An edge computing and cloud-based collaborative method for environmental monitoring at soil remediation sites

Abstract

This invention relates to the field of soil environmental monitoring technology and discloses an edge computing-cloud collaborative method for environmental monitoring of soil remediation sites. The method includes deploying edge computing nodes at the remediation site to collect real-time data on soil moisture, heavy metal concentration, volatile organic compound content, and meteorological parameters through multi-source sensors; inputting the data into a spatiotemporal feature extraction network to generate a spatiotemporal fusion feature matrix using hierarchical convolution and cross-channel attention mechanisms; based on this matrix, using a deep probabilistic network to calculate the probability of index anomalies and generating a risk feature sequence with uncertainty measurement; constructing a dynamic risk field model based on the risk feature sequence and predicting pollution diffusion paths using a spatiotemporal propagation algorithm; and combining the equipment operating status with a collaborative decision engine to generate edge-side local control commands and cloud-based global optimization strategies to meet the needs of efficient and precise environmental management at the site.

Description

Technical Field

[0001] This invention relates to the field of soil environmental monitoring technology, specifically to a method for monitoring the environmental conditions of soil remediation sites using edge computing and cloud collaboration. Background Technology

[0002] With the acceleration of industrial development, soil pollution problems are becoming increasingly prominent, making soil remediation projects increasingly important in the field of ecological and environmental protection. Currently, environmental monitoring at soil remediation sites mainly relies on traditional manual sampling and laboratory analysis, a method with significant limitations. Manual sampling often requires staff to enter the remediation site, facing not only the risk of pollution exposure but also a low sampling frequency. This makes it difficult to continuously monitor key indicators such as soil moisture, heavy metal concentration, and volatile organic compound content, resulting in data gaps over time and an inability to accurately reflect the dynamic changes in pollution levels.

[0003] In terms of data processing, traditional monitoring methods mostly adopt a centralized data processing model, requiring the transmission of large amounts of monitoring data to a remote cloud for analysis and calculation. However, soil remediation sites are usually located in remote areas with poor network communication conditions, and data transmission is prone to delays and packet loss, severely affecting the real-time performance of data processing. Furthermore, under the centralized processing model, the cloud server bears a huge data processing load; when multiple remediation sites transmit data simultaneously, it can easily cause server overload, further reducing data processing efficiency.

[0004] Existing monitoring systems often focus on analyzing data at the level of independent judgment of single indicators, lacking in-depth exploration of the correlations between data across time and space. Soil pollution spread exhibits distinct temporal and spatial characteristics, and complex interactions exist between different monitoring indicators. Independent analysis of a single indicator cannot fully reflect the true state and spread trend of pollution, leading to inaccurate assessments of pollution risk. Furthermore, existing systems struggle to effectively integrate monitoring data with remediation process parameters and equipment operating status, failing to adjust remediation strategies promptly based on real-time monitoring results. This lack of dynamic control over the remediation process not only affects remediation effectiveness but may also result in resource waste.

[0005] In terms of risk prediction and decision-making, traditional methods lack scientific model support, and predictions of pollution diffusion paths are mostly based on empirical judgments, resulting in low accuracy and reliability. Simultaneously, there is a lack of effective coordination mechanisms between the edge and cloud environments. The edge cannot quickly respond to local monitoring data, and the cloud struggles to develop optimization strategies by integrating global data. This leads to a disconnect between the entire monitoring and decision-making process, failing to meet the needs of soil remediation projects for efficient and precise environmental management. Summary of the Invention

[0006] The purpose of this invention is to provide an edge computing and cloud-based collaborative method for environmental monitoring of soil remediation sites, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for environmental monitoring of soil remediation sites using edge computing and cloud collaboration, the method comprising:

[0008] Edge computing nodes are deployed at soil remediation sites to collect real-time monitoring data on soil moisture, heavy metal concentration, volatile organic compound content, and meteorological parameters through multi-source environmental sensors.

[0009] Real-time monitoring data is input into a spatiotemporal feature extraction network, and a spatiotemporal fusion feature matrix is ​​generated through hierarchical convolution and cross-channel attention mechanism.

[0010] Based on the spatiotemporal fusion feature matrix, the anomaly probability of each monitoring indicator is calculated using a deep probabilistic network to generate a risk feature sequence with uncertainty measurement.

[0011] A dynamic risk field model is constructed based on the risk characteristic sequence, and the pollution diffusion path is predicted through a spatiotemporal propagation algorithm.

[0012] By combining a dynamic risk field model with the operational status data of repair equipment, a collaborative decision engine is used to generate local control commands at the edge and global optimization strategies in the cloud.

[0013] Preferably, the spatiotemporal feature extraction network performs the following operations:

[0014] Time alignment and normalization are performed on real-time monitoring data collected by multi-source environmental sensors to eliminate dimensional differences between sensors.

[0015] A three-dimensional convolutional kernel is used to extract the temporal continuity and spatial correlation of the monitoring data, and output the initial spatiotemporal feature block.

[0016] Inter-channel feature enhancement is performed on the initial spatiotemporal feature block using a cross-channel attention weight matrix to suppress noise interference;

[0017] The enhanced feature blocks are input into the residual compression module, and after removing redundant information, the spatiotemporal fusion feature matrix is ​​output.

[0018] Preferably, the deep probabilistic network performs the following operations:

[0019] The spatiotemporal fusion feature matrix is ​​divided into feature subsequences within a continuous time window;

[0020] For each feature subsequence, calculate its probability density distribution in the high-dimensional latent space;

[0021] The probability of anomalies in the current monitoring indicators is quantified based on the probability density distribution and the KL divergence of the historical normal operating condition database.

[0022] By integrating the anomaly probability and uncertainty measure of each time window, a risk feature sequence is generated.

[0023] Preferably, the construction process of the dynamic risk field model includes:

[0024] Analyze the probability of soil moisture and heavy metal concentration anomalies in the risk feature sequence and map them to the three-dimensional geographic coordinates of the construction site;

[0025] An adaptive interpolation algorithm is used to generate the spatial concentration gradient field of pollutants;

[0026] By combining wind direction and speed data from meteorological parameters, the concentration gradient field is corrected through fluid dynamics simulation.

[0027] The output is a dynamic risk field model that includes the time dimension.

[0028] Preferably, the execution steps of the spatiotemporal propagation algorithm are as follows:

[0029] The initial diffusion direction and rate of pollutants are extracted based on the dynamic risk field model;

[0030] Based on soil permeability coefficient and groundwater flow direction data, the migration trajectory of pollutants in the unsaturated zone is calculated.

