Adjustable soil harmful component detection system based on environment monitoring

Abstract

The application discloses a controllable harmful component detection system in soil based on environmental monitoring and belongs to the technical field of soil environmental monitoring and remediation. The system comprises a data acquisition module, a feature construction module, a migration simulation module and a control execution module. The data acquisition module acquires the soil in-situ spectral reflectance sequence in the monitoring area and synchronously acquires the environmental temperature fluctuation and precipitation intensity data; the feature construction module fuses the above data to construct a soil spectral time sequence feature matrix; the migration simulation module inputs the matrix into a soil harmful component dynamic extraction model to simulate the migration and accumulation process of heavy metal ions in soil pore solution, and outputs the predicted heavy metal ion concentration gradient of different depth soil layers; and the control execution module generates a control instruction according to the difference between the concentration gradient and the safety threshold value to control the start and stop time and output power of the electromagnetic induction heating device and the ultrasonic assisted extraction device.

Description

Technical Field

[0001] This invention relates to the field of soil environmental monitoring and remediation technology, specifically to an adjustable soil harmful component detection system based on environmental monitoring. Background Technology

[0002] Chinese Patent Publication No. CN114443982A discloses a method and system for detecting and analyzing the spatiotemporal distribution characteristics of heavy metals in soil over a large area. This method acquires multi-source remote sensing images and ground sample data from different periods to construct a spatiotemporal distribution model, enabling the analysis of the spatial distribution and temporal trends of heavy metal content in soil at a regional scale. While this method reveals the spatiotemporal evolution of pollution distribution at a macroscopic scale, it has the following limitations: its technical approach focuses on statistical regression or spatial interpolation based on the spectral indices of remote sensing images. The resulting spatiotemporal distribution is essentially a spatial statistical extrapolation of static detection data from different time phases, failing to integrate real-time environmental driving factors such as ambient temperature fluctuations and precipitation infiltration into the dynamic simulation of the migration and accumulation process of heavy metal ions in soil pore solution at the physical mechanism level; its spatial resolution is limited by the pixel scale of remote sensing images, making it difficult to characterize the vertical concentration gradient changes along the soil depth direction within the monitoring point.

[0003] In the remediation phase, technologies such as electromagnetic induction heating to promote desorption and ultrasound-assisted extraction to accelerate mass transfer have been developed for heavy metal contaminated soils. In practical applications, the start-up and shutdown conditions and power output of these remediation devices are usually set according to preset fixed procedures or human experience, without dynamic matching based on real-time pollution migration risks on-site. Detection and remediation are two independent stages; even if detection reveals excessive concentrations at a certain moment, it cannot automatically trigger precise intervention as needed; this easily leads to energy waste from over-remediation or residual risks due to insufficient remediation intensity. Faced with the rapid leaching window of pollutants caused by natural precipitation, ultrasound extraction with fixed parameters cannot actively adapt to changes in pore fluid viscosity, weakening the cavitation effect's contribution to improving ion extraction efficiency. Summary of the Invention

[0004] The present invention aims to provide an adjustable soil harmful component detection system based on environmental monitoring, in order to solve the problem that existing technologies cannot integrate environmental factors to dynamically predict the migration of soil harmful components and adaptively regulate the remediation device.

[0005] The objective of this invention can be achieved through the following technical solutions: This invention relates to an adjustable soil hazardous component detection system based on environmental monitoring. This system, through the deep integration of multi-source environmental sensing and spectral analysis technologies, achieves accurate simulation and proactive intervention of the migration and accumulation process of hazardous components in soil, effectively improving the real-time performance and accuracy of environmental soil safety management.

[0006] As a technical solution of the present invention, the system includes a data acquisition module, which is used to acquire the in-situ spectral reflectance sequence of soil samples in the monitoring area within a preset sampling period, and simultaneously acquire the environmental temperature fluctuation data and environmental precipitation intensity data of the monitoring area within the preset sampling period. The feature construction module is used to construct a soil spectral temporal feature matrix that integrates environmental factors based on the in-situ spectral reflectance sequence, the environmental temperature fluctuation data, and the environmental precipitation intensity data. The migration simulation module is used to input the soil spectral time-series feature matrix into a pre-constructed dynamic extraction model of soil harmful components, simulate the migration and accumulation process of heavy metal ions in soil pore solution through the dynamic extraction model of soil harmful components, and output the predicted heavy metal ion concentration gradient of soil layers at different depths in the monitoring area. The control execution module is used to generate control commands based on the difference between the predicted heavy metal ion concentration gradient and the preset soil safety threshold. The control commands are used to control the start-up and shutdown times and output power of the electromagnetic induction heating device and the ultrasonic-assisted extraction device arranged in the monitoring area.

[0007] The data acquisition module can capture the transient change characteristics of harmful components under dynamic environmental conditions by collecting high-frequency, synchronized in-situ spectra and environmental parameters, thus avoiding information loss caused by time delay or environmental fluctuations from a single signal source.

[0008] Preferably, obtaining the in-situ spectral reflectance sequence of soil samples within the monitoring area during a preset sampling period includes: The system controls a multispectral sensor array deployed within the monitoring area to collect raw reflectance data at preset sampling time intervals. Multiple characteristic wavelength reflectance values ​​are extracted from the original reflectance data to form a characteristic reflectance vector for each sampling time. The feature reflectance vectors at each sampling time are arranged in the order of sampling to construct the in-situ spectral reflectance sequence. This method can efficiently compress sensitive band features from massive spectral information, reducing the dimensionality of subsequent processing while ensuring data representativeness.

[0009] Preferably, the synchronous acquisition of environmental temperature fluctuation data and environmental precipitation intensity data of the monitoring area within the preset sampling period includes: Read the time-series temperature data output by the temperature sensor arranged in the monitoring area and exposed to the atmosphere, and use it as the ambient temperature fluctuation data; Read the time-series data of precipitation intensity output by the rain gauges arranged in the monitoring area, and use it as the environmental precipitation intensity data; The temperature time-series data and the precipitation intensity time-series data are aligned to a unified time axis of the preset sampling period according to their respective timestamps. Thus, environmental factors and spectral characteristics are incorporated into the same time reference, laying the foundation for constructing multimodal fusion features.

[0010] Preferably, the construction of the soil spectral temporal feature matrix integrating environmental factors includes: performing wavelet decomposition on the in-situ spectral reflectance sequence to extract the intensity values ​​of spectral absorption characteristic peaks in different frequency bands; simultaneously, aligning the environmental temperature fluctuation data and environmental precipitation intensity data in time, and calculating the temperature change gradient and precipitation accumulation respectively; then, tensor concatenating the spectral absorption characteristic peak intensity values, temperature change gradient, and precipitation accumulation to generate the soil spectral temporal feature matrix. This matrix uniformly expresses spectral absorption characteristics, thermal driving force, and water replenishment, ensuring that the input features upon which the simulation of harmful component migration is based possess physical consistency and environmental sensitivity.

[0011] Preferably, during simulation, the soil spectral temporal feature matrix is ​​first divided into sub-feature matrices of multiple time steps, which are then sequentially input into the model's input layer, and each sub-feature matrix is ​​mapped to a migration potential scalar field. The migration potential scalar field is then convolved with a pre-defined soil pore structure topology network to obtain the ion flux distribution matrix corresponding to each time step. Finally, the ion flux distribution matrices are accumulated in chronological order to obtain the predicted heavy metal ion concentration gradient. This process simulates the gradual transport and enrichment effects of ions in porous media driven by temperature, moisture, and electrochemical potential fields, enabling the concentration gradient prediction to possess physical interpretability and temporal evolution continuity.

