Evaluation and management method based on solid waste classification recycling

Abstract

The application discloses an evaluation and management method based on solid waste classification recycling, and belongs to the technical field of solid waste evaluation and management. The method comprises the following steps: obtaining a solid waste thermodynamic diagram sequence, a component spectrum sequence and a recycling logistics record; generating a solid waste spatial accumulation feature map set by using a first feature extraction network, and generating a solid waste material component feature matrix by using a second feature extraction network; performing cross-modal attention fusion on the two to generate a solid waste resource enrichment distribution map; inputting the distribution map into a recycling value evaluation network to obtain a recycling comprehensive potential index; and combining a historical recycling efficiency sequence and a logistics cost sequence, and outputting a differentiated recycling scheduling strategy through a management strategy generation network. The scheme realizes deep fusion of spatial accumulation features and material component features, dynamically generates a scheduling strategy based on potential and cost, and improves recycling evaluation accuracy and resource scheduling efficiency.

Description

Technical Field

[0001] This invention relates to the field of solid waste classification, recycling, assessment, and management technology, specifically to assessment and management methods based on solid waste classification, recycling, and utilization. Background Technology

[0002] In the assessment and management of solid waste recycling, it is necessary to obtain regional solid waste distribution data to evaluate recycling potential and formulate scheduling plans. Traditional technologies typically obtain solid waste thermal distribution maps or component spectral data independently. Thermal maps reflect bulk density and its spatiotemporal variations, while spectral data identifies material categories. However, these two types of data lack an effective information correlation mechanism. The assessment process can only determine the bulk density based on thermal distribution or the component type based on spectral data, failing to accurately correlate the category of recyclable materials with their actual spatial enrichment location. This leads to inaccurate identification of high-value material enrichment areas during recycling value assessment, resulting in deviations in recycling resource planning.

[0003] In terms of scheduling and management, existing solutions mostly rely on fixed-cycle route planning or static scheduling rules based on historical average recycling volumes, failing to organically integrate the real-time recycling potential, historical recycling efficiency, and transportation route costs of each site. When different sites experience differentiated recycling needs or changes in transportation conditions, this scheduling method lacks dynamic adjustment capabilities, easily leading to waste accumulation at some sites and idle capacity on some routes, resulting in low overall recycling system operating efficiency and difficulty in controlling logistics costs.

[0004] The aforementioned deficiencies mean that the solid waste recycling process must address two key issues: how to achieve deep interaction between spatial stacking characteristics and material composition characteristics to accurately generate resource enrichment distribution, and how to jointly model recycling potential, recycling efficiency, and logistics costs to generate differentiated scheduling strategies. Summary of the Invention

[0005] This paper provides an assessment and management method based on solid waste classification and recycling, which solves the problems in the existing technology where the spatial accumulation information and material composition information of solid waste are isolated from each other, resulting in insufficient accuracy of recycling value assessment, and the failure of recycling scheduling strategies to integrate potential, efficiency and cost to dynamically generate differentiated solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an assessment and management method based on solid waste classification and recycling, the method comprising: acquiring a solid waste heat map sequence, a solid waste component spectral sequence, and a recycling logistics record from multiple monitoring stations in an area to be assessed within a preset time window; inputting the solid waste heat map sequence into a first feature extraction network to generate a solid waste spatial accumulation feature map set of the area to be assessed; inputting the solid waste component spectral sequence into a second feature extraction network to generate a solid waste material component feature matrix of the area to be assessed; performing cross-modal attention fusion processing on the solid waste spatial accumulation feature map set and the solid waste material component feature matrix to generate a solid waste resource enrichment distribution map of the area to be assessed; inputting the solid waste resource enrichment distribution map into a recycling value assessment network to output a comprehensive recycling potential index of the area to be assessed within the preset time window; acquiring a historical recycling efficiency sequence and a logistics cost sequence of the area to be assessed within the preset time window; jointly inputting the comprehensive recycling potential index, the historical recycling efficiency sequence, and the logistics cost sequence into a management strategy generation network to output a differentiated recycling scheduling strategy for the area to be assessed. This technical solution integrates the spatial distribution thermal characteristics of solid waste with the spectral characteristics of material components across modes, which can accurately characterize the enrichment and distribution state of solid waste resources. Based on this, it can conduct recycling value assessment and generate scheduling strategies, realizing end-to-end intelligent management from solid waste status perception to recycling strategy formulation.

