Offshore pollutant concentration prediction method and system based on double-channel multi-layer perceptron
By using a dual-channel multilayer perceptron model, combined with feature interpretation and regularization techniques, the problems of data scarcity and environmental complexity in nearshore pollutant prediction were solved, achieving accurate prediction of pollutant concentrations and high model adaptability.
Patent Information
- Application Number
- CN202510922660.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-04
- Publication Date
- 2025-11-25
AI Technical Summary
Existing technologies for predicting pollutants in nearshore waters face challenges such as data scarcity, insufficient model generalization ability due to environmental complexity, and regional heterogeneity, making it difficult to achieve accurate pollutant concentration predictions.
A dual-channel multilayer perceptron approach is adopted. The first multilayer perceptron model is used for preliminary feature screening, and a second multilayer perceptron model with main channel residual dual channels is designed for further prediction. By combining feature interpretation methods and regularization techniques, the differential characteristics of marine environmental parameters and pollutant concentrations are processed, and the gradient propagation path is optimized.
It achieves accurate prediction of pollutant concentrations under small sample conditions, reduces dependence on monitoring resources, improves the model's adaptability and prediction accuracy to complex marine environments, and reduces the risk of overfitting.
Smart Images
Figure CN121009290A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of marine pollutant monitoring, and particularly relates to a near-shore pollutant concentration prediction method and system based on a double-channel multilayer perceptron. BACKGROUND
[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.
[0003] As a representative technology in the field of machine learning with high iteration efficiency and strong adaptability, neural networks have shown remarkable advantages in the field of environmental pollutant prediction in recent years. By simulating the working mechanism of the biological nervous system, the model constructs a multi-layer connection structure composed of an input layer, a hidden layer and an output layer, and realizes data nonlinear transformation and high-order feature extraction with the help of an activation function. Its core advantage lies in dynamically optimizing network parameters through a back propagation algorithm and automatically establishing a complex input-output mapping relationship using a gradient descent strategy. In particular, as a classic feedforward network, the multilayer perceptron (MLP) has become an important infrastructure in the field of environmental modeling due to its simple hierarchical structure and strong nonlinear fitting capability, and has provided an effective tool for dealing with complex environmental problems such as the spatiotemporal distribution of pollutants.
[0004] However, in the actual application process of neural networks in the prediction of near-shore pollutants, the following technical bottlenecks are faced: On the one hand, many pollutants exhibit trace diversity characteristics, and due to the sensitivity threshold of traditional detection technology, large-volume sampling and pre-processing enrichment (concentration factor > 1000 times) are required to realize quantitative analysis, resulting in high monitoring data acquisition cost and limited spatiotemporal resolution. On the other hand, the near-shore environment has significant regional heterogeneity - land-sea interaction, tidal circulation dynamic processes and multiphase medium migration mechanisms jointly form a complex pollutant transport network. The existing monitoring system has problems such as single data dimension and low fusion degree of heterogeneous data. In particular, it is worth noting that the characteristics of the mainstream atmospheric prediction model relying on large-scale training data and the sparsity of near-shore monitoring data form a sharp contradiction, resulting in a significant decrease in the regional adaptability of the model, making it difficult to accurately analyze the coupling rules of multi-source pollutants.
[0005] In summary, how to solve the problem of nonlinear mapping of complex marine monitoring system data and realize pollutant level inversion under extreme conditions such as limited analysis resources or missing monitoring data has become a problem to be solved in the prior art. SUMMARY
[0006] In view of the insufficient model generalization ability and regional heterogeneity problems caused by data scarcity and environmental complexity in existing offshore pollutant prediction, the purpose of the present application is to provide an offshore pollutant concentration prediction method and system based on a double-channel multilayer perceptron, which uses a first multilayer perceptron model to preliminarily extract features of offshore pollutants, and designs a second multilayer perceptron model with a main channel residual double channel for further prediction, realizes accurate prediction of pollutant concentration under small sample conditions, quantifies the coupling mechanism of environmental parameters and pollutant monomers, and reduces the dependence on monitoring resources.