[0031] Predict pollution diffusion paths within future time intervals using particle tracking models;

[0032] The prediction results are compared with the boundary conditions of the repair area to generate an accessibility analysis report for the diffusion path.

[0033] Preferably, the operating logic of the collaborative decision-making engine includes:

[0034] Receive the pollutant diffusion path prediction results output by the dynamic risk field model;

[0035] Obtain the reagent dosage, stirring frequency, and energy consumption parameters from the operating status data of the repair equipment;

[0036] Real-time control commands are generated at the edge to target local pollution hotspots;

[0037] After integrating data from multiple edge nodes in the cloud, a global optimization strategy for repair process parameters is output.

[0038] Preferably, the generation process of the edge-side local control command is as follows:

[0039] Compare the spatial matching degree between the current pollutant concentration gradient and the coverage area of ​​the remediation equipment;

[0040] Adjust the start-stop sequence and spray angle of the pesticide spraying device according to the matching degree deviation;

[0041] Dynamically adjust the pesticide penetration depth control parameters based on soil moisture data;

[0042] Generate an instruction set that includes the device number, control parameters, and execution timestamp.

[0043] Preferably, the generation process of the cloud-based global optimization strategy is as follows:

[0044] Aggregate the pollutant diffusion path prediction results and equipment operation logs uploaded by each edge node;

[0045] Construct a multi-objective optimization function for repair effects and energy consumption;

[0046] The optimal combination of process parameters is calculated using a distributed solution algorithm;

[0047] The output includes a strategy file containing drug ratios, equipment scheduling plans, and monitoring frequency adjustments.

[0048] Preferably, the method further includes a real-time feedback calibration mechanism:

[0049] Collect actual pollutant concentration change data after the remediation equipment executes control commands, calculate the error rate between the actual concentration change and the prediction model, and trigger online update of model parameters when the error rate exceeds the threshold;

[0050] The updated parameters will be synchronized to the edge nodes and the cloud decision-making layer.

[0051] Preferably, the specific steps for online updating of the model parameters are as follows:

[0052] Extract the original monitoring data and equipment response records within the time interval when the error rate exceeds the standard;

[0053] Retrain the cross-channel attention weight matrix in the spatiotemporal feature extraction network;

[0054] Adjust the way the latent space dimensions are partitioned in deep probabilistic networks;

[0055] Update the interpolation algorithm parameters and hydrodynamic coefficients of the dynamic risk field model.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] By deploying edge computing nodes at soil remediation sites and using multi-source environmental sensors to collect multi-dimensional real-time monitoring data, the traditional manual sampling and monitoring mode has been changed. This enables continuous and comprehensive monitoring of soil moisture, heavy metal concentration, volatile organic compound content, and meteorological parameters. This allows staff to obtain accurate monitoring data without frequently entering the contaminated site, reducing the risk of personnel exposure to the contaminated environment. It also avoids the time gap problem caused by manual sampling, enabling the monitoring data to more accurately reflect the dynamic changes in the pollution situation.

[0058] In the data processing stage, real-time monitoring data is input into a spatiotemporal feature extraction network. Hierarchical convolution and cross-channel attention mechanisms are used to generate a spatiotemporal fusion feature matrix, overcoming the limitations of traditional independent analysis of single indicators. Hierarchical convolution effectively uncovers the spatial correlation information of monitoring data, while the cross-channel attention mechanism highlights the important correlation features between different monitoring indicators. The spatiotemporal fusion feature matrix formed by the combination of these two mechanisms comprehensively presents the complex relationships of monitoring data in the spatiotemporal dimension, providing a more comprehensive and accurate data foundation for subsequent risk assessment. This allows the analysis of pollution status to move beyond isolated numerical values ​​and grasp the patterns of pollution change from a holistic spatiotemporal perspective.

[0059] Based on the spatiotemporal fusion feature matrix, deep probabilistic networks are used to calculate the anomaly probability of each monitoring indicator and generate a risk feature sequence with uncertainty measurement. Compared with the traditional experience-based risk assessment method, deep probabilistic networks can conduct in-depth analysis of monitoring data through complex algorithm models, accurately identify the anomalies of each indicator, and the introduction of uncertainty measurement can enable staff to clearly understand the reliability of the risk assessment results, providing a more scientific reference for subsequent decision-making and avoiding decision-making errors caused by the bias of a single assessment result.

[0060] A dynamic risk field model is constructed based on risk characteristic sequences, and a spatiotemporal propagation algorithm is used to predict pollution diffusion paths, providing an effective means for early prevention and control of pollution risks. The dynamic risk field model can integrate monitoring data and process parameters in real time to accurately simulate the diffusion process of pollution in the spatiotemporal dimensions. The spatiotemporal propagation algorithm can quickly and accurately predict the path and range of pollution diffusion. Based on these prediction results, staff can take preventive measures in advance to prevent further pollution spread and reduce the harm caused by pollution.

[0061] By combining a dynamic risk field model with operational data from remediation equipment, a collaborative decision engine generates local control commands at the edge and global optimization strategies in the cloud, achieving efficient collaboration between the edge and cloud. The edge can quickly generate local control commands based on local monitoring data and equipment operational status, adjusting the operating parameters of remediation equipment within the area to address localized pollution issues. The cloud, on the other hand, combines monitoring data from all edge nodes, the risk model, and the overall remediation progress to formulate global optimization strategies, coordinating the work at various remediation sites, avoiding resource waste, and improving overall remediation efficiency. This collaborative mechanism breaks down information barriers between the edge and cloud, creating a complete closed loop of monitoring, analysis, decision-making, and control. This makes environmental management of soil remediation projects more efficient and precise, better adapting to the complex and ever-changing needs of remediation sites. Attached Figure Description

[0062] Figure 1 This is a schematic diagram illustrating the working principle of the edge computing cloud-based collaborative method for monitoring the environmental conditions at soil remediation sites as described in this invention.

