Pollutant feature inversion and dynamic tracking method for non-continuous time-varying flow field

Abstract

The application discloses a pollution source feature inversion and dynamic tracking method for a non-continuous time-varying flow field, comprising the following steps: acquiring fluorescence intensity, flow velocity field and suspended sediment concentration data, performing non-uniform sampling based on the fluorescence intensity probability density, and generating a particle point set with medium state identification; constructing a physical information neural network, embedding fluid transport constraints and adsorption-desorption kinetics constraints in the loss function, and outputting a three-dimensional concentration field; converting the concentration field into a random point mode, estimating the number of pollution sources based on the Bayesian information criterion, and using a Gaussian mixture model to invert the pollution source position, size and source strength; constructing a tracking framework integrating a data-driven model and a fluid dynamics model, dynamically adjusting the fusion weight through a logistic function according to the flow field spatial gradient; calculating the tracking confidence to exclude abnormalities and updating the neural network parameters with high confidence results. The application is suitable for accurate inversion and stable tracking of pollution sources in a complex hydrodynamic environment.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and intelligent information processing technology, specifically to a method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields. Background Technology

[0002] The monitoring, source tracing, and dynamic tracking of pollutants are key technical issues in water environment protection and emergency response to sudden pollution incidents. With the increasing demands for watershed ecological governance and water resource security, how to achieve accurate identification and stable tracking of pollution sources under complex hydrodynamic conditions has become a critical technical challenge that urgently needs to be addressed.

[0003] Current pollutant analysis methods primarily rely on a combination of numerical simulation and monitoring data inversion. Numerical simulation methods based on fluid dynamics models typically characterize pollutant diffusion processes by solving fluid transport equations. These methods are highly dependent on boundary conditions and model parameters, and require fine mesh generation in large-scale watersheds, resulting in high computational complexity and long solution times, making them unsuitable for the real-time and large-scale analysis requirements of sudden pollution events. Meanwhile, while some data-driven inversion methods can reduce computational complexity to some extent, they lack effective constraints on fluid physics processes, making them prone to prediction biases and exhibiting insufficient stability under complex environmental conditions.

[0004] In real-world aquatic environments, pollutant migration is influenced by a variety of complex factors. On one hand, in high-sediment-laden waters, adsorption and desorption occur between pollutants and sediment particles, making the spatial distribution of pollutants not only related to the current flow field state but also closely linked to their historical migration processes, exhibiting a clear history-dependent characteristic. Existing models typically treat pollutants as a single state, making it difficult to distinguish between free-diffusion and adsorbed states, resulting in insufficiently refined descriptions of the adsorption-desorption process. Traditional diffusion models based on instantaneous states struggle to accurately characterize this process. On the other hand, under conditions such as dam and gate regulation, flow field boundary conditions may abruptly change within a short period, forming discontinuous time-varying flow characteristics. This leads to sudden changes in pollutant migration trajectories, and existing methods are prone to prediction instability or trajectory deviation under such conditions.

[0005] Furthermore, under the combined influence of multiple pollution sources such as industrial point source emissions and agricultural non-point source pollution, pollutant concentration fields often exhibit spatial overlap, with mutual interference between different pollution sources increasing the difficulty of pollution source identification and parameter inversion. Existing methods typically struggle to effectively separate and independently characterize each pollution source when dealing with multi-source superposition problems.

[0006] In summary, existing technologies still struggle to simultaneously achieve physical consistency, real-time performance, and multi-source resolution in complex hydrodynamic environments. There is an urgent need to propose a pollutant inversion and tracking method that can adapt to discontinuous flow field conditions, characterize the historical dependence of pollutants, and possess multi-source decoupling capabilities. Summary of the Invention

[0007] To address the shortcomings of existing technologies in complex hydrodynamic environments, this invention provides a method for pollutant characteristic inversion and dynamic tracking in discontinuous time-varying flow fields. The aim is to use the pollutant concentration evolution process as an information carrier, and through particle-based representation, physical constraint modeling, multi-source decoupling, and adaptive tracking collaborative processing, achieve accurate inversion and stable tracking of pollution sources. To achieve the above objectives, this invention adopts the following technical solution.