[0007] In order to achieve the above-mentioned purpose, the present application is realized by the following technical solutions: The present application provides an offshore pollutant concentration prediction method based on a double-channel multilayer perceptron, comprising the following steps: Obtain and integrate pollutant occurrence data in offshore waters, and simultaneously collect marine environmental parameters to construct a multi-dimensional training data set; Use a first multilayer perceptron model combined with a feature interpretation method to screen and process key pollutant concentrations of the training data set; Establish a second multilayer perceptron model based on a double-channel multilayer perceptron with a main channel residual connection; Use the screened key pollutant concentrations to train the second multilayer perceptron model; Use the trained second multilayer perceptron model to predict the offshore pollutant concentration.
[0008] Further, the pollutant occurrence data includes dissolved phase data in offshore waters, and the marine environmental parameters include water depth, temperature, salinity, horizontal flow rate, vertical flow rate, and concentrations of dissolved oxygen, dissolved inorganic carbon, dissolved iron, nitrate, phosphate, silicate and chlorophyll-a.
[0009] Further, the specific steps of using a first multilayer perceptron model combined with a feature interpretation method to screen and process key pollutant concentrations of the training data set are: Use the first multilayer perceptron model to select features; Use the feature interpretation method to analyze and optimize the selected features.
[0010] Further, the specific steps of using the feature interpretation method to analyze and optimize the selected features are: Use SHAP values to quantify the marginal contribution of pollutant concentrations in different samples to the output of the first multilayer perceptron model, and use recursive feature elimination method to iteratively remove variables with low marginal contribution, and screen out key pollutant monomers with high marginal contribution.
[0011] Further, the structure of the second multi-layer perception model comprises a double-channel multi-layer perception, the double-channel comprises a main channel and an auxiliary channel, the main channel adopts a deep feature extractor to process the screened key pollutant concentration data and embeds a cross-layer residual connection structure, and the auxiliary channel adopts a lightweight module to process the marine environment parameters, and the output features of the double-channel are fused in a hidden layer through splicing.
[0012] Further, the specific steps for training the second multi-layer perception model by using the screened key pollutant concentration are as follows: The second multi-layer perception model is simply designed. Regularization and batch normalization operations are introduced in the training process. The training effect is verified in a cross-validation manner. The key pollutant concentration is screened again based on the verification result of the evaluation index by using a feature explanation method.
[0013] The second aspect of the present application provides a nearshore pollutant concentration prediction system based on a double-channel multi-layer perception, comprising: A data acquisition module is configured to acquire and integrate pollutant occurrence data in a nearshore sea area, and simultaneously collect marine environment parameters to build a multi-dimensional training data set. A first model construction module is configured to screen and process the key pollutant concentration of the training data set by using a first multi-layer perception model combined with a feature explanation method. A second model construction module is configured to establish a second multi-layer perception model based on a double-channel multi-layer perception with main channel residual connection. A model training module is configured to train the second multi-layer perception model by using the screened key pollutant concentration. A concentration prediction module is configured to predict the nearshore pollutant concentration by using the trained second multi-layer perception model.
[0014] The third aspect of the present application provides a computer readable storage medium, which stores a computer program, the computer program is suitable for being loaded and executed by a processor to perform the steps in the nearshore pollutant concentration prediction method based on a double-channel multi-layer perception as described in the first aspect of the present application.
[0015] The fourth aspect of the present application provides a computer device, which comprises: A processor is suitable for executing a computer program. A computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the nearshore pollutant concentration prediction method based on a double-channel multi-layer perception as described in the first aspect of the present application.
[0016] The fifth aspect of the present application provides a computer program product or computer program, which comprises computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the computer device to perform the steps in the offshore pollutant concentration prediction method based on a double-channel multi-layer perception machine according to the first aspect of the present application.