[0012] By mapping the predicted concentration gradient to the engineering quantities of heating and ultrasound, the system achieves a self-consistent closed loop of "detection-prediction-control".

[0013] Preferably, when generating the control command, the predicted heavy metal ion concentration gradient is first multiplied by a preset depth weighting coefficient vector to obtain a weighted over-limit risk value. When the weighted over-limit risk value is greater than the first power adjustment threshold, the first start-up time and first running time of the electromagnetic induction heating device are determined; when the weighted over-limit risk value is greater than the second power adjustment threshold and the cumulative amount of environmental precipitation intensity data within a preset time window is greater than the precipitation trigger threshold, the second start-up time and second running time of the ultrasonic-assisted extraction device are determined; then the start-up time and running time of the two devices are packaged into a control command. The setting of differentiated thresholds and precipitation trigger conditions ensures the orderly intervention of heating field control and ultrasonic-assisted extraction under different risk levels and precipitation infiltration conditions, avoiding ineffective energy consumption and equipment idling.

[0014] For the regulation of the electromagnetic induction heating device, as a specific scheme, the maximum heating power of the device and the soil thermal diffusivity are obtained. The required heating energy per unit volume is calculated based on the difference between the weighted excess risk value and the first power adjustment threshold. The heating energy per unit volume is divided by the maximum heating power to obtain the basic heating duration. The current moment is taken as the first start-up moment, and the basic heating duration is taken as the first running duration. More preferably, the heating energy per unit volume is further corrected based on the soil specific heat capacity and the target temperature rise, thereby improving the accuracy of matching the heat supply and demand of the soil system.

[0015] For the control of the ultrasonic-assisted extraction device, its operating frequency array and the viscosity-temperature characteristic curve of the soil pore liquid can be obtained; the viscosity-temperature characteristic curve can be queried according to the ambient temperature fluctuation data to obtain the dynamic viscosity value of the pore liquid at the current temperature; based on the viscosity value and the weighted over-limit risk value, a matching ultrasonic operating frequency can be selected from the operating frequency array; the preset extraction time corresponding to the matching ultrasonic operating frequency is used as the second running time, and the time after the current time is delayed by a preset precipitation response lag interval is used as the second start time.

[0016] Preferably, the precipitation response lag interval is dynamically calculated based on the soil saturated permeability coefficient and the deployment depth of the ultrasonic-assisted extraction device, so that the start time of the ultrasonic action coincides with the time when the precipitation infiltration front reaches the target depth, thereby maximizing the extraction efficiency.

[0017] As another technical solution of the present invention, after generating the control command, the system further enhances the model's adaptability through a feedback correction mechanism. Specifically, the system acquires the feedback soil spectral data after the electromagnetic induction heating device and the ultrasonic-assisted extraction device execute the control command, performs a difference operation between the feedback soil spectral data and the in-situ spectral reflectance sequence to obtain the environmental disturbance residual matrix, and then inputs the environmental disturbance residual matrix into the reverse correction module of the dynamic extraction model of harmful soil components. This module outputs the model parameter correction amount and updates the internal weight coefficients of the model. In this way, the actual spectral response after control excitation is used to drive the online update of the model parameters, enabling the model to continuously approximate the actual soil-pollutant-environment coupling characteristics.

[0018] Preferably, when the back-correction module outputs the model parameter correction, it back-propagates the environmental perturbation residual matrix along the time dimension to calculate the error gradient tensor of each convolution kernel in the model; it then performs a Hadamard product operation between the error gradient tensor and a preset learning rate tensor to obtain the initial correction tensor; finally, it performs a weighted fusion of the initial correction tensor with the historical correction tensor from the previous time step to obtain the model parameter correction. This method introduces a momentum effect into the parameter update, which helps to suppress noise fluctuations in the residual signal and improves the stability and accuracy of the model update process.

[0019] Through the above modules and their collaborative work, this invention realizes non-contact real-time sensing of harmful soil components, process migration simulation, and on-demand differentiated control, forming an integrated technical solution that combines environmental information sensing, dynamic modeling, and proactive treatment.

[0020] The beneficial effects of this invention are: By constructing a soil spectral temporal feature matrix that integrates environmental factors and inputting this matrix into a dynamic extraction model of harmful soil components, a dynamic simulation of heavy metal ion transport under environmental disturbances was achieved. Specifically, wavelet decomposition was performed on the in-situ spectral reflectance sequence to extract the intensity values ​​of spectral absorption characteristic peaks in different frequency bands. Simultaneously, environmental temperature fluctuation data and environmental precipitation intensity data were aligned to a unified time axis to calculate the temperature change gradient and precipitation accumulation. Tensor concatenation of spectral absorption characteristic peak intensities, temperature change gradients, and precipitation accumulation resulted in a tight coupling between spectral features and environmental physical quantities in the time dimension. The resulting soil spectral temporal feature matrix simultaneously carries the spectroscopic response information of pollutants and the driving information of ion activity changes and convective transport. This feature matrix was divided according to the time step and mapped sequentially to a migration potential scalar field. Convolution operations were then performed with the soil pore structure topology network to generate the ion flux distribution matrix corresponding to each step. The predicted heavy metal ion concentration gradients at different soil depths were obtained by accumulating these matrices. This processing method embeds the temperature-controlled molecular diffusion enhancement effect and the precipitation-induced downward seepage driving effect into the computational graph of the migration model. It can continuously output the longitudinal envelope of heavy metal concentration that reflects the real environmental evolution, no longer limited to the static content at isolated moments, and significantly improves the dynamism and accuracy of the spatiotemporal distribution prediction of harmful components.

[0021] A weighted exceedance risk value is obtained by multiplying the predicted heavy metal ion concentration gradient with the depth weighting coefficient vector. Based on this value and its comparison with the power adjustment threshold, control commands for the electromagnetic induction heating device and the ultrasonic-assisted extraction device are generated collaboratively. When determining the start and stop of electromagnetic induction heating, the required heating energy per unit volume is calculated from the weighted exceedance risk value, taking into account the soil thermal diffusivity, specific heat capacity, and target temperature rise, and the heating duration is converted to achieve spatial matching between heat injection intensity and pollution risk. For ultrasonic-assisted extraction, the dynamic viscosity value of the pore fluid is obtained by querying the soil pore fluid viscosity-temperature characteristic curve at the current ambient temperature. A matching ultrasonic operating frequency is selected from the operating frequency array, and a precipitation response lag interval is set based on the soil saturated permeability coefficient and the device deployment depth to delay the start of ultrasonic extraction. Therefore, by applying an optimal frequency ultrasonic field during the time window when precipitation fully infiltrates and the viscosity of the pore liquid decreases, the cavitation and microjet effects are amplified by utilizing the wetting conditions and low viscosity characteristics formed by natural precipitation, thereby enhancing the migration of heavy metals from soil particles to the extraction zone. This effectively reduces the dependence on the amount of added liquid and heating energy consumption, and enables the remediation device's operating parameters to be adaptively adjusted according to migration risks and weather conditions. Attached Figure Description

[0022] The invention will now be further described with reference to the accompanying drawings.

