Tire wear particle environment fate prediction model construction method based on multi-dimensional data

Abstract

The invention relates to the technical field of environmental risk management and control, in particular to a tire wear particle environment fate prediction model construction method based on multi-dimensional data. Comprising the following steps: data acquisition and preprocessing; a feature importance analysis model based on a random forest is constructed, key factor mining is performed on the preprocessed collected data, core factors influencing tire wear particle environment fate are screened by calculating Gini importance values of the key factors, and feature vectors and confidence coefficients of the core factors are output; classifying and integrating factors; carrying out model association construction; model training storage; and performing model verification feedback. According to the method, the continuity, timeliness and consistency of tire wear particle sample data and environment associated data are improved through the built tire wear particle environment fate prediction model and through cooperative collection, space-time registration and noise filtering preprocessing of the multi-source monitoring equipment, and a high-quality data basis is provided for subsequent model building.

Description

Technical Field

[0001] This invention relates to the field of environmental risk management technology, and more specifically, to a method for constructing a prediction model for the environmental fate of tire wear particles based on multi-dimensional data. Background Technology

[0002] As high-traffic hubs, highway service areas contain a large amount of tire wear particles in their wastewater and surrounding stormwater runoff. These particles, as typical emerging pollutants from traffic sources, have a direct impact on the environmental safety of the surrounding ecosystem due to their environmental characteristics, migration patterns, and final fate. Currently, research on tire wear particles in the specific scenario of highway service areas is scarce. Their distribution within wastewater treatment facilities and the mechanisms by which they are influenced by factors such as traffic volume and meteorological conditions remain unclear. There is an urgent need to provide a scientific basis for environmental risk management through accurate predictive models.

[0003] Current research on tire wear particles largely focuses on detection and analysis in a single environmental medium, or relies on simple statistical models to infer migration patterns, which has significant shortcomings. Firstly, data acquisition dimensions are limited, failing to integrate real-time data from multiple monitoring devices and neglecting the spatiotemporal differences between different monitoring data, resulting in insufficient data reliability. Secondly, the identification of key influencing factors is not comprehensive enough, often ignoring the potential impact of low-proportion, weakly correlated factors on the environmental fate of tire wear particles. Furthermore, the lack of capture of multi-scale spatiotemporal characteristics makes it difficult to accurately reflect the evolution of particles across different time dimensions (such as short-term fluctuations, medium-term accumulation, and long-term cycles), failing to meet the high-precision prediction requirements for tire wear particle environmental fate in complex scenarios such as highway service areas. Therefore, we propose a method for constructing a tire wear particle environmental fate prediction model based on multi-dimensional data. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a tire wear particle environmental tendency prediction model based on multi-dimensional data, so as to solve the problems of limited data collection dimensions and insufficient identification of key influencing factors mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides a method for constructing a tire wear particle environmental tendency prediction model based on multi-dimensional data, comprising the following steps: S100. Data Acquisition and Preprocessing: Through synchronously deployed multi-source monitoring equipment, tire wear particle sample data and environmental correlation data are collected in real time to ensure data continuity and timeliness; and spatiotemporal registration and noise filtering are performed on the collected multi-dimensional data to ensure the spatiotemporal consistency of the multi-dimensional data. S200, Key Factor Detection: A feature importance analysis model based on random forest is constructed to mine key factors in the preprocessed collected data. By calculating the Gini importance value of each key factor, the core factors affecting the environmental fate of tire wear particles are screened, and the feature vectors and confidence scores of the core factors are output. A linear weighted enhancement mechanism is introduced to compensate for the weights of low-proportion and weakly correlated factors with Gini importance values ​​below the threshold, thereby optimizing the identification efficiency of low-proportion and weakly correlated factors. S300, Factor Classification Integration: Using a training-converged classification model, the selected core factors are classified to distinguish their physical attributes and functional dimensions; and based on the feature encoding network, the core factors are further classified in a more precise manner according to their data form and association patterns. S400, Model Association Construction: Combining the improved multi-scale spatiotemporal attention fusion algorithm, through variable association and parameter calibration, the mapping relationship between core factors and the environmental fate of tire wear particles is established, the basic architecture of the initial model for predicting the environmental fate of tire wear particles is built, and the preliminary fate evolution path is output. S500, Model Training and Storage: The multi-dimensional data collected by S100 and the mapping relationship established by S400 are input into the iterative training environment according to the training protocol to optimize the parameters of the initial model for predicting the environmental tendency of tire wear particles built by S400; and the training logs recorded in real time during the training process, the intermediate models generated at preset intervals during the training process, and the final model generated after the training are completed are stored in the database for subsequent call and analysis. S600 Model Validation Feedback: Based on the output of the intermediate model generated during the S500 training process, when the prediction error exceeds the preset threshold or the evolution logic is abnormal, a closed-loop optimization mechanism is triggered. The model parameters, training set, or algorithm module of the intermediate model are adjusted according to the validation rules until the environmental regression prediction accuracy requirements are met.

[0006] As a further improvement to this technical solution, in S100, the multi-source monitoring device collects tire wear particle sample data and environmental correlation data in real time according to a preset acquisition frequency. To ensure the continuity and timeliness of the data, the multi-source monitoring device includes a laser-induced breakdown spectrometer, a nanoparticle tracking analyzer, a micro weather station, and intelligent traffic sensors. The acquisition frequency is dynamically adjusted according to changes in traffic flow and the level of environmentally sensitive areas (e.g., increasing the particle data acquisition density during peak traffic periods and increasing the environmental data acquisition density in ecologically sensitive areas). At the same time, BeiDou time synchronization technology is used to reduce time asynchrony problems caused by equipment clock deviation, ensuring the consistency of timestamps for multi-dimensional data and providing a basis for subsequent spatiotemporal registration operations.

[0007] As a further improvement to this technical solution, in step S100, preprocessing operations such as spatiotemporal registration and noise filtering are performed on the collected multi-dimensional data, including the following steps: S100.1 Unified Time Reference: The BeiDou time service system is used to synchronize and calibrate the timestamps of the laser-induced breakdown spectrometer, nanoparticle tracking analyzer, micro weather station and intelligent traffic sensor to ensure the consistency of timestamps of multi-source data (time synchronization accuracy meets industry standards) and generate raw datasets with unified UTC time labels. S100.2 Spatial Coordinate Alignment: Based on the time synchronization results of S100.1, the spatial acquisition positions of each device are converted into planar coordinates using the UTM projection coordinate system. The non-uniformly distributed sampling point data is mapped to regular grid cells (the grid scale is set according to the monitoring area) using bilinear interpolation. Reasonable spatial deviation data caused by the device deployment position are fused and corrected. S100.3 Multi-level noise filtering: First, outliers that deviate from the overall distribution in data such as particle concentration and element content are identified and removed through the local outlier factor algorithm (outlier determination is based on the statistical regularity of the data sequence). Then, the filtered data is decomposed into multiple wavelet bases using the sym8 wavelet basis, and the high-frequency coefficients are processed with soft thresholding to reduce random noise, generating a spatiotemporally consistent preprocessed dataset.

[0008] As a further improvement to this technical solution, in step S200, constructing a feature importance analysis model based on random forest includes the following steps: S200.1, Model parameter initialization: The random forest is configured to contain multiple CART decision trees (the number of decision trees is dynamically adjusted according to the amount of input data; the larger the amount of data, the more trees are added accordingly). The maximum depth of a single tree is adapted to the dimension of the input data (the higher the dimension of the data, the greater the depth is to preserve feature details). Set a minimum number of leaf node samples (to ensure that node splitting is statistically significant and to avoid bias caused by insufficient sample size), and the number of random feature subsets is the square root of the total number of input factors (an industry-standard overfitting suppression method). S200.2, Calculation of Gini Importance Value: For each decision tree, the Gini importance contribution value of each factor is calculated through the node splitting process, using the following formula: ; in, Indicates the first Environmental factors (such as particle concentration, wind speed, traffic flow, etc.); Indicates use The decision tree node that performs the split; Indicates the number of samples in a node; Indicates the total number of samples; Represents a node The impurity of the gin; , This represents the Gini impurity of the left and right child nodes after a node splits. Indicates the first Environmental factors At the node The Gini significance contribution value at the location; The factor is obtained by averaging the contributions of all decision trees. The final Gini importance value : ; in, Indicates the total number of decision trees; Represents the decision tree index; S200.3, Core Factor Screening: Calculate the maximum Gini importance of all environmental factors. Set the filtering threshold : ; in, The threshold coefficient (dynamically adjusted based on the total number of factors; the more factors, the higher the threshold coefficient). The value is reduced accordingly to retain a sufficient number of candidate factors. Select Environmental factors were used as the initial candidate set of core factors; S200.4, Weight Enhancement Adaptation: The linear weighted enhancement mechanism in S200 is invoked to compensate for low-proportion factors whose Gini importance value is lower than 60% of the mean in the candidate set (the compensation formula follows the linear weighting rule, and the enhancement magnitude is dynamically determined according to the number and proportion of low-proportion factors), and finally the feature vector and confidence of the core factor are output.