[0007] As a technical solution of the present invention, the step of inputting the solid waste heat map sequence into a first feature extraction network to generate a solid waste spatial accumulation feature map set for the region to be evaluated includes: performing spatial resolution normalization processing on each frame of the solid waste heat map sequence to unify the pixel size of all solid waste heat maps to a preset pixel size; inputting each frame of the solid waste heat map after spatial resolution normalization processing into the spatiotemporal convolutional layer of the first feature extraction network in chronological order, and extracting the local spatial accumulation intensity features and the accumulation diffusion trend features between adjacent frames of the solid waste heat map through the spatiotemporal convolutional layer; inputting the local spatial accumulation intensity features and the accumulation diffusion trend features into the spatial attention layer of the first feature extraction network, and performing weighted enhancement processing on the accumulation intensity at different spatial locations in the solid waste heat map through the spatial attention layer to generate an enhanced spatial accumulation feature map; inputting the enhanced spatial accumulation feature map into the feature pyramid structure of the first feature extraction network, and performing multi-level fusion of spatial accumulation features at different scales through the feature pyramid structure to output the solid waste spatial accumulation feature map set. This scheme effectively enhances the perception of hotspots in solid waste accumulation and their evolution trends through spatiotemporal convolution and spatial attention mechanisms. After multi-scale fusion of feature pyramids, the generated feature map set can simultaneously retain fine spatial details and global distribution semantics.

[0008] As a technical solution of the present invention, the step of inputting the solid waste component spectral sequence into a second feature extraction network to generate a solid waste material component feature matrix of the region to be evaluated includes: performing baseline correction and noise filtering on each spectral data in the solid waste component spectral sequence to generate a corrected spectral data sequence; dividing the corrected spectral data sequence into band subsets according to the known characteristic bands of the solid waste material category to generate sub-spectral sequences corresponding to multiple band subsets; inputting each of the sub-spectral sequences in parallel into a multi-branch convolutional layer of the second feature extraction network, and extracting the material type response features corresponding to each band subset through each branch convolutional layer; inputting all material type response features output by each branch convolutional layer into a feature splicing layer of the second feature extraction network, and splicing all material type response features along the feature dimension through the feature splicing layer to output the solid waste material component feature matrix. Preferably, the size of each branch convolutional kernel of the multi-branch convolutional layer is inversely proportional to the characteristic spectral bandwidth of the solid waste material category corresponding to each band subset. This scheme divides the material characteristic bands into sub-spectral sequences and uses a parallel branch network with adaptive convolution kernel size, which can specifically capture the specific spectral responses of different solid waste materials, effectively improving the accuracy and robustness of material component identification.

[0009] As a technical solution of the present invention, the step of performing cross-modal attention fusion processing on the solid waste spatial accumulation feature map and the solid waste material component feature matrix to generate a solid waste resource enrichment distribution map of the area to be evaluated includes: expanding the solid waste spatial accumulation feature map into a spatial feature vector sequence along the channel dimension, and expanding the solid waste material component feature matrix into a component feature vector sequence along the row direction; performing a dot product operation on each spatial feature vector in the spatial feature vector sequence and each component feature vector in the component feature vector sequence to generate a cross-modal correlation weight matrix; performing normalization processing on the cross-modal correlation weight matrix along the direction of the component feature vector sequence to generate an attention weight distribution; and performing weighted reconstruction of the spatial accumulation features at each position in the solid waste spatial accumulation feature map according to the attention weight distribution to generate the solid waste resource enrichment distribution map that integrates spatial accumulation information and material component information. Through the cross-modal attention mechanism, the thermal accumulation information in the spatial domain and the material composition information in the spectral domain can be deeply aligned and coupled, so that the generated enrichment distribution map can accurately locate the solid waste accumulation location and characterize its material composition, providing rich and accurate input information for subsequent value assessment.

[0010] As a technical solution of the present invention, the step of inputting the solid waste resource enrichment distribution map into the recycling value assessment network and outputting the comprehensive recycling potential index of the area to be assessed within the preset time window includes: inputting the solid waste resource enrichment distribution map into the first convolutional block of the recycling value assessment network, and extracting the spatial location and enrichment density of the recyclable component enrichment area in the solid waste resource enrichment distribution map through the first convolutional block; inputting the spatial location and enrichment density of the recyclable component enrichment area into the second convolutional block of the recycling value assessment network, and performing multi-scale feature extraction of the recyclable component enrichment area with different receptive fields through the second convolutional block to generate multi-scale enrichment features; inputting the multi-scale enrichment features into the fully connected regression layer of the recycling value assessment network, and mapping the multi-scale enrichment features into the potential score of the area to be assessed in each recycling category through the fully connected regression layer; performing a weighted summation operation on the potential score of the area to be assessed in each recycling category, wherein the weight coefficient of each recycling category is associated with the historical recycling price sequence of the area to be assessed, and using the weighted summation result as the comprehensive recycling potential index. Preferably, the weighting coefficient of each recycling category is the normalized value of the average market price of each recycling category in the region to be evaluated within the preset time window. This scheme, through a weighting mechanism that combines multi-scale feature extraction with market price correlation, ensures that the assessed comprehensive potential index not only reflects the physical enrichment of solid waste but also dynamically reflects its market value orientation, making the assessment results more relevant for practical decision-making.