[0023] Figure 1 This is a schematic diagram of a controllable soil harmful component detection system based on environmental monitoring. Figure 2 This is a flowchart of data acquisition and time alignment; Figure 3 This is a flowchart for predicting the concentration gradient of harmful components in soil based on migration simulation; Figure 4 This is a flowchart of the control instruction generation process of the control execution module; Figure 5 This is a flowchart of the reverse calibration module parameter correction process; Figure 6 This is a flowchart of the process for generating model parameter correction values ​​for the reverse correction module. Detailed Implementation

[0024] 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.

[0025] See Figure 1 The present invention provides an adjustable soil harmful component detection system based on environmental monitoring. The system includes a data acquisition module, a feature construction module, a migration simulation module and a control execution module.

[0026] The data acquisition module obtains the in-situ spectral reflectance sequence of soil samples within the monitoring area during a preset sampling period, and simultaneously acquires environmental temperature fluctuation data and environmental precipitation intensity data for the same monitoring area during the same preset sampling period. The feature construction module receives the in-situ spectral reflectance sequence, environmental temperature fluctuation data, and environmental precipitation intensity data output by the data acquisition module, and constructs a soil spectral temporal feature matrix incorporating environmental factors. The migration simulation module has a pre-built dynamic extraction model for harmful soil components. This module inputs the soil spectral temporal feature matrix into this model, simulates the migration and accumulation process of heavy metal ions in the soil pore solution, and finally outputs the predicted heavy metal ion concentration gradient at different depths within the monitoring area. The control and execution module generates control commands based on the difference between the predicted heavy metal ion concentration gradient and a preset soil safety threshold. These control commands are used to control the start / stop times and output power of the electromagnetic induction heating device and the ultrasonic-assisted extraction device deployed within the monitoring area.

[0027] In practice, the data acquisition module obtains the in-situ spectral reflectance sequence of soil samples within a preset sampling period within the monitoring area through a multispectral sensor array deployed within the monitoring area. The multispectral sensor array contains multiple spectral acquisition units, which are arranged vertically along the soil profile or in a grid-like planar arrangement, enabling the multispectral sensor array to simultaneously sense soil samples at different locations within the monitoring area.

[0028] See Figure 2 The multispectral sensor array collects raw reflectance data at preset sampling time intervals. These intervals are determined in advance based on the transport rate and spectral response characteristics of heavy metal ions in the soil of the monitoring area, and range from 5 to 30 minutes. At each sampling moment, each spectral acquisition unit in the multispectral sensor array outputs raw reflectance data covering a continuous spectral range, which is recorded in the form of wavelength-reflectance pairs.

[0029] The process of extracting reflectance values ​​from multiple characteristic wavelengths from raw reflectance data is as follows: A set of characteristic wavelengths corresponding to the spectral absorption characteristics of the heavy metal element to be detected is pre-determined. This set is selected based on the positions of the characteristic absorption peaks of the heavy metal element in the visible to near-infrared band, and includes wavelengths of 485 nm, 650 nm, 720 nm, and 850 nm. For the raw reflectance data acquired at each sampling time, the reflectance value corresponding to each characteristic wavelength is extracted by interpolation. The reflectance values ​​of all characteristic wavelengths extracted at the same sampling time are combined to form the characteristic reflectance vector for that sampling time. Each element in the characteristic reflectance vector corresponds to the reflectance value of a characteristic wavelength, and the order of the elements is consistent with the increasing order of the wavelengths in the characteristic wavelength set.

[0030] Arrange the characteristic reflectance vectors at each sampling time according to the sampling order to form a time series, which is the in-situ spectral reflectance sequence. The in-situ spectral reflectance sequence can be represented as a two-dimensional matrix, where the rows of the matrix correspond to the sampling time and the columns correspond to the characteristic wavelengths.

[0031] The data acquisition module synchronously collects ambient temperature fluctuation data and ambient precipitation intensity data, accomplished through an environmental sensing component independent of the spectral acquisition. Temperature sensors exposed to the atmosphere are deployed within the monitoring area. These sensors employ platinum resistance temperature probes or thermocouples, with their sensing ends directly contacting the surrounding air. Temperature values ​​are recorded using a clock synchronized with the multispectral sensor array. The time-series temperature data output by the temperature sensors is read in real time and stored as ambient temperature fluctuation data. Each data point in the ambient temperature fluctuation data includes a timestamp and the corresponding Celsius temperature value.

[0032] Rain gauges, either tipping bucket or weighing type, are deployed within the monitoring area to measure the intensity of natural precipitation. The time-series precipitation intensity data output by the rain gauges is read in real time and stored as environmental precipitation intensity data. Each data point in the environmental precipitation intensity data includes a timestamp and a precipitation intensity value in millimeters per hour.

[0033] The process of aligning temperature time-series data and precipitation intensity time-series data onto a unified time axis of a preset sampling period according to their respective timestamps is as follows: The time range of the preset sampling period is determined, and a unified time axis is established with the start point of the preset sampling period as zero. The time resolution of the unified time axis is consistent with the sampling time interval of the multispectral sensor array. For temperature time-series data, if a temperature value with a matching timestamp exists at a certain moment on the unified time axis, that temperature value is directly used; otherwise, linear interpolation is performed using two adjacent real sampled temperature values ​​to obtain the temperature value corresponding to that moment on the unified time axis. For precipitation intensity time-series data, the same nearest neighbor interpolation or linear interpolation method is used to map the original precipitation intensity values ​​to discrete moments on the unified time axis, obtaining environmental precipitation intensity data synchronized with the environmental temperature fluctuation data. After time alignment, the environmental temperature fluctuation data and the environmental precipitation intensity data form two numerical sequences that correspond one-to-one with the moments on the unified time axis, which are then used by the feature construction module.

[0034] In specific implementation, please refer to Figure 3 The feature construction module receives the in-situ spectral reflectance sequence, environmental temperature fluctuation data, and environmental precipitation intensity data output by the data acquisition module, and constructs a soil spectral temporal feature matrix that integrates environmental factors through digital signal processing and tensor operations.

[0035] The process of wavelet decomposition of the in-situ spectral reflectance sequence is as follows: A compactly supported orthogonal wavelet basis function is selected to perform a discrete wavelet transform on the time series corresponding to each spectral characteristic wavelength of the in-situ spectral reflectance sequence. The wavelet basis function uses a fourth-order variant of the Dobermann wavelet, and the number of wavelet decomposition levels is set to 5. The discrete wavelet transform decomposes the time series of each characteristic wavelength into a low-frequency approximation coefficient sequence and a high-frequency detail coefficient sequence, where the high-frequency detail coefficient sequence includes the first to fifth levels of high-frequency detail coefficient sequences. The method for extracting the spectral absorption characteristic peak intensity values ​​of different frequency bands is as follows: For each characteristic wavelength corresponding to the first level of high-frequency detail coefficient sequence, the average value of the absolute value of each coefficient in the first level of high-frequency detail coefficient sequence is calculated, and this average value is used as the spectral absorption characteristic peak intensity value of the characteristic wavelength in the first frequency band; for each characteristic wavelength corresponding to the second level of high-frequency detail coefficient sequence, the average value of the absolute value of each coefficient in the second level of high-frequency detail coefficient sequence is calculated, and this average value is used as the spectral absorption characteristic peak intensity value of the characteristic wavelength in the second frequency band; and so on, to obtain the spectral absorption characteristic peak intensity values ​​of each characteristic wavelength in five frequency bands. The intensity values ​​of the spectral absorption characteristic peaks of all characteristic wavelengths in all frequency bands constitute the set of spectral absorption characteristic peak intensities. The dimension of the set of spectral absorption characteristic peak intensities is the number of characteristic wavelengths multiplied by the number of wavelet decomposition layers.