[0017] The above one or more technical solutions have the following beneficial effects: The present application discloses an offshore pollutant concentration prediction method and system based on a double-channel multi-layer perception machine. A first multi-layer perception machine model is used to preliminarily screen the features of offshore pollutants, and a second multi-layer perception machine model with a main channel residual double channel is designed for further prediction, thereby realizing accurate prediction of pollutant concentration under small sample conditions, quantifying the coupling mechanism of environmental parameters and pollutant monomers, and reducing the dependence on monitoring resources. The present application separates the differentiated features of pollutant concentration and environmental parameters through a double-channel architecture, avoids feature confusion, and improves the adaptability of the model to complex marine environments. The main channel residual connection optimizes the gradient propagation path, alleviates the gradient vanishing problem in deep network training, and improves the convergence speed and stability of the model. Based on the feature contribution decoupling technology, the overfitting risk under small sample data is significantly reduced, providing high toughness technical support for offshore pollution dynamic regulation.
[0018] The present application proposes a solution of double-channel architecture and residual connection collaborative optimization to solve the problems of insufficient model generalization ability and regional heterogeneity caused by data sparsity and environmental complexity in the prior art. The main channel deep feature extractor strengthens the learning of pollutant concentration data, the auxiliary channel lightweight module processes marine environmental parameters, feature fusion is performed in the middle section of the hidden layer, cross-layer residual connection is introduced to optimize the gradient propagation path, and regularization and other optimization techniques are used to suppress overfitting. The present application significantly improves the prediction accuracy of pollutant concentration by analyzing the coupling mechanism of multiple sources of pollutants.
[0019] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS
[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.
[0021] Figure 1 The flow chart of the offshore pollutant concentration prediction method based on the double-channel multi-layer perception machine in the embodiment one of the present application is shown in Figure 2 The prediction result schematic diagram of the MCRDC-MLP model in the embodiment one of the present application is shown in DETAILED DESCRIPTION
[0022] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
[0023] It is to be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments consistent with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof. The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0024] Embodiment one: The embodiment one of the present application provides a offshore pollutant concentration prediction method based on a double-channel multi-layer perception machine, as shown in Figure 1 The method comprises the following steps: Step S1: Obtain and integrate the pollutant occurrence data in the offshore sea area, and synchronously collect the marine environmental parameters to construct a multi-dimensional training data set.
[0025] The pollutant occurrence data includes the dissolved phase data of the offshore sea area, and the data is expanded to obtain a dissolved phase sample set.
[0026] The data expansion step comprises: The vertical profile data of each spatial sampling point is analyzed by depth layering, and the dissolved phase concentration data of different depth layers in the vertical direction of a single point are analyzed as independent samples, so as to expand the training set density in the spatial dimension and form a three-dimensional data set coupled with multiple depths.
[0027] The marine environmental parameters include water depth, temperature (T), salinity (Sal), horizontal flow velocity (V h ), vertical flow velocity (Vv ) and the concentrations of dissolved oxygen (DO), dissolved inorganic carbon (DIC), dissolved iron (DI), nitrate (N), phosphate (P), silicate (Si), and chlorophyll-a (Chl).
[0028] In a specific embodiment, a set of offshore seawater PAHs (Polycyclic aromatic hydrocarbons) dissolved phase samples are determined. The samples are pretreated and detected by using accelerated solvent extraction combined with gas chromatography mass spectrometry technology to obtain PAHs concentration data in each sample. The PAHs concentration data includes the content of naphthalene (Nap), acenaphthylene (Acy), acenaphthene (Ace), fluorene (Flu), phenanthrene (Phe), anthracene (Ant), fluoranthene (Flt), pyrene (Pyr), benzo[a]anthracene (BaA), chrysene (Chry), benzo[b]fluoranthene (BbF), benzo[k]fluoranthene (BkF), benzo[a]pyrene (BaP), indeno[1,2,3-cd]pyrene (IcdP), dibenzo[a,h]anthracene (DahA), and benzo[ghi]perylene (BghiE) in the above-mentioned samples. All 16 PAHs determined are divided into low molecular weight (LMW), medium molecular weight (MMW), and high molecular weight (HMW) according to the structure.