[0011] As a technical solution of the present invention, the steps of obtaining the historical recycling efficiency sequence and logistics cost sequence of the area to be evaluated within a preset time window, and jointly inputting the comprehensive recycling potential index, the historical recycling efficiency sequence, and the logistics cost sequence into a management strategy generation network to output a differentiated recycling scheduling strategy for the area to be evaluated include: spatially back-projecting the comprehensive recycling potential index according to each monitoring station within the preset time window to generate a station potential distribution vector for each monitoring station; splitting the historical recycling efficiency sequence according to each monitoring station and each type of recycled material to generate a historical efficiency matrix for each monitoring station in each recycling category; splitting the logistics cost sequence according to the transportation paths between each monitoring station to generate a path cost matrix between each monitoring station; concatenating the station potential distribution vector, the historical efficiency matrix, and the path cost matrix to generate a joint input feature vector; inputting the joint input feature vector into the graph attention layer of the management strategy generation network; iteratively updating the scheduling priority between each monitoring station through the graph attention layer; and outputting the differentiated recycling scheduling strategy. Preferably, the number of iterations of the graph attention layer is positively correlated with the sparsity of the path cost matrix between each monitoring station. Further, the step of iteratively updating the scheduling priority among monitoring stations through the graph attention layer includes: treating each monitoring station as a graph node of the graph attention layer, and using the negative correlation value of the path cost matrix between each monitoring station as the initial weight of the edges between graph nodes; using the site potential distribution vector as the initial node feature of each graph node, and using the historical efficiency matrix as the auxiliary node feature of each graph node; for each graph node, calculating the attention coefficient between the graph node and all adjacent graph nodes, wherein the attention coefficient is determined based on the similarity between the initial node feature and auxiliary node feature of the graph node and the initial node feature and auxiliary node feature of adjacent graph nodes; performing weighted aggregation on the initial node feature of each adjacent graph node based on the attention coefficient to generate the updated node feature of each graph node; and determining the scheduling priority ranking of each monitoring station based on the difference between the updated node feature of each graph node and the initial node feature of each graph node. This scheme utilizes graph attention networks to model the relationship between recycling potential, historical efficiency, and transportation costs among multiple sites as a graph structure. Through the propagation and iterative updating of attention weights between nodes, it can adaptively determine the differentiated scheduling priorities of each site from a global perspective, thereby achieving the optimal balance between transportation costs and recycling efficiency.

[0012] As a technical solution of the present invention, the method further includes the step of inputting the differentiated recycling scheduling strategy into a scheduling simulation environment for feasibility verification. The step of inputting the differentiated recycling scheduling strategy into the scheduling simulation environment for feasibility verification includes: constructing a digital twin map of the scheduling simulation environment based on the geographic information data of the area to be evaluated; the digital twin map includes the location of each monitoring station, the distance of each transportation route, and road condition parameters; converting the differentiated recycling scheduling strategy into a simulation scheduling instruction sequence in the scheduling simulation environment; the simulation scheduling instruction sequence includes the material collection time window of each monitoring station and the material transfer volume of each transportation route; executing the simulation scheduling instruction sequence in the scheduling simulation environment, recording the material backlog of each monitoring station and the cumulative transportation time of each transportation route during the execution process; determining whether the material backlog of each monitoring station exceeds a preset backlog threshold and whether the cumulative transportation time of each transportation route exceeds a preset time threshold; if neither exceeds, the differentiated recycling scheduling strategy is determined to have passed the feasibility verification; if either exceeds, the corresponding scheduling parameter in the differentiated recycling scheduling strategy is corrected. By performing pre-verification of scheduling strategies in a digital twin simulation environment, it is possible to effectively identify the risks of material backlog or transportation delays that may be caused by the strategies under actual road and logistics conditions, and to make forward-looking corrections to the parameters to ensure the practical feasibility of the strategies.

[0013] As a technical solution of the present invention, the method further includes a step of performing time-series tracking correction on the comprehensive recycling potential index. This step includes: obtaining a sequence of tracking solid waste heat maps and a sequence of tracking solid waste component spectral sequences for multiple consecutive tracking time windows following a preset time window for the area to be evaluated; sequentially performing feature extraction and cross-modal fusion on the tracking solid waste heat map sequence and the tracking solid waste component spectral sequences according to the processing flow of the preset time window to generate a tracking solid waste resource enrichment distribution map for each tracking time window; sequentially inputting the tracking solid waste resource enrichment distribution maps for each tracking time window into the recycling value assessment network to generate a tracking comprehensive recycling potential index for each tracking time window; calculating the trend change between the comprehensive recycling potential index of the preset time window and the tracking comprehensive recycling potential index of each tracking time window; and performing a moving average correction on the comprehensive recycling potential index of the preset time window based on the trend change to generate a corrected comprehensive recycling potential index. This scheme continuously tracks data from subsequent time windows and uses a moving average to correct the initial assessment potential index. This approach can promptly mitigate assessment biases caused by short-term sudden fluctuations, making the recovery potential assessment value more closely reflect the actual evolution trend of regional solid waste generation.