[0036] The time alignment of ambient temperature fluctuation data and ambient precipitation intensity data is completed by the data acquisition module. The feature construction module directly uses the aligned ambient temperature fluctuation data and ambient precipitation intensity data. The temperature change gradient of the ambient temperature fluctuation data is calculated as follows: for each sampling moment on a unified time axis, the temperature value of the next sampling moment is subtracted from the temperature value of the current sampling moment. The difference is divided by the sampling time interval to obtain the temperature change gradient value for that sampling moment, with the unit being degrees Celsius per hour. The cumulative precipitation of the ambient precipitation intensity data is calculated as follows: for each sampling moment on a unified time axis, a 24-hour sliding time window is traced backward. All precipitation intensity values ​​within this sliding time window are multiplied by their respective time lengths and summed to obtain the cumulative precipitation corresponding to that sampling moment, with the unit being millimeters.

[0037] The process of generating the soil spectral temporal feature matrix by tensor splicing is as follows: At each sampling time along a unified time axis, the set of spectral absorption characteristic peak intensities, temperature change gradient values, and accumulated precipitation at that sampling time are arranged into a one-dimensional vector. The first half of the one-dimensional vector contains the spectral absorption characteristic peak intensities of all characteristic wavelengths in the five frequency bands, the temperature change gradient values, and the accumulated precipitation values, respectively. All one-dimensional vectors from the sampling times are stacked in chronological order to form a two-dimensional matrix. The row dimension of the two-dimensional matrix is ​​the number of sampling times, and the column dimension is the length of the one-dimensional vector. This two-dimensional matrix is ​​the soil spectral temporal feature matrix.

[0038] The migration simulation module includes a pre-built dynamic extraction model for harmful soil components. This model simulates the migration and accumulation of heavy metal ions in soil pore solution. The core architecture of the dynamic extraction model for harmful soil components consists of an input layer, a potential energy field mapping layer, a convolutional transport layer, and a cumulative output layer. The input layer receives the sub-feature matrix and performs dimensionality reshaping. The potential energy field mapping layer consists of three cascaded fully connected hidden layers, used to map spectral and environmental information into a migration potential energy scalar field. The convolutional transport layer contains a pre-defined soil pore structure topology network as the convolution kernel. The cumulative output layer performs a time-dimension accumulation operation.

[0039] The training process of the dynamic extraction model for harmful soil components before deployment is as follows: Soil monitoring data samples with known heavy metal ion concentration gradient labels are collected. Each sample includes historical in-situ spectral reflectance sequences, historical environmental temperature fluctuation data, historical environmental precipitation intensity data, and real heavy metal ion concentration gradients at different soil depths obtained through indoor soil column leaching experiments or on-site sampling and testing. For each sample, the same wavelet decomposition, feature extraction, gradient calculation, and concatenation operations as in the online phase are performed to construct a historical soil spectral temporal feature matrix as model input, with the corresponding real heavy metal ion concentration gradient as the supervision signal. Model training uses a batch stochastic gradient descent optimizer with the root mean square error loss function, an initial learning rate of 0.001, 2000 training iterations, and a batch size of 32. After training, the fully connected weights and bias parameters of the potential field mapping layer are determined, and the convolutional kernel weights in the soil pore structure topology network are optimized, resulting in the dynamic extraction model for harmful soil components.

[0040] During the online operation phase, the soil spectral time-series feature matrix is ​​divided into multiple sub-feature matrices with different time steps as follows: the soil spectral time-series feature matrix is ​​non-overlapped along the row dimension with a fixed step size, which is set to the number of sampling rows corresponding to a 1-hour time period. Each sub-feature matrix obtained by the division has the same time step size in the row dimension, and the column dimension of the sub-feature matrix is ​​consistent with the column dimension of the soil spectral time-series feature matrix.

[0041] Each sub-feature matrix is ​​sequentially input into the input layer of the dynamic extraction model for harmful components in soil. The process by which the input layer maps each sub-feature matrix into a migration potential scalar field is as follows: the input layer expands the sub-feature matrix into a one-dimensional input vector, which is then sequentially fed into three fully connected hidden layers of the potential field mapping layer. The number of neurons in the three fully connected hidden layers are 256, 128, and 64, respectively, and each fully connected hidden layer is followed by a modified linear unit activation function. The final output of the potential field mapping layer is a two-dimensional matrix. The number of rows in the two-dimensional matrix corresponds to the number of grid cells in the horizontal plane of the monitoring area, and the number of columns corresponds to the number of layers in the soil depth direction. Each element value in the two-dimensional matrix represents the migration potential scalar value at the corresponding spatial location. The migration potential scalar value indicates the strength of the tendency of heavy metal ions to detach from the soil matrix and enter the pore solution at that spatial location. The formula for calculating the migration potential scalar value is: ; in, The horizontal grid position index is Depth layer index is The scalar value of the migration potential energy at the spatial location. The value ranges from 1 to the total number of horizontal grid cells in the monitoring area. The value ranges from 1 to the number of layers divided along the soil depth direction; This represents the sigmoid activation function, which maps the output value to the range of 0 to 1. This indicates the third fully connected hidden layer in the potential energy field mapping layer. The output value of each neuron The value ranges from 1 to 64; This indicates that the third fully connected hidden layer is connected. Individual neurons and spatial location The weight parameters of the migration potential output node are determined through backpropagation during model training. Indicates spatial location The corresponding bias parameters are also determined through backpropagation during model training. The scalar values ​​of the migration potential energy at all spatial locations together constitute the migration potential energy scalar field.

[0042] The method for obtaining the ion flux distribution matrix for each time step by convolving the migration potential energy scalar field with a pre-defined soil pore structure topology network is as follows: The soil pore structure topology network is a three-dimensional convolutional kernel group containing 15 3x3x3 convolutional kernels. The weight value of each convolutional kernel reflects the connectivity probability and ion conduction capacity of soil pores in different spatial directions. The weight values ​​of the soil pore structure topology network are determined through graph structure analysis of micron-level computed tomography images of a large number of soil samples and joint constraints from actual migration data during the training phase. The convolution operation adopts an effective padding mode with a stride of 1. The migration potential energy scalar field is used as the input feature map and is subjected to three-dimensional discrete convolution with each convolutional kernel, outputting 15 two-dimensional feature maps. Each element value of the feature map represents the ion flux density at the corresponding spatial location within a unit time step. The 15 two-dimensional feature maps are stacked along the depth direction to form the ion flux distribution matrix. The dimension of the ion flux distribution matrix is ​​the number of horizontal grid cells multiplied by the number of depth layers multiplied by the number of convolutional kernels.

[0043] The process of accumulating the ion flux distribution matrices in chronological order to obtain the predicted heavy metal ion concentration gradient is as follows: Initialize a zero-based accumulation matrix with the same spatial size as the ion flux distribution matrix. After processing the sub-feature matrix for each time step and obtaining the corresponding ion flux distribution matrix, sum and compress the ion flux distribution matrix along the depth dimension. Then, convert the ion flux density to mass concentration according to the preset heavy metal element molar mass conversion factor and accumulate it to the corresponding depth layer position of the accumulation matrix. After processing all time steps, the accumulation matrix is ​​arranged along the depth direction, with each depth layer corresponding to a cumulative heavy metal concentration value. The vector formed by connecting the cumulative heavy metal concentration values ​​of all depth layers is the predicted heavy metal ion concentration gradient.