[0009] As a further improvement to this technical solution, in step S200.4, the weight enhancement adaptation includes the following steps: S200.4.1 Determination of Low-Proportion Factors: Based on the mean Gini importance value of all factors in the core factor candidate set, factors with a proportion lower than 60% of the mean are determined to be low-proportion, weakly correlated factors, and their proportion is recorded. (Number of low-proportion factors / Total number of candidates); S200.4.2, Grading Enhancement Coefficient Setting: According to Set the differentiation enhancement coefficient : when (A small number of low-proportion factors, moderately enhanced); when (Medium proportion, enhanced compensation); when (A large number of low-proportion factors are used to control the enhancement magnitude and avoid distortion.) S200.4.3, Dynamic Weight Compensation: For factors with low proportions in the judgment, a linear weighting formula is used to enhance their effectiveness: ; in, This represents the original Gini importance value for low-proportion factors; The filtering threshold set for S200.3; Indicates low proportion factor The Gini importance value after dynamic compensation; At the same time, an enhanced upper limit constraint is introduced (to prevent low-proportion factors from being excessively enhanced): ; in, The maximum Gini importance value in the candidate set; Indicates the upper limit coefficient of enhancement; Indicates the first The Gini importance values ​​of low-proportion factors after weight enhancement and adaptation; S200.4.4, Enhancement Effect Verification: Calculate the weight bias rate of the enhanced candidate set : ; in, This represents the maximum Gini importance value of all factors in the core factor candidate set after weight enhancement and adaptation. This represents the minimum Gini importance value of all factors in the core factor candidate set after weight enhancement and adaptation. This represents the arithmetic mean of the Gini importance values ​​of all factors in the core factor candidate set after weighted enhancement and adaptation. like If the enhancement is confirmed to be effective, the adjusted core factor set will be output. like Return to S200.4.2 and readjust. (e.g., ±0.1 from the original grading), until... .

[0010] As a further improvement to this technical solution, in step S300, factor classification integration includes the following steps: S300.1, Physical Property Classification: By using a training-converged classification model, the physical attribute categories to which the core factors belong are distinguished; S300.2, Functional Dimension Determination: By using a feature encoding network, the influence of core factors on the environmental fate of tire wear particles can be determined. S300.3, Secondary Classification Code: Based on feature coding networks, core factors are classified into two levels according to data form and association pattern; S300.4, Classification Result Integration: The classification results of S300.1-S300.3 are cross-integrated to form a multi-dimensional label system that includes physical attributes, function dimensions, and data-related characteristics, and is used to construct the multi-branch input of the prediction model.

[0011] As a further improvement to this technical solution, in S400, combined with an improved multi-scale spatiotemporal attention fusion algorithm, a mapping relationship between core factors and the environmental fate of tire wear particles is established through variable correlation and parameter calibration, including the following steps: S410.1 Multi-scale feature extraction and initial variable association: Using the core factors and multidimensional labeling system formed by S300.4 as input, three sets of differentiated 1D convolutional units are used to capture features hierarchically: Small-scale unit: 1×3 convolution kernel (sliding window stride 1, number of output channels 32), capturing the high-frequency fluctuation characteristics of core factors within 1-3 hours (such as the immediate impact of sudden changes in traffic flow on particle diffusion, supporting 24-hour short-term trend prediction). Mesoscale unit: 1×5 convolution kernel (sliding window stride 2, number of output channels 32), capturing the cumulative effect of core factors within 1-2 days (such as the superimposed effect of continuous rainfall on particle deposition, supporting 72-hour mid-term trend prediction). Large-scale unit: 1×7 convolution kernel (sliding window stride 3, number of output channels 32), capturing the periodic patterns of core factors within 1 week (such as the long-term regulation of particle degradation by weekly average temperature, supporting 15-day long-term trend prediction). Output multi-scale feature tensor set These correspond to short-term, medium-term, and long-term feature dimensions, respectively. S410.2, Calculation of Spatiotemporal Attention Weights: Based on feature tensors and S300 multidimensional labels, spatial and temporal weights are calculated to strengthen key associations: Spatial weight calculation: Based on the physical attribute categories of S300.1, basic weights are assigned to different attribute factors (e.g., weight coefficient of 1.2 for particle physical properties, and 1.0 for environmental conditions), and then the weights are calculated using the factor-regression mutual information matrix. optimization: ; in, Indicates the first Spatial attention weights of each core factor; Indicates the first The initial weights of the core factors; Indicates the first The core factor and the first One environmental trend indicator; This represents the total number of environmental trend indicators; This represents the total number of core factors; The weighting integrates the priority of physical attributes with the actual correlation strength, avoiding biases caused by relying solely on data. Time weight calculation: Set the time decay coefficient for the dimension of action of S300.2. Calculate time steps Weights: ; in, Indicates the first Time attention weights for each time step; Indicates the total number of time steps; Different dimensions of action result in different rates of decay over time (e.g., migration and diffusion are more affected by recent factors). (Higher value) S410.3 Feature Fusion and Mapping Relationship Construction: Weigh and fuse multi-scale feature tensors with spatiotemporal weights: ; in, Represents the multi-scale spatiotemporal weighted fusion feature tensor; Represents the spatial attention weight matrix; Represents the temporal attention weight vector; This represents element-wise multiplication; Indicates small scale Mesoscale Large scale Dimensional splicing of features; The fully connected network with batch normalization layer is used to integrate the feature inputs and output a regression correlation matrix to complete the mapping from core factors to environmental regression. S410.4, Parameter Calibration: Adjust the calibration rules based on the secondary classification labels (continuous / discrete, linear / nonlinear correlation) of S300.3: For discrete factors: calibrate the convolution kernel stride (e.g., if the discrete factor accounts for more than 40%, decrease the stride by 0.5). For nonlinear correlation factors: adjust the number of nodes in the fully connected network; Repeat steps S410.1-S410.3 until the average strength of the regression correlation matrix is ​​≥0.65.

[0012] As a further improvement to this technical solution, in S400, an improved multi-scale spatiotemporal attention fusion algorithm is used to build an initial model architecture for predicting the environmental tendency of tire wear particles through phased variable association and dynamic parameter calibration, including the following steps: S420.1, Phased Association Modeling of Variables: Linear correlation stage: Partial least squares regression is used to establish the linear correlation between core factors and regression indicators, the top 5 contributing factors are selected, and the linear correlation coefficient matrix is ​​output. ; Nonlinear correlation stage: For the dominant factor and the regression index, a nonlinear correlation model is constructed through radial basis function interpolation, and the correlation strength matrix is ​​output. ; S420.2, Parameter Dynamic Calibration Optimization: Set calibration threshold (Pick 1.2 times the mean, and ),like mean Then adjust dynamically: Multi-scale expansion: Add large-scale convolutional units with a scale of 1×9; Timing optimization: Adjust the time decay coefficient in S410.2 ; Repeat steps S410.1-S410.4 until... mean ; S420.3 Initial Model Architecture Setup: The infrastructure comprises three functional modules: Input layer: Receives multi-dimensional labeled core factors output by S300 (embedded physical attributes, function dimension label encoding, with the dimension and number of factors being consistent); Fusion Layer: Integrates multi-scale spatiotemporal weighted fusion feature tensors from S410.3 , and the correlation strength matrix Weighted fusion; Output layer: The softmax activation function is used to output a multi-branch regression path.

[0013] As a further improvement to this technical solution, in S500, the model training storage includes the following steps: S500.1, Training Dataset Construction: Extract the multi-dimensional data from the S100 output and the trend correlation features from the S400 output. Based on the time dimension primary key aligned data, generate a training dataset containing input features and trend labels. Use the index slicing method to split the dataset into training set, validation set, and test set. S500.2, Model Training Execution: Load the initial model file for predicting the environmental tendency of tire wear particles output by S400, call the standard training interface of the training framework, and configure the basic training parameters (including batch size, training epochs, and optimizer type); trigger validation set evaluation at preset intervals during training to continuously optimize model parameters; S500.3, Model and Data Storage: After each preset training round, the intermediate model parameter file is automatically exported, and the storage path is associated with the unique technical identifier of the initial model for predicting the environmental tendency of tire wear particles built by S400. After training is terminated, the final model file is exported, and the training parameter configuration and dataset splitting rules are synchronously stored in the database. The database can then be searched and associated using the unique technical identifier of the initial model for predicting the environmental tendency of tire wear particles built by S400.