[0063] Figure 2 A flowchart illustrating the operation of a spatiotemporal feature extraction network;

[0064] Figure 3 Flowchart for constructing a dynamic risk field model;

[0065] Figure 4 This is a diagram illustrating the spatiotemporal evolution and dynamic parameters of pollutant diffusion. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Please see Figure 1This invention provides an edge computing-cloud collaborative method for environmental monitoring at soil remediation sites. The method includes: deploying edge computing nodes at the soil remediation site and using multi-source environmental sensors to collect real-time monitoring data on soil moisture, heavy metal concentration, volatile organic compound content, and meteorological parameters. These data are input into a spatiotemporal feature extraction network, which processes the data using hierarchical convolution and cross-channel attention mechanisms to generate a spatiotemporal fusion feature matrix. Based on this matrix, a deep probabilistic network is used to calculate the anomaly probability of each monitoring indicator and generate a risk feature sequence with uncertainty measurement. A dynamic risk field model is constructed based on the risk feature sequence. A spatiotemporal propagation algorithm is used to predict the pollution diffusion path. Finally, combining the dynamic risk field model with the operating status data of the remediation equipment, a collaborative decision engine is used to generate edge-side local control instructions and cloud-side global optimization strategies to achieve efficient and accurate environmental monitoring and control.

[0068] Example 1: See Figure 2 The operation of the spatiotemporal feature extraction network begins with the preprocessing stage of the raw data collected by multi-source environmental sensors. These sensors are deployed at different spatial locations on the construction site, continuously generating readings on soil moisture, heavy metal concentration, volatile organic compound content, and meteorological parameters. These data streams are not completely synchronized in terms of timestamps and each has different dimensions and numerical ranges. The time alignment operation uses a unified clock signal to buffer and resample all input data streams, so that each data point corresponds to a precise moment on a global time axis. Normalization is performed independently for each sensor data type. For example, the soil moisture percentage is scaled to the [0,1] range, and the heavy metal concentration ppm value is processed using the maximum-minimum normalization method. This aims to eliminate numerical differences caused by different physical units of the sensors and create a dimensionally unified numerical environment for subsequent feature extraction. The preprocessed data is organized into a multidimensional data cube with three dimensions: time, space, and channel. The core component of the spatiotemporal feature extraction network, the three-dimensional convolutional kernel, begins to perform sliding window calculations on this data cube. The three-dimensional convolutional kernel moves not only in the two-dimensional plane space (i.e., the latitude and longitude coordinates of the construction site) but also in the time dimension, thereby capturing the spatial correlation and temporal continuity of the monitoring data. For example, a larger convolutional kernel with a span in the time dimension can perceive the slow changing trend of pollutant concentration over several hours, while a smaller convolutional kernel may be sensitive to sudden fluctuations within a few minutes. The output of the convolution operation generates an initial spatiotemporal feature block containing a variety of primary features, which preserves the local patterns of the data in spatiotemporal space.

[0069] While the initial spatiotemporal feature block contains rich information, it is also mixed with various noises from the sensor itself or the environment. Furthermore, the importance of features from different channels (i.e., different monitoring indicators) is not equal. A cross-channel attention mechanism is introduced to address this issue. This mechanism performs global pooling on all channel values ​​of the initial feature block at each spatial-temporal location, generating a vector representing global information. Subsequently, a set of attention weights is learned through a small feedforward neural network. This set of weights constitutes a cross-channel attention weight matrix, where each element reflects the importance of the corresponding feature channel. During the feature enhancement stage, each channel of the initial spatiotemporal feature block is multiplied by its corresponding attention weight. Feature channels that are highly correlated with abnormal pollution events are given higher weights and thus enhanced, while some background noise or channels with gradual changes are relatively suppressed. Even after attention-enhanced feature blocks, their dimensionality may still be high and contain redundant information. The residual compression module is responsible for refining and reducing their dimensionality. This module typically consists of several consecutive residual convolutional layers. Each residual layer contains a 3D convolution, an activation function, and a skip connection. The skip connection directly adds the module's input to the output, effectively alleviating the gradient vanishing problem in deep networks. This allows the network to be built deeper and learn more complex features. Convolutional operations gradually fuse feature information from adjacent spatiotemporal regions. At the same time, pooling layers or stride convolution operations reduce the spatial resolution of the feature map, thereby compressing the amount of data while retaining the most critical information. Finally, this module outputs a highly condensed spatiotemporal fusion feature matrix. This matrix is ​​a high-level summary of the useful spatiotemporal patterns in the original monitoring data, laying a solid foundation for subsequent probability assessment.

[0070] The deep probabilistic network receives the aforementioned spatiotemporal fusion feature matrix as input. Its primary task is to segment this continuous matrix stream into a series of overlapping or continuous feature subsequences within time windows. The length of each time window needs to be set according to the specific monitoring scenario and decision-making timeliness requirements; for example, it might be set to include subsequences containing data from the past hour. The sliding step size of the window determines the temporal resolution of the risk assessment. The segmentation operation transforms continuous spatiotemporal information into a series of discrete, independently analyzable time segments, facilitating serialized anomaly detection. For each feature subsequence, the goal of the deep probabilistic network is to estimate its probability density distribution in a certain high-dimensional latent space. This is usually achieved by constructing a probabilistic model, such as a deep Bayesian network or a normalized flow model. This model maps the input feature subsequence to a predefined latent space and fits its probability distribution within this space. During the training phase, the model uses a large amount of data collected under historical normal operating conditions for training, thus learning the distribution pattern of "normal" data in this latent space. Once training is complete, for a newly incoming feature subsequence, the model can calculate the probability that it belongs to this "normal" distribution. To quantify the probability of anomalies, it is necessary to calculate the difference between the probability distribution of the current feature subsequence and the baseline distribution represented by the historical normal operating condition database. KL divergence is a mathematical tool that measures the degree of difference between two probability distributions. The more similar the distribution of the current subsequence is to the historical normal distribution, the smaller the KL divergence value, indicating that the current operating condition is more likely to be normal. Conversely, when the KL divergence value increases, it means that the current data shows a significant deviation from the historical normal pattern, suggesting an increased probability of anomalies. This KL divergence value, after being transformed by a calibration function (Sigmoid function), yields an anomaly probability value between 0 and 1. Specifically, deep probabilistic networks calculate the probability density distribution of feature subsequences in a high-dimensional latent space and compare it with the distribution of the historical normal operating condition database, using KL divergence to quantify the difference. The KL divergence value itself is a non-negative real number, reflecting the degree of deviation between the current data distribution and the historical normal distribution; the larger the value, the higher the probability of anomalies. To convert the KL divergence values ​​into anomaly probability values ​​between 0 and 1 for easier subsequent risk decision-making, the network uses the Sigmoid function as a calibration function. The Sigmoid function is an S-shaped curve function that maps any real-valued input to the (0,1) interval, and the output value can be interpreted as a probability. Specifically, the network normalizes the KL divergence values ​​to eliminate the influence of dimensions and ensure they fit the input range of the Sigmoid function. The Sigmoid function is applied for transformation: an output value close to 0 indicates a low anomaly probability (i.e., similar distributions), while a value close to 1 indicates a high anomaly probability (i.e., large distribution differences).This transformation process also incorporates uncertainty metrics, calibrating the parameters of the sigmoid function (such as slope and translation) to maintain its stability across different monitoring scenarios, ensuring the reliability and interpretability of the anomaly probability values. The generated risk feature sequence not only contains anomaly probabilities but also includes uncertainty information, providing accurate input for dynamic risk field models.