[0008] A method for pollutant characteristic inversion and dynamic tracking in discontinuous time-varying flow fields includes the following steps: Step 1: Acquire fluorescence intensity data, flow velocity field data, suspended sediment concentration data, initial conditions, and boundary conditions of the monitoring area; construct a probability density function based on fluorescence intensity and perform non-uniform sampling to generate a particle point set; Step 2: Construct a physical information neural network and use the parameters of the particle point set as input and training target. Use suspended sediment concentration data to participate in the adsorption-desorption kinetic constraint calculation, and use initial conditions and boundary conditions to participate in the initial boundary condition constraint calculation. Train the physical information neural network by minimizing the loss function that includes fluid transport constraints, adsorption-desorption kinetic constraints and initial boundary condition constraints, and output a three-dimensional concentration field. Step 3: Convert the three-dimensional concentration field into a random point pattern according to the sampling threshold, estimate the number of pollution sources using the model selection criterion, and use a mixed probability distribution model to cluster the sampling points to retrieve the spatial center coordinates, spatial scale parameters, and source strength parameters of each pollution source, collectively referred to as pollution source parameters. Step 4: Initialize the pollution source state with pollution source parameters, and construct a tracking framework that integrates the data-driven model and the fluid dynamics model. The data-driven model predicts based on historical state sequences, and the fluid dynamics model predicts based on flow velocity field data. The fusion weights of the two models are dynamically adjusted according to the flow field spatial gradient through the logistic function to obtain the fused predicted state as the tracking result. Step 5: Calculate the deviation between the spatial center coordinates of the pollution source in the fused prediction state and the spatial center coordinates of the pollution source parameters to obtain the tracking confidence. Eliminate abnormal results with confidence below a preset threshold, retain high-confidence tracking results and store them in chronological order to form a dynamic evolution trajectory, which together with the pollution source parameters constitutes the final output. At the same time, accumulate and store the high-confidence tracking results to retrain the physical information neural network at a set period.

[0009] Furthermore, each particle in the particle set includes spatial coordinates, a timestamp, a concentration weight, and a medium state identifier. The concentration weight is used to characterize the relative concentration of pollutants, and the medium state identifier is used to distinguish between free diffusion and adsorption by sediment particles. The initial conditions include the initial concentration field of pollutants across the entire monitoring area at the initial moment. The boundary conditions include time-varying boundary data at the four boundaries of the monitoring area, the river inlet and outlet, and the dam control section. The concentration weight is calculated based on the fluorescence intensity value corresponding to the particle sampling location. The medium state identifier is determined based on the suspended sediment concentration at the particle location and the critical adsorption threshold of sediment. During the training of the physical information neural network, the coordinates and timestamp of the particle set are used as input, and the relative concentration of pollutants characterized by the concentration weight is used as the training target.

[0010] Furthermore, the loss function mentioned in step 2 is defined as follows: in, For data fitting terms, The initial and boundary condition constraints are broken down into initial residuals. With boundary residuals , This refers to the physical mechanism constraint terms that include fluid transport constraints and adsorption-desorption kinetic constraints; , , These are the constraints for the corresponding loss terms.

[0011] Furthermore, the physical mechanism constraint term is: for: in, For pollutant concentration, For the velocity field, Where is the diffusion coefficient. The concentration of suspended sediment. and These are the adsorption coefficient and the desorption coefficient, respectively. The concentration of pollutants adsorbed on the surface of sediment particles is calculated using the following formula: , The saturated adsorption coefficient of sediment. For identifying the state of particle media in a local area The arithmetic mean.

[0012] Furthermore, the sampling threshold mentioned in step 3 adopts a sampling strategy that combines a lower concentration limit threshold with sediment zoning: 15% of the average pollutant concentration across the entire monitoring area is used as the minimum screening threshold; in high sediment-laden areas... Intensive sampling was conducted using 50% of the average concentration of pollutants across the entire area as the sampling threshold; in the clear water area... The lowest screening threshold is used as the sampling threshold for sparse sampling; wherein To obtain the suspended sediment concentration in the monitoring area, The critical adsorption threshold for sediment is determined through adsorption experiments; points with concentrations lower than the corresponding sampling threshold in their respective areas are removed to obtain the sampling point set. , This represents the number of sampling points.

[0013] Furthermore, in step 3, the model selection criterion is the Bayesian information criterion, and the calculation formula is: in, The maximum likelihood value for the mixed probability distribution model. The number of model parameters, The total number of particle points; compare the corresponding values ​​for different numbers of pollution sources. The value is selected based on the minimum number of pollution sources; the mixed probability distribution model is a Gaussian mixture model, with the optimal number of pollution sources. As the clustering component number of the Gaussian mixture model, the inverted set of parameters for each pollution source is: , For position parameters, For spatial scale parameters, For source strength parameters.

[0014] Furthermore, the data-driven model mentioned in step 4 is any one of a Long Short-Term Memory network, a gated recurrent unit network, or a temporal convolutional network, and its input is the historical state sequence of the pollution source. The output is the predicted state. This is used to capture the nonlinear evolution and history-dependent characteristics of pollution source states in complex flow fields, where... This represents the time step index at the current moment. For parameterized nonlinear mapping functions, The length of the historical window; the fluid dynamics model is based on the advection transport equation. predict The location of the pollution source at any given time is used to provide physical constraints during abrupt changes in the flow field. For the velocity field, For time step, This indicates the flow velocity at the current location; Indicating a data-driven model in Predicted state at any given time Indicating fluid dynamics model in The predicted state at any given time.