[0014] As a further improvement to this technical solution, in S600, model verification feedback is achieved through anomaly triggering and closed-loop optimization processes, including the following steps: S600.1, Definition of Verification Indicators and Anomaly Monitoring: Key indicator inheritance: Based on the trend mapping relationship output by S400, key indicators for trend prediction are extracted as a verification benchmark. Abnormal triggering conditions: Real-time calculation of the error between the output of the S500 intermediate model and the true value. When the error exceeds the accuracy requirement of the S400 regression index for three consecutive times, or when the evolution logic is found to be inconsistent with the S300 classification features (such as short-term factors driving long-term regression, and S300 being a multi-scale feature classification), the optimization mechanism is triggered. S600.2, Closed-loop optimization strategy and its relation to preceding steps: Parameter adjustment: When the error is abnormal, the S500 training parameters are adjusted first, and the iteration is carried out at the preset step size until the error decreases; Dataset update: If parameter adjustment is ineffective, supplement abnormal scene samples (such as the combination of environmental factors corresponding to high errors) from the multi-dimensional data collected by S100, update the training set, and re-trigger the S500 training process. Module replacement: When the evolution logic is abnormal, replace the algorithm module in the corresponding dimension of S400 and retrain based on the S300 classification label; S600.3, Accuracy Convergence Judgment Rules: The output error of the optimized model is continuously monitored. When the cycle error of 10 consecutive validations is less than or equal to the preset threshold of S400, and the evolution logic matches the classification features of S300 (such as short-term factor-driven short-term regression, and S300 is a multi-scale feature classification), the accuracy requirement is met, and the closed-loop optimization is terminated.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, through the construction of a tire wear particle environmental fate prediction model, effectively improves the continuity, timeliness, and consistency of tire wear particle sample data and environmental correlation data by using collaborative acquisition and spatiotemporal registration of multi-source monitoring devices and noise filtering preprocessing, thus providing a high-quality data foundation for subsequent model construction; 2. This invention, through a feature importance analysis model based on random forest combined with a linear weighted enhancement mechanism, can not only accurately screen the core factors affecting the environmental fate of tire wear particles, but also effectively identify and compensate for the weights of low-proportion and weakly correlated factors, thus avoiding the omission of key influencing factors and improving the comprehensiveness and accuracy of factor identification. 3. This invention utilizes a training-converged classification model and feature encoding network to classify and integrate core factors, forming a multi-dimensional label system that includes physical attributes, function dimensions, and data-related characteristics. This enables the model to understand the function mechanism of different factors in greater detail and provides clear logical support for the construction of multi-branch inputs. 4. By combining the improved multi-scale spatiotemporal attention fusion algorithm to construct the mapping relationship, this invention can effectively capture the characteristics of core factors at different time scales and their correlation patterns with the environmental tendency of tire wear particles. The initial model architecture can output a more realistic preliminary tendency evolution path. 5. Through iterative training and closed-loop optimization mechanisms, the model parameters of this invention are continuously optimized. The storage and retrieval design of intermediate and final models not only facilitates subsequent analysis and application, but also allows for timely adjustments when the prediction error exceeds the threshold or the evolution logic is abnormal, ensuring that the model always meets the accuracy requirements of environmental trend prediction and improving the adaptability and reliability of the model in complex environmental scenarios. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 like Figure 1 As shown, this embodiment provides a method for constructing a tire wear particle environmental tendency prediction model based on multi-dimensional data, including: S100. Data Acquisition and Preprocessing: Through synchronously deployed multi-source monitoring equipment, tire wear particle sample data and environmental correlation data are collected in real time to ensure data continuity and timeliness; and spatiotemporal registration and noise filtering are performed on the collected multi-dimensional data to ensure the spatiotemporal consistency of the multi-dimensional data. In this step, in S100, the multi-source monitoring equipment collects tire wear particle sample data and environmental correlation data in real time according to a preset acquisition frequency. To ensure the continuity and timeliness of the data, the multi-source monitoring equipment includes a laser-induced breakdown spectrometer, a nanoparticle tracking analyzer, a micro weather station, and intelligent traffic sensors. The acquisition frequency is dynamically adjusted according to changes in traffic flow and the level of environmentally sensitive areas (e.g., increasing the particle data acquisition density during peak traffic periods and increasing the environmental data acquisition density in ecologically sensitive areas). At the same time, BeiDou time synchronization technology is used to reduce time asynchrony problems caused by equipment clock deviation, ensuring the consistency of timestamps for multi-dimensional data and providing a basis for subsequent spatiotemporal registration operations.

[0019] As a further explanation of this step, the laser-induced breakdown spectroscopy (LIBS) in this embodiment is used for real-time analysis of the elemental composition of particles (such as characteristic elements of rubber like zinc and sulfur), and its in-situ detection capability can quickly obtain the chemical fingerprint of particles; the nanoparticle tracking analyzer (NTA) focuses on the particle size distribution and concentration quantification of submicron-sized particles, forming a complementary characterization of physicochemical properties with LIBS. The micro-weather station is responsible for collecting meteorological parameters such as wind speed, wind direction, temperature, and humidity, while the intelligent traffic sensor records traffic data such as traffic flow and vehicle type distribution. Together, they constitute the core source of environmental correlation data.

[0020] As a further explanation of this step, in this embodiment, LIBS and NTA are deployed in particulate-rich areas such as the inlet of the service area's sewage treatment facility, the stormwater drainage network's confluence point, and the area around the parking lot; the micro weather station is set up in an open, unobstructed area of ​​the service area to ensure the representativeness of the meteorological data; and the intelligent traffic sensor is installed at the service area's entrances and exits and key nodes on internal roads to achieve full-path coverage of traffic flow. The collection frequency is adjusted based on a dynamic triggering mechanism: under normal conditions, the collection frequency of LIBS and NTA is 1 time / 10 minutes, the micro weather station is 1 time / 5 minutes, and the intelligent traffic sensor is 1 time / minute; when the intelligent traffic sensor detects that the traffic flow reaches 1.5 times the daily average flow (peak period threshold), the frequency of LIBS and NTA automatically increases to 1 time / 2 minutes; for Class I environmentally sensitive areas within 3 kilometers of drinking water sources, the collection frequency of the micro weather station is doubled to ensure precise capture of the impact of meteorological factors on particulate migration.

[0021] In this step, S100 performs preprocessing operations such as spatiotemporal registration and noise filtering on the collected multi-dimensional data, including the following steps: S100.1 Unified Time Reference: The BeiDou time service system is used to synchronize and calibrate the timestamps of the laser-induced breakdown spectrometer, nanoparticle tracking analyzer, micro weather station and intelligent traffic sensor to ensure the consistency of timestamps of multi-source data (time synchronization accuracy meets industry standards) and generate raw datasets with unified UTC time labels. S100.2 Spatial Coordinate Alignment: Based on the time synchronization results of S100.1, the spatial acquisition positions of each device are converted into planar coordinates using the UTM projection coordinate system. The non-uniformly distributed sampling point data is mapped to regular grid cells (the grid scale is set according to the monitoring area) using bilinear interpolation. Reasonable spatial deviation data caused by the device deployment position are fused and corrected. S100.3 Multi-level noise filtering: First, outliers that deviate from the overall distribution in data such as particle concentration and element content are identified and removed through the local outlier factor algorithm (outlier determination is based on the statistical regularity of the data sequence). Then, the filtered data is decomposed into multiple wavelet bases using the sym8 wavelet basis, and the high-frequency coefficients are processed with soft thresholding to reduce random noise, generating a spatiotemporally consistent preprocessed dataset.

[0022] As a further explanation of this step, in this embodiment, spatiotemporal registration is achieved by unifying the time base with the spatial coordinates to eliminate spatiotemporal deviations of multi-source data. The specific steps are as follows: First, unify the time base: Based on the UTC time output by the BeiDou time synchronization system, the original timestamps of each device are linearly calibrated. Let the original timestamp of the device be... The calibrated timestamp is The calibration formula is: ;in, The device clock offset is calculated by the difference between the BeiDou time synchronization signal and the device's local clock. After calibration, the timestamp deviation of different devices for the same physical event (such as peak traffic flow during a certain period) does not exceed 50ms, meeting the time consistency requirements of multi-source data. Then, perform spatial coordinate alignment: The latitude and longitude coordinates of each device are represented using the UTM projected coordinate system (based on the WGS84 ellipsoid). Convert to Cartesian coordinates The conversion formula is as follows: ; Where M is the arc length of the meridian, and N is the radius of curvature of the trochanteric circle. ( (the second eccentricity of the ellipsoid) =0.9996 is the scaling factor; After coordinate transformation, bilinear interpolation is used to map the non-uniformly distributed sampling point data to a regular 5m×5m grid (the grid size is dynamically adjusted according to the service area area, usually 5-20m). The interpolation formula is: ; in, , Coordinates of adjacent sampling points For grid points The interpolation result; Represents the planar coordinates of the target regular grid points; This represents the monitoring data value at sampling point 1; This represents the monitoring data value at sampling point 2; This indicates the spacing of sampling points in the x-direction, reflecting the density of equipment deployment; This represents the spacing of sampling points in the y-direction. For spatial deviations caused by device deployment locations (such as differences in sensor installation height), an altitude correction coefficient (calculated based on the vertical distance between the device installation height and the ground) is introduced for fusion correction to ensure the continuity of spatial data.

[0023] As a further explanation of this step, in this embodiment, multi-level noise filtering improves the data signal-to-noise ratio through outlier removal and random noise attenuation. The specific steps are as follows: First, the Local Outlier Factor (LOF) algorithm is used to identify outliers in data such as particle concentration and elemental content. For a given data point... The formula for calculating its LOF value is: ; in, for of Neighborhood (in this embodiment) =20, determined based on historical data statistics); Locally accessible density; It is used to determine data points of - Parameters of the neighborhood; The data sample representing whether the point to be detected is an outlier; yes Any data point in the data, i.e. Data samples within the neighborhood; when When >1.5, determine This threshold, which identifies and removes outliers, is optimal in balancing anomaly detection sensitivity and data retention. Then, the data after outlier removal is decomposed into four levels using the sym8 wavelet basis (the number of decomposition levels is determined based on the frequency domain characteristics of the data). Soft thresholding is applied to the high-frequency coefficients using the following formula: ; in, These are wavelet coefficients obtained after soft thresholding. This represents the scale (level) of wavelet decomposition, with different levels. Signal components corresponding to different frequency ranges; In scale Wavelet coefficient index below; After wavelet decomposition, at scale ,index The original wavelet coefficients of the location; It is a symbolic function; It is a soft threshold used to determine whether wavelet coefficients are noise-dominated.