[0014] Furthermore, the method also includes a step of triggering incremental learning updates of categories based on the modified comprehensive recycling potential index. This step includes: comparing the modified comprehensive recycling potential index with a preset potential threshold; when the modified comprehensive recycling potential index exceeds the preset potential threshold, marking the current time window as a category incremental learning trigger window; within the category incremental learning trigger window, acquiring component spectral samples and spatial accumulation samples of the newly added solid waste category; inputting the component spectral samples of the newly added solid waste category into the second feature extraction network, generating a material component feature vector for the newly added category through forward propagation of the second feature extraction network; embedding the material component feature vector of the newly added category as a prototype of the newly added category into the category prototype library of the fully connected regression layer of the recycling value assessment network, freezing the existing parameters of the first feature extraction network and the second feature extraction network, and performing incremental updates only on the classification weights of the fully connected regression layer of the recycling value assessment network to generate an updated recycling value assessment network. This scheme monitors changes in the potential index and automatically triggers an incremental learning mechanism when new solid waste categories emerge. While retaining existing knowledge, it quickly absorbs the characteristic prototypes of new categories, enabling the assessment model to continuously adapt to the dynamic evolution of solid waste types and avoiding the decline in assessment accuracy due to the emergence of new categories.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0016] Cross-modal attention fusion processing is performed on the spatial accumulation feature atlas of solid waste and the feature matrix of solid waste material components. A cross-modal correlation weight matrix is ​​generated by calculating the dot product between the spatial feature vector sequence and the component feature vector sequence, and then normalized along the component feature direction to obtain the attention weight distribution. This attention weight is used to reconstruct the spatial accumulation features at each location in the spatial accumulation feature atlas, generating a solid waste resource enrichment distribution map that integrates spatial accumulation information and material component information. This processing allows for precise spatial location of the enrichment density of each material component. The distribution map simultaneously carries spatial location attributes and material category attributes, overcoming the information blind spots in value assessment caused by single heat maps or single spectral features. Subsequent recycling value assessment networks, based on this distribution map, can focus on the actual enrichment areas of recyclable components, extracting more discriminative multi-scale enrichment features. Therefore, the output comprehensive recycling potential index more accurately reflects the actual regional distribution of recyclable resources, reducing the distortion of the potential index caused by modal separation.

[0017] The comprehensive recycling potential index, historical recycling efficiency sequences, and logistics cost sequences are jointly input into the management strategy generation network to construct site potential distribution vectors for each monitoring station, historical efficiency matrices for each recycling category, and path cost matrices between stations. These vectors are then concatenated to form joint input features. A graph attention layer abstracts monitoring stations as graph nodes, using path cost correlation values ​​as initial edge weights. Attention coefficients between nodes are calculated using node potential and historical efficiency features, and adjacent node features are weighted and aggregated. During iteration, the scheduling priority of each station is dynamically updated, outputting a differentiated recycling scheduling strategy. This strategy integrates multi-dimensional information on potential distribution, recycling capacity, and transportation costs, enabling the scheduling scheme to dynamically adapt to the immediate recycling needs and cost differences of different stations. This guides transportation capacity towards high-potential, low-cost paths, alleviates station backlog, reduces ineffective transportation, and improves the overall spatiotemporal matching efficiency and logistical economy of the recycling system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 It is a flowchart of the assessment and management methods based on solid waste classification, recycling and utilization;

[0020] Figure 2 This is a flowchart of the process for generating a solid waste spatial accumulation feature atlas;

[0021] Figure 3 This is a flowchart of cross-modal attention fusion processing;

[0022] Figure 4 It is the curve of the change of training loss and validation loss during the training process of the first feature extraction network;

[0023] Figure 5 This is a schematic diagram of cross-modal attention weight distribution;