[0015] Further, the flow field spatial gradient scalar mentioned in step 4 The logistic function is: in, To integrate weights, For threshold coefficient, This is the sensitivity coefficient; The fusion prediction formula is: When the flow field is stable , The data-driven model is dominant at times; when the flow field changes abruptly... , At that time, fluid dynamics models were the dominant approach.

[0016] Furthermore, the tracking confidence level mentioned in step 5 is: in The coordinates of the spatial center of the pollution source are used to integrate the predicted state; The coordinates of the spatial center of the pollution source are the parameters of the pollution source. For scale parameters; The preset threshold is an adaptive threshold: it calculates the lower quartile of the confidence levels of all pollution sources at the current time. Upper quartiles Interquartile range Adaptive threshold .

[0017] Further, the cycle set in step 5 is 24 hours; in the subsequent dynamic tracking process, at each moment, based on the three-dimensional pollutant concentration field updated in real time in step 2, the pollution source parameters at the current moment are recalculated using the same pollution source parameter inversion method as in step 3, and the obtained pollution source parameters are used as the pollution source observation status at the current moment for the tracking confidence assessment in step 5.

[0018] The beneficial effects of this invention are: by combining laser-induced fluorescence with random point patterns, complex fluid visual signals are converted into easily processed particle point cloud distributions, avoiding the limitations of traditional computational fluid dynamics with its strict mesh division and high computational cost, and better meeting the needs of rapid and comprehensive investigation of sudden pollution events in large-scale watersheds; In terms of technical performance, this application explicitly embeds the sediment adsorption kinetic equation into the physical information neural network, encodes the "memory effect" of suspended particulate matter on pollutants as a physical residual constraint, corrects the evolution distortion of the purely data-driven model in complex media from a physical mechanism perspective, and substantially improves the accuracy and robustness of concentration field reconstruction in high sediment-laden waters. To address complex water conservancy scheduling scenarios, this application introduces a dual-model-assisted tracking framework at the tracking level. When faced with millisecond-level flow step changes caused by the instantaneous opening and closing of dams, the model can dynamically adjust the evolution weights through adaptive balance coefficients to prevent trajectory bifurcation and instability caused by discontinuous boundary conditions, thus ensuring the continuity of long-term monitoring. At the feature inversion end, a random point pattern hybrid distribution model is used to deeply decouple the overlapping concentration fields. Combined with the Bayesian information criterion, the number of pollution sources is dynamically estimated, which breaks through the bottleneck of fingerprint ambiguity in the traditional gradient source tracing method when facing "industrial / agricultural multi-source superposition" and improves the robustness of the mapping of parameters such as location and scale. In terms of engineering application value, this solution relies on an end-to-end architecture of "physical perception → point cloud mapping → physical constraint reconstruction → statistical decoupling inversion → adaptive tracking" to achieve high-precision three-dimensional source tracing under conditions of data sparsity and strong environmental interference, providing a feasible quantitative assessment method for emergency supervision in watershed ecological protection and complex fluid-structure interaction environments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 Structure diagram of the three-dimensional pollutant concentration field reconstruction model that incorporates physical constraints; Figure 3 This is a framework diagram for the integrated inversion of pollution source characteristics and dynamic tracking. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings. Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The directional terms such as left, center, right, top, and bottom in the embodiments of the present invention are only relative concepts or referenced to the normal use state of the product, and should not be considered restrictive.

[0021] This embodiment uses pollutant monitoring and source tracing in a high-sediment-content river environment as an application scenario. In this scenario, water flow is affected by dam regulation, resulting in a significantly discontinuous flow field. Simultaneously, the suspended sediment concentration is high, and pollutants exhibit adsorption and desorption behaviors with sediment particles during migration, thus demonstrating a clear historical dependence characteristic.

[0022] Methods for pollutant characteristic inversion and dynamic tracking in discontinuous time-varying flow fields, such as Figure 1 As shown, it includes the following steps: Step 1: Particle-based representation of pollutant multi-source sensing data Step 1.1: Multi-source data acquisition and preprocessing: A multi-source observation system was deployed in the target waters, including observation platforms such as drones, unmanned vessels, and fixed hydrological monitoring stations. Each platform was equipped with a laser-induced fluorescence sensor to acquire pollutant fluorescence intensity data. The sampling frequency is preferably no less than 100 Hz to ensure high temporal resolution for pollutant diffusion dynamics. Simultaneously, flow velocity field data is acquired through hydrological monitoring equipment. and suspended sediment concentration data Simultaneous collection of monitoring data in the monitoring area The corresponding initial and boundary conditions, and the entire data space, are all limited to the monitoring area. Inside: ① Initial conditions: Monitoring area Inner initial time Initial concentration field of pollutants across the entire region This is used to define the background pollutant distribution at the initial moment of the model and is a subsequent loss term. Initial residual The input data source; ② Boundary conditions: Monitoring area Time-varying boundary data at the surrounding boundaries, river inlets and outlets, and dam control sections, adapted to the discontinuous time-varying flow field boundary abrupt changes caused by dam opening and closing, constitute subsequent loss terms. Middle boundary residual The input data source.