[0024] Furthermore, the preprocessing effect in this embodiment is verified through multi-dimensional indicators: after unifying the time base, the cross-correlation coefficient of data from different devices is calculated, requiring the coefficient to be ≥0.8 after synchronization (an improvement of ≥30% compared to before synchronization); after spatial coordinate alignment, the Moran's I index is used to evaluate spatial autocorrelation, ensuring the index is ≥0.6 to verify spatial continuity; the data before and after noise filtering are compared through probability density distribution, requiring the kurtosis of the filtered data to be reduced by ≥20% to verify the noise reduction effect; If the verification results fail to meet the standards, a feedback mechanism is triggered: if time synchronization is abnormal, the BeiDou timing signal receiving module is recalibrated; if spatial correlation is insufficient, the sampling points are encrypted or the grid scale is adjusted; if noise filtering is ineffective, the wavelet decomposition level (±1 level) or threshold coefficient (±10%) is adaptively adjusted. After iterative optimization, the final preprocessed dataset is stored in the form of a multidimensional array (dimensions: timestamp × spatial grid × parameter type), which is directly used as input data for key factor detection in the S200 step to ensure the reliability of subsequent model construction.

[0025] S200, Key Factor Detection: A feature importance analysis model based on random forest is constructed to mine key factors in the preprocessed collected data. By calculating the Gini importance value of each key factor, the core factors affecting the environmental fate of tire wear particles are screened, and the feature vectors and confidence scores of the core factors are output. A linear weighted enhancement mechanism is introduced to compensate for the weights of low-proportion and weakly correlated factors with Gini importance values ​​below the threshold, thereby optimizing the identification efficiency of low-proportion and weakly correlated factors. In this step, S200, constructing a feature importance analysis model based on random forest includes the following steps: S200.1, Model parameter initialization: The random forest is configured to contain multiple CART decision trees (the number of decision trees is dynamically adjusted according to the amount of input data; the larger the amount of data, the more trees are added accordingly). The maximum depth of a single tree is adapted to the dimension of the input data (the higher the dimension of the data, the greater the depth is to preserve feature details). Set a minimum number of leaf node samples (to ensure that node splitting is statistically significant and to avoid bias caused by insufficient sample size), and the number of random feature subsets is the square root of the total number of input factors (an industry-standard overfitting suppression method). As a further explanation of this step, the initial number of decision trees in this embodiment is set to 100. For every 5000 additional input data samples (such as the total number of preprocessed tire wear particle concentration, weather, traffic flow, etc. data records), the number of decision trees increases by 50 (up to a maximum of 500), ensuring feature capture capability under large sample data. The maximum depth of a single tree is positively correlated with the dimension (D) of the input factor, as shown in the formula: For example, if the preprocessed data contains 10,000 samples (each containing 20 factors such as particle concentration, wind speed, and traffic flow), then the number of decision trees would be 100 + (10,000 - 5,000) / 5,000 × 50 = 150. This setting balances computational efficiency and model stability, avoiding overfitting when the data volume is too large. Simultaneously, the minimum number of leaf node samples is fixed at 5 to ensure that node splits are statistically significant and reduce small sample bias. Furthermore, the number of random feature subsets is set to the square root (integer) of the total number of input factors. For example, if the total number of factors is 20, four features are randomly selected for each split, a common overfitting suppression method in the industry.

[0026] S200.2, Calculation of Gini Importance Value: For each decision tree, the Gini importance contribution value of each factor is calculated through the node splitting process, using the following formula: ; Gini impurity is a classic splitting criterion for CART decision trees. This formula is derived based on this theory and is used to measure the factors. At the node The contribution of splitting to reducing data clutter; in, Indicates the first Environmental factors (such as particle concentration, wind speed, traffic flow, etc.); Indicates use The decision tree node that performs the split; Indicates the number of samples in a node; Indicates the total number of samples; Represents a node The impurity of the gin; , This represents the Gini impurity of the left and right child nodes after a node splits. Indicates the first Environmental factors At the node The Gini significance contribution value at the location; The environmental factor is obtained by averaging the contribution values ​​of all decision trees. The final Gini importance value : ; in, Indicates the total number of decision trees; Represents the decision tree index; for each decision tree (M trees in total), factors... All nodes involved in the split , cumulatively its Gini importance contribution value ( (For the decision tree index), then take the average of all decision trees to quantify the factor. The overall impact on the environmental fate of tire wear particles.

[0027] S200.3, Core Factor Screening: Calculate the maximum Gini importance of all environmental factors. Set the filtering threshold : ; in, The threshold coefficient (dynamically adjusted based on the total number of factors; the more factors, the higher the threshold coefficient). The value is reduced accordingly to retain a sufficient number of candidate factors. This formula references the conventional strategy of "feature importance threshold screening," and in this embodiment, the coefficient is dynamically adjusted based on the number of environmental factors. It is suitable for tire wear particle monitoring scenarios.

[0028] Select Environmental factors were used as the initial candidate set of core factors; S200.4, Weight Enhancement Adaptation: The linear weighted enhancement mechanism in S200 is invoked to compensate for low-proportion factors whose Gini importance value is lower than 60% of the mean in the candidate set (the compensation formula follows the linear weighting rule, and the enhancement magnitude is dynamically determined according to the number and proportion of low-proportion factors), and finally the feature vector and confidence of the core factor are output.

[0029] In this step, S200.4, the weight enhancement adaptation includes the following steps: S200.4.1 Determination of Low-Proportion Factors: Based on the mean Gini importance value of all factors in the core factor candidate set, factors with a proportion lower than 60% of the mean are determined to be low-proportion, weakly correlated factors, and their proportion is recorded. This rule is based on the environmental data analysis experience that "weak factors may have a small individual impact, but may have a significant synergistic effect with other factors." It is established by calculating the proportion of low-proportion factors. ( =Number of low-proportion factors / Total number of candidates), dynamically adjust the enhancement strategy.

[0030] S200.4.2, Grading Enhancement Coefficient Setting: According to Set the differentiation enhancement coefficient : when (A small number of low-proportion factors, moderately enhanced); when (Medium proportion, enhanced compensation); when (A large number of low-proportion factors are used to control the enhancement magnitude and avoid distortion.) S200.4.3, Dynamic Weight Compensation: For factors with low proportions in the judgment, a linear weighting formula is used to enhance their effectiveness: ; in, This represents the original Gini importance value for low-proportion factors; The filtering threshold set for S200.3; Indicates low proportion factor The Gini importance value after dynamic compensation; this design uses linear interpolation to adjust the Gini importance value of weak factors towards the screening threshold. Proximity can both increase the weight of weak factors and avoid over-enhancement.

[0031] At the same time, an enhanced upper limit constraint is introduced (to prevent low-proportion factors from being excessively enhanced): ; in, The maximum Gini importance value in the candidate set; Indicates the upper limit coefficient of enhancement; Indicates the first The Gini importance values ​​of low-proportion factors after weight enhancement and adaptation; S200.4.4, Enhancement Effect Verification: Calculate the weight bias rate of the enhanced candidate set : ; in, This represents the maximum Gini importance value of all factors in the core factor candidate set after weight enhancement and adaptation. This represents the minimum Gini importance value of all factors in the core factor candidate set after weight enhancement and adaptation. This represents the arithmetic mean of the Gini importance values ​​of all factors in the core factor candidate set after weighted enhancement and adaptation. like If the enhancement is confirmed to be effective, the adjusted core factor set will be output. like Return to S200.4.2 and readjust. (e.g., ±0.1 from the original grading), until... .

[0032] Understandably, through the above steps, S200 ultimately outputs the feature vectors of the core factors (including the enhanced Gini importance value) and confidence scores (calculated based on the out-of-bag error of the random forest model), providing accurate input for the subsequent factor classification and integration of S300. This ensures that key influencing factors (including weakly correlated factors) are effectively incorporated into the tire wear particle environmental fate prediction model, supporting the model's adaptability and prediction accuracy to complex environmental scenarios.

[0033] S300, Factor Classification Integration: Using a training-converged classification model, the selected core factors are classified to distinguish their physical attributes and functional dimensions; and based on the feature encoding network, the core factors are further classified in a more precise manner according to their data form and association patterns. In this step, in S300, factor classification integration includes the following steps: S300.1, Physical Property Classification: By using a training-converged classification model, the physical attribute categories to which the core factors belong are distinguished; As a further explanation of this step, this embodiment uses a Support Vector Machine (SVM) classifier to classify physical attributes. This model is suitable for scenarios with small samples and high-dimensional features. The training data consists of historical environmental factors and manually labeled data. For example, the particle size and concentration of tire wear particles are labeled as "particle feature factors," wind speed and humidity are labeled as "meteorological environmental factors," traffic flow and vehicle type ratio are labeled as "traffic parameter factors," and wastewater treatment process parameters (such as residence time) are labeled as "process control factors." Furthermore, the model parameters are set as follows: the kernel function is the radial basis function. ,in =0.1, adaptability factor feature dimension; penalty coefficient C=10 (balancing classification accuracy and generalization ability); the loss function uses hinge loss. ,in For tags, Model output; training continues until 100 iterations or loss convergence; Furthermore, during classification, the core factor feature vectors output by S200 are input into the model, which outputs the corresponding physical attribute category and confidence level (a probability value ≥ 0.7 is considered a valid classification), thus clarifying the essential characteristics of the factors.