[0024] Figure 6 It is the loss curve for network training to assess recovery value. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] See Figure 1 This invention provides an assessment and management method based on solid waste classification and recycling, comprising: acquiring solid waste heat map sequences, solid waste component spectral sequences, and recycling logistics records from multiple monitoring stations in the area to be assessed within a preset time window; inputting the solid waste heat map sequences into a first feature extraction network to generate a solid waste spatial accumulation feature map set for the area to be assessed; inputting the solid waste component spectral sequences into a second feature extraction network to generate a solid waste material component feature matrix for the area to be assessed; performing cross-modal attention fusion processing on the solid waste spatial accumulation feature map set and the solid waste material component feature matrix to generate a solid waste resource enrichment distribution map for the area to be assessed; inputting the solid waste resource enrichment distribution map into a recycling value assessment network to output a comprehensive recycling potential index for the area to be assessed within the preset time window; acquiring historical recycling efficiency sequences and logistics cost sequences for the area to be assessed within the preset time window, and jointly inputting the comprehensive recycling potential index, historical recycling efficiency sequences, and logistics cost sequences into a management strategy generation network to output a differentiated recycling scheduling strategy for the area to be assessed.

[0027] Example 1:

[0028] In specific implementation, please refer to Figure 2 The specific steps for generating a spatial accumulation feature map of solid waste in the area to be evaluated by inputting the solid waste heat map sequence into the first feature extraction network based on the assessment and management method of solid waste classification and recycling are as follows.

[0029] Spatial resolution normalization is performed on each frame of the solid waste heatmap sequence to unify the pixel size of all solid waste heatmaps to a preset pixel size. The preset pixel size is determined based on the fixed input size required by the input layer of the first feature extraction network and the sampling resolution distribution range of the original solid waste heatmap acquisition device, and is set to 512×512 pixels. The normalization process uses a bilinear interpolation algorithm to interpolate and map the original pixels of each frame of solid waste heatmap to a 512×512 grid, generating solid waste heatmaps with consistent spatial resolution.

[0030] After spatial resolution normalization, each frame of the solid waste heatmap is arranged chronologically into a temporal sequence and input into the spatiotemporal convolutional layer of the first feature extraction network. The spatiotemporal convolutional layer of the first feature extraction network consists of multiple stacked 3D convolutional kernels, each with a spatial dimension of 3×3 and a temporal dimension of 3. The spatiotemporal convolutional layer includes a first 3D convolutional layer and a second 3D convolutional layer; the first 3D convolutional layer has 64 3D convolutional kernels, and the second 3D convolutional layer has 128 3D convolutional kernels. Each convolutional layer is followed by a batch normalization layer and a linear rectified activation function. The solid waste heatmap sequence is input into the spatiotemporal convolutional layer in the form of dimensions T×H×W×C, where T represents the number of frames in the sequence, H and W represent the normalized height (512) and width (512), and C represents the number of channels in the solid waste heatmap. The spatiotemporal convolutional layer performs a three-dimensional convolution operation on the solid waste heat map sequence, extracts the local spatial accumulation intensity features of each frame of the solid waste heat map, and extracts the accumulation diffusion trend features between adjacent frames, outputting a spatiotemporal feature tensor. The spatiotemporal feature tensor contains the spatial location information and inter-frame diffusion change information corresponding to each frame.

[0031] The local spatial stacking intensity features and stacking diffusion trend features output from the spatiotemporal convolutional layer are input into the spatial attention layer of the first feature extraction network. The spatial attention layer consists of an attention weight generation branch and a feature weighting branch. The attention weight generation branch contains a 1×1 convolutional layer, a batch normalization layer, and a sigmoid activation function. The 1×1 convolutional layer compresses the number of input feature channels to 1, generating the original attention score for each spatial location. The sigmoid function then maps the original attention score to the [0,1] interval, generating the spatial attention weight distribution. The feature weighting branch multiplies the spatial attention weight distribution element-wise with the input local spatial stacking intensity features and stacking diffusion trend features along the spatial location, performing weighted enhancement processing on the stacking intensity at different spatial locations in the solid waste heat map. Regions with high stacking intensity have weights close to 1 and are maintained, while regions with low stacking intensity have weights close to 0 and are suppressed, thus generating an enhanced spatial stacking feature map. In specific implementation, the spatial attention layer performs weighted enhancement on the features corresponding to each frame output by the spatiotemporal convolutional layer, obtaining the enhanced spatial stacking feature map for each frame.

[0032] The enhanced spatially stacked feature map is input into the feature pyramid structure of the first feature extraction network. The feature pyramid structure includes a bottom-up forward propagation path, a top-down upsampling path, and lateral connections. The bottom-up path consists of a series of convolutional layers with a stride of 2, which progressively downsample the enhanced spatially stacked feature map to form feature maps at four scales, denoted as the second-scale, third-scale, fourth-scale, and fifth-scale feature maps, respectively. The spatial resolution of each scale feature map is half that of the previous scale. The top-down path starts with the fifth-scale feature map, increases the feature map resolution through a 2x upsampling operation, and merges it with the feature maps of the same spatial resolution in the forward propagation path through lateral connections. The lateral connections use 1×1 convolutions to unify the number of channels in the feature maps of the forward propagation path, and then add the upsampling result element-wise with the feature map processed by the 1×1 convolution. The feature pyramid structure sequentially upsamples the fifth-scale feature map and fuses it with the fourth-scale feature map to obtain a fused fourth feature map. The fused fourth feature map is then upsampled and fused with the third-scale feature map to obtain a fused third feature map. Finally, the fused third feature map is upsampled and fused with the second-scale feature map to obtain the final multi-level fused feature map. The final output set is the solid waste spatial accumulation feature map set. This set contains multiple spatial accumulation feature maps fused at multiple scales, preserving both local details and global semantics.