[0023] All multi-source data, including fluorescence intensity, flow velocity field, suspended sediment concentration, initial conditions, and boundary conditions, were processed in a unified time synchronization manner. Outlier removal, filtering and denoising, and normalization were then performed to eliminate dimensional differences between different data sources and improve the stability and consistency of subsequent modeling.

[0024] The initial and boundary conditions serve as input data for the initial boundary condition constraint terms during the physical information neural network training process in step 2.

[0025] The above multi-source data are processed with unified time synchronization, and outlier removal, filtering and denoising are performed to eliminate the dimensional differences between different data sources and improve the stability and consistency of subsequent modeling.

[0026] Step 1.2, Non-uniform probability sampling and particle point set construction: To transform continuous pollutant concentration information into a discrete representation suitable for modeling, a spatial probability density function is constructed based on fluorescence intensity data: in, Indicates the monitoring area.

[0027] Guided by the probability density function, non-uniform random sampling is performed on the spatial region to generate a particle point set. Among them, the number of particles It can be adaptively set according to the monitoring area, sensor resolution, and computing resource conditions, for example, by setting the sampling density per unit area.

[0028] The state vector of each particle is defined as: in, Indicates the spatial position of the particle; For timestamps; The concentration weight of the particles is calculated based on the fluorescence intensity value corresponding to the particle sampling location. This serves as a medium state indicator, used to distinguish whether pollutants are in a free-diffusion state or adsorbed onto sediment particles. It is determined based on the suspended sediment concentration at the particle's location and the critical adsorption threshold of the sediment: when the suspended sediment concentration is less than the critical adsorption threshold... This indicates that the pollutants are in a free diffusion state; when the suspended sediment concentration is greater than or equal to the critical adsorption threshold of the sediment... This indicates that the pollutants are in a state of being adsorbed by sediment.

[0029] The aforementioned particle-based representation enables a discretized representation of the continuous pollutant concentration field, thereby reducing computational complexity and improving the efficiency of subsequent modeling. The probability density function is used to determine the sampling probability of a particle, while the concentration weight represents the relative magnitude of the pollutant concentration at the corresponding particle location.

[0030] Step 2: Reconstruct the concentration field by fusing physical constraints; Step 2.1: Constructing a reconstructed model with embedded physical constraints: Combination Figure 2 As shown, a spatiotemporal coordinate system is constructed. For input, pollutant concentration The neural network outputs physical information. This is used to achieve a continuous representation of the pollutant concentration field. After the physical information neural network is trained, it can output the corresponding pollutant concentration value at any spatiotemporal coordinate in the monitoring area, forming a spatiotemporally continuous three-dimensional pollutant concentration field.

[0031] In one embodiment, the model is implemented using a multi-layer feedforward neural network structure, which includes an input layer, several hidden layers, and an output layer, wherein: the input layer is used to receive spatiotemporal coordinates. The hidden layer consists of multiple fully connected layers, each mapping the input features through a non-linear activation function, which can be a ReLU function, a Sigmoid function, or a Tanh function. The output layer outputs the pollutant concentration value at the corresponding location. .

[0032] The forward propagation process of the physical information neural network can be represented as follows: in, Indicates the first Layer output; , These represent weights and biases, respectively. Nonlinear activation function.

[0033] The final output layer yields predicted pollutant concentrations: During model training, the model parameters are adjusted by minimizing the loss function. To optimize, the loss function is defined as: in, , , These are the constraints for the corresponding loss terms; This is the data fitting term, used to measure the deviation between the model output and the observed values ​​of the particle point set; The initial and boundary condition constraints can be broken down into initial residuals. With boundary residuals The input data for the two parts and two residuals are all taken from the initial and boundary conditions completed in step 1.1: ① Initial residual : The monitoring area collected in step 1 Internal initial concentration field As a baseline, only at the initial time Calculate the residuals and force the model output to determine the solubility concentration. With adsorbed state concentration The total concentration after superposition is equal to the measured initial concentration. ; ② Boundary residuals The monitoring area collected in step 1 Based on all boundary flux data, The residuals calculated under full simulation time series and all spatial boundaries are used to ensure that the flow velocity, sediment, and pollutant flux at the boundary locations are consistent with the measured boundary conditions. By minimizing This can prevent the model from outputting distorted results that contradict the measured hydrological environment at the initial moment and at the water boundary; The physical mechanism constraint term is mathematically expressed by the residual terms of the fluid transport process and the adsorption-desorption kinetics process. This constraint term achieves physical constraints on the transport and adsorption coupling behavior of pollutants by joint residual modeling of the complete convection-diffusion-adsorption-desorption process. The specific mathematical expression is given in step 2.2.