[0034] S300.2, Functional Dimension Determination: By using a feature encoding network, the influence of core factors on the environmental fate of tire wear particles can be determined. As a further explanation of this step, this embodiment uses a shallow neural network to determine the influence dimension. The network is a two-layer fully connected structure: the input layer is the factor time-series feature vector (such as continuous 24-hour monitoring values), and the hidden layer contains 32 nodes (the activation function is...). The output layer corresponds to four action dimensions (the activation function is softmax). Furthermore, the dimension of action is defined based on the particulate environmental behavior mechanism: Migration-driven dimensions: those affecting particle spatial migration (e.g., wind speed); Dimensions affecting sedimentation: Influences on particle sedimentation / suspension (e.g., particle size); Transformation facilitation dimension: Influences on particle physical / chemical transformation (e.g., temperature); Retention control dimension: affecting particle retention in the facility (e.g., filter media pore size); Furthermore, during the discrimination process, after the factor time-series feature vectors are processed by the network, the output layer calculates the probabilities of each dimension using the softmax function. : ;in, For the first The network output value for the class dimension. The dimension corresponding to the maximum probability is taken as the discrimination result.

[0035] S300.3, Secondary Classification Code: Based on feature coding networks, core factors are classified into two levels according to data form and association pattern; As a further explanation of this step, the secondary classification specifically includes data form classification and association pattern classification, wherein: Data format classification: Continuous types (such as wind speed and particle concentration): Standardized treatment is adopted. ( The mean, (Standard deviation) Discrete types (such as vehicle type, weather type): use one-hot encoding (e.g., "sunny day = 001, rainy day = 010"); Time-series data (e.g., hourly traffic flow): retain the time-series array format and input it after standardization; Association pattern classification: Linear correlation (e.g., traffic flow and particulate concentration): determined by Pearson correlation coefficient. ; Nonlinear correlations (such as the effect of temperature on particle transformation): determined by Spearman's rank correlation coefficient; Time-lag correlation (e.g., changes in particle concentration after rainfall): The duration of the time lag is determined through cross-correlation analysis.

[0036] The association pattern is transformed into a 3D embedding vector (e.g., linear association = [1,0,0]) by the network.

[0037] S300.4, Classification Result Integration: The classification results of S300.1-S300.3 are cross-integrated to form a multi-dimensional label system that includes physical attributes, function dimensions, and data-related characteristics, and is used to construct the multi-branch input of the prediction model.

[0038] As a further explanation of this step, this embodiment integrates the classification results of physical attributes, function dimensions, and data-association characteristics into a four-tuple label system of "physical attributes + function dimensions + data form + association pattern", for example: Wind speed: "Meteorological environmental factors + migration-driven dimension + continuous type + linear correlation"; Particle size: "Particle characteristic factors + sedimentation influence dimension + continuous type + nonlinear correlation"; Hourly traffic flow: "Traffic parameter factors + migration-driven dimension + time-series type + time-delay correlation"; This labeling system is directly used to construct multi-branch inputs for the S400 prediction model. For example, "granular feature factor" data is input into one branch, and "transfer-driven dimension" data is input into another branch, enabling the model to learn the mechanism of action of different types of factors in a targeted manner and reduce feature confusion.

[0039] S400, Model Association Construction: Combining the improved multi-scale spatiotemporal attention fusion algorithm, through variable association and parameter calibration, the mapping relationship between core factors and the environmental fate of tire wear particles is established, the basic architecture of the initial model for predicting the environmental fate of tire wear particles is built, and the preliminary fate evolution path is output. In this step, within S400, an improved multi-scale spatiotemporal attention fusion algorithm is used to establish a mapping relationship between core factors and the environmental fate of tire wear particles through variable correlation and parameter calibration. This includes the following steps: S410.1 Multi-scale feature extraction and initial variable association: Using the core factors and multidimensional labeling system formed by S300.4 as input, three sets of differentiated 1D convolutional units are used to capture features hierarchically: Small-scale unit: 1×3 convolution kernel (sliding window stride 1, number of output channels 32), capturing the high-frequency fluctuation features of core factors within 1-3 hours (such as the immediate impact of sudden changes in traffic flow on particle diffusion, supporting 24-hour short-term trend prediction); the kernel size of 3 is chosen because the feature period of short-term fluctuations is usually 2-3 hours, and the stride of 1 ensures that details are not lost, adapting to 24-hour short-term trend prediction.

[0040] Mesoscale unit: 1×5 convolution kernel (sliding window stride 2, number of output channels 32), capturing the cumulative effect of core factors within 1-2 days (such as the superimposed effect of continuous rainfall on particle sedimentation, supporting 72-hour medium-term trend prediction); the kernel size of 5 corresponds to the key time period within 24 hours of each day (such as morning peak and evening peak), and the stride of 2 balances computational efficiency and feature preservation, supporting 72-hour medium-term prediction.

[0041] Large-scale unit: 1×7 convolution kernel (sliding window stride 3, number of output channels 32), capturing the periodic patterns of core factors within 1 week (such as the long-term regulation of particle degradation by weekly average temperature, supporting 15-day long-term trend prediction); the kernel size of 7 matches the weekly cycle, and the stride of 3 compresses redundant information, adapting to 15-day long-term prediction.

[0042] Output multi-scale feature tensor set These correspond to short, medium, and long-term feature dimensions, respectively, and the tensor dimension is "time step × feature channel" (e.g., the short-term feature tensor dimension is 24×32, corresponding to 24 hours and 32 feature channels). S410.2, Calculation of Spatiotemporal Attention Weights: Based on feature tensors and S300 multidimensional labels, spatial and temporal weights are calculated to strengthen key associations: Spatial weight calculation: Based on the physical attribute categories of S300.1, basic weights are assigned to different attribute factors (e.g., weight coefficient of 1.2 for particle physical properties, and 1.0 for environmental conditions), and then the weights are calculated using the factor-regression mutual information matrix. optimization: ; in, Indicates the first Spatial attention weights of each core factor; Indicates the first The initial weights of the core factors; Indicates the first The core factor and the first One environmental trend indicator; This represents the total number of environmental trend indicators; This represents the total number of core factors; As a further explanation of this step, in this embodiment... The calculation formula is: ; in, For joint probability; Marginal probability; The weighting integrates the priority of physical attributes with the actual correlation strength, avoiding biases caused by relying solely on data. Time weight calculation: Set the time decay coefficient for the dimension of action of S300.2. Calculate time steps Weights: ; in, Indicates the first Time attention weights for each time step; This represents the total number of time steps; the formula is based on an exponential decay model (common in temporal weight allocation, such as the temporal forgetting mechanism in RNNs). This embodiment uses... Adaptation to the dimensional characteristics.

[0043] Different dimensions of action result in different rates of decay over time (e.g., migration and diffusion are more affected by recent factors). (Higher value) As a further explanation of this step, in this embodiment... The rules for determining the value are as follows: Migration-driven dimensions (such as wind speed): =0.3 (The impact is greater in the near term and decays faster); Transformation facilitation dimensions (e.g., temperature): =0.1 (long-term cumulative effect, slow decay); Dimensions of Settlement Impact / Retention Control: =0.2 (moderate attenuation); pass The temporal effects of differentiating factors were analyzed. For example, the effect of wind speed on particle migration was most significant within 1 hour, while the effect of temperature on particle degradation required more than 3 days to accumulate. The weighting was matched with the actual mechanism of action.

[0044] S410.3 Feature Fusion and Mapping Relationship Construction: Weigh and fuse multi-scale feature tensors with spatiotemporal weights: ; in, Represents the multi-scale spatiotemporal weighted fusion feature tensor; Represents the spatial attention weight matrix; Represents the temporal attention weight vector; This represents element-wise multiplication; Indicates small scale Mesoscale Large scale The dimensions of features are concatenated (e.g., short-term 24×32, medium-term 72×32, long-term 120×32, the concatenated dimensions are 216×32). The fully connected network with batch normalization layer is used to integrate the feature inputs and output a regression correlation matrix to complete the mapping from core factors to environmental regression. As a further explanation of this step, this embodiment inputs the fused features into a fully connected network, and the specific network configuration is as follows: Network structure: Two fully connected layers are set up, with 128 hidden nodes, and batch normalization layers are used. The activation function is ReLU. Output: Generate a regression correlation matrix (dimension: ). ,in For the number of factors, (The matrix represents the influence strength of the corresponding factor on the regression index, with values ​​ranging from 0 to 1, and higher values ​​indicating stronger correlations).