[0033] The first feature extraction network was trained using supervised learning. Training data included historical solid waste heatmap sequences and corresponding manually labeled solid waste spatial accumulation distribution maps as labels. The mean squared error loss function was used, and the Adam optimizer was employed. The initial learning rate was set to 0.001, and the training batch size was set to 8. During training, forward propagation yielded the predicted solid waste spatial accumulation feature map set, and the mean squared error loss between the predicted map and the labels was calculated. Backpropagation was then used to update the parameters of each layer of the first feature extraction network until the loss converged.

[0034] See Figure 4 The figure shows the curves of training loss and validation loss as a function of the number of training epochs during the training of the first feature extraction network. The horizontal axis represents the number of training epochs, ranging from 0 to 200, and the vertical axis represents the mean squared error loss value, ranging from 0 to 0.55. The solid curve in the figure represents the training loss, and the dashed curve represents the validation loss.

[0035] As shown in the figure, both training and validation losses exhibit a rapid decreasing trend in the first 10 training epochs, with the training loss decreasing from approximately 0.35 to approximately 0.15 and the validation loss from approximately 0.45 to approximately 0.25, indicating that the network quickly learns effective features in the initial stage. Between 10 and 50 training epochs, the rate of loss decreases, with the training and validation losses decreasing to approximately 0.05 and 0.1, respectively, and the loss gap gradually narrows, indicating that the model gradually converges without significant overfitting. Between 50 and 200 training epochs, both training and validation losses tend to stabilize, remaining at low levels. The training loss eventually drops to near 0, while the validation loss remains around 0.015, demonstrating that the network possesses good generalization ability.

[0036] Overall, the training loss curve is smooth and the validation loss decreases synchronously with the training loss, which meets the design requirements of the first feature extraction network using the mean squared error loss function and Adam optimizer for supervised training. This verifies the effectiveness and stability of the network in the task of extracting spatial accumulation features from solid waste heat map sequences.

[0037] Example 2:

[0038] In specific implementation, the steps for inputting the spectral sequence of solid waste components into the second feature extraction network to generate the solid waste material component feature matrix of the area to be evaluated, based on the assessment and management method of solid waste classification and recycling, are as follows.

[0039] Baseline correction was performed on each spectral data point in the solid waste component spectral sequence. The baseline correction employed an asymmetric least squares smoothing algorithm, which estimates the baseline of the spectral data by minimizing an objective function with a penalty term. The objective function consists of a residual sum of squares term and a baseline second-order difference smoothing term, with the smoothing parameter λ set to 1 × 10⁻⁶. 5 The asymmetric weighting factor p is set to 0.001. For each spectral data point, an asymmetric least squares smoothing algorithm is used iteratively to fit the baseline. In each iteration, the weighting factor is adjusted according to the direction of the residual: a smaller weight is given to points with positive residuals, and a normal weight is given to points with negative residuals, so that the fitted baseline gradually approaches the bottom envelope of the spectral data. The number of iterations is set to 10, and the estimated baseline curve for each spectral data point is output. Subtracting the estimated baseline curve from the original spectral data yields the baseline-corrected spectral data.

[0040] Noise filtering is performed on the baseline-corrected spectral data to generate corrected spectral data. Savitzky-Golay filtering is used, with a window width of 15 data points and a polynomial order of 3. Savitzky-Golay convolution smoothing is applied to each baseline-corrected spectral data point. A 3rd-order polynomial is then fitted using least-squares with the data points within the 15-point window. The original values ​​are then replaced with the value of the fitted polynomial at the center point of the window, thus suppressing high-frequency random noise and preserving spectral peak characteristics, resulting in corrected spectral data. All corrected spectral data constitute a corrected spectral data sequence.