[0044] In specific implementation, please refer to Figure 4 The regulation execution module receives the predicted heavy metal ion concentration gradient output by the migration simulation module and generates regulation instructions based on a preset soil safety threshold. The regulation execution module internally comprises a risk quantification unit, a heating regulation decision unit, and an ultrasonic regulation decision unit; these three units work together sequentially to complete the generation of regulation instructions.

[0045] The weighted risk value is obtained by multiplying the predicted heavy metal ion concentration gradient with a preset depth weighting coefficient vector. The predicted heavy metal ion concentration gradient is a one-dimensional vector, where each element corresponds to the predicted heavy metal concentration value at a given depth layer. The depth layer division is consistent with the number of soil depth layers in the migration simulation module, set to 10 layers, with each layer being 5 cm thick from the surface downwards. The preset depth weighting coefficient vector is a constant vector determined in advance through farmland ecological risk assessment. Each weighting coefficient in the preset depth weighting coefficient vector corresponds to a depth layer, reflecting the contribution of heavy metal ions at that depth layer to crop root absorption. In the preset depth weighting coefficient vector, the weighting coefficients for depth layers corresponding to the main distribution areas of crop roots are higher, while the weighting coefficients for depth layers that are difficult for roots to reach are lower. The dot product operation is as follows: multiply the predicted heavy metal concentration value of each depth layer in the predicted heavy metal ion concentration gradient by the weight coefficient of the same depth layer index in the preset depth weight coefficient vector, and sum the multiplications of all depth layers to obtain a scalar value. This scalar value is the weighted over-limit risk value, and the unit of the weighted over-limit risk value is milligrams per kilogram.

[0046] When the weighted exceedance risk value exceeds the first power adjustment threshold, the first start-up time and first running duration of the electromagnetic induction heating device are determined. The first power adjustment threshold is a preset start-up threshold parameter, set based on 150% of the heavy metal risk screening value specified in the soil environmental quality standard. When the weighted exceedance risk value exceeds the first power adjustment threshold, it indicates that the migration activity of heavy metal ions in the soil pore solution has reached a level requiring thermal intervention.

[0047] The method for determining the first start-up time and first running duration of the electromagnetic induction heating device is as follows: Obtain the maximum heating power of the electromagnetic induction heating device and the soil thermal diffusivity. The maximum heating power of the electromagnetic induction heating device is the rated output power value indicated on the device's nameplate, in kilowatts. The soil thermal diffusivity is obtained through in-situ measurement using a thermal pulse probe deployed within the monitoring area. The measurement process involves simultaneously burying the thermal pulse probe into the target soil depth during the installation of the electromagnetic induction heating device, applying a short-duration thermal pulse, and recording the temperature decay curve. The soil thermal diffusivity is calculated using the half-life of the temperature decay curve. The unit of the soil thermal diffusivity is square meters per second.

[0048] The formula for calculating the required heating energy per unit volume based on the difference between the weighted over-limit risk value and the first power adjustment threshold is as follows: ; in, This indicates the required heating energy per unit volume, expressed in joules per cubic meter. The soil bulk density is expressed in kilograms per cubic meter and is obtained by sampling and measuring within the monitoring area using the ring sampler method. It represents the specific heat capacity of soil, with units of joules per kilogram per Kelvin, and is obtained by differential scanning calorimetry from soil samples collected within the monitoring area. This represents the target temperature increase, expressed in Kelvin. The target temperature increase is set at 15 Kelvin. The basis for setting 15 Kelvin is the Arrhenius equation fitting curve of the relationship between the desorption rate of heavy metal ions in soil pore liquid and temperature. Increasing the temperature by 15 Kelvin can increase the desorption rate constant to 2.5 to 3 times that under room temperature conditions, thus achieving a balance between energy consumption and effectiveness. This represents the weighted over-limit risk value; This represents the first power adjustment threshold. The heating energy per unit volume increases linearly with the increase of the weighted over-limit risk value, and the heating energy per unit volume is zero when the weighted over-limit risk value is exactly equal to the first power adjustment threshold.

[0049] The basic heating time is obtained by dividing the heating energy per unit volume by the maximum heating power. The basic heating time is calculated as follows: multiply the effective working volume of the electromagnetic induction heating device by the heating energy per unit volume to obtain the total heating energy requirement, and then divide the total heating energy requirement by the maximum heating power. The effective working volume is determined based on the geometric dimensions of the induction coil and the magnetic field penetration depth of the electromagnetic induction heating device. The effective working volume is measured through calibration experiments during device installation. The current moment is taken as the first start-up moment, and the basic heating time is taken as the first running time. Upon receiving the first start-up moment and the first running time, the electromagnetic induction heating device immediately starts and operates at maximum heating power, automatically shutting down after the first running time is reached.

[0050] When the weighted risk value exceeds the second power adjustment threshold and the cumulative amount of environmental precipitation intensity data within the preset time window exceeds the precipitation trigger threshold, the second start-up time and second running duration of the ultrasonic-assisted extraction device are determined. The second power adjustment threshold is another preset start-up threshold parameter, set to 80% of the first power adjustment threshold. The basis for setting the second power adjustment threshold lower than the first power adjustment threshold is that the mechanism of ultrasonic-assisted extraction does not require reaching the high-risk level required for thermal desorption. The preset time window is a 6-hour period backward from the current time point. The precipitation trigger threshold is 15 mm, set based on the fact that when the cumulative precipitation reaches 15 mm within 6 hours, the rainwater infiltration front can basically reach the deployment depth of the ultrasonic-assisted extraction device, and the pore solution mobile phase begins to form.

[0051] The method for determining the second start-up time and second running time of the ultrasonic-assisted extraction device is as follows: The operating frequency array of the ultrasonic-assisted extraction device and the viscosity-temperature characteristic curve of the soil pore fluid are obtained. The operating frequency array of the ultrasonic-assisted extraction device is a set of discrete frequency values ​​pre-existing within the control execution module, including four frequencies: 20 kHz, 28 kHz, 40 kHz, and 68 kHz. These four frequencies correspond to pore fluids with different viscosity ranges; 20 kHz is suitable for high-viscosity pore fluids, and 68 kHz is suitable for low-viscosity pore fluids. The viscosity-temperature characteristic curve of the soil pore fluid is a lookup table composed of experimental calibration data. The lookup table records the dynamic viscosity value of the pore fluid at 1-degree Celsius intervals within the range of 5°C to 45°C. The lookup table was obtained by measuring the soil pore leachate collected in the monitoring area using a rotational viscometer in a laboratory constant temperature bath. The measurement temperature was gradually increased from 5°C to 45°C, and the dynamic viscosity value after stabilization at each temperature point was recorded.