[0045] S410.4, Parameter Calibration: Adjust the calibration rules based on the secondary classification labels (continuous / discrete, linear / nonlinear correlation) of S300.3: For discrete factors: calibrate the convolution kernel stride (e.g., when the discrete factor accounts for more than 40%, reduce the stride by 0.5) to improve the ability to capture discrete mutations. For nonlinear correlation factors: adjust the number of nodes in the fully connected network to enhance the model's nonlinear fitting ability; Repeat steps S410.1-S410.3 until the average strength of the regression correlation matrix is ​​≥0.65 (the average strength is the mean of all elements in the matrix, and 0.65 is an empirically verified threshold to ensure that the correlation is significant).

[0046] As a further explanation of this step, the calibration rules in this embodiment include: In this step, within S400, an improved multi-scale spatiotemporal attention fusion algorithm is used to build an initial model architecture for predicting the environmental tendency of tire wear particles through phased variable association and dynamic parameter calibration. This includes the following steps: S420.1, Phased Association Modeling of Variables: Linear correlation stage: Partial least squares regression is used to establish the linear correlation between core factors and regression indicators, the top 5 contributing factors are selected, and the linear correlation coefficient matrix is ​​output. ; Nonlinear correlation stage: For the dominant factor and the regression index, a nonlinear correlation model is constructed through radial basis function interpolation, and the correlation strength matrix is ​​output. ; As a further explanation of this step, in order to further improve the mapping accuracy, the association is refined in two stages: linear and nonlinear. In the linear correlation stage, partial least squares regression (PLS, theoretically derived from WoldS et al.'s 2001 study in "PLS-regression: an abscissa of chemometrics") is used. The core factor eigenvectors are input, and the predicted regression index values ​​are output, thereby generating the linear correlation coefficient matrix. (dimension is) The matrix elements represent the linear contribution of factors to the trend, and the top 5 factors in terms of contribution are selected as the dominant factors (such as traffic flow, wind speed, particle concentration, temperature, humidity, etc.). In the nonlinear correlation stage, a nonlinear model is constructed using radial basis function (RBF) interpolation for the dominant factor and the regression index, as shown in the following formula: ,in: This is the predicted value of the regression index. As input factors, For sample points, For interpolation weights, =0.1, adapting to environmental factor scale; generating correlation strength matrix. (dimensions and) Consistent (the element represents the non-linear correlation strength, with values ​​ranging from 0 to 1), thereby enabling accurate modeling of different correlation characteristic factors.

[0047] S420.2, Parameter Dynamic Calibration Optimization: Set calibration threshold ,like mean Then adjust dynamically: Multi-scale expansion: Add large-scale convolutional units with a scale of 1×9; used to capture ultra-long-term features of more than 15 days, enriching the model's ability to extract features at different time scales; Timing optimization: Adjust the time decay coefficient in S410.2 For example, migration-driven dimensions Increase by 0.1, conversion promotion dimension Reduce by 0.05 to adjust the time decay characteristics of different impact dimensions, making them more consistent with the actual impact patterns.

[0048] Repeat steps S410.1-S410.4 until... mean This ensures that the correlation strength of the model reaches the expected level. As a further explanation of this step, the calibration threshold is described in this embodiment. The formula is: ,in for The mean, This indicates a dynamic improvement target based on the current accuracy (ensuring substantial optimization of correlation strength and avoiding meaningless adjustments). This is a minimum threshold; this threshold ensures that the correlation strength is not lower than the baseline level, thus guaranteeing the accuracy of the model.

[0049] S420.3 Initial Model Architecture Setup: The infrastructure comprises three functional modules: Input layer: Receives multi-dimensional labeled core factors output by S300 (embedded physical attributes, function dimension label encoding, with the dimension and number of factors being consistent); Fusion Layer: Integrates multi-scale spatiotemporal weighted fusion feature tensors from S410.3 , and the correlation strength matrix Weighted fusion; Output layer: The softmax activation function is used to output a multi-branch regression path.

[0050] As a further explanation of this step, in this embodiment, to build an initial model for predicting the environmental fate of tire wear particles, S420.3 integrates the outputs of multiple modules through a hierarchical architecture: the input layer receives the multi-dimensional labeled core factors output by S300 (embedded with one-hot encoding of physical attributes and action dimensions, with the dimension being "number of core factors × label dimension"), achieving seamless connection between "classification and association"; the fusion layer simultaneously accesses multi-scale spatiotemporal fusion features. (Covering trend characteristics from 1–3 hours to very long duration) and correlation strength matrix (Quantifying the linear / nonlinear correlation between factors and migration paths), this architecture strengthens highly correlated features (such as the synergistic effect of "traffic flow-wind speed" on particle migration) through element-wise multiplication. The output layer uses a Softmax function to output multi-branch migration paths (such as the probability distribution of "short-term migration to water bodies, medium-term retention in soil, and long-term degradation"), supporting time-series environmental decision-making. This architecture improves training efficiency by 25% through feature reuse, enhances prediction interpretability with multi-branch output (traceable factor contribution weights, such as wind speed contributing 0.62 to short-term migration), and its modular design supports future factor / index expansion, laying the foundation for model iteration.

[0051] S500, Model Training and Storage: The multi-dimensional data collected by S100 and the mapping relationship established by S400 are input into the iterative training environment according to the training protocol to optimize the parameters of the initial model for predicting the environmental tendency of tire wear particles built by S400; and the training logs recorded in real time during the training process, the intermediate models generated at preset intervals during the training process, and the final model generated after the training are completed are stored in the database for subsequent call and analysis. In this step, the model training storage in S500 includes the following steps: S500.1, Training Dataset Construction: Extract the multi-dimensional data from the S100 output and the trend correlation features from the S400 output. Based on the time dimension primary key aligned data, generate a training dataset containing input features and trend labels. Use the index slicing method to split the dataset into training set, validation set, and test set. As a further explanation of this step, this embodiment uses "monitoring point ID + timestamp (format: YYYY-MM-DDHH:MM)" as the joint primary key to align the multi-dimensional data output by S100 (such as tire wear particle concentration, meteorological parameters, and traffic flow) with the directional correlation features output by S400 (such as the physical attribute labels and action dimension labels of factors). The directional labels directly extract the actual monitoring results of S100 (such as water migration, soil retention rate, bioaccumulation concentration, degradation rate, etc.) to ensure the temporal consistency between the input features and the labels.

[0052] Furthermore, the dataset split follows the time-order priority principle (to avoid future data leakage), dividing it into training, validation, and test sets in a 7:2:1 ratio: the aligned data is sorted by timestamp to generate continuous indexes, the training set selects the first 70% of the indexes (e.g., 0–6999), the validation set selects the middle 20% (7000–8999), and the test set selects the last 10% (9000–9999), ensuring that each subset covers a complete time period (e.g., seasons, daily cycles) and maintaining the temporal correlation of the data.

[0053] S500.2, Model Training Execution: Load the initial model file for predicting the environmental tendency of tire wear particles output by S400, call the standard training interface of the training framework, and configure the basic training parameters (including batch size, training epochs, and optimizer type); trigger validation set evaluation at preset intervals during training to continuously optimize model parameters; As a further explanation of this step, this embodiment loads the initial model file output by S400 and calls the standard training interface based on the PyTorch or TensorFlow framework. Specifically: Basic parameter configuration: Batch size is set to 32 (to adapt to regular GPU memory; can be adjusted to 64 when the data volume exceeds 100,000 records), initial training epochs are set to 100 (dynamically adjusted through the validation set); the optimizer is Adam optimizer, the initial learning rate is set to 0.001, and it decays to 1 / 10 of the original value every 20 epochs. Validation and parameter optimization: Validation set evaluation is triggered every 10 rounds, and the mean squared error (MSE) is calculated. If the MSE of the validation set increases for 5 consecutive rounds, an early stopping mechanism is triggered, and the historical best model is loaded; the learning rate is adjusted synchronously, and the gradient update direction is optimized in conjunction with the default momentum parameter of the Adam optimizer.

[0054] S500.3, Model and Data Storage: After each preset training round, the intermediate model parameter file is automatically exported, and the storage path is associated with the unique technical identifier of the initial model for predicting the environmental tendency of tire wear particles built by S400. After training is terminated, the final model file is exported, and the training parameter configuration and dataset splitting rules are synchronously stored in the database. The database can then be searched and associated using the unique technical identifier of the initial model for predicting the environmental tendency of tire wear particles built by S400.

[0055] As a further explanation of this step, in order to achieve traceability and reuse of the model training process in this embodiment, the initial S400 model, training data, and process records are associated with a unique technical identifier. The specific implementation is as follows: A unique identifier is generated using "S400 initial model version + training start timestamp + model parameter hash value" to ensure that each training instance corresponds one-to-one with the initial model architecture, training timing and parameter status built by S400, providing a basis for cross-version model comparison; After each preset number of training rounds (e.g., 20 rounds), intermediate model parameter files are automatically exported, and the storage logic is bound to a unique identifier. This rule allows for the retrospective analysis of the dynamic evolution of model parameters during training (e.g., changes in prediction bias across different rounds), aiding in the assessment of training convergence. After training terminates, the final model file is exported; training parameters (batch size, training epochs, optimizer configuration) and dataset splitting rules are recorded synchronously and stored in the database using unique identifiers. This design enables integrated retrieval of "model file-training parameters-data rules," supporting subsequent model reproduction and parameter tuning. Key metrics during the training process (loss value per round, learning rate changes, hardware resource usage, etc.) are recorded in real time and stored with unique identifiers. Log data provides direct evidence for model performance analysis (such as the reasons for fluctuations in validation set loss) and troubleshooting training failures.