[0071] Simple anomaly probability estimation may contain errors due to model uncertainty. Therefore, introducing uncertainty measures is crucial. For probabilistic models, the output itself can contain uncertainty information. For example, in Bayesian neural networks, the variance or confidence interval of the predicted value can be estimated through Monte Carlo Dropout or ensemble learning. This uncertainty measure reflects the model's confidence in the anomaly probability value it gives. High uncertainty suggests that the current judgment may be unreliable and requires caution. Finally, the anomaly probability calculated for each time window is fused with its corresponding importance measure to form a complete risk feature sequence. This sequence not only marks the probability of an anomaly occurring at each time point but also adds the reliability of the judgment, providing richer and more reliable information input for constructing a dynamic risk field.

[0072] Example 2: See Figure 3The construction of the dynamic risk field model begins with the analysis of the risk feature sequence, which is derived from the output of the deep probabilistic network. This sequence contains the anomaly probability values ​​of each monitoring point in the time dimension. The model analysis focuses on the anomaly probabilities of soil moisture and heavy metal concentration, as these two indicators are directly related to the migration capacity of pollutants in the environment and ecological risks. The analysis operation requires mapping the abstract anomaly probability values ​​to their locations in the real physical space. Each probability value is accompanied by the geographic coordinate information of its acquisition sensor, including latitude, longitude, and altitude. Through the coordinate mapping interface, these probability data with spatial attributes are transformed into the three-dimensional geographic coordinate system of the construction site, forming a discrete set of spatial points. Each point carries the risk intensity information of a specific location at a specific time. Based on a discrete set of spatial risk points, the model employs an adaptive interpolation algorithm to generate a continuous spatial concentration gradient field. Unlike traditional fixed-parameter interpolation methods, this algorithm automatically adjusts the interpolation radius and weighting function according to the spatial distribution density of the data in the point set. In areas with dense monitoring points, the interpolation algorithm uses a smaller neighborhood range to preserve local details, while in sparse edge areas, it appropriately expands the search radius to ensure the integrity of the interpolation surface. The algorithm iteratively calculates the distance-weighted average between each grid point and its surrounding monitoring points, while also considering the uncertainty measure of the anomaly probability of each point. Data points with lower reliability are assigned smaller weights, ultimately outputting a spatially continuous pollutant concentration gradient field covering the entire construction site area. This gradient field is presented in a grid format, with each grid cell containing a standardized concentration intensity value.

[0073] After generating the initial concentration gradient field, the model incorporates wind direction and speed data from meteorological parameters for correction. This is because atmospheric flow significantly affects the diffusion path and deposition pattern of surface pollutants. Wind direction data is used to determine the dominant direction of pollutant migration, while wind speed data affects the rate and extent of diffusion. The fluid dynamics simulation module is activated, which combines site elevation data, surface roughness parameters, and real-time meteorological data to construct a simplified near-surface airflow field model. The simulation process is based on computational fluid dynamics principles, considering factors such as air viscosity, pressure gradient, and topographic disturbance to calculate the distribution of the wind field in the three-dimensional space of the site. Subsequently, this wind field is coupled with the initial concentration gradient field for calculation. The convection effect of the wind field changes the spatial distribution of pollutants, causing the concentration gradient field to be stretched and shifted along the wind direction, thus obtaining a corrected concentration gradient field that better conforms to the actual physical propagation laws.

[0074] The concentration gradient field corrected by hydrodynamics is still a static snapshot. The dynamic risk field model introduces a time dimension to make it evolve. The model organizes multiple corrected concentration fields from a continuous time series in chronological order. Each time slice corresponds to a complete spatial concentration distribution. The time interval is consistent with the data acquisition frequency, forming a four-dimensional data volume structure. By analyzing the change patterns of the concentration field between adjacent time slices, the movement direction, diffusion speed, and intensity decay trend of the pollution plume can be identified. The final output dynamic risk field model is a spatiotemporally continuous data product. It can intuitively display the dynamic diffusion process of pollutants in the construction site and surrounding areas in the form of animation or sequence diagrams, providing accurate input for subsequent path prediction. The spatiotemporal propagation algorithm uses a dynamic risk field model as its core input. Its primary task is to extract the initial diffusion direction and rate of pollutants. The algorithm analyzes the concentration gradient field at the current time step and determines the direction of maximum gradient by calculating the spatial derivative of the concentration value at each grid point. This direction is identified as the main direction of instantaneous diffusion of pollutants. The calculation of the diffusion rate depends on the concentration field data of two consecutive time steps. By comparing the rate of change of concentration at a specific location with time and combining the grid spacing information, the movement speed of the pollutant front is derived. These initial parameters provide a starting point for subsequent detailed simulations.

[0075] After obtaining the initial diffusion parameters, the algorithm further integrates soil permeability coefficient and groundwater flow direction data to simulate the migration trajectory of pollutants in the unsaturated zone. The soil permeability coefficient comes from the pre-construction geological survey report and is marked in the form of a spatial distribution map to indicate the water conductivity of different soil layers. The groundwater flow direction data is obtained by periodically measuring through monitoring wells placed around the site. The algorithm couples the surface concentration field with the underground soil parameters in three dimensions and uses numerical methods to solve the transport equation of pollutants in porous media. The simulation process considers the combined effects of molecular diffusion, mechanical dispersion, and convection with water flow. For volatile organic compounds, a simplified calculation of the gas phase mass transfer process is also added to depict the complex path of pollutants infiltrating from the surface to the ground and moving with soil moisture. Based on the results of the above physical process simulation, the algorithm initiates a particle tracking model to predict the pollution diffusion path in the future time interval. This model conceptualizes pollutants as a large number of discrete tracer particles, each carrying a certain mass attribute. The particles are deployed in the currently identified high-concentration pollution areas, and their trajectories are driven by the previously calculated three-dimensional flow velocity field. The algorithm updates the position of the particles in each time step and considers random walk processes to simulate diffusion phenomena. By running the model to simulate scenarios for the next few hours to days, a large set of particle movement paths can be obtained. By performing statistical cluster analysis on these paths, high-probability pollution diffusion channels and possible aggregation areas can be identified.

