross-layer shared latent orthogonal basis;

[0031] Step S33: Under the constraint of the shared potential orthogonal basis across layers, the remaining components of the linear mapping weight parameters of each layer are orthogonally decomposed to obtain the orthogonal direction components of the corresponding layer and the axial strength parameters along the potential orthogonal basis direction; based on the axial strength parameters, the strength adjustment along the shared potential direction is performed on the linear mapping weight parameters of each layer to form the axial strength update result.

[0032] Step S34: Based on the axial strength parameter, a sparse non-axial perturbation matrix is ​​introduced to establish a restricted cross-coupling relationship between different potential dimensions under the shared potential orthogonal basis across layers; based on the sparse non-axial perturbation matrix, non-axial perturbation updates are performed on the linear mapping weight parameters of each layer, and rotational parameter adjustments relative to the original potential direction are realized in the potential representation space to form a rotational update result.

[0033] Step S35: Combine the axial strength update result with the rotational update result to generate a parameter update form that integrates axial strength adjustment and rotational degree of freedom extension; based on the combined update result, replace the corresponding linear mapping weight parameters in the Chronos-T5 model to obtain a rotationally enhanced Chronos-T5 model.

[0034] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0035] First, this invention constructs a cross-media collaborative characterization method for the migration mechanism of heavy metals in soil and groundwater in karst areas, realizing unified modeling and collaborative assessment of soil and groundwater pollution status. This avoids the problem of fragmented assessment caused by separate modeling of different media in existing technologies, improves the overall characterization of the migration process of heavy metals from soil to groundwater, and solves the technical problem that pollution evolution is difficult to accurately identify due to the development of karst fissures and strong media connectivity in karst areas. This enhances the credibility and consistency of collaborative pollution risk identification results under complex hydrogeological conditions.

[0036] Second, by introducing an enhanced time-series modeling method with efficient parameter adaptation capabilities and combining it with a composite dynamic loss function designed for non-stationary pollution evolution characteristics, this invention achieves an adaptive balance between classification accuracy and continuous concentration prediction stability at different hydrological stages. This improves the prediction stability for sudden changes and phased fluctuations in heavy metal concentrations caused by external disturbances such as rainfall events and water level fluctuations. It also solves the problems of bias accumulation, response lag, and prediction instability that existing intelligent assessment methods are prone to in long-term time-series predictions, thereby enhancing the forward-looking prediction capability of this invention for pollution trends and risk changes under complex karst hydrological conditions.

[0037] Third, this invention combines collaborative prediction results with spatial gridded risk analysis, realizing the transformation from pollution state prediction to spatial risk zoning and hierarchical early warning. This improves the interpretability and operability of the assessment results at the regional scale and solves the problem that existing intelligent technologies cannot directly support environmental management decisions such as monitoring optimization, remediation prioritization, and groundwater extraction control. Thus, it enhances the practical application value and environmental management support capability of this invention in the prevention and control of heavy metal pollution in karst areas, and provides a reliable technical basis for implementing precise monitoring, dynamic early warning, and hierarchical control. Attached Figure Description

[0038] Figure 1 is a schematic diagram of the overall structure of an AI-based karst area soil-groundwater co-contamination assessment system provided by the present invention.

[0039] Figure 2 is a gridded collaborative pollution risk zoning map proposed in Example 6. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] Example 1: According to Figure 1, the present invention proposes an AI-based soil-groundwater collaborative pollution assessment system for karst areas, which is applied to the monitoring and risk assessment of heavy metal inorganic pollution in karst watersheds. The karst watershed includes overburden soil, karst fissure zone and karst aquifer, and is equipped with soil monitoring sampling points and groundwater monitoring wells. The system includes a field monitoring subsystem, a central data server cluster and an environmental management terminal.