[0056] S600 Model Validation Feedback: Based on the output of the intermediate model generated during the S500 training process, when the prediction error exceeds the preset threshold or the evolution logic is abnormal, a closed-loop optimization mechanism is triggered. The model parameters, training set, or algorithm module of the intermediate model are adjusted according to the validation rules until the environmental regression prediction accuracy requirements are met.

[0057] In this step, in S600, model verification feedback is achieved through anomaly triggering and closed-loop optimization processes, including the following steps: S600.1, Definition of Verification Indicators and Anomaly Monitoring: Key indicator inheritance: Based on the trend mapping relationship output by S400, key indicators for trend prediction are extracted as a verification benchmark. Abnormal triggering conditions: Real-time calculation of the error between the output of the S500 intermediate model and the true value. When the error exceeds the accuracy requirement of the S400 regression index for three consecutive times, or when the evolution logic is found to be inconsistent with the S300 classification features (such as short-term factors driving long-term regression, and S300 being a multi-scale feature classification), the optimization mechanism is triggered. As a further explanation of this step, the abnormal triggering conditions in this embodiment include two types of scenarios: Error exceeding threshold: Real-time calculation of the error between the prediction results of the S500 intermediate model and the actual monitoring values ​​of S100 (such as the deviation between the predicted and measured values ​​of water migration, and the prediction deviation of soil retention rate). When the error exceeds the accuracy requirements of the S400 trend index (such as the reasonable deviation range set based on the trend correlation matrix) for three consecutive times, it is judged as an error anomaly and optimization is triggered. Evolutionary logic contradiction: Based on the classification features of S300 (physical attributes, influence dimension, and time scale label), a verification rule is established to determine whether the regression evolution path output by the model is reasonable. For example, S300 has classified traffic flow as a "short-term high-frequency fluctuation factor" (influence dimension is migration-driven, time scale 1-3 hours). If the model predicts that traffic flow dominates the long-term regression for more than 72 hours (such as "a surge in traffic flow causes particles to continue migrating 15 days later"), then through manual review or procedural rules (comparing the factor time scale label with the regression cycle), it is determined that there is a contradiction between the evolutionary logic and the classification features, triggering optimization.

[0058] S600.2, Closed-loop optimization strategy and its relation to preceding steps: Parameter adjustment: When the error is abnormal, the S500 training parameters are adjusted first, and the iteration is carried out at the preset step size until the error decreases; Dataset update: If parameter adjustment is ineffective, supplement abnormal scene samples (such as the combination of environmental factors corresponding to high errors) from the multi-dimensional data collected by S100, update the training set, and re-trigger the S500 training process. Module replacement: When the evolution logic is abnormal, replace the algorithm module in the corresponding dimension of S400 and retrain based on the S300 classification label; As a further explanation of this step, in the parameter adjustment of this embodiment, the training parameters of S500 are adjusted first, and the iteration is performed according to a preset step size: If the error is caused by the model not converging, increase the number of training rounds (e.g., from 100 rounds to 150 rounds). If the error is caused by gradient oscillations, reduce the learning rate (e.g., decrease it from 0.001 in steps of 0.5). Adjust the batch size (e.g., increase from 32 to 64 to enhance gradient stability).

[0059] After each adjustment, the training process of S500.2 is re-executed until the error falls back to within the threshold.

[0060] Furthermore, in the dataset update of this embodiment, samples matching high-error scenarios (such as time intervals where errors exceed the standard, factor combination features) are extracted from the multi-dimensional data collected in S100 and added to the training set (such as adding 20% ​​more similar samples), while keeping the proportions of the training set, validation set, and test set unchanged. By strengthening the data representation of abnormal scenarios, the model's adaptability to complex environmental conditions is improved, and S500.1-S500.2 are re-executed after the update.

[0061] Furthermore, in the module replacement in this embodiment, the root cause of the contradiction is located and the corresponding algorithm module in S400 is replaced: if the logical contradiction originates from the calculation of time weights (e.g., short-term factors are given too high long-term weights), the time decay coefficient in S410.2 is redesigned. (For example, the migration-driven dimension λ is increased from 0.3 to 0.5, strengthening the weight of recent data). If the problem originates from the feature fusion logic (such as multi-scale feature splicing errors), adjust the fusion formula in S410.3 (such as increasing the weight ratio of mesoscale features). After the module is replaced, the model building process of S410-S420 is re-executed based on the classification labels of S300, and then the training of S500 begins.

[0062] S600.3, Accuracy Convergence Judgment Rules: The output error of the optimized model is continuously monitored. When the cycle error of 10 consecutive validations is less than or equal to the preset threshold of S400, and the evolution logic matches the classification features of S300 (such as short-term factor-driven short-term regression, and S300 is a multi-scale feature classification), the accuracy requirement is met, and the closed-loop optimization is terminated.

[0063] As a further explanation of this step, this embodiment continuously monitors the output of the optimized model and terminates the closed-loop optimization when the following conditions are met: Error stability meets the target: Within 10 consecutive validation periods (each period corresponds to 1 round of training), the prediction error is ≤ the accuracy threshold of the regression index set by S400. Logical consistency verification: The evolution path matches the S300 classification features (e.g., short-term factors only dominate the 0-24 hour trend, while long-term trends are driven by large-scale factors such as weekly average temperature), and consistency is confirmed through manual review combined with S300 classification label rules.

[0064] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[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 preferred examples and are not intended to limit 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 building a tire wear particle environment fate prediction model based on multi-dimensional data, characterized in that, Comprise the following steps: S100, data acquisition and pretreatment: through the synchronous deployment of multi-source monitoring equipment, real-time acquisition of tire wear particle sample data and environmental correlation data; and the collected multi-dimensional data are subjected to time and space registration, noise filtering pretreatment operation; S200, key factor detection: a feature importance analysis model based on random forest is constructed, the key factors of the pretreated collected data are mined, the core factors affecting the environmental fate of tire wear particles are screened by calculating the Gini importance value of each key factor, and the feature vector and confidence of the core factors are output; a linear weighted enhancement mechanism is introduced to compensate the weights of low proportion and weakly associated factors with Gini importance values below the threshold, and the identification efficiency of low proportion and weakly associated factors is optimized; S300, factor classification and integration: using the trained and converged classification model, the types of the screened core factors are distinguished, and the physical properties and action dimensions of the core factors are distinguished; And based on the feature coding network, the data form and correlation mode of the core factors are further accurately classified; S400, model correlation construction: combining the improved multi-scale spatio-temporal attention fusion algorithm, the mapping relationship between the core factors and the environmental fate of tire wear particles is established through variable correlation and parameter calibration, the basic framework of the initial model for predicting the environmental fate of tire wear particles is built, and the preliminary evolution path is output; S500, model training and storage: the multi-dimensional data collected in S100 and the mapping relationship established in S400 are input into the iterative training environment according to the training protocol, the parameters of the initial model for predicting the environmental fate of tire wear particles built in S400 are optimized; and the training logs recorded in real time during the training process, the intermediate models generated at preset intervals during the training process, and the final model generated after the training are stored in the database for subsequent call analysis; S600, model verification and feedback: according to the output results of the intermediate model generated in the training process of S500, when the prediction error exceeds the preset threshold or the evolution logic is abnormal, the closed-loop optimization mechanism is triggered, the model parameters, training set or algorithm module of the intermediate model are adjusted according to the verification rules, until the environmental fate prediction accuracy requirement is met.

2. The method of claim 1, wherein the method further comprises: In the S100, the multi-source monitoring equipment acquires tire wear particle sample data and environmental correlation data in real time according to the preset acquisition frequency; in order to ensure the continuity and timeliness of the data, the multi-source monitoring equipment includes a laser-induced breakdown spectrometer, a nanoparticle tracking analyzer, a micro weather station and an intelligent traffic sensor, and the acquisition frequency is dynamically adjusted according to the traffic flow change and the grade of the environmental sensitive area; at the same time, the Beidou time synchronization technology is adopted to reduce the time asynchronous problem caused by the clock deviation of the equipment, and to ensure the consistency of the time stamp of the multi-dimensional data, and to provide basic conditions for subsequent time and space registration operation.

3. The method of claim 2, wherein the method further comprises: In the S100, the collected multi-dimensional data are subjected to time and space registration, noise filtering pretreatment operation, comprising the following steps: S100.1, Time reference unification: The time stamps of the laser-induced breakdown spectrometer, nanoparticle tracking analyzer, micro weather station, and intelligent traffic sensor are synchronized and calibrated using the Beidou time system to ensure the consistency of the time stamps of the multi-source data and generate an original data set with a unified UTC time label; S100.2, Spatial coordinate alignment: Based on the time synchronization result of S100.1, the spatial collection positions of each device are converted to plane coordinates using the UTM projection coordinate system, and the non-uniformly distributed sampling point data is mapped to a regular grid element by a bilinear interpolation method to fuse and correct the reasonable spatial deviation data caused by the device deployment position; S100.3, Multi-level noise filtering: First, the local outlier factor algorithm is used to identify and remove abnormal values in the particle concentration and element content data that deviate from the overall distribution, and then the sym8 wavelet basis is used to perform multi-level wavelet decomposition on the filtered data, and the high-frequency coefficients are processed by a soft threshold to weaken random noise, generating a spatio-temporally consistent preprocessed data set.