[0076] After the particle tracking simulation is completed, the algorithm compares and analyzes the predicted pollution diffusion path with the boundary conditions of the remediation area. The boundary conditions include the physical fence of the site, the groundwater monitoring red line, and the location information of downstream sensitive targets. The comparison process uses spatial overlay analysis technology to calculate the probability and time of the pollution path intersecting with these boundaries. For paths that may break through the boundaries, the algorithm assesses their environmental risk level and marks the key time periods and areas that require priority intervention. The final accessibility analysis report is presented in a combination of text and graphics, including a path probability distribution map, a boundary breakthrough time prediction table, and a list of risk hotspot areas, providing managers with intuitive and quantitative basis for decision-making.

[0077] See Figure 4 In the visualization architecture for predicting pollutant diffusion paths, the spatiotemporal evolution process is represented based on multi-time-point contour overlay technology. Specifically, the diffusion simulation output is represented as a scalar field of concentration distribution over discrete time series (5 / 10 / 15 / 20 hours), with the repaired region boundary box and sensitive targets embedded in the coordinate system as spatial constraints. Contour generation is achieved through radial basis function interpolation: spatial interpolation is performed on the concentration data at monitoring points to construct a continuous concentration surface, and closed contours at specific thresholds (0.2-0.6 mg / kg) are extracted. The annular diffusion model reveals the anisotropic migration pattern of pollutants from the pollution source to the surrounding area, and the contour spacing reflects the intensity of concentration gradient changes. The diffusion distance versus peak concentration time variation map adopts a dual-axis coordinate system design. The diffusion distance curve is fitted with a linear regression model to measure the spatial expansion rate, and the slope parameter characterizes the average diffusion velocity. The peak concentration curve is processed using a moving average filter to highlight the decay trend, and the right sub-axis is dimensionally scaled in milligrams per kilogram. During the quantification of kinetic parameters, the time axis is configured with full coverage sampling points from 0 to 25 hours, the diffusion distance axis uses a linear scale to indicate displacement increments, and the concentration axis uses logarithmic scaling to optimize the display accuracy of high-value areas. Cross-analysis of the two curves can deduce the coupling relationship between concentration decay and migration range during diffusion.

[0078] Example 3: The operation of the collaborative decision engine begins with receiving the pollutant diffusion path prediction results output by the dynamic risk field model. These results are continuously input in the form of a structured data stream. The data packet not only contains the sequence of predicted path coordinate points, but also includes the risk level estimate, concentration prediction value, and uncertainty measure of each path point. The engine's data interface layer is responsible for parsing this data, converting it into an internally unified spatiotemporal data format, and synchronizing it with a real-time clock to ensure that subsequent decisions are based on the latest prediction scenario. The path prediction results actually depict multiple possible future scenarios. The engine needs to quickly evaluate these scenarios and select paths with high probability of occurrence and serious consequences as key focus objects. While receiving diffusion path information, the engine acquires operational status data of the remediation equipment through the device communication network. The status data covers key operational indicators such as reagent dosage, stirring frequency, and energy consumption parameters. This data is collected in real time from sensors installed on the remediation equipment, and uploaded after preliminary aggregation by the edge gateway. The reagent dosage reflects the intensity of chemical remediation, the stirring frequency affects the uniformity of mixing between the reagent and the contaminated soil, and the energy consumption parameters are an important basis for assessing the economic efficiency and sustainability of the remediation operation. The engine's built-in device status analysis module verifies the integrity and rationality of the data, eliminates obviously abnormal readings, and standardizes the different data formats of different devices to prepare for subsequent collaborative analysis.

[0079] Based on the integrated path prediction and equipment status information, the collaborative decision engine generates real-time control instructions for local pollution hotspots at the edge. Edge-side decision-making emphasizes timeliness, requiring rapid response to sudden high-risk situations. The decision logic performs spatial overlay analysis of the latest pollution diffusion path with the physical area governed by the edge computing node to identify pollution plumes that are about to enter or are already in the area. Then, the engine retrieves the real-time status and spatial coverage of available remediation equipment in the area, calculates the degree of matching between the equipment's intervention capability and the pollution threat, and the instruction generation process involves a simple decision function designed to maximize the immediate risk control effect, in the following form:

[0080]

[0081] in: This represents the overall utility value of a decision, which needs to be maximized. This represents the total number of controllable repair devices at this edge node. It is aimed at the first The device is expected to reduce the local risk value by comparing the predicted pollutant concentrations under the two scenarios of device intervention and no intervention. Is to mobilize the first The cost incurred by the equipment includes the depreciated value of pharmaceutical consumption, energy consumption, and equipment depreciation. It is a normalized benchmark risk value. It is a normalized baseline cost. and These are risk weight and cost weight, respectively. These two weight coefficients are dynamically distributed by the cloud based on the global strategy. The core of edge-side decision-making is to solve for a set of equipment operating parameters (such as start-up and shutdown time, workload) under resource constraints, so that the utility value... To achieve optimal performance, the final real-time control command includes the specific device number, action type, parameter settings, and precise execution timestamp.