[0042] The on-site monitoring subsystem, deployed in the karst watershed, includes: a multi-point soil monitoring unit, located in the overburden and topsoil / subsurface soil of the karst area, with multiple soil monitoring points at different soil depths; the multi-point soil monitoring unit includes soil sampling probes and soil physicochemical parameter sensors, used to collect soil pH, moisture content, conductivity, and heavy metal content data, reflecting the occurrence state and migration conditions of heavy metals in the soil; and a groundwater monitoring unit, located in monitoring wells in the karst area, including a water level gauge, multi-parameter water quality probe, and groundwater sampling device; the groundwater monitoring unit collects groundwater level, temperature, conductivity, redox potential, and heavy metal concentration data, reflecting groundwater heavy metal pollution. The monitoring index data characterizes the heavy metal pollution features and dynamic changes in karst aquifers; it integrates soil physicochemical and heavy metal pollution monitoring index data with groundwater heavy metal pollution monitoring index data to obtain soil-groundwater heavy metal synergistic monitoring data; the field monitoring subsystem also includes a laboratory analysis interface unit, used to receive offline collected soil and groundwater samples collected by inductively coupled plasma mass spectrometry (ICP-MS) to obtain heavy metal content analysis results; the heavy metal content analysis results are numerical detection results of heavy metal element concentrations, including the concentration values ​​of heavy metal elements and the corresponding sampling point identifiers and sampling time information, wherein the heavy metal concentration of soil samples is expressed in the form of mass fraction, and the heavy metal concentration of groundwater samples is expressed in the form of volume concentration;

[0043] The edge acquisition terminal, deployed near soil monitoring sampling points and observation wells, is equipped with an industrial-grade ARM processor, analog-to-digital converter chip, NB-IoT communication module, and local storage module. The edge acquisition terminal is used to perform timestamp alignment, basic noise reduction, and local caching on soil-groundwater heavy metal collaborative monitoring data to obtain soil-groundwater heavy metal pollution acquisition data. The data and heavy metal content analysis results are then sent to the central data server cluster via the NB-IoT communication module through a wireless network.

[0044] The central data server cluster is deployed in the data center of the environmental monitoring and management center. It is a rack-mounted server equipped with a 64-core CPU, GPU accelerator card and RAID disk array. It is used to run the collaborative pollution assessment software platform, which includes: data preprocessing module, feature construction module, collaborative pollution assessment module and risk zoning module.

[0045] The data preprocessing module calibrates the soil-groundwater heavy metal pollution collection data based on the heavy metal content analysis results to obtain high-precision heavy metal concentration label data. It then performs missing value imputation, outlier detection, unit unification, spatial coordinate transformation, and time-scale resampling on the soil-groundwater heavy metal pollution collection data and the corresponding high-precision heavy metal concentration label data to construct a soil-groundwater heavy metal time series dataset.

[0046] The feature construction module constructs a multi-channel input feature set based on the soil-groundwater heavy metal time series dataset;

[0047] The collaborative pollution assessment module constructs a Chronos-T5 model with enhanced rotational parameters; it introduces a composite dynamic loss function as a constraint mechanism for model training and prediction, and performs temporal encoding and joint modeling on the multi-channel input feature set based on the Chronos-T5 model with enhanced rotational parameters to obtain the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater. Based on the prediction results (distribution of heavy metal content in soil and distribution of heavy metal concentration in groundwater), it outputs a collaborative pollution risk index that characterizes the degree of soil-groundwater collaborative pollution.

[0048] The risk zoning module maps the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater to a unified geospatial grid in the karst region, and generates a continuous spatial distribution field of heavy metal concentration through spatial interpolation algorithms. Based on the spatial distribution field of heavy metal concentration and synergistic pollution risk indicators, it performs threshold determination and exceeds the probability calculation, dividing the study area into safe zones, controllable risk zones, key prevention and control zones, and high-risk zones, and generates synergistic pollution risk zoning results. When the synergistic pollution risk indicator corresponding to any spatial grid unit exceeds the preset threshold, synergistic pollution early warning information is automatically generated and pushed to the environmental management terminal.

[0049] The environmental management terminal includes desktop workstations connected to a central data server cluster. Each workstation is equipped with a monitor, a human-computer interface, and a geographic information system (GIS) client to receive collaborative pollution early warning information. Based on the received collaborative pollution early warning information and collaborative pollution risk zoning results, the environmental management terminal generates corresponding environmental management recommendations. These recommendations include: suggestions for adjusting monitoring frequency, delineating key monitoring areas, prioritizing soil remediation, and restricting groundwater extraction. These recommendations assist in decision-making regarding heavy metal pollution prevention and control and environmental management in karst areas. The environmental management terminal also visualizes the collaborative pollution risk zoning results in the form of a gridded collaborative pollution risk zoning map, pollution evolution curves at typical monitoring points, and statistical analysis reports, thus providing intuitive technical support for decision-making regarding heavy metal pollution prevention and control and environmental management in karst areas.