[0029] It should be particularly noted that PAHs in the present embodiment represents the polycyclic aromatic hydrocarbon category or the plural form, and monomer PAH represents a single polycyclic aromatic hydrocarbon compound, such as naphthalene, benzoperylene, and the like specific substances.
[0030] Step S2: using the first multi-layer perception model combined with the feature explanation method to perform key pollutant concentration screening processing on the training data set.
[0031] In actual environmental monitoring, due to the low concentration of target compounds, instrument sensitivity limitations, small sample size, and sample loss during transportation, etc. Challenges often lead to incomplete data sets. This problem is particularly prominent in marine pollution monitoring. To address this problem, the present embodiment constructs a first multi-layer perception model based on a feature explanation method, which aims to use the smallest subset of feature PAHs to predict the total concentration of 16 PAHs.
[0032] Step S201: using the first multi-layer perception model for feature selection.
[0033] In a specific embodiment, in order to facilitate training and verification, a simple MLP is first used to perform feature selection on LMW PAHs, MMW PAHs, and HMW PAHs.
[0034] The first multi-layer perception model (simple-MLP) of the embodiment is a common MLP structure, the MLP architecture includes three hidden layers, each layer has 32 neurons, adopts a ReLU activation function, L2 regularization (λ = 0.01), batch normalization and a dropout rate of 0.35. The model is optimized for the regression task using the Adam optimizer with a learning rate of 0.005. The input features include monomeric PAH concentrations classified into LMW, MMW and HMW groups, and the output is the corresponding total concentration of each group. The dissolved phase sample set contains 392 samples, which are divided into a training subset (68%), a validation subset (12%) and a test subset (20%), the training set and the validation set are standardized by Z-score, and the test set is standardized by the obtained scalar (standardizer). Three evaluation indicators are used to evaluate the performance of the model: the coefficient of determination (R²), the root mean square error (RMSE) and the mean absolute error (MAE).
[0035] , , .
[0036] wherein, represents the number of samples, represents the sample serial number, represents the true value, represents the predicted value, represents the average value of all true values.
[0037] As shown in Table 1, the simple-MLP model shows strong performance in both training and validation tasks.
[0038] Table 1. Performance parameters of the simple-MLP model in training and validation tasks
[0039] Step S202: The selected features are analyzed and optimized using a feature interpretation method, so that they can eliminate redundant inputs while maintaining prediction accuracy, thereby optimizing the model to adapt to the situation of limited monitoring resources.
[0040] The feature interpretation method of the embodiment selects SHAP (SHapley Additive exPlanations).
[0041] In one specific embodiment, SHAP values were calculated to quantify the contribution of each input variable for the purpose of interpreting the model and selecting relevant features. SHAP provides a consistent and interpretable way to attribute model predictions to individual features. Feature selection was performed and ranked by calculating the mean absolute SHAP value for each individual PAH (mean |SHAP Value|), with lower values indicating less impact and being potential candidates for removal in the model optimization process. The SHAP value formula for a feature is as follows: .
[0042] where, S denotes the union of S and the current jth feature, is the SHAP value for a feature j, F is the set of all features, S is a subset of features that does not include feature j, and the notation denotes the cardinality operator (used to measure the size of a set), and f( ) is the prediction function.
[0043] In one specific embodiment, SHAP values were used to quantify the marginal contribution of the concentration of pollutants in different samples to the output of the first multilayer perceptron model, and recursive feature elimination was used to iteratively remove variables with low marginal contribution, and to select key pollutants monomers with high marginal contribution.
[0044] SHAP analysis quantified and visualized the contribution of corresponding monomer PAHs to the total concentration of LMW, MMW, and HMW PAHs in the dissolved phase. In the dissolved phase, the contribution components of each group of PAHs were ranked in descending order of mean |SHAP Value| as follows: LMW PAHs were Nap, Phe, Flu, Acy, Ace, and Ant. MMW PAHs were Pyr, Flt, BaA, and Chry. HMW PAHs were BaP, DahA, BghiE, BbF, IcdP, and BkF.
[0045] Step S3: Establishing a second multilayer perceptron model based on a dual-channel multilayer perceptron with a main channel residual connection.