[0052] The method for obtaining the dynamic viscosity value of pore liquid at the current temperature by querying the viscosity-temperature characteristic curve based on ambient temperature fluctuation data is as follows: read the temperature value at the current moment on the same time axis from the ambient temperature fluctuation data, use this temperature value as the query key value, and look up the viscosity value corresponding to the closest temperature value in the viscosity-temperature characteristic curve lookup table. If the current temperature value falls between two recorded temperature points, then the viscosity value of the intermediate temperature point is calculated by linear interpolation between the two recorded temperature points. Based on the dynamic viscosity of the pore fluid at the current temperature and the weighted over-limit risk value, the method for selecting a matching ultrasonic working frequency from the working frequency array is as follows: A pre-set rule corresponds between viscosity segment thresholds and frequencies. When the dynamic viscosity of the pore fluid is greater than 50 mPa·s, a 20 kHz frequency is matched; when the dynamic viscosity is between 20 and 50 mPa·s, a 28 kHz frequency is matched; when the dynamic viscosity is between 6 and 20 mPa·s, a 40 kHz frequency is matched; and when the dynamic viscosity is less than 6 mPa·s, a 68 kHz frequency is matched. If the weighted over-limit risk value exceeds the second power adjustment threshold by more than 50%, the frequency will be downgraded by one level to a lower level to enhance the cavitation effect.

[0053] The preset extraction time corresponding to the matched ultrasonic operating frequency is used as the second running time. The preset extraction time for each ultrasonic operating frequency is determined in the factory calibration: 45 minutes for 20 kHz, 35 minutes for 28 kHz, 25 minutes for 40 kHz, and 15 minutes for 68 kHz. The second start time is the time after a preset precipitation response lag interval is delayed from the current time. The precipitation response lag interval is dynamically calculated based on the soil saturated permeability coefficient and the deployment depth of the ultrasonic-assisted extraction device. The calculation method is as follows: divide the deployment depth of the ultrasonic-assisted extraction device by the soil saturated permeability coefficient, and then multiply by a reduction factor. The reduction factor is set to 0.7. The basis for setting the reduction factor of 0.7 is that the actual advance rate of the wetting front under unsaturated seepage conditions is approximately 70% of the saturated permeability coefficient. The soil saturated permeability coefficient is obtained by on-site measurement in the monitoring area using a dual-ring infiltrator, and the unit is centimeters per hour. The first start-up time, the first running time, the second start-up time, and the second running time are concatenated into a control command frame in chronological order. A cyclic redundancy check code is added to the control command frame, and then it is packaged into a control command. This command is then sent to the respective execution controllers of the electromagnetic induction heating device and the ultrasonic-assisted extraction device via an industrial bus or wireless communication module.

[0054] In specific implementation, please refer to Figure 5 After the control and execution module sends the control commands to the execution controllers of the electromagnetic induction heating device and the ultrasonic-assisted extraction device, the soil hazardous component detection system enters the feedback correction phase. This feedback correction phase is executed by the reverse correction module, which is integrated within the migration simulation module and shares parameter storage space with the dynamic extraction model of soil hazardous components.

[0055] The process of acquiring feedback soil spectral data after the electromagnetic induction heating device and ultrasonic-assisted extraction device execute control commands is as follows: After the first and second running durations specified in the control commands are completed, a preset stabilization waiting time is performed. The preset stabilization waiting time is set to 2 hours. The 2-hour setting is based on the time required for the soil temperature to naturally drop back to the ambient temperature after heating stops and for the pore fluid disturbance caused by ultrasonic extraction to recover to a steady state. After this time, the soil spectral characteristics tend to stabilize. After the stabilization waiting time ends, the data acquisition module controls the multispectral sensor array deployed in the monitoring area to perform a spectral acquisition using the same sampling parameters as the in-situ spectral reflectance sequence. The acquisition process lasts for a complete preset sampling cycle, obtaining feedback soil spectral data. The format of the feedback soil spectral data is consistent with the format of the in-situ spectral reflectance sequence, both being two-dimensional matrix forms. The rows of the matrix correspond to the sampling time, and the columns of the matrix correspond to the characteristic wavelengths.

[0056] The environmental disturbance residual matrix is ​​obtained by performing a difference operation between the feedback soil spectral data and the in-situ spectral reflectance sequence. The difference operation is performed under time alignment conditions. Specifically, the feedback soil spectral data and the in-situ spectral reflectance sequence are aligned on a unified time axis. Since the sampling start time of the feedback soil spectral data is after the completion of the control command, while the sampling start time of the in-situ spectral reflectance sequence is before the execution of the control command, the time axes of the two sequences do not directly coincide. By using the last sampling period in the in-situ spectral reflectance sequence as a reference window, the time axis of the feedback soil spectral data is shifted to match the sampling time of the reference window, so that the two sequences have the same sampling time index. At each identical sampling time, the characteristic reflectance vector of that sampling time in the feedback soil spectral data is subtracted from the corresponding characteristic reflectance vector in the in-situ spectral reflectance sequence. The result of the subtraction is still a vector, where each element is the reflectance difference before and after processing the corresponding characteristic wavelength. The reflectance difference vectors from all sampling times are arranged in chronological order to form an environmental perturbation residual matrix. The number of rows in the environmental perturbation residual matrix equals the number of sampling times within a preset sampling period, and the number of columns equals the number of characteristic wavelengths. The physical meaning of the environmental perturbation residual matrix is ​​the net change in soil spectral reflectance before and after electromagnetic induction heating and ultrasound-assisted extraction interventions. This net change includes the spectral response signal caused by the change in the migration and accumulation state of heavy metal ions during the control process, while filtering out background environmental noise.

[0057] The environmental perturbation residual matrix is ​​input into the backpropagation module of the dynamic extraction model of soil hazardous components. The core architecture of the backpropagation module includes a time backpropagation unit, a gradient tensor calculation unit, a correction fusion unit, and a weight update unit, which are cascaded in sequence. The time backpropagation unit receives the environmental perturbation residual matrix and the intermediate activation values ​​saved by the dynamic extraction model of soil hazardous components during the forward inference process. The gradient tensor calculation unit calculates the error gradient tensor based on the output of the time backpropagation unit. The correction fusion unit fuses the error gradient tensor with the learning rate tensor and the historical correction tensor. The weight update unit is responsible for writing the model parameter corrections into the convolutional kernel weight matrix of the dynamic extraction model of soil hazardous components.

[0058] The process of outputting model parameter corrections by the backpropagation module is as follows: the environmental perturbation residual matrix is ​​backpropagated along the time dimension, and the error gradient tensor of each convolutional kernel in the dynamic extraction model of soil harmful components is calculated. The environmental perturbation residual matrix serves as the initial loss signal for backpropagation, starting from the cumulative output layer of the dynamic extraction model of soil harmful components and propagating layer by layer from back to front along the time dimension. In the backpropagation of the cumulative output layer, each row of the environmental perturbation residual matrix is ​​considered as the instantaneous residual for the corresponding time step, and the instantaneous residual is propagated to the convolutional transport layer. In the convolutional transport layer, for each convolutional kernel contained in the soil pore structure topology network, the error gradient component of the convolutional kernel at the corresponding time step is obtained by cross-correlation between the instantaneous residual and the input migration potential scalar field corresponding to the convolutional kernel during forward inference. The error gradient components of all time steps are summed along the time dimension to obtain the error gradient tensor of the convolutional kernel. The dimension of the error gradient tensor is the same as the dimension of the convolutional kernel itself, both being 3x3x3. The dynamic extraction model for harmful components in soil contains 15 convolutional kernels, thus generating 15 error gradient tensors, each corresponding to a convolutional kernel.