[0050] Example 2, based on Example 1, describes the process of constructing a multi-channel input feature set based on a soil-groundwater heavy metal time-series dataset. The specific steps include:

[0051] Step B1: Based on the soil-groundwater heavy metal time series dataset, combined with the spatial pairing relationship between soil monitoring points and underlying groundwater monitoring wells, calculate the heavy metal concentration difference between soil and groundwater, the magnitude of concentration gradient change per unit time, the apparent migration rate of heavy metals from soil to groundwater, and the heavy metal enrichment coefficient, and generate migration intensity feature data to characterize the migration intensity of heavy metals between soil and groundwater.

[0052] Step B2: Based on the soil-groundwater heavy metal time series dataset and the corresponding rainfall and groundwater level data, calculate the groundwater level rise rate after the rainfall event, the response amplitude of heavy metal concentration in groundwater, the time lag between rainfall and heavy metal concentration change, and the correlation index between rainfall intensity and heavy metal concentration change, and generate heavy metal response coupling characteristic data to characterize the external hydrological driving effect.

[0053] Step B3: Based on the soil-groundwater heavy metal time series dataset, calculate the time fluctuation coefficient, frequency of abnormal fluctuations, concentration decline rate, and long-term trend drift amplitude of heavy metal concentration in groundwater, and generate pollution stability characteristic data to characterize the response characteristics of the groundwater system to external disturbances.

[0054] Step B4: Based on the soil-groundwater heavy metal time series dataset, calculate the synchronicity of concentration changes, consistency of migration trends, and peak response order of different heavy metal elements in soil and groundwater, and generate co-migration consistency feature data to characterize the co-migration behavior of multiple heavy metals; the different heavy metal elements include cadmium (Cd), lead (Pb), arsenic (As), mercury (Hg), and hexavalent chromium (Cr(VI)).

[0055] Step B5: Based on the soil-groundwater heavy metal time series dataset, calculate the time difference of the occurrence of the peak concentration of heavy metals in soil and groundwater, the peak response decay rate, and the time interval between consecutive peak events to generate time-delay transport characteristic data to characterize the time delay characteristics of heavy metal transport from soil to groundwater.

[0056] Step B6: Organize, align, and integrate the migration intensity feature data, heavy metal response coupling feature data, pollution stability feature data, cooperative migration consistency feature data, and time-delayed transport feature data at a unified time scale to form a multi-channel input feature set for model input, so as to jointly characterize the heavy metal migration process of soil-groundwater in karst areas from multiple dimensions such as migration intensity, external driving force, system stability, cooperative effect, and transport time delay.

[0057] Example 3, based on Example 2, describes the construction of a composite dynamic loss function as follows: A cross-entropy loss function is constructed by introducing a time-oscillation weighting factor based on a triangular wave to dynamically modulate the class gradient contribution and loss landscape curvature characteristics of the cross-entropy loss function, resulting in a dynamic cross-entropy loss function. A mean square error loss function is also constructed by introducing a time-oscillation weighting factor based on a square wave to time-varyingly modulate the target side of the mean square error loss function, achieving dynamic shaping of the class gradient distribution and second-order geometry (curvature / valley width), resulting in a dynamic mean square error loss function. A composite dynamic loss function is constructed by combining the dynamic cross-entropy loss function and the dynamic mean square error loss function. This composite dynamic loss function is used to optimize the temporal feature modeling capability and prediction accuracy of the rotationally enhanced Chronos-T5 model.

[0058] In conventional techniques, the Chronos-T5 model's temporal feature modeling capability and prediction accuracy are enhanced by optimizing the rotation parameters using the cross-entropy loss function or the mean squared error loss function.