[0041] The corrected spectral data sequence is divided into band subsets according to the known characteristic bands of solid waste material categories, generating multiple sub-spectral sequences corresponding to these band subsets. The known characteristic bands of solid waste material categories are provided by a material composition spectral database. Typical solid waste material categories include polyethylene, polypropylene, polyethylene terephthalate, aluminum, iron, and paper fiber, each with specific characteristic absorption or reflection bands. For example, the characteristic bands for polyethylene are in the 1650nm-1750nm range, for polypropylene in the 1150nm-1250nm range, for aluminum in the 700nm-900nm range, and for paper fiber in the 1900nm-2100nm range. For each corrected spectral data point in the corrected spectral data sequence, spectral data points within the characteristic band range corresponding to each solid waste material category are extracted to form a band subset. The spectral data point fragments corresponding to all band subsets constitute a sub-spectral sequence. After division, multiple sub-spectral sequences are obtained, equal in number to the number of solid waste material categories.

[0042] The sub-spectral sequences are input in parallel to the multi-branch convolutional layer of the second feature extraction network. The multi-branch convolutional layer of the second feature extraction network consists of multiple independent one-dimensional convolutional branches, the number of which equals the number of solid waste material categories. Each one-dimensional convolutional branch specifically processes a sub-spectral sequence of one band. Each one-dimensional convolutional branch contains a one-dimensional convolutional layer, a batch normalization layer, and a modified linear unit activation function. The kernel size of the one-dimensional convolutional layer is determined by the characteristic spectral bandwidth of the corresponding band subset, and the kernel size of each branch of the multi-branch convolutional layer is inversely proportional to the characteristic spectral bandwidth of the solid waste material category corresponding to each band subset. Let the solid waste material category index be p, and the characteristic spectral bandwidth of the p-th band subset be... Characteristic spectral bandwidth Defined as the wavelength range width of the p-th band subset, in nanometers; the size of the one-dimensional convolutional kernel of the p-th one-dimensional convolutional branch is denoted as... kernel size Determine according to the following formula:

[0043]

[0044] in, This is a size adjustment constant, in nanometers, with a value set to 2000 nanometers; This represents the floor operation. When the characteristic spectral bandwidth... When larger, the kernel size Relatively small, used to capture overall profile features over a wider band; when the characteristic spectral bandwidth When smaller, the kernel size The kernel size is relatively large, used to finely resolve local spectral details within narrow bands. The number of kernels in each one-dimensional convolutional layer is set to 64, the stride is set to 1, and the padding method is set to the same padding.

[0045] Each sub-spectral sequence is input in parallel to a one-dimensional convolutional layer of the corresponding branch. The convolutional kernel of the one-dimensional convolutional layer slides along the spectral wavelength dimension to extract the material type response features corresponding to each band subset. A batch normalization layer follows the one-dimensional convolutional layer to normalize the features, and then a modified linear unit activation function is used to introduce nonlinearity, resulting in the primary material type response feature map output by each one-dimensional convolutional branch. The length of the primary material type response feature map is the same as the length of the input sub-spectral sequence, and the number of channels is 64. Global max pooling is performed on the primary material type response feature map output by each one-dimensional convolutional branch, taking the maximum value along the spectral wavelength dimension. The maximum value of each channel is used as the aggregated feature of that channel, thus compressing all channel features corresponding to a band subset into a 64-dimensional material type response feature vector.

[0046] The material type response feature vectors output from each branch of the convolutional layer are input into the feature concatenation layer of the second feature extraction network. The feature concatenation layer receives the material type response feature vectors from all one-dimensional convolutional branches and concatenates them end-to-end along the feature dimension to form a high-dimensional concatenated feature vector. The dimension of the concatenated feature vector is 64 times the number of solid waste material categories. The above process from band division to feature concatenation is performed on each corrected spectral data point in the corrected spectral data sequence, generating a concatenated feature vector for each corrected spectral data point. The concatenated feature vectors corresponding to all corrected spectral data points in the corrected spectral data sequence are arranged row-wise, outputting a solid waste material component feature matrix. The number of rows in the solid waste material component feature matrix equals the number of spectral data points in the corrected spectral data sequence, and the number of columns equals the dimension of the concatenated feature vector.

[0047] The training process of the second feature extraction network is as follows: A training dataset is constructed, consisting of a large number of spectral samples containing known solid waste material category labels. These spectral samples undergo the same baseline correction, noise filtering, and band subset partitioning as described above, and are then input into the second feature extraction network for forward propagation. After the feature concatenation layer, the second feature extraction network connects to a fully connected classification layer. The number of output nodes in the fully connected classification layer equals the number of solid waste material categories to be classified, and the classification probability is output using the Softmax activation function. The loss function is the classification cross-entropy loss function. The optimization algorithm uses the Adam optimizer, with an initial learning rate of 0.0005, a training batch size of 32, and 100 training epochs. In each training iteration, the cross-entropy loss between the predicted classification probability and the true category label is calculated. The parameters of the multi-branch convolutional layer, feature concatenation layer, and fully connected classification layer are updated through error backpropagation until the loss function converges. After training, the fully connected classification layer is removed, and the multi-branch convolutional layer and feature concatenation layer are retained as the fixed feature extraction network for extracting the feature matrix of solid waste material components.