[0059] The initial correction tensor is obtained by performing a Hadamard product operation between the error gradient tensor and the preset learning rate tensor. The preset learning rate tensor and the error gradient tensor have the same dimensional structure. The value of each element in the learning rate tensor is not a globally uniform value, but is set separately for different spatial directions and different convolution kernels. The learning rate tensor value scheme is as follows: for each convolution kernel, the learning rate tensor element corresponding to the error gradient element along the soil depth direction is set to 0.5 times the base learning rate, and the learning rate tensor elements corresponding to the error gradient elements along the two orthogonal directions of the horizontal plane are set to 1.0 times the base learning rate. The base learning rate is set to 0.0005. The base learning rate of 0.0005 is based on the convergence stability test of soil spectral data in multiple feedback iterations, where 0.0005 can achieve a balance between the correction speed and the parameter oscillation amplitude, ensuring that the prediction error fluctuation after three consecutive feedback corrections does not exceed 5%. The Hadamard product operation is performed by multiplying each element of the error gradient tensor with the corresponding element of the learning rate tensor element by element, and the product forms the initial correction tensor.

[0060] The initial correction tensor is weighted and fused with the historical correction tensor from the previous time step to obtain the model parameter correction. The historical correction tensor is stored in the correction fusion unit of the back-correction module and updated after each feedback correction. The weighted fusion is calculated as follows: each element of the initial correction tensor is multiplied by a first-order momentum coefficient of 0.9, and the corresponding element of the historical correction tensor is multiplied by the complement of the first-order momentum coefficient of 0.9 (i.e., 0.1). The sum of the two is the model parameter correction. The reason for using the first-order momentum coefficient for weighted fusion is that the introduction of momentum mechanism can smooth the parameter correction trajectory and avoid drastic changes in model parameters due to random perturbations in the soil spectral data in a single feedback. After fusion, the model parameter correction is sent to the weight update unit, which adds the model parameter correction element by element to the original weight parameters of the corresponding convolution kernel in the dynamic extraction model of soil harmful components, thus updating the weight coefficients within the dynamic extraction model of soil harmful components. The updated weight coefficients take effect immediately and are used to predict the heavy metal ion concentration gradient in the next sampling period.

[0061] In specific implementation, the process of the back-correction module outputting the model parameter correction amount is completed collaboratively by the time backpropagation unit, the gradient tensor calculation unit, the correction fusion unit, and the weight update unit. This process is the core computational part of the feedback correction stage in the above embodiment. The specific generation steps of the model parameter correction amount are described in detail below.

[0062] See Figure 6 The environmental perturbation residual matrix is ​​backpropagated along the time dimension to calculate the error gradient tensor of each convolutional kernel in the dynamic extraction model of soil hazardous components. The environmental perturbation residual matrix, as the loss signal, enters the time backpropagation unit from the cumulative output layer. The time backpropagation unit maintains intermediate tensor snapshots of the dynamic extraction model of soil hazardous components at each time step during forward inference. These intermediate tensor snapshots include the migration potential scalar field input to the convolutional transport layer at each time step and the ion flux distribution matrix output by the convolutional transport layer. The time backpropagation unit decomposes the environmental perturbation residual matrix row-wise into a sequence of residual vectors corresponding to each time step. The row order of the residual vector sequence is the reverse of the input order of the sub-feature matrix during forward inference, i.e., it recursively advances from the last time step.

[0063] In each backward recursion time step, the residual vector is transformed into the error signal of the convolutional transport layer through the gradient transfer function of the cumulative output layer. The gradient transfer function of the cumulative output layer is a unit transfer function, meaning the error signal is equal to the residual vector itself, because the cumulative output layer performs a parameterless accumulation operation during forward inference. After receiving the error signal, the convolutional transport layer calculates the error gradient individually for each convolutional kernel in the soil pore structure topology network. The soil pore structure topology network contains 15 convolutional kernels, denoted by their indices. , The value can be an integer from 1 to 15. Each convolutional kernel has a size of 3x3x3, corresponding to the depth direction index. Horizontal first direction index and horizontal second direction index ,in, The values ​​are 1, 2, and 3. The values ​​are 1, 2, and 3. The values ​​are 1, 2, and 3. For index ... The method for calculating the error gradient tensor of the convolution kernel is as follows: The error signal matrix corresponding to the convolution kernel at the current time step is multiplied element-wise by a slice of the migration potential scalar field of the local receptive field region in which the convolution kernel acts during forward inference. Then, the multiplications at all local receptive field locations are summed to obtain the gradient contribution tensor of the convolution kernel at a single time step. The convolution kernel is obtained by summing the gradient contribution tensors along the time dimension at all backward recursion time steps. The error gradient tensor is denoted as... . It is a three-dimensional tensor with dimensions 3x3x3. Medium depth directional index is The first horizontal direction index is The second horizontal direction index is The element is denoted as , Represents the convolution kernel In spatial location The error gradient value at each point is calculated. The above calculation process is repeated to obtain the error gradient tensor for each of the 15 convolutional kernels.

[0064] The initial correction tensor is obtained by performing a Hadamard product operation between the error gradient tensor and the preset learning rate tensor. The preset learning rate tensor is a three-dimensional parameter array pre-stored in the gradient tensor computation unit, denoted as […]. , The dimensions are 3x3x3. Medium depth directional index is The first horizontal direction index is The second horizontal direction index is The element is denoted as The preset learning rate tensor The values ​​of the elements in the array are not uniform, but are set separately according to the differences in the influence of spatial direction. The specific assignment rule is: when the depth direction index When it is 2, all the corresponding The element has the value When the depth direction index When it is 1 or 3, all the corresponding Elements and The element has the value .in, Indicates the basic learning rate. The value is set to 0.0005. The value of 0.0005 is chosen based on the convergence stability test results of soil spectral data in multiple feedback iterations. At this value, the prediction error fluctuation after three consecutive feedback corrections does not exceed 5%, and the parameter correction direction remains monotonically convergent without oscillating divergence. The preset learning rate tensor... It is applicable to all 15 convolutional kernels, and the same method is used for each convolutional kernel. Perform the Hadamard product operation. The process of the Hadamard product operation is as follows: For the convolution kernel... error gradient tensor ,Will Each element in With the preset learning rate tensor element at the corresponding position Element-by-element multiplication, product Constructing convolution kernels initial correction tensor The element at the corresponding position in Initial correction tensor Size and error gradient tensor They are exactly the same, and are 3x3x3.

[0065] The initial correction tensor is weighted and fused with the historical correction tensor from the previous time step to obtain the model parameter correction. The historical correction tensors are stored in the non-volatile storage area within the correction fusion unit; each convolutional kernel corresponds to one historical correction tensor. The historical correction tensor is denoted as superscript This represents the version identifier of the history correction tensor at each time step. When the inverse correction module performs feedback correction for the first time, the history correction tensors of all convolutional kernels are... It is initialized as an all-zero tensor, meaning all elements have a value of 0. The weighted fusion calculation process uses a first-order momentum exponential moving average method for the convolution kernel. , the initial correction tensor History correction tensor with the previous moment The updated history correction tensor is obtained by fusing according to the following formula. and will Output as a correction value for model parameters: ; in, Represents the convolution kernel The updated historical correction tensor at the current moment is also the output model parameter correction, with a size of 3x3x3; This represents the first-order momentum coefficient. The value is set to 0.9; Represents the convolution kernel The historical correction tensor of the previous moment; Represents the convolution kernel The initial correction tensor calculated within the current feedback correction period. The value of 0.9 is set based on the introduction of a momentum mechanism to continuously accumulate historical gradient directions, which maintains the smoothness of the correction direction and avoids the violent oscillation of model parameters caused by random disturbances in single feedback soil spectral data. The value of 0.9 makes the weight of the historical correction tensor in the moving average account for 90%, and the weight of the current initial correction tensor account for 10%, thus achieving a balance between smoothness and response speed.