[0059] Example 4, based on Example 3, describes a process of temporal encoding and joint modeling of multi-channel input feature sets using a rotationally enhanced Chronos-T5 model to obtain the distribution of heavy metal content in soil and the concentration of heavy metals in groundwater, and outputting a synergistic pollution risk indicator. The specific steps include:

[0060] Step S1: Normalize and quantize the multi-channel input feature set to form a token sequence;

[0061] Step S2: Build a Chronos-T5 model. Input the token sequence into the encoder part of the Chronos-T5 model. Use the multi-head self-attention mechanism and residual connection structure to perform contextual modeling on historical time segments and extract cross-time period latent feature representations.

[0062] Step S3: Based on the cross-time period latent feature representation, perform efficient parameter adaptation processing on the Chronos-T5 model based on rotational degrees of freedom. Specifically, this includes: introducing an efficient parameter adaptation mechanism based on cross-layer joint decomposition and rotational degrees of freedom, adjusting the axial strength and optimizing the linear mapping weight parameters of the encoder and decoder in the Chronos-T5 model with sparse rotational updates, and constructing a Chronos-T5 model with enhanced rotational parameters.

[0063] Step S4: Input the cross-time period latent feature representation into the decoder part of the Chronos-T5 model with rotational parameter enhancement, and use an autoregressive sampling mechanism to predict the output step by step; constrain the probability distribution of the predicted output according to the composite dynamic loss function, generate a predicted token sequence, and obtain a set of predicted trajectories to characterize the uncertainty of changes in heavy metal content in soil and heavy metal concentration in groundwater in multiple future time steps by sampling the predicted output multiple times.

[0064] Introducing the dynamic cross-entropy loss function formula based on the time oscillation weighting factor of triangular waves:

[0065] , ;

[0066] in, This indicates that in the training steps Next, the The time oscillation weighting factor corresponds to each heavy metal pollution status category; this weighting factor changes periodically with the training process in the form of a triangular wave to simulate the phased fluctuation characteristics of the heavy metal pollution response in the karst soil-groundwater system with seasonal changes, rainfall cycles, and groundwater recharge and discharge processes. This represents the oscillation amplitude coefficient, which is used to adjust the influence of the time oscillation weight factor on the gradient update intensity of different pollution state categories, so as to avoid a certain pollution state from dominating for a long time during the model training process and causing model bias. Represents the trigonometric wave function. The total number of categories indicates the heavy metal pollution status participating in the collaborative pollution assessment. The period length (in training steps) represents the weighting factor for time oscillation. Indicates the first The phase offset corresponding to each pollution state category is used to stagger the weight enhancement periods of different categories during the training process, so as to improve the model's ability to distinguish multiple metal co-contamination states.

[0067] ;

[0068] in, Indicates training steps The calculated dynamic cross-entropy loss is used to constrain the ability of the enhanced Chronos-T5 model to distinguish different soil-groundwater co-contamination states. This indicates the total number of training entries in the current batch. Indicates the index of the training entry. Indicates the first The input vector of each training item, i.e., the latent feature representation across time periods. Indicates the first The actual contamination state label corresponding to each training item Indicates training steps Model parameters at time step (Chronos-T5 decoder weights). This indicates that the Chronos-T5 is performing well in training steps. At that time, use the current parameters For the sample In the calculated logit vector, the true class The indexation result of the corresponding component; Indicates the sample In training steps At that time, the summation of the exponentialized results of all class logits output by the Chronos-T5 model;

[0069] Introducing a dynamic mean square error loss function based on a square wave time oscillation weighting factor:

[0070] ;

[0071] in, Indicating in training steps Next, the Each prediction channel corresponds to a time-oscillation weighting factor based on square waves; each prediction channel corresponds to a prediction variable for different soil or groundwater heavy metal concentrations. This represents the square wave amplitude coefficient, used to periodically increase or decrease the penalty intensity of the model on prediction errors during training, in order to adapt to the significant non-stationary characteristics of the karst groundwater system between the wet and dry seasons; Represents a square wave function;

[0072] ;

[0073] in, Indicates training steps The calculated dynamic mean squared error loss is used to constrain the fitting accuracy of the enhanced Chronos-T5 model for continuous heavy metal concentration prediction results. This indicates that the enhanced Chronos-T5 decoder supports the first... Input The output of the first Predicted values ​​for channel-like structures Indicates the first The one-hot label of each entry in the category dimension The amount;

[0074] Composite dynamic loss function:

[0075] ;

[0076] in, Indicates training steps The calculated composite dynamic loss, , These represent the tradeoff coefficient, which adjusts the dynamic cross-entropy term, and the tradeoff coefficient, which adjusts the dynamic mean square error term, respectively. Represents the regularization coefficient. This represents a regularization term that imposes structural constraints on the parameters;

[0077] Step S5: Perform dequantization and descaling operations on the predicted token sequence to restore it to a continuous real-valued predicted sequence, thus obtaining the dequantized predicted sequence; and calculate statistical features based on the predicted trajectory set.