[0046] In one specific embodiment, the structure of the second multi-layer perceptron model comprises a dual-channel multi-layer perceptron, specifically comprising an input layer, a dual-channel hidden layer, a merged hidden layer, and an output layer. The present embodiment applies a main-channel residual dual-channel multi-layer perceptron (MCRDC-MLP) model to predict the total PAHs concentration in the dissolved phase. The model is specifically designed to handle two heterogeneous data types through two independent computational paths - channel A for feature PAHs (X PAHs ) and channel B for marine environmental variables (X Envs ). By introducing residual connections, the architecture ensures stable gradient propagation and deepens the model depth. The outputs of the two channels are integrated through a fully connected layer to generate the final prediction. This dual-channel structure enables the model to effectively capture different nonlinear relationships between dominant chemical components and secondary environmental factors, thereby improving fitting accuracy, especially under limited data conditions.
[0047] Specifically, the dual-channel includes a main channel and an auxiliary channel. The main channel uses a deep feature extractor to process the screened key pollutant concentration data and embeds a cross-layer residual connection structure to achieve effective gradient propagation through an identity mapping path. The cross-layer residual connection only acts on the main channel, and the skip connection across two layers directly transmits the features of the previous layer to the subsequent layer. The cross-layer residual connection structure of the main channel directly adds the input features and the hidden layer outputs without additional trainable parameters, which can effectively alleviate the gradient vanishing problem in deep network training.
[0048] The auxiliary channel uses a lightweight module to process marine environmental parameters (temperature, salinity, pH, dissolved oxygen, suspended particulate matter concentration, and sampling depth, etc.). The output features of the dual-channel are fused through concatenation in the hidden layer.
[0049] The present embodiment adopts a dual-channel MLP structure to adapt to different mechanisms of these different types of variables affecting the results. This structure can extract nonlinear features separately and improve prediction accuracy through cross-channel integration. Residual connections are introduced in the main channel to solve the gradient vanishing problem, allowing information to flow directly through the identity mapping. These "shortcuts" help to preserve basic information from PAHs features and reduce the risk of deep feature degradation. To further enhance the generalization ability and prevent overfitting, L2 regularization and kernel initialization strategies are applied. The model processes PAHs features (X PAHs ) and marine environmental variables (X Envs ) in parallel and combines their outputs through a fully connected layer for the final concentration prediction.
[0050] The complete formulation of the MCRDC-MLP model is as follows: .
[0051] This model aims to encode the effects of dominant PAHs features and secondary marine environmental parameters separately to reflect their different underlying mechanisms. By doing so, this model avoids the inefficiency of learning known physicochemical relationships from scratch, thereby improving the utilization of available data. Unlike standard MLPs that apply fully connected layers indiscriminately, this customized architecture is particularly advantageous when training data is limited.
[0052] Step S4: Train the second multi-layer perception model using the screened key pollutant concentrations.
[0053] This embodiment accelerates convergence through architecture simplification, regularization constraints, and batch normalization, and reduces data bias through K-fold cross-validation, constructing a second multi-layer perception model that has efficient training, precise learning ability, and strong anti-overfitting performance.
[0054] Step S401: Simplify the design of the second multi-layer perception model.
[0055] The simplification design includes constraining the MLP hidden layer dimension. Specifically, a MLP architecture with no more than 64 hidden layer neurons is used for complexity control.
[0056] This embodiment combines heterogeneous channel separation processing and residual connection mechanism, focusing on core feature extraction in the main channel and supplementing environmental background information in the auxiliary channel, effectively compensating for information loss caused by dimension reduction while maintaining lightweight hidden layers, achieving the unity of "high precision-low complexity-strong generalization".
[0057] Step S402: Introduce regularization and batch normalization operations during training.
[0058] In one specific embodiment, this embodiment introduces regularization operations during training: by modifying the loss function, incorporating L2 regularization terms and Dropout random neuron dropout strategies into the loss function, and imposing constraints on model weights to suppress complex fitting patterns: (1) L2 regularization: add a weight penalty term to the loss function: .