4. The method of claim 3, wherein the method further comprises: In the S200, the feature importance analysis model based on random forest includes the following steps: S200.1, Model parameter initialization: Set the random forest to contain multiple CART decision trees; The maximum depth of a single tree is adapted according to the dimension of the input data; Set the minimum leaf node sample size and the number of random feature subsets to the square root of the total number of input factors; S200.2, Gini importance value calculation: For each decision tree, calculate the Gini importance contribution value of each factor through the node splitting process, the formula is: ; wherein, represents the th environmental factor; represents a decision tree node for splitting using ; represents the number of samples of a node; represents the total number of samples; represents the Gini impurity of a node ; , represents the Gini impurity of the left and right child nodes after splitting of a node; represents the th environmental factor 's Gini importance contribution value at node ; the contribution values of all decision trees are averaged to obtain the final Gini importance value of the environmental factor :​ S200.3, Core factor screening: calculating the maximum gini importance of all environmental factors setting a screening threshold : Selecting environmental factors as initial core factor candidate set; S200.4, Weight enhancement adaptation: Call the linear weighting enhancement mechanism in S200 to compensate for low-occupancy factors with a Gini importance value less than 60% of the average value, and finally output the feature vector and confidence of the core factor.

5. The method of claim 4, wherein the method further comprises: In S200.4, the weight enhancement adaptation includes the following steps: S200.4.1, low proportion factor determination: take the average of the Gini importance values of all factors in the core factor candidate set as the benchmark, and determine the factors below 60% of the average as low proportion, weakly correlated factors, and record the proportion of their number ; S200.4.2, bin enhancement factor setting: according to Setting a differentiated enhancement factor : When ; When ; When ; S200.4.3, Dynamic weight compensation: For the determined low-occupancy factors, use the linear weighting formula for enhancement: ; wherein, represents the original Gini importance value of the low proportion factor; the screening threshold set for S200.3; represents the low proportion factor the Gini importance value after dynamic compensation; At the same time, introduce an upper limit constraint for enhancement: ； wherein, is the maximum Gini importance value in the candidate set; represents an enhanced upper bound coefficient; represents the Gini importance value of the i-th low proportion factor after weight enhancement adaptation; low proportion factor after weight enhancement adaptation; S200.4.4, Enhancement effect verification: Computing a weight bias rate for a post-augmentation candidate set : If , it is confirmed that the enhancement is effective, and the adjusted core factor set is output; if , S200.4.2 is returned to re-adjust , until .

6. The method of claim 5, wherein the method further comprises: In the S300, the factor classification integration includes the following steps: S300.1, Physical property classification: Use the trained and converged classification model to distinguish the physical property category to which the core factor belongs; S300.2, Action dimension discrimination: Determine the action dimension of the core factor on the tire wear particle environmental fate through the feature encoding network; S300.3, Two-level classification coding: Based on the feature encoding network, perform two-level classification on the core factor according to the data form and association mode; S300.4, Classification result integration: Cross-integrate the classification results of S300.1-S300.3 to form a multi-dimensional label system containing physical properties, action dimensions, and data-association characteristics, and use it for multi-branch input construction of the prediction model.

7. The method of claim 6, wherein the method further comprises: determining a tire wear particle environment fate prediction model based on the multi-dimensional data. In the S400, combined with the improved multi-scale spatio-temporal attention fusion algorithm, the mapping relationship between the core factor and the tire wear particle environmental fate is established through variable association and parameter calibration, including the following steps: S410.1, Multi-scale feature extraction and variable initial association: Take the core factors and multi-dimensional label system formed in S300.4 as input, use 3 groups of differentiated 1D convolution units to capture features at different scales: Small-scale unit: 1x3 convolution kernel, capture the high-frequency fluctuation characteristics of core factors within 1-3 hours; Medium-scale unit: 1x5 convolution kernel, capture the cumulative effect of core factors within 1-2 days; Large-scale unit: 1x7 convolution kernel, capture the periodicity of core factors within 1 week; Output multi-scale feature tensor set corresponding to short, medium, and long-term feature dimensions, respectively; S410.2, Spatio-temporal attention weight calculation: Based on the feature tensor and the multi-dimensional label of S300, calculate the spatial weight and temporal weight to strengthen the key association: Spatial weight calculation: Based on the physical attribute categories of S300.1, base weights are assigned to different attribute factors, which are then multiplied by the factor-destination mutual information matrix Optimization: ； wherein, represents the spatial attention weight of the th core factor; represents the initial weight of the th core factor; represents the initial weight of the th core factor and the th environmental destiny indicator; represents the total number of environmental destiny indicators; represents the total number of core factors; Time weight calculation: set a time decay coefficient for the action dimension of S300.2 , calculate the weight of the time step . ； wherein, denotes the time attention weight for the time step; denotes the total number of time steps; S410.3, Feature fusion and mapping relationship construction: Weighted fusion of multi-scale feature tensor and spatio-temporal weight: ； wherein, denotes a multi-scale spatio-temporal weighted fusion feature tensor; denotes a spatial attention weight matrix; denotes a temporal attention weight vector; denotes an element-wise multiplication; denotes a small-scale , a medium-scale , a large-scale dimension concatenation of features; Input the fused features into the fully connected network with batch normalization layer, output the trend association matrix, and complete the mapping from core factors to environmental trend; S410.4, Parameter calibration: Adjust the calibration rules according to the secondary classification label of S300.3: For discrete factors: calibrate the convolution kernel step; For non-linear association factors: adjust the number of nodes in the fully connected network; Repeat S410.1-S410.3 until the average intensity of the trend association matrix is greater than or equal to 0.

65.

8. The method of claim 7, wherein the method further comprises: In the S400, an initial model architecture for predicting the environmental trend of tire wear particles is built by combining the improved multi-scale spatio-temporal attention fusion algorithm, variable stage association, and parameter dynamic calibration, including the following steps: S420.1, Variable stage association modeling: Linear correlation stage: using partial least squares regression, establish the linear correlation between core factors and the target index, filter the TOP5 dominant factors, output the linear correlation coefficient matrix ; Nonlinear correlation stage: for the dominant factor and the target index, a nonlinear correlation model is constructed by radial basis function interpolation, and an association strength matrix is output ; S420.2, Parameter dynamic calibration optimization: Setting calibration threshold , if mean , then dynamically adjust: Multi-scale expansion: add a large-scale convolution unit with a 1x9 scale; Timing optimization: adjust time decay coefficient in S410.2 ; Repeat S410.1-S410.4 until the mean ; S420.3, Initial model architecture construction: The basic architecture includes 3 functional modules: Input layer: receive the multi-dimensional labeled core factors output by S300; Fusion layer: integrated multi-scale spatio-temporal weighted fusion feature tensor of S410.3 with the correlation strength matrix weighted fusion; Output layer: output multiple branch trend evolution paths through the Softmax activation function.

9. The method of claim 8, wherein the method further comprises: In the S500, model training and storage includes the following steps: S500.1, Training data set construction: Extract the multi-dimensional data output by S100 and the trend association features output by S400, align the data based on the time dimension primary key, and generate a training data set containing input features and trend labels; use the index slicing method to split the data set into training set, validation set, and test set; S500.2, Model training execution: Load the tire wear particle environmental trend prediction initial model file output by S400, call the standard training interface of the training framework, and configure the basic training parameters; trigger the validation set evaluation at the preset interval during training, and continuously optimize the model parameters; S500.3, Model and data storage: After completing the preset training rounds, automatically export the intermediate model parameter file, and store the path associated with the unique technical identifier of the tire wear particle environmental trend prediction initial model built by S400; After training is completed, export the final model file, synchronize the training parameter configuration and data set splitting rules to the database, and associate with the unique technical identifier of the tire wear particle environmental trend prediction initial model built by S400 for retrieval.

10. The method of claim 9, wherein the method further comprises: In the S600, the model verification feedback is achieved through the abnormal trigger and closed-loop optimization process, including the following steps: S600.1, verification index definition and abnormal monitoring: Key indicator inheritance: based on the convergence mapping relationship output by S400, extract the convergence prediction key indicators as the verification benchmark; Abnormal trigger condition: real-time calculation of the error value between the intermediate model output result in S500 and the true value, when the error is continuously 3 times higher than the accuracy requirement of S400 convergence index, or the evolution logic is monitored to be contradictory to the S300 classification features, trigger the optimization mechanism; S600.2, closed-loop optimization strategy and presequence association: Parameter adjustment: when the error is abnormal, adjust the S500 training parameters first, iterate by the preset step size until the error decreases; Dataset update: if the parameter adjustment is invalid, supplement the abnormal scene samples from the multi-dimensional data collected in S100 to update the training set, and trigger the S500 training process again; Module replacement: when the evolution logic is abnormal, replace the algorithm module of the corresponding action dimension in S400, and retrain based on the S300 classification label; S600.3, precision convergence judgment rule: Continuously monitor the output error of the optimized model, when the error is ≤ the preset threshold of S400 for 10 consecutive verification periods, and the evolution logic matches the S300 classification features, it is judged that the precision requirement is met, and the closed-loop optimization is terminated.

Images