[0078] Step S6: Summarize the inverse quantification prediction sequence and statistical characteristics, determine the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater in multiple future time steps as the prediction results, and calculate the synergistic pollution risk index to characterize the degree of soil-groundwater synergistic pollution based on the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater.

[0079] In the conventional technology field, the process of performing temporal encoding and joint modeling on multi-channel input feature sets based on the Chronos-T5 model to obtain the distribution of heavy metal content in soil and the concentration distribution of heavy metals in groundwater, and outputting synergistic pollution risk indicators, specifically includes the following steps:

[0080] Step E1: Normalize and quantize the multi-channel input feature set to form a token sequence;

[0081] Step E2: Build a Chronos-T5 model. Input the token sequence into the encoder part of the Chronos-T5 model. Use the multi-head self-attention mechanism and residual connection structure to perform contextual modeling on historical time segments and extract cross-time period latent feature representations.

[0082] Step E3: Input the cross-time period latent feature representation into the decoder part of the Chronos-T5 model, and use an autoregressive sampling mechanism to predict the output step by step; constrain the probability distribution of the predicted output according to the composite dynamic loss function, generate the predicted token sequence, and obtain the predicted trajectory set to characterize the uncertainty of the changes in heavy metal content in soil and heavy metal concentration in groundwater in multiple future time steps by sampling the predicted output multiple times.

[0083] Step E4: Perform dequantization and descaling operations on the predicted token sequence to restore it to a continuous real-valued predicted sequence, thus obtaining the dequantized predicted sequence; and calculate statistical features based on the predicted trajectory set.

[0084] Step E5: Summarize the inverse quantification prediction sequence and statistical characteristics, determine the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater in multiple future time steps as the prediction results, and calculate the synergistic pollution risk index to characterize the degree of soil-groundwater synergistic pollution based on the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater.

[0085] Example 5, this example is based on Example 4. In this example, step S3 specifically includes the following steps:

[0086] Step S31: Based on the cross-time period latent feature representation, extract the linear mapping weight parameters corresponding to the encoder and decoder in the Chronos-T5 model; perform cross-layer weight projection and structure alignment processing on each linear mapping weight parameter to map the weight parameters of different layers to a unified latent representation space, so as to eliminate the scale difference of parameters of different layers and align their structural features, and obtain the aligned cross-layer weight projection result, which provides a unified representation basis for subsequent joint decomposition.

[0087] Step S32: Based on the aligned cross-layer weight projection results, perform cross-layer joint decomposition on the linear mapping weight parameters of multiple layers in the Chronos-T5 model; during the joint decomposition process, extract the latent orthogonal basis that is applicable to multiple layers to form a cross-layer shared latent orthogonal basis, which is used to characterize the cross-layer consistent temporal modeling structure features in the Chronos-T5 model.

[0088] Step S33: Under the constraint of the shared potential orthogonal basis across layers, the remaining components of the linear mapping weight parameters of each layer are orthogonally decomposed to obtain the orthogonal direction components of the corresponding layer and the axial strength parameters along the potential orthogonal basis direction; based on the axial strength parameters, the strength adjustment along the shared potential direction is performed on the linear mapping weight parameters of each layer to form the axial strength update result; wherein, the axial strength update result is used to finely adjust the response amplitude of the model weights without changing the potential orthogonal basis direction, thereby maintaining the stability of the dominant structure and temporal modeling characteristics of the Chronos-T5 model;