[0059] where, is the regularized loss function, is the initial loss function, is the weight, j is the jth feature, is the intensity coefficient, forcing the weight to tend to zero, used to reduce the sensitivity of the model to specific features.
[0060] (2) Dropout: randomly drop some neurons during training, forcing the network to learn redundant feature representations, simulating the effect of ensemble learning.
[0061] This embodiment introduces a batch normalization operation during training: batch normalization is performed on the input data of each layer, and the formula is: .
[0062] wherein, represents the standardized data, represents the data before standardization, and is the mean and variance of the current batch, is a data stability constant.
[0063] The use of batch normalization operation can reduce internal covariate shift, stabilize the distribution of each layer input, accelerate convergence; allow the use of a larger learning rate and improve the training speed, while the gradient smoothness reduces the risk of local minimum.
[0064] Step S403: verify the training effect using the evaluation index in a cross-validation manner.
[0065] In one specific embodiment, to ensure consistency and repeatability, this embodiment adopts a unified data division scheme: 68% for training, 12% for validation, and 20% for testing. Model performance is evaluated using evaluation indexes R², RMSE and MAE. Considering the relatively small data set, K-fold cross-validation is used to improve the robustness of the model. In this method, the data is divided into K subsets, of which K-1 are used for training and the remaining one is used for validation. This process is repeated K times to ensure that each subset acts as a validation set once. This technique reduces variance, helping to prevent overfitting and maximize the use of limited data. To further stabilize training and alleviate problems such as gradient explosion or disappearance, an early stopping mechanism is integrated. Specifically, 5-fold cross-validation is used, the model is trained for a maximum of 250 rounds, and if the validation loss does not improve for 50 consecutive rounds, early stopping is triggered.
[0066] K-fold cross-validation reduces data partitioning bias by dividing the training and validation sets in a loop, and ensures that each sample can participate in at least one validation, thereby improving data utilization to nearly 100%. During this process, the calculation results are averaged multiple times (such as 5-fold cross-validation), which can significantly reduce the random bias of data partitioning, making the model evaluation index have higher statistical confidence.
[0067] Step S404: using the feature interpretation method to perform secondary screening on the key pollutant concentration based on the verification result of the evaluation index, and further narrow the key pollutant concentration range.
[0068] In actual environmental monitoring, the low concentration of target compounds, the limitation of instrument sensitivity, the small sample size, and the sample loss during transportation often lead to incomplete data sets. This problem is particularly prominent in marine pollution monitoring. To address this issue, a minimum subset of characteristic PAHs was used to predict the total concentration of 16 priority control PAHs. The SHAP analysis combined with evaluation indicators was used to optimize the prediction results of the second multilayer perceptron model, enabling it to eliminate redundant inputs while maintaining prediction accuracy, thereby optimizing the model to adapt to limited monitoring resources. The evaluation indicators include R², RMSE, and MAE.
[0069] Specifically, the SHAP value combined with the verification result of the evaluation index quantifies the marginal contribution of different samples of marine environmental parameters and pollutant concentrations to the output of the second multilayer perceptron model, and the recursive feature elimination method is used to iteratively remove variables with low marginal contribution. On the basis of the first screening, key pollutant monomers with high marginal contribution are further screened out.
[0070] For example, in the dissolved phase, removing Phe from the input set significantly increases MAE and RMSE, indicating that it plays a crucial role in representing LMW PAHs together with Nap. Similar assessments were conducted for the MMW and HMW groups, and the PAHs with the highest SHAP scores were selected for each group. Verification confirmed that the selected PAHs are crucial to model performance, as excluding any single retained feature would result in a significant decrease in prediction accuracy. The final optimized PAH feature set contains only four key PAHs: Nap, Phe, Pyr, and BaP. Notably, the optimized MCRDC-MLP model achieves reliable prediction of total PAH concentration in the dissolved phase using only four key PAHs as inputs.
[0071] Step S5: using the trained second multilayer perceptron model to predict the concentration of pollutants in the nearshore area.