[0082] While rapidly responding locally at the edge, the cloud platform asynchronously generates a global optimization strategy. The cloud receives data packets uploaded from all edge nodes within the region, including their respective pollutant diffusion path predictions, executed control command records, equipment operation logs, and energy consumption details. The data aggregation module performs spatiotemporal alignment and redundancy removal on the multi-source data, eliminating potential duplicate calculations or information conflicts between different nodes, forming a unified global situational view covering the entire construction site. Based on this global view, the cloud decision-making module constructs a multi-objective optimization function for remediation effectiveness and energy consumption. This function is far more complex than the edge-side decision function, needing to consider cross-regional pollution transfer effects, the complementarity of equipment collaborative operations, and the balance between long-term remediation goals and short-term operating costs. Optimization objectives typically include minimizing the overall average pollutant concentration over a certain period, minimizing total energy consumption, and balancing the load on key equipment. This is a typical multi-objective optimization problem with complex constraints. Specifically, the cloud decision-making module constructs a multi-objective optimization function for remediation effectiveness and energy consumption based on the aggregated data from each edge node (including pollutant diffusion path predictions, equipment operation logs, and remediation process parameters). The function is built by cleaning and aligning the global view through a data integration module, forming a unified spatiotemporal data cube containing key variables such as pollutant concentration, equipment energy consumption, and load status. The optimization function design considers multiple conflicting objectives: the main objectives include minimizing the overall average pollutant concentration over a certain period (to maximize remediation effectiveness), minimizing total energy consumption (covering electricity, reagents, and equipment wear), and balancing the load of key equipment (avoiding local overload). The function incorporates complex constraints, such as cross-regional pollution transfer effects (simulating pollutant migration through a dynamic risk field model), the complementarity of equipment collaborative operations (e.g., the timing coordination of reagent spraying and mixing equipment), and the balance between long-term remediation goals and short-term operating costs (adjusted using trade-off weights). A distributed solution algorithm is used in the cloud to decompose the global problem into spatiotemporal sub-problems (e.g., divided by region or time slice), iteratively solving local optima through parallel computing nodes, and then integrating and processing boundary coupling effects (e.g., inter-regional pollution interactions) through a coordinator. The solution outputs the optimal combination of process parameters, including reagent ratio, equipment scheduling scheme and monitoring frequency adjustment, and distributes them to edge nodes in the form of a strategy file to ensure the coordination between local control and global optimization.

[0083] The cloud-based distributed algorithm calculates the optimal combination of process parameters. Due to the massive scale of the problem, centralized solutions are inefficient. Therefore, the algorithm decomposes the global problem into multiple sub-problems, such as by time period or geographical region, and assigns them to different computing nodes for parallel processing. After solving their local sub-problems, each computing node exchanges intermediate results through a coordinator, iteratively approximating the global optimum. The solution process not only considers the current state but also predicts pollution spread trends and equipment performance degradation over a future period, thus generating a forward-looking strategy. The final output global optimization strategy file is a comprehensive solution that details the recommended reagent ratios for each edge node within a future scheduling cycle, the scheduling scheme for key equipment (such as the optimal deployment location for mobile equipment), and the adjustment recommendations for monitoring frequencies in different areas. This strategy file is generated in a standardized configuration file format and distributed to each edge node through a secure channel. When generating subsequent local instructions, the decision engines of the edge nodes prioritize this global strategy framework, thereby achieving a synergy between rapid local response and the global optimal goal.

[0084] Example 4: Assume that in area A of a soil remediation site, an edge computing node detects a local pollution hotspot with an abnormally high concentration of the heavy metal cadmium. This hotspot is located near coordinates (X105, Y220). The concentration gradient shows that the pollution plume is spreading southeastward. The generation process of the edge-side local control command is then initiated. The system compares the spatial matching degree between the current pollutant concentration gradient and the coverage area of ​​the remediation equipment. Three intelligent chemical spraying devices are deployed in this area, numbered SP-01, SP-02 and SP-03. The system overlays the spatial distribution raster map of the pollutants with the effective spraying range polygon of each device and calculates the degree of coverage of each raster unit by the device. The matching degree deviation is quantified by the ratio of the spatial intersection area to the total area of ​​the pollution plume. The analysis shows that the location of device SP-02 can best cover the core area of ​​the pollution plume, but its current spray angle is biased to due south, which is about 15 degrees away from the direction of the pollution plume spreading southeastward.

[0085] Based on the analysis results of the matching degree deviation, the system began to adjust the start-up and shutdown sequence and spray angle of the chemical spraying device. For the SP-02 device, the system calculated that adjusting its spray direction by 15 degrees clockwise could better align it with the migration path of the pollutants. At the same time, in order to form an interception barrier, the system instructed the SP-03 device, which is located downstream, to start 5 seconds in advance to cover the edge area that the SP-02 device may not have fully covered after adjustment. The SP-01 device, which is located upstream and off the main path, was instructed not to start temporarily to save chemical consumption. The start-up and shutdown sequence was set based on the estimation of the movement speed of the pollution plume to ensure that the device is in working condition just when the pollutant front reaches its effective range. While adjusting the equipment's spatial operating parameters, the system dynamically corrects the pesticide penetration depth control parameters based on real-time collected soil moisture data. Sensor data shows that the soil volumetric water content in the target area is 22%, which is a medium humidity level. Excessive humidity may cause the pesticide to dilute and infiltrate prematurely, while excessively low humidity may hinder sufficient contact between the pesticide and the pollutants. Based on a preset soil moisture-penetration depth relationship model, the system increases the spraying pressure of SP-02 and SP-03 equipment from the standard value of 0.5 MPa to 0.55 MPa, and instructs to increase the concentration of the sprayed pesticide solution by 3% accordingly. The aim is to ensure that, under the existing humidity conditions, the remediation agent can reach the soil layer at a depth of 20-40 cm where the pollutants are mainly concentrated, and maintain a sufficient effective reaction concentration. Referring to Table 1, after completing all parameter calculations, the system generates an instruction set containing the equipment number, control parameters, and execution timestamp. This instruction set exists in a structured data format.

[0086] Table 1: Local Control Instruction Set on the Edge Side

[0087]

[0088] The instruction set is immediately distributed to the corresponding field controllers via the industrial wireless network. The equipment executes the above operations at the specified time, thereby achieving precise and rapid response to local pollution hotspots. While the edge nodes generate and execute local instructions, the generation process of the global optimization strategy in the cloud is also carried out in parallel. The cloud platform continuously aggregates data packets uploaded from all edge nodes in Area A, Area B of the construction site, and even the entire project. These data packets not only contain real-time pollutant diffusion path prediction results, but also include detailed equipment operation logs, such as the execution records of the SP series equipment in Area A, the speed and energy consumption of the soil mixing equipment in Area B, and the flow rate and inlet and outlet concentration monitoring values ​​of the extraction and treatment system in Area C. The data aggregation center cleans, aligns, and merges this massive amount of heterogeneous data to build a unified view of the remediation operation situation covering the entire area.