[0089] Under the constraint of a shared potential orthogonal basis across layers, the linear mapping weight parameters of each layer are decomposed to obtain the orthogonal direction components within each layer and their corresponding axial strength parameters, expressed as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] in, The layer index represents the number of layers in the Chronos-T5 model. One linear mapping layer is selected to participate in decomposition / adaptation; Indicates the first The projection / alignment matrix of layer weights. Represents the orthogonal matrix obtained from the cross-layer structural decomposition; Indicates the first Local decomposition representation of layers in a shared orthogonal coordinate system; Indicates the first Axial strength parameter matrix (diagonal matrix) of the layer. Represents the diagonalization operator. Indicates the first Layer local decomposition representation The L2 norms of each axial component calculated in the shared potential orthogonal coordinate system are used to characterize the axial strength of the layer weight along each potential orthogonal direction. express The inverse matrix; Indicates the first Orthogonal direction components of the layer;

[0094] When only axial updates are performed, the first The update form of the layer linear mapping weights satisfies:

[0095] ;

[0096] in, Indicates the first The axial strength update result obtained by adjusting the weight strength only along the cross-layer shared potential orthogonal basis direction without introducing rotational degrees of freedom; Represents a potential orthogonal basis shared across layers. Indicates transpose;

[0097] Step S34: Based on the axial strength parameters, a sparse non-axial perturbation matrix is ​​introduced to establish a restricted cross-coupling relationship between different potential dimensions under the shared potential orthogonal basis across layers; based on the sparse non-axial perturbation matrix, non-axial perturbation updates are performed on the linear mapping weight parameters of each layer, realizing rotational parameter adjustments relative to the original potential direction in the potential representation space, forming a rotational update result; wherein, by constraining the number of non-zero elements and the perturbation amplitude of the sparse non-axial perturbation matrix, the rotational update result only introduces rotational degrees of freedom within a controlled range, so as to avoid damaging the original dominant structure and temporal modeling stability of the Chronos-T5 model;

[0098] The linear mapping weight update form after introducing rotational degrees of freedom is:

[0099] ;

[0100] in, Indicates the first The result of rotational enhancement update after superimposing rotational degrees of freedom on the axial strength update layer; This represents a sparse non-axial perturbation matrix;

[0101] Step S35: Combine the axial strength update results with the rotational update results to generate a parameter update form that integrates axial strength adjustment and rotational degree of freedom extension; based on the combined update results, replace the corresponding linear mapping weight parameters in the Chronos-T5 model to obtain a rotationally enhanced Chronos-T5 model, which is used for subsequent autoregressive prediction and synergistic pollution risk assessment calculations.

[0102] Example 6, according to Figure 2, this example is based on Example 5. In this example,

[0103] The risk zoning module maps the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater to a unified geospatial grid in the karst region, and generates a continuous spatial distribution field of heavy metal concentration through spatial interpolation algorithms. Based on the spatial distribution field of heavy metal concentration and synergistic pollution risk indicators, it performs threshold determination and exceeds the probability calculation, dividing the study area into safe zones, controllable risk zones, key prevention and control zones, and high-risk zones, and generates synergistic pollution risk zoning results. When the synergistic pollution risk indicator corresponding to any spatial grid unit exceeds the preset threshold, synergistic pollution early warning information is automatically generated and pushed to the environmental management terminal.

[0104] In this embodiment, the collaborative pollution risk index for each grid is calculated, as shown in Table 1:

[0105] Table 1

[0106] ;

[0107] Collaborative pollution early warning information:

[0108] Warning information generated for G(8,12):

[0109] Warning number: ALM-20260108-001;

[0110] Trigger time: 2026-01-08 09:15;

[0111] Grid positioning: G(8,12) (corresponding to S-0812 / W-0812);

[0112] Main pollutants: Cd, Pb;

[0113] Synergistic pollution risk indicator: R=0.71 (high-risk area);

[0114] Predicted probability of exceeding the threshold: P(R≥0.65)=0.78 (for the next 7 days);

[0115] Summary of triggering causes (system generated): 1. The apparent migration rate from soil to groundwater increased by approximately +32% compared to the previous period; 2. The Cd response in groundwater was significant after rain (24h) (from 4.9 to 6.8 µg / L), with a lag of approximately 18h; 3. Cd and Pb showed a synchronous upward trend in both soil and groundwater (increased consistency index).

[0116] Warning level: High (Immediate action recommended).