[0072] In a specific embodiment, the trained second multilayer perceptron model outputs the predicted value of the nearshore pollutant concentration and the feature contribution analysis report.
[0073] Specifically, the trained second multilayer perceptron model outputs the predicted value of PAH concentration in the nearshore dissolved phase, and simultaneously generates a feature contribution analysis report based on SHAP values, quantifying the model contribution weight of each pollutant monomer and marine environmental parameters (temperature, salinity, dissolved oxygen, etc.), and clarifying the coupling mechanism of key pollutant monomers and environmental parameters.
[0074] After training the MCRDC-MLP model, the predicted total PAHs concentration was highly consistent with the measured value. Figure 2 The correlation between the MCRDC-MLP model prediction value and the analysis value was represented. For the dissolved phase dataset, the R², RMSE and MAE were 0.86, 4.13 and 2.73, respectively. The model effectively captured the distribution pattern of total PAHs in the dissolved phase under the influence of marine environmental factors. To further study the contribution of each input variable to the model prediction, a SHAP-based sensitivity analysis was performed. The analysis results showed that the key PAHs contribution ranking was Nap, Pyr, Phe and BaP.
[0075] The experimental results clearly demonstrated that the MCRDC-MLP model had strong generalization ability, which achieved accurate prediction of PAHs concentration in the nearshore, thereby confirming its reliability. This performance was largely due to the unique dual-channel residual architecture of the model, which could effectively capture the nonlinear patterns in the pollutant data, while enhancing the representation ability of complex environmental information through multi-level feature fusion. By incorporating more widely sourced environmental data, the adaptability of the model to different spatial and pollution conditions can be further enhanced. In addition, the ability to predict total PAHs concentration using only a subset of representative PAHs makes this method particularly valuable when instrument sensitivity is limited or monitoring data is partially missing. Overall, the MCRDC-MLP model provides a robust and scalable solution for future applications in marine environmental monitoring and pollutant prediction.
[0076] Embodiment Two: The embodiment two of the present application provides a nearshore pollutant concentration prediction system based on a dual-channel multilayer perceptron, comprising: A data acquisition module configured to acquire and integrate pollutant occurrence data in the nearshore sea area, and simultaneously collect marine environmental parameters to build a multi-dimensional training dataset; A first model construction module configured to use a first multilayer perceptron model combined with a feature interpretation method to perform key pollutant concentration screening processing on the training dataset; A second model construction module configured to establish a second multilayer perceptron model based on a dual-channel multilayer perceptron with main channel residual connection; A model training module configured to train the second multilayer perceptron model using the screened key pollutant concentration; A concentration prediction module configured to predict the nearshore pollutant concentration using the trained second multilayer perceptron model.
[0077] Embodiment Three: The embodiment three of the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is suitable for being loaded and executed by a processor to perform steps in the offshore pollutant concentration prediction method based on a double-channel multi-layer perception machine as described in the embodiment one of the present application.
[0078] Embodiment four: The embodiment four of the present application provides a computer device, the device comprises: a processor suitable for executing a computer program; a computer readable storage medium, the computer readable storage medium stores a computer program, when the computer program is executed by the processor, the steps in the offshore pollutant concentration prediction method based on a double-channel multi-layer perception machine as described in the embodiment one of the present application are realized.
[0079] Embodiment five: The embodiment five of the present application provides a computer program product or a computer program, the computer program product or the computer program comprises computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the offshore pollutant concentration prediction method based on a double-channel multi-layer perception machine as described in the embodiment one of the present application.
[0080] The steps and methods in the above embodiments two, three, four and five correspond to the embodiment one, and the specific embodiments can refer to the related description part of the embodiment one.
[0081] The units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized in electronic hardware or a combination of electronic hardware and computer software. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The professional object can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.
[0082] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network or other programmable device. The computer instructions can be stored in or transmitted by a computer readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (for example, coaxial cable, optical fiber, digital line (DSL)) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data processing device such as a server, data center, etc. integrated with one or more available media. The available media can be a magnetic medium (for example, floppy disk, hard disk, magnetic tape), an optical medium (for example, DVD) or a semiconductor medium (for example, solid state disk (SSD)) and the like.