[0089] Based on this panoramic view, the cloud-based optimization engine constructs a multi-objective optimization function for remediation effect and energy consumption. The core objective of this function is not to pursue the ultimate removal of pollution in a single area, but to seek the best balance between overall remediation effect and total resource consumption throughout the entire project cycle and spatial scope. The effect indicator may be the rate of decrease in the average concentration of pollutants across the entire area within a certain period or the percentage of points that reach the remediation target value. The energy consumption indicator integrates electricity consumption, reagent consumption, equipment depreciation, and labor costs. The optimization function needs to handle complex constraints, such as the coordination requirements of remediation progress in different areas, the availability of working hours for key equipment, and the synergistic treatment effect between different pollutants. Faced with such a complex optimization problem, the cloud employs a distributed solution algorithm for computation. The algorithm decomposes the massive global problem into time slices and spatial blocks. For example, the optimization problem for the next 24 hours is decomposed into 96 sub-problems with 15-minute time slots. The construction site is divided into several logical regions and distributed to different nodes in the cloud computing cluster for parallel solution. Each node is responsible for calculating the approximate optimal solution within its assigned time slot and region. A central coordinator is responsible for aggregating these local solutions and handling the coupling relationships between them (e.g., the treatment of region A may affect the groundwater flow field of region B). Through multiple iterations and the exchange of boundary information, a feasible combination of process parameters that is close to the global optimum is finally obtained.

[0090] After the calculations are completed, a detailed strategy file is output from the cloud. This file is not an instantaneous command directly issued to the equipment for execution, but rather a framework plan guiding edge node decisions over a period of time. The strategy file uses standard formats such as XML or JSON, and its content may include: it is recommended that the overall chemical dosing intensity in Area A be maintained at 90% of the current level within the next 6 hours to reserve chemical supply for the upcoming high-intensity remediation work in Area B; it is recommended that the SP-01 device be put into sleep mode in the afternoon, as the forecast shows that the risk level of its main coverage area is low during this period, and power resources can be allocated to the more energy-intensive soil thermal desorption equipment; it is recommended to increase the sampling frequency of the groundwater monitoring wells in Area C from once per hour to once every half hour to more accurately monitor the cross-regional pollution migration effect. This strategy file is securely distributed to each edge node. When the collaborative decision-making engine on the edge side generates local control instructions, it will use this global strategy as the primary constraint to ensure that local actions are consistent with the overall goal.

[0091] Example 5: The operation of the real-time feedback calibration mechanism begins with continuous monitoring of the environmental response after the control command is executed on the repair equipment. This mechanism relies on a multi-source environmental sensor network deployed throughout the construction site. After the control command is generated and issued at the edge side, such as for the SP-02 and SP-03 equipment in the implementation case, the system does not enter a waiting state, but immediately starts tracking the effect of the command execution area. The soil moisture sensor, heavy metal concentration monitoring probe and other equipment located within the equipment's range of action collect environmental data at a preset high frequency (such as once every 30 seconds). This data is transmitted to the edge computing node in real time. The data processing module in the node quickly extracts the changing trends of key indicators, especially the instantaneous value and moving average value of the heavy metal cadmium concentration, which is directly related to the control target. The system calculates the error rate between the actual concentration change and the prediction model. This process involves aligning and comparing time series data. Taking the implementation case as an example, suppose that 10 minutes after the command is executed, the cadmium concentration data returned by monitoring point P-09 shows a decrease of 15%, while the dynamic risk field model previously predicted that the concentration decrease at that time point should be 22%. The calculation of the error rate is not a simple difference, but uses a weighted relative error method that takes into account the uncertainty of the data. The model prediction value itself comes with a confidence interval, and the actual measurement value also has sensor accuracy error. The error rate calculation module will integrate these uncertainty information to obtain a statistically significant error metric value, which reflects the degree of deviation between the model prediction and the actual environmental response. When the calculated error rate exceeds the preset threshold, the system automatically triggers the online update process of the model parameters. The threshold setting is not fixed, but dynamically adjusted according to the specific monitoring indicators, environmental background values, and remediation stage. For example, the warning threshold for key pollutants may be set more sensitive in the early stage of remediation, while in the later stable stage of remediation, the threshold can be appropriately relaxed to avoid excessively frequent model updates. In the case, assuming the threshold is set to a relative error of 15%, and the actual calculated error rate reaches 20%, the triggering condition is met, and the system will generate a model update event. This event marks the start time, spatial location, and monitoring indicators and model modules involved in the error occurrence.

[0092] After the update process is initiated, the first step is to extract the original monitoring data and equipment response records within the time interval of the error rate exceeding the standard. The system will backtrack all relevant data from a period of time before the error began to appear (e.g., half an hour) up to the current moment. This includes the original reading time series of all sensors in the area, complete copies of all control commands issued by the edge nodes, and the actual operating status logs of the controlled repair equipment (such as motor current, valve opening, etc., to verify that the commands were executed correctly). This data is packaged into a calibration dataset for subsequent model parameter retraining. The specific steps for online model parameter updates then unfold. The system optimizes the spatiotemporal feature extraction network, focusing on retraining the cross-channel attention weight matrix. The calibration dataset is input into the network to perform a forward propagation, calculating the loss between the current network output and the true observations. Through the backpropagation algorithm, the gradient of the loss function with respect to each parameter in the attention weight matrix is ​​calculated. Subsequently, optimization algorithms (such as a variant of stochastic gradient descent) are used to update these weight parameters. This aims to adjust the network's judgment on the importance of different sensor channels. For example, it may be found that insufficient attention was previously paid to a certain auxiliary meteorological parameter, resulting in a failure to fully capture the impact of wind on pollutant diffusion. Through this update, the network can better focus on the feature channels most relevant to the current error.

[0093] Adjusting the dimensional partitioning of the latent space in a deep probabilistic network (DPN) is crucial for evaluating anomaly probabilities. DPN maps high-dimensional features to the latent space, and the dimensional structure of the latent space (such as the meaning and scale of each dimension) directly affects the accuracy of probability density estimation. When systematic errors occur, it indicates that the current latent space structure may not be able to effectively capture the complex patterns of actual working conditions. The update process analyzes the distribution characteristics of error data in the existing latent space, identifies dimensions with insufficient discriminative power or redundancy, and adjusts the technical parameters of the encoder-decoder structure, such as changing the number of dimensions in the bottleneck layer or modifying the parameters of the prior distribution, to enable the latent space to more clearly separate "normal" and "abnormal" data patterns, thereby improving the reliability of subsequent anomaly probability calculations. The system updates the interpolation algorithm parameters and hydrodynamic coefficients of the dynamic risk field model. Parameters controlling smoothness and the range of the variogram in the interpolation algorithm directly affect the smoothness and detail of the generated concentration field. The system compares the concentration field output by the current interpolation algorithm with the spatially distributed measured concentration points in the calibration dataset. By minimizing the mean square error and other criteria, the system automatically adjusts the internal parameters of the interpolation algorithm to make its output more closely match the new observation data. Simultaneously, key coefficients used in the hydrodynamic simulation to describe the interaction between air and the surface, such as drag coefficients or diffusion coefficients, are fine-tuned based on the matching degree between wind speed and direction observation data and pollutant diffusion patterns, making the simulated diffusion process more physically realistic. After all model parameters are updated, the system synchronizes the updated parameter set to the edge nodes and the cloud decision layer. The synchronization process uses a version control mechanism to ensure that each update is traceable. Edge nodes may only receive local model parameter update packages related to their jurisdiction, such as updating the model parts affecting diffusion prediction in area A in this case, while the cloud receives complete, global model parameter updates.