[0117] The environmental management terminal includes desktop workstations connected to a central data server cluster. Each workstation is equipped with a monitor, a human-computer interface, and a geographic information system (GIS) client to receive collaborative pollution early warning information. Based on the received collaborative pollution early warning information and collaborative pollution risk zoning results, the environmental management terminal generates corresponding environmental management recommendations. These recommendations include: recommendations for adjusting monitoring frequency, delineating key monitoring areas, prioritizing soil remediation, and restricting groundwater extraction. These recommendations assist in decision-making regarding heavy metal pollution prevention and control and environmental management in karst areas. The environmental management terminal also visualizes the collaborative pollution risk zoning results in the form of a gridded collaborative pollution risk zoning map, pollution evolution curves at typical monitoring points, and statistical analysis reports, thus providing intuitive technical support for decision-making regarding heavy metal pollution prevention and control and environmental management in karst areas.

[0118] Figure 2 is a gridded collaborative pollution risk zoning map proposed in this embodiment; in Figure 2: horizontal axis: grid column index, vertical axis: grid row index; the four regions are the safe zone, the controllable risk zone, the key prevention and control zone and the high-risk zone; the high-risk zone in the figure is locally clustered and is used to determine the key monitoring areas and priority treatment units.

[0119] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. An AI-based soil-groundwater co-contamination assessment system for karst areas, characterized in that: The system includes an on-site monitoring subsystem, a central data server cluster, and an environmental management terminal. The on-site monitoring subsystem collects soil-groundwater heavy metal co-monitoring data and heavy metal content analysis results. The edge acquisition terminal performs timestamp alignment, basic noise reduction, and local caching processing on the soil-groundwater heavy metal co-monitoring data to obtain soil-groundwater heavy metal pollution acquisition data. The central data server cluster includes: a data preprocessing module, a feature construction module, a co-polluting assessment module, and a risk zoning module. The data preprocessing module calibrates soil-groundwater heavy metal pollution data based on heavy metal content analysis results, constructing a soil-groundwater heavy metal time-series dataset. The feature construction module constructs a multi-channel input feature set based on the soil-groundwater heavy metal time-series dataset. The co-polluting assessment module constructs a rotationally enhanced Chronos-T5 model; it introduces a composite dynamic loss function as a model training and prediction constraint mechanism, and performs time-series encoding and joint modeling of the multi-channel input feature set based on the rotationally enhanced Chronos-T5 model to obtain the distribution of heavy metal content in soil and the concentration distribution of heavy metals in groundwater, and outputs a co-polluting risk index. The partitioning module generates a collaborative pollution early warning message and pushes it to the environmental management terminal when the collaborative pollution risk index exceeds a preset threshold. The environmental management terminal receives and displays the collaborative pollution early warning message. The composite dynamic loss function is constructed as follows: a cross-entropy loss function is constructed, and a time oscillation weighting factor based on a triangular wave is introduced to dynamically modulate the category gradient contribution and loss landscape curvature characteristics of the cross-entropy loss function, resulting in a dynamic cross-entropy loss function; a mean square error loss function is constructed, and a time oscillation weighting factor based on a square wave is introduced to time-varyingly modulate the target side of the mean square error loss function, resulting in a dynamic mean square error loss function; a composite dynamic loss function is constructed by combining the dynamic cross-entropy loss function and the dynamic mean square error loss function; the formula for the dynamic cross-entropy loss function based on the time oscillation weighting factor based on a triangular wave is introduced. ， ;in, Indicates training steps Next, the Time oscillation weighting factor corresponding to each prediction channel; Indicates the oscillation amplitude coefficient; Represents the trigonometric wave function. Indicates the total number of categories; The period length of the time oscillation weighting factor is indicated. Indicates the first Phase offset corresponding to each prediction channel; ;in, Indicates training steps The calculated dynamic cross-entropy loss; This indicates the total number of training entries in the current batch. Indicates the index of the training entry. Indicates the first The input vector of each training entry. Indicates the first The actual contamination state label corresponding to each training item Indicates training steps Model parameters at time 10:00 Indicates the indexed result; This represents the summation of the exponentialized results; a dynamic mean square error loss function based on a square wave time oscillation weighting factor is introduced: ;in, Indicating in training steps Next, the Each prediction channel corresponds to a time-oscillation weighting factor based on square waves; Indicates the amplitude coefficient of the square wave; Represents a square wave function; ;in, Indicates training steps The calculated dynamic mean square error loss; Indicates to The output of the first The predicted values ​​for each prediction channel, Indicates the first The one-hot tag of the first entry is in the... Components in each prediction channel dimension; composite dynamic loss function: ;in, Indicates training steps The calculated composite dynamic loss, 、 These represent the trade-off coefficients; Represents the regularization coefficient. This represents the regularization term.