[0066] After the fusion unit completes the weighted fusion, the model parameters corresponding to all 15 convolutional kernels are adjusted. The data is transmitted in parallel to the weight update unit. The weight update unit internally maintains the current weight tensors of the 15 convolutional kernels in the convolutional transport layer of the dynamic extraction model for harmful soil components. The current weight tensor is denoted as , The dimensions are 3x3x3. The weight update unit adjusts the model parameters. With the current weight tensor Add elements one by one, and replace the original with the result. As convolution kernel The updated weight tensor. Simultaneously, the correction fusion unit updates the historical correction tensor at the current time. Write to the non-volatile memory area, overwriting the original. This is used for the weighted fusion operation in the next feedback correction cycle.

[0067] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A controllable soil harmful component detection system based on environmental monitoring, characterized in that, include: The data acquisition module is used to acquire the in-situ spectral reflectance sequence of soil samples within the monitoring area within a preset sampling period, and simultaneously acquire the environmental temperature fluctuation data and environmental precipitation intensity data of the monitoring area within the preset sampling period. The feature construction module is used to construct a soil spectral temporal feature matrix that integrates environmental factors based on the in-situ spectral reflectance sequence, the environmental temperature fluctuation data, and the environmental precipitation intensity data. The migration simulation module is used to input the soil spectral time-series feature matrix into a pre-constructed dynamic extraction model of soil harmful components, simulate the migration and accumulation process of heavy metal ions in soil pore solution through the dynamic extraction model of soil harmful components, and output the predicted heavy metal ion concentration gradient of soil layers at different depths in the monitoring area. The control execution module is used to generate control commands based on the difference between the predicted heavy metal ion concentration gradient and the preset soil safety threshold. The control commands are used to control the start-up and shutdown times and output power of the electromagnetic induction heating device and the ultrasonic-assisted extraction device arranged in the monitoring area.

2. The adjustable soil harmful component detection system based on environmental monitoring according to claim 1, characterized in that, The acquisition of the in-situ spectral reflectance sequence of soil samples within the monitoring area during a preset sampling period includes: The system controls a multispectral sensor array deployed within the monitoring area to collect raw reflectance data at preset sampling time intervals. Multiple characteristic wavelength reflectance values ​​are extracted from the original reflectance data to form a characteristic reflectance vector for each sampling time. The feature reflectance vectors at each sampling time are arranged in the order of sampling to construct the in-situ spectral reflectance sequence.

3. The adjustable soil harmful component detection system based on environmental monitoring according to claim 1, characterized in that, The synchronous collection of environmental temperature fluctuation data and environmental precipitation intensity data in the monitoring area within the preset sampling period includes: Read the time-series temperature data output by the temperature sensor arranged in the monitoring area and exposed to the atmosphere, and use it as the ambient temperature fluctuation data; Read the time-series data of precipitation intensity output by the rain gauges arranged in the monitoring area, and use it as the environmental precipitation intensity data; The temperature time series data and the precipitation intensity time series data are aligned to a unified time axis of the preset sampling period according to their respective timestamps.

4. The adjustable soil harmful component detection system based on environmental monitoring according to claim 1, characterized in that, The construction of the soil spectral temporal feature matrix integrating environmental factors includes: Wavelet decomposition was performed on the in-situ spectral reflectance sequence to extract the intensity values ​​of spectral absorption characteristic peaks in different frequency bands; After aligning the ambient temperature fluctuation data with the ambient precipitation intensity data over time, the temperature change gradient of the ambient temperature fluctuation data and the cumulative precipitation of the ambient precipitation intensity data are calculated respectively. The soil spectral temporal feature matrix is ​​generated by tensor splicing the intensity values ​​of the spectral absorption characteristic peaks, the temperature change gradient, and the cumulative precipitation.

5. The adjustable soil harmful component detection system based on environmental monitoring according to claim 4, characterized in that, The dynamic extraction model of harmful components in soil simulates the migration and accumulation process of heavy metal ions in soil pore solution, including: The soil spectral temporal feature matrix is ​​divided into multiple sub-feature matrices with multiple time steps; Each of the sub-feature matrices is sequentially input into the input layer of the dynamic extraction model of harmful soil components, and the input layer maps each of the sub-feature matrices into a migration potential scalar field; The migration potential energy scalar field is convolved with a preset soil pore structure topology network to obtain the ion flux distribution matrix corresponding to each time step. The predicted heavy metal ion concentration gradient is obtained by accumulating the ion flux distribution matrices in chronological order.

6. The adjustable soil harmful component detection system based on environmental monitoring according to claim 5, characterized in that, The generation of control instructions includes: The predicted heavy metal ion concentration gradient is multiplied by a preset depth weighting coefficient vector to obtain a weighted over-limit risk value. When the weighted over-limit risk value is greater than the first power adjustment threshold, the first start-up time and the first running time of the electromagnetic induction heating device are determined. When the weighted over-standard risk value is greater than the second power adjustment threshold and the cumulative amount of the environmental precipitation intensity data within the preset time window is greater than the precipitation trigger threshold, the second start-up time and the second running time of the ultrasonic-assisted extraction device are determined. The first start time, the first runtime, the second start time, and the second runtime are packaged into the control command.

7. The adjustable soil harmful component detection system based on environmental monitoring according to claim 6, characterized in that, Determining the first start-up time and first running duration of the electromagnetic induction heating device includes: Obtain the maximum heating power of the electromagnetic induction heating device and the soil thermal diffusivity; The required heating energy per unit volume is calculated based on the difference between the weighted over-limit risk value and the first power adjustment threshold. Divide the heating energy per unit volume by the maximum heating power to obtain the basic heating time; The current time is taken as the first start time, and the basic heating duration is taken as the first running duration.

8. The adjustable soil harmful component detection system based on environmental monitoring according to claim 7, characterized in that, The heating energy per unit volume is further adjusted based on the soil specific heat capacity and the target temperature rise.

9. The adjustable soil harmful component detection system based on environmental monitoring according to claim 6, characterized in that, Determining the second start-up time and second running duration of the ultrasound-assisted extraction device includes: Obtain the operating frequency array of the ultrasonic-assisted extraction device and the viscosity-temperature characteristics curve of the soil pore fluid; The dynamic viscosity value of the pore fluid at the current temperature is obtained by querying the viscosity-temperature characteristic curve based on the ambient temperature fluctuation data. Based on the dynamic viscosity value of the pore fluid at the current temperature and the weighted risk value of exceeding the standard, a matching ultrasonic working frequency is selected from the working frequency array; The preset extraction time corresponding to the matched ultrasonic working frequency is taken as the second running time, and the time after the current time is delayed by a preset precipitation response lag interval is taken as the second start time.

10. The adjustable soil harmful component detection system based on environmental monitoring according to claim 9, characterized in that, The precipitation response lag interval is dynamically calculated based on the soil saturation permeability coefficient and the deployment depth of the ultrasonic-assisted extraction device.

Images