[0034] By designing the loss function as described above, the model can satisfy the physical laws of pollutant transport while fitting the observed data, thereby improving its robustness under conditions of sparse data or noise interference.

[0035] Step 2.2, Adsorption-Desorption Kinetics Constraint Modeling: To characterize the coupling behavior between pollutants and sediment, an adsorption-desorption kinetic model is introduced into the physical mechanism constraint term. Represented as: In the formula, Corresponding to constraints in fluid transport processes (including convection and diffusion terms). Corresponding to the constraints of the adsorption-desorption kinetic process (including adsorption and desorption terms); The diffusion coefficient can be calculated using empirical formulas or obtained through experimental measurements based on water flow conditions. and These are the adsorption coefficient and desorption coefficient, respectively, which can be determined through experimental calibration, historical data fitting, or empirical models. The real-time sediment concentration is taken from the measured suspended sediment data of multi-source monitoring in step 1.1; This refers to the concentration of pollutant components adsorbed individually on the surface of sediment particles at any given time and space, which is distinct from the initial concentration field of total pollutants across the entire region at the initial moment collected in step 1. The calculation formula is as follows: The saturated adsorption coefficient of sediment is determined by experiments in specific scenarios. The concentration of pollutants; For identifying the state of particle media in a local area The arithmetic mean.

[0036] By introducing the aforementioned physical constraints, we can achieve joint modeling of pollutant transport processes and adsorption kinetics, thereby obtaining a stable concentration field distribution that conforms to physical laws even under conditions of sparse data or noise.

[0037] The final output is a spatiotemporally continuous three-dimensional pollutant concentration field.

[0038] Furthermore, the time-varying boundary conditions of the dam collected in step 1 will continue to participate in the calculation of the fluid transport term, corresponding to Boundary residuals Adaptable monitoring area The sudden, discontinuous flow field caused by the instantaneous opening and closing of the inner dam further improves the simulation accuracy of the adsorption-desorption process in high-sediment-laden water.

[0039] Step 3: Decoupling of multi-source pollution characteristics and parameter inversion Step 3.1, Representation of the random point pattern: Combination Figure 3 As shown, to facilitate multi-source pollution separation, the reconstructed three-dimensional concentration field is discretized into a point set according to a sampling strategy combining a lower concentration threshold and sediment zoning: 15% of the average pollutant concentration across the entire monitoring area is used as the minimum screening threshold; in high sediment-laden areas... Intensive sampling was conducted using 50% of the average concentration of pollutants across the entire area as the sampling threshold; in the clear water area... The lowest screening threshold (i.e., 15% of the average concentration of pollutants across the entire region) is used as the sampling threshold for sparse sampling; wherein To obtain the suspended sediment concentration in the monitoring area, The critical adsorption threshold for sediment is determined through adsorption experiments. Points with concentrations lower than the corresponding sampling threshold in their respective areas are removed, and finally, valid coordinates are extracted to form a point set. , The number of sampling points; the sampling involves discretizing the entire three-dimensional concentration field into a point set for multi-source pollution separation and parameter inversion.

[0040] Step 3.2, Estimation of the number of pollution sources: Construct a multi-model candidate set and select the model using the Bayesian information criterion: in, The maximum likelihood value for the mixed probability distribution model. The number of model parameters, The total number of particle points generated in step 1; by comparing the Bayesian information criterion values ​​corresponding to different numbers of pollution sources, the number with the smallest value is selected as the optimal number of pollution sources.

[0041] Step 3.3, Pollution source parameter inversion: The mixed probability distribution model is a Gaussian mixture model, using the optimal number of pollution sources obtained in step 3.2. As the number of clustering components in the Gaussian mixture model, it is adopted Cluster analysis of the point set using component Gaussian mixture probability distribution yields parameters for each pollution source: location parameters. , indicating the spatial center of the pollution source; spatial scale parameters Characterizing the diffusion range of pollutants; source strength parameters By weighting the particle concentration Solve using a weighted integral method.

[0042] The final set of pollution source parameters is obtained as follows: This process effectively decouples multi-source overlapping pollution fields, and determines the spatial center coordinates of the pollution sources. Spatial scale parameters Source strength parameters Subscript With the number of pollution sources Correspondingly.

[0043] The above set of pollution source parameters is used as the initial state input to the subsequent dynamic tracking model to realize the continuous evolution prediction of pollution sources in time-varying flow fields.

[0044] In the subsequent dynamic tracking process, at each moment, based on the three-dimensional pollutant concentration field updated in real time in step 2, the pollution source parameters at the current moment are recalculated using the same pollution source parameter inversion method as in this step. The obtained pollution source parameters are used as the pollution source observation status at the current moment and are used for the tracking confidence assessment in step 5.