[0083] The above description is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A method for offshore pollutant concentration prediction based on dual-channel multi-layer perceptron, characterized in that, The method comprises the following steps: acquiring and integrating pollutant occurrence data in offshore waters, and synchronously collecting marine environmental parameters to construct a multi-dimensional training data set; using a first multi-layer perception model combined with a feature interpretation method to screen key pollutant concentrations from the training data set; establishing a second multi-layer perception model based on a double-channel multi-layer perception model with main channel residual connection; training the second multi-layer perception model using the screened key pollutant concentrations; using the trained second multi-layer perception model to predict offshore pollutant concentrations.
2. The dual-channel multi-layer perceptron based offshore pollutant concentration prediction method of claim 1, wherein, The marine environmental parameters include water depth, temperature, salinity, horizontal flow velocity, vertical flow velocity, and the concentrations of dissolved oxygen, dissolved inorganic carbon, dissolved iron, nitrate, phosphate, silicate, and chlorophyll-a.
3. The dual-channel multi-layer perceptron based offshore pollutant concentration prediction method of claim 1, wherein, The specific steps of using the first multi-layer perception model combined with the feature interpretation method to screen key pollutant concentrations from the training data set are: using the first multi-layer perception model to select features; using the feature interpretation method to analyze and optimize the selected features.
4. The dual-channel multi-layer perceptron based offshore contaminant concentration prediction method of claim 3, wherein, The specific steps of using the feature interpretation method to analyze and optimize the selected features are: using SHAP values to quantify the marginal contribution of pollutant concentrations in different samples to the output of the first multi-layer perception model, and using recursive feature elimination to iteratively remove variables with low marginal contribution, thereby screening key pollutant monomers with high marginal contribution.
5. The dual-channel multi-layer perceptron based offshore contaminant concentration prediction method of claim 1, wherein, The structure of the second multi-layer perception model includes a double-channel multi-layer perception model, with the double channel including a main channel and an auxiliary channel. The main channel uses a deep feature extractor to process the screened key pollutant concentration data and embeds a cross-layer residual connection structure. The auxiliary channel uses a lightweight module to process marine environmental parameters, and the output features of the double channel are fused in the hidden layer through splicing.
6. The dual-channel multi-layer perceptron based offshore contaminant concentration prediction method of claim 1, wherein, The specific steps of training the second multi-layer perception model using the screened key pollutant concentrations are: simplifying the design of the second multi-layer perception model; introducing regularization and batch normalization operations during training; using cross-validation to verify the training effect; using the feature interpretation method to perform secondary screening of key pollutant concentrations based on the evaluation index verification results.
7. A dual-channel multi-layer perceptron based offshore pollutant concentration prediction system, characterized in that, It comprises: a data acquisition module configured to acquire and integrate pollutant occurrence data in offshore waters, and synchronously collect marine environmental parameters to construct a multi-dimensional training data set; a first model construction module configured to use a first multi-layer perception model combined with a feature interpretation method to screen key pollutant concentrations from the training data set; a second model construction module configured to establish a second multi-layer perception model based on a double-channel multi-layer perception model with main channel residual connection; a model training module configured to train the second multi-layer perception model using the screened key pollutant concentrations; a concentration prediction module configured to use the trained second multi-layer perception model to predict offshore pollutant concentrations.
8. A computer program product, characterised in that, The computer program product comprises a computer program that, when executed by a processor, implements the offshore pollutant concentration prediction method based on the double-channel multi-layer perception model as claimed in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, the computer program being adapted to be loaded and executed by the processor to implement the offshore pollutant concentration prediction method based on the dual-channel multi-layer perception machine as claimed in any one of claims 1-6.
10. A computer device, comprising: Comprise: a processor adapted to execute the computer program; a computer readable storage medium, the computer readable storage medium storing a computer program, the computer program being executed by the processor to implement the offshore pollutant concentration prediction method based on the dual-channel multi-layer perception machine as claimed in any one of claims 1-6.