2. The AI-based soil-groundwater co-contamination assessment system for karst areas according to claim 1, characterized in that: The multi-channel input feature set includes migration intensity feature data, heavy metal response coupling feature data, pollution stability feature data, cooperative migration consistency feature data, and time-delayed transport feature data.

3. The AI-based soil-groundwater co-contamination assessment system for karst areas according to claim 1, characterized in that: The Chronos-T5 model, enhanced by rotational parameters, performs temporal encoding and joint modeling on a multi-channel input feature set to obtain the distribution of heavy metal content in soil and the concentration of heavy metals in groundwater, and outputs a synergistic pollution risk index. The process includes the following steps: Step S1: Normalize and quantize the multi-channel input feature set to form a token sequence; Step S2: Build a Chronos-T5 model, input the token sequence into the encoder part of the Chronos-T5 model, and use a multi-head self-attention mechanism and residual connection structure to perform contextual modeling of historical time segments and extract cross-time period latent feature representations; Step S3: Based on the cross-time period latent feature representations, perform efficient parameter adaptation processing on the Chronos-T5 model based on rotational degrees of freedom. Specifically, this includes introducing an efficient parameter adaptation mechanism based on cross-layer joint decomposition and rotational degrees of freedom to optimize the Chronos-T5 model. In the T5 model, the linear mapping weight parameters of the encoder and decoder are optimized by adjusting the axial strength and sparse rotational update to construct a rotational parameter-enhanced Chronos-T5 model; Step S4: The cross-time period latent feature representation is input into the decoder part of the rotational parameter-enhanced Chronos-T5 model, and the output is predicted step by step using an autoregressive sampling mechanism; The probability distribution of the predicted output is constrained according to the composite dynamic loss function to generate a predicted token sequence, and the predicted trajectory set is obtained by sampling the predicted output multiple times; Step S5: The predicted token sequence is dequantized and descaled to restore it to a continuous real value prediction sequence, and the dequantized prediction sequence is obtained; Statistical features are calculated based on the predicted trajectory set; Step S6: The dequantized prediction sequence and statistical features are summarized to determine the distribution of heavy metal content in soil and the distribution of heavy metal concentration in groundwater as the prediction results, and the synergistic pollution risk index is calculated.

4. The AI-based soil-groundwater co-contamination assessment system for karst areas according to claim 3, characterized in that: Step S3 specifically includes the following steps: Step S31: Based on the cross-time period latent feature representation, extract the linear mapping weight parameters corresponding to the encoder and decoder in the Chronos-T5 model; perform cross-layer weight projection and structure alignment processing on each linear mapping weight parameter to obtain the aligned cross-layer weight projection result; Step S32: Based on the aligned cross-layer weight projection result, perform cross-layer joint decomposition processing on the linear mapping weight parameters of multiple layers in the Chronos-T5 model; during the joint decomposition process, extract the latent orthogonal basis to form a cross-layer shared latent orthogonal basis; Step S33: Under the constraint of the cross-layer shared latent orthogonal basis, process the remaining components of the linear mapping weight parameters of each layer... Perform orthogonal decomposition to obtain axial strength parameters; based on the axial strength parameters, perform strength adjustment along the shared potential direction on the linear mapping weight parameters of each layer to form axial strength update results; Step S34: Based on the axial strength parameters, introduce a sparse non-axial perturbation matrix to establish a restricted cross-coupling relationship between different potential dimensions under the shared potential orthogonal basis across layers; based on the sparse non-axial perturbation matrix, perform non-axial perturbation update to form rotational update results; Step S35: Combine the axial strength update results with the rotational update results to replace the corresponding linear mapping weight parameters in the Chronos-T5 model to obtain the rotationally enhanced Chronos-T5 model.

Images