[0045] Step 4: Dynamic tracking of pollution sources for discontinuous flow fields; The pollution source state is initialized using the pollution source parameters obtained from step 3, wherein the location parameter is used to initialize the spatial location of the pollution source, and the source strength parameter is used to initialize the intensity of the pollution source; the spatial scale parameter is used to initialize reference information for the pollutant diffusion range, serving as the basis for adaptive adjustment of the diffusion coefficient in the subsequent fluid dynamics model—when the spatial scale of the pollution source is large, the diffusion coefficient increases accordingly to reflect the changing trend of the pollution source diffusion range; specifically: Step 4.1, Construction of the state evolution model: To address the step response in flow caused by dam regulation, a dual-model-assisted tracking architecture is constructed. The pollution source state vector is defined as follows: in, It indicates that the source is strong.

[0046] Building a fusion model: in, To integrate weights, The data-driven model predicts the nonlinear evolution and history-dependent characteristics of pollution source states in complex flow fields. Calculated by a fluid dynamics model, it is used to provide physical constraints when the flow field changes abruptly.

[0047] The data-driven model models the complex nonlinear evolution characteristics of pollution sources by performing nonlinear mapping on historical state sequences; the fluid dynamics model models the physical migration process of pollutants in the flow field by introducing flow velocity field data to update the location of pollution sources.

[0048] Step 4.1.1, Data-driven model construction: The data-driven model can be any one of a Long Short-Term Memory network, a gated recurrent unit network, or a temporal convolutional network, and its input is a historical state sequence of the pollution source. Each state vector in the historical state sequence contains the coordinates of the spatial center location of the pollution source. Source strength parameters and spatial scale parameters The output is the predicted state: in, This represents the time step index at the current moment. For parameterized nonlinear mapping functions, it is limited to parameterized mappings constructed by three types of temporal neural networks: Long Short-Term Memory Networks, Gated Recurrent Unit Networks, or Temporal Convolutional Networks. The historical window length can be set according to actual needs. This model is trained by minimizing prediction error to capture the nonlinear drift, random disturbances, and historical dependence characteristics of pollutants in complex flow fields. At the same time, it improves the ability to predict the evolution of the pollution source diffusion range by learning the changing trends of spatial scale parameters in historical sequences.

[0049] The model is trained by minimizing the prediction error to capture the nonlinear drift, random disturbances and history dependence characteristics of pollutants in complex flow fields.

[0050] Step 4.1.2, Construction of the fluid dynamics model: The fluid dynamics model is used to describe the physical migration process of pollutants in the flow field. It is based on velocity field data. The location of pollution sources is predicted using an advection transport model. in, For time step; This indicates the flow velocity at the current location.

[0051] When considering diffusion effects, a diffusion term can also be introduced. : in This represents the random diffusion noise term, the diffusion term The diffusion coefficient is adaptively adjusted based on the spatial scale parameters of the pollution source obtained from step 3. When the spatial scale parameters are large, the diffusion coefficient is increased to accurately reflect the changes in the diffusion range of the pollution source during the tracking process.

[0052] This model is used to provide predictions that satisfy physical laws when there are abrupt changes in the flow field or insufficient observations.

[0053] Step 4.2, Adaptive Weight Adjustment Mechanism: To address abrupt changes caused by discontinuous flow fields, the spatial gradient scalar of the flow velocity within the monitoring area is calculated in real time. This parameter characterizes the intensity of discontinuity or abrupt change in the flow field.

[0054] Utilizing fusion weights The logistic function for dynamically calculating the spatial gradient scalar can be expressed as: in, , where is the threshold coefficient, representing the critical sensitivity point of flow field fluctuations. When hour, That is, the data-driven model and the fluid dynamics model have equal weights. This is a sensitivity coefficient used to control the speed of weight switching. The larger the value, the more drastic the model's response to sudden changes in the flow field; When the value is smaller, the switching process is smoother.

[0055] The model dynamically adjusts the weights based on the gradient magnitude when the flow field is stable. ), The system primarily relies on data-driven models to capture fine-grained nonlinear diffusion characteristics. This is especially important when the dam is opened or closed, or when there are sudden changes in the flow field. ), The system automatically switches to a fluid dynamics model as the primary model, using physical laws to forcefully correct the tracking trajectory and prevent drift.

[0056] The fusion prediction formula is: in, The fused prediction state represents the predicted state of the pollution source at the current moment, which includes parameters such as the spatial center coordinates of the pollution source, spatial scale, and source strength. As time progresses, the fused prediction states at each moment are continuously recorded in chronological order, forming the dynamic evolution trajectory of the pollution source.

[0057] The fusion prediction state serves as the prediction state input in step 5, which involves calculating the real-time tracking confidence level.