[0094] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0095] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for environmental monitoring at soil remediation sites using edge computing and cloud collaboration, characterized in that, include: Edge computing nodes are deployed at soil remediation sites to collect real-time monitoring data on soil moisture, heavy metal concentration, volatile organic compound content, and meteorological parameters through multi-source environmental sensors. Real-time monitoring data is input into a spatiotemporal feature extraction network, and a spatiotemporal fusion feature matrix is ​​generated through hierarchical convolution and cross-channel attention mechanism. Based on the spatiotemporal fusion feature matrix, the anomaly probability of each monitoring indicator is calculated using a deep probabilistic network to generate a risk feature sequence with uncertainty measurement. A dynamic risk field model is constructed based on the risk characteristic sequence, and the pollution diffusion path is predicted through a spatiotemporal propagation algorithm. By combining a dynamic risk field model with the operating status data of repair equipment, a collaborative decision engine is used to generate local control instructions at the edge and global optimization strategies in the cloud. The construction process of the dynamic risk field model includes: The probability of soil moisture and heavy metal concentration anomalies in the risk feature sequence is analyzed and mapped to the three-dimensional geographic coordinates of the construction site to form a discrete set of spatial risk points. Based on a discrete set of spatial risk points, an adaptive interpolation algorithm is used to generate a spatial concentration gradient field for pollutants. By combining wind direction and speed data from meteorological parameters, the concentration gradient field is corrected through fluid dynamics simulation. Multiple modified concentration fields on a continuous time series are organized in chronological order to output a dynamic risk field model that includes the time dimension.

2. The method for monitoring the environmental conditions of a soil remediation site using edge computing and cloud collaboration as described in claim 1, characterized in that, The spatiotemporal feature extraction network performs the following operations: Time alignment and normalization are performed on real-time monitoring data collected by multi-source environmental sensors to eliminate dimensional differences between sensors. A three-dimensional convolutional kernel is used to extract the temporal continuity and spatial correlation of the monitoring data, and output the initial spatiotemporal feature block. Inter-channel feature enhancement is performed on the initial spatiotemporal feature block using a cross-channel attention weight matrix to suppress noise interference; The enhanced feature blocks are input into the residual compression module, and after removing redundant information, the spatiotemporal fusion feature matrix is ​​output.

3. The method for monitoring the environmental conditions of a soil remediation site using edge computing and cloud collaboration as described in claim 2, characterized in that, The deep probabilistic network performs the following operations: The spatiotemporal fusion feature matrix is ​​divided into feature subsequences within a continuous time window; For each feature subsequence, calculate its probability density distribution in the high-dimensional latent space; The probability of anomalies in the current monitoring indicators is quantified based on the probability density distribution and the KL divergence of the historical normal operating condition database. By integrating the anomaly probability and uncertainty measure of each time window, a risk feature sequence is generated.

4. The method for monitoring the environmental conditions of a soil remediation site using edge computing and cloud collaboration as described in claim 3, characterized in that, The execution steps of the spatiotemporal propagation algorithm are as follows: The initial diffusion direction and rate of pollutants are extracted based on the dynamic risk field model; Based on soil permeability coefficient and groundwater flow direction data, the migration trajectory of pollutants in the unsaturated zone was simulated. Based on the results of the above physical process simulation, the pollution diffusion path in future time intervals is predicted using a particle tracking model. The prediction results are compared with the boundary conditions of the repair area to generate an accessibility analysis report for the diffusion path.

5. The method for monitoring the environmental conditions of a soil remediation site using edge computing and cloud collaboration as described in claim 4, characterized in that, The operational logic of the collaborative decision-making engine includes: Receive the pollutant diffusion path prediction results output by the dynamic risk field model; Obtain the reagent dosage, stirring frequency, and energy consumption parameters from the operating status data of the repair equipment; Real-time control commands are generated at the edge to target local pollution hotspots; After integrating data from multiple edge nodes in the cloud, a global optimization strategy for repair process parameters is output.

6. The method for monitoring the environmental conditions of a soil remediation site using edge computing and cloud collaboration as described in claim 5, characterized in that, The generation process of the edge-side local control command is as follows: Compare the spatial matching degree between the current pollutant concentration gradient and the coverage area of ​​the remediation equipment; Adjust the start-stop sequence and spray angle of the pesticide spraying device according to the matching degree deviation; Dynamically adjust the pesticide penetration depth control parameters based on soil moisture data; Generate an instruction set that includes the device number, control parameters, and execution timestamp.

7. The method for monitoring the environmental conditions of a soil remediation site using edge computing and cloud collaboration as described in claim 6, characterized in that, The generation process of the cloud-based global optimization strategy is as follows: Aggregate the pollutant diffusion path prediction results and equipment operation logs uploaded by each edge node; Construct a multi-objective optimization function for repair effects and energy consumption; The optimal combination of process parameters is calculated using a distributed solution algorithm; The output includes a strategy file containing drug ratios, equipment scheduling plans, and monitoring frequency adjustments.

8. The method for monitoring the environmental conditions of a soil remediation site using edge computing and cloud collaboration as described in claim 7, characterized in that, It also includes a real-time feedback calibration mechanism: Collect actual pollutant concentration change data after the remediation equipment executes control commands, calculate the error rate between the actual concentration change and the prediction model, and trigger online update of model parameters when the error rate exceeds the threshold; The updated parameters are synchronized to the edge nodes and the cloud decision-making layer.

9. A method for monitoring the environmental conditions of a soil remediation site using edge computing and cloud collaboration as described in claim 8, characterized in that, The specific steps for online updating of the model parameters are as follows: Extract the original monitoring data and equipment response records within the time interval when the error rate exceeds the standard; Retrain the cross-channel attention weight matrix in the spatiotemporal feature extraction network; Adjust the way the latent space dimensions are partitioned in deep probabilistic networks; Update the interpolation algorithm parameters and hydrodynamic coefficients of the dynamic risk field model.

Images