[0058] This mechanism enhances the constraints of the physical model during sudden changes in the flow field and strengthens the data-driven capability under stable conditions, thereby ensuring the stability and accuracy of the tracking results.

[0059] Step 5: Result optimization and adaptive model update; Step 5.1, Confidence Assessment and Anomaly Screening: To evaluate the reliability of the dynamic tracking results and eliminate abnormal tracking results caused by sudden changes in the flow field or model prediction errors, the consistency between the fused prediction state obtained in step 4 and the pollution source parameters obtained by re-inversion in step 3 is assessed. Define real-time tracking confidence: in, For the first The real-time tracking confidence level of the pollution source is measured in the range of (0,1]. The larger the value, the more reliable the prediction result. The coordinates of the spatial center of the pollution source obtained from the fusion prediction in step 4; The coordinates of the spatial center of the pollution source are obtained by re-inverting the three-dimensional pollutant concentration field output in step 2 in step 3. This is the scale parameter.

[0060] The spatial scale parameter is used to normalize the location error. When the pollution source has a large diffusion range, a larger location deviation is allowed; when the pollution source has a small diffusion range, a more stringent location consistency evaluation is adopted, thereby improving the consistency of confidence evaluation between pollution sources of different scales.

[0061] Since both the position error term and the spatial scale parameter have the dimension of length, the exponent term satisfies the dimensionless requirement, ensuring that the confidence calculation has a unified physical meaning.

[0062] An adaptive thresholding method based on box statistics is used to filter out outlier data: the confidence scores of all pollution sources at the current time are aggregated, and the quartiles of the confidence scores are calculated. Upper quartiles Interquartile range Take the robustness coefficient Adaptive threshold: Remove those with confidence levels lower than 10%. The anomaly tracking results are retained, and high-confidence tracking results are stored continuously in chronological order to form a dynamic evolution trajectory of the pollution source. This dynamic evolution trajectory, together with the spatial center coordinates, spatial scale parameters, and source strength parameters of the pollution source obtained in step 3, constitutes the final output of this invention. The spatial center coordinates, spatial scale parameters, and source strength parameters of the pollution source are collectively referred to as pollution source characteristic parameters. This threshold is automatically updated in real time according to the distribution of each batch of monitoring data, without the need for manual pre-assignment, adapting to discontinuous abrupt flow fields and variable operating conditions in high-sediment-laden water bodies.

[0063] Step 5.2, Online Model Update Mechanism: The high-confidence tracking results are accumulated and stored as empirical sample data for subsequent retraining, and the physical information neural network in step 2 is retrained or its parameters are updated at a set period (e.g., 24 hours).

[0064] This mechanism enables the model to adapt to environmental changes, such as variations in sediment distribution or seasonal flow field changes, thereby improving long-term operational stability.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for pollutant characteristic inversion and dynamic tracking in discontinuous time-varying flow fields, characterized in that, Includes the following steps: Step 1: Acquire fluorescence intensity data, flow velocity field data, suspended sediment concentration data, initial conditions, and boundary conditions of the monitoring area; construct a probability density function based on fluorescence intensity and perform non-uniform sampling to generate a particle point set; Step 2: Construct a physical information neural network and use the parameters of the particle point set as input and training target. Use suspended sediment concentration data to participate in the adsorption-desorption kinetic constraint calculation, and use initial conditions and boundary conditions to participate in the initial boundary condition constraint calculation. Train the physical information neural network by minimizing the loss function that includes fluid transport constraints, adsorption-desorption kinetic constraints and initial boundary condition constraints, and output a three-dimensional concentration field. Step 3: Convert the three-dimensional concentration field into a random point pattern according to the sampling threshold, estimate the number of pollution sources using the model selection criterion, and use a mixed probability distribution model to cluster the sampling points to retrieve the spatial center coordinates, spatial scale parameters, and source strength parameters of each pollution source, collectively referred to as pollution source parameters. Step 4: Initialize the pollution source state with pollution source parameters, and construct a tracking framework that integrates the data-driven model and the fluid dynamics model. The data-driven model predicts based on historical state sequences, and the fluid dynamics model predicts based on flow velocity field data. The fusion weights of the two models are dynamically adjusted according to the flow field spatial gradient through the logistic function to obtain the fused predicted state as the tracking result. Step 5: Calculate the deviation between the spatial center coordinates of the pollution source in the fused prediction state and the spatial center coordinates of the pollution source parameters to obtain the tracking confidence. Eliminate abnormal results with confidence below a preset threshold, retain high-confidence tracking results and store them in chronological order to form a dynamic evolution trajectory, which together with the pollution source parameters constitutes the final output. At the same time, accumulate and store the high-confidence tracking results to retrain the physical information neural network at a set period.

2. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 1, characterized in that: Each particle in the particle set contains spatial coordinates, a timestamp, a concentration weight, and a medium state identifier. The concentration weight is used to characterize the relative concentration of pollutants, and the medium state identifier is used to distinguish between free diffusion and adsorption by sediment particles. The initial conditions include the initial concentration field of pollutants across the entire monitoring area at the initial moment. The boundary conditions include time-varying boundary data at the four boundaries of the monitoring area, river inlets and outlets, and dam control sections. The concentration weight is calculated based on the fluorescence intensity value corresponding to the particle sampling location. The medium state identifier is determined based on the suspended sediment concentration at the particle location and the critical sediment adsorption threshold. During the training of the physical information neural network, the coordinates and timestamp of the particle set are used as input, and the relative concentration of pollutants characterized by the concentration weight is used as the training target.

3. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 2, characterized in that, The loss function described in step 2 is defined as follows: in, For data fitting terms, The initial and boundary condition constraints are broken down into initial residuals. With boundary residuals , The physical mechanism constraint term includes both fluid transport constraints and adsorption-desorption kinetic constraints; , , These are the constraints for the corresponding loss terms.

4. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 3, characterized in that, The physical mechanism constraint for: in, For pollutant concentration, For the velocity field, Where is the diffusion coefficient. The concentration of suspended sediment. and These are the adsorption coefficient and the desorption coefficient, respectively. The concentration of pollutants adsorbed on the surface of sediment particles is calculated using the following formula: , The saturated adsorption coefficient of sediment. For identifying the state of particle media in a local area The arithmetic mean.

5. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 4, characterized in that, The sampling threshold mentioned in step 3 adopts a sampling strategy that combines the lower limit of concentration threshold with sediment zoning: 15% of the average pollutant concentration across the entire monitoring area is used as the minimum screening threshold; in high sediment-laden areas... Intensive sampling was conducted using 50% of the average concentration of pollutants across the entire area as the sampling threshold; in the clear water area... The lowest screening threshold is used as the sampling threshold for sparse sampling. in To obtain the suspended sediment concentration in the monitoring area, The critical adsorption threshold for sediment is determined through adsorption experiments; points with concentrations lower than the corresponding sampling threshold in their respective areas are removed to obtain the sampling point set. , This represents the number of sampling points.

6. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 1, characterized in that, In step 3, the model selection criterion is the Bayesian information criterion, and the calculation formula is as follows: in, The maximum likelihood value for the mixed probability distribution model. The number of model parameters, The total number of particle points; compare the corresponding values ​​for different numbers of pollution sources. The value is selected based on the minimum number of pollution sources; the mixed probability distribution model is a Gaussian mixture model, with the optimal number of pollution sources. As the clustering component number of the Gaussian mixture model, the inverted set of parameters for each pollution source is: , For position parameters, For spatial scale parameters, For source strength parameters.

7. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 1, characterized in that, The data-driven model mentioned in step 4 can be any one of a Long Short-Term Memory network, a gated recurrent unit network, or a temporal convolutional network, and its input is the historical state sequence of the pollution source. The output is the predicted state. This is used to capture the nonlinear evolution and history-dependent characteristics of pollution source states in complex flow fields, where... This represents the time step index at the current moment. For parameterized nonlinear mapping functions, The length of the historical window; the fluid dynamics model is based on the advection transport equation. predict The location of the pollution source at any given time is used to provide physical constraints during abrupt changes in the flow field. For the velocity field, For time step, This indicates the flow velocity at the current location; Indicating a data-driven model in Predicted state at any given time Indicating fluid dynamics model in The predicted state at any given time.

8. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 7, characterized in that, The flow field spatial gradient scalar mentioned in step 4 The logistic function is: in, To integrate weights, For threshold coefficient, This is the sensitivity coefficient; The fusion prediction formula is: When the flow field is stable , The data-driven model is dominant at times; when the flow field changes abruptly... , At that time, fluid dynamics models were the dominant approach.

9. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 1, characterized in that, The tracking confidence level mentioned in step 5 is: in The coordinates of the spatial center of the pollution source are used to integrate the predicted state; The coordinates of the spatial center of the pollution source are the parameters of the pollution source. For scale parameters; The preset threshold is an adaptive threshold: it calculates the lower quartile of the confidence levels of all pollution sources at the current time. Upper quartiles Interquartile range Adaptive threshold .

10. The method for pollutant characteristic inversion and dynamic tracking for discontinuous time-varying flow fields according to claim 1, characterized in that, The set period in step 5 is 24 hours. In the subsequent dynamic tracking process, at each moment, based on the three-dimensional pollutant concentration field updated in real time in step 2, the pollution source parameters at the current moment are recalculated using the same pollution source parameter inversion method as in step 3. The obtained pollution source parameters are used as the pollution source observation status at the current moment and are used for the tracking confidence assessment in step 5.

Images