[0048] Example 3:

[0049] In specific implementation, please refer to Figure 3 The specific steps for generating a solid waste resource enrichment distribution map of the area to be evaluated by performing cross-modal attention fusion processing on the solid waste spatial accumulation feature atlas and the solid waste material component feature matrix are as follows.

[0050] The solid waste spatial accumulation feature map set is unfolded along the channel dimension into a sequence of spatial feature vectors. The solid waste spatial accumulation feature map set consists of feature maps at K spatial scales output by the feature pyramid structure. The total number of feature maps is denoted as S, and the spatial dimension of each feature map is uniformly represented by its height. and width The number of channels in each feature map is uniformly represented as For each feature map, it is unfolded into a one-dimensional sequence along the height and width dimensions, that is, at each spatial location (u,v), a sequence of length is extracted. The channel feature vectors are used to transform a feature map into L spatial feature vectors, where L equals Multiply The spatial feature vectors of all S feature maps are concatenated end-to-end along the sequence direction to obtain a spatial feature vector sequence. The total length of the spatial feature vector sequence is... Equals S multiplied by L, where the dimension of each spatial feature vector in the sequence is... The spatial eigenvector sequence is denoted as... The index m ranges from 1 to 1. , Let m represent the m-th spatial feature vector.

[0051] The solid waste material component feature matrix is ​​expanded into a sequence of component feature vectors along the row direction. The solid waste material component feature matrix has a shape of M rows multiplied by D columns, where M represents the number of spectral data points in the corrected spectral data sequence, and D represents the dimension of the concatenated feature vectors. Each row of the solid waste material component feature matrix is ​​considered as a component feature vector, and the dimension of the component feature vector is D. All rows of the solid waste material component feature matrix are sequentially extracted along the row direction and arranged into a sequence of component feature vectors. The total length of the component feature vector sequence is... The sequence is M, and the dimension of each component feature vector in the sequence is D. The sequence of component feature vectors is denoted as... The index n ranges from 1 to 1. , This represents the eigenvector of the nth component.

[0052] The cross-modal correlation weight matrix is ​​generated by performing a dot product operation between each spatial feature vector in the spatial feature vector sequence and each component feature vector in the component feature vector sequence. The cross-modal correlation weight matrix is ​​constructed as follows: for the m-th spatial feature vector in the spatial feature vector sequence... The nth component eigenvector in the component eigenvector sequence ,calculate and The dot product of the values ​​is used as the element value at position (m,n) in the cross-modal correlation weight matrix. All element values, after calculation, form a matrix of shape... Multiply The cross-modal correlation weight matrix of the column is denoted as ,in The element in the m-th row and n-th column Represented as and The inner product of.

[0053] The cross-modal correlation weight matrix is ​​normalized along the direction of the component eigenvector sequence to generate the attention weight distribution. The normalization process uses the Softmax function along the cross-modal correlation weight matrix. For each line of execution, that is, for each fixed spatial feature vector index m, calculate... All of the m-th row The Softmax normalized values ​​between each element. The formula for generating the attention weight distribution is as follows:

[0054]

[0055] in, This represents the attention weight distribution value for the m-th spatial feature vector and the n-th component feature vector, with a value range of (0,1). Represents the cross-modal correlation weight matrix The value of the element in the m-th row and n-th column is the dot product of the m-th spatial eigenvector and the n-th component eigenvector. Represents the cross-modal correlation weight matrix The value of the element in the m-th row and q-th column, where q is the summation index variable, is obtained by traversing all the component feature vector sequences. One location; This represents an exponential function with the natural constant e as its base. This represents the total length of the component feature vector sequence. This is equal to the number of rows M in the solid waste material component feature matrix. After Softmax normalization, for each spatial feature vector index m, the distribution of attention weights corresponding to all n is... The sum of these is 1, forming a probability distribution.

[0056] Based on the attention weight distribution, the spatial accumulation features at each location in the solid waste spatial accumulation feature map set are weighted and reconstructed to generate a solid waste resource enrichment distribution map that integrates spatial accumulation information and material composition information. The specific method of weighted reconstruction is as follows: for the spatial location corresponding to the m-th spatial feature vector in the spatial feature vector sequence, the attention weight distribution is used... The weighted sum of all component feature vectors in the component feature vector sequence is used to generate the reconstructed fusion feature vector. The m-th reconstructed fusion feature vector is denoted as... The calculation method is as follows:

[0057]

[0058] in, This represents the m-th fused feature vector, with dimensions and component feature vectors. Same, so it's D; These are the attention weight distribution values; Let n be the eigenvector of the nth component; the summation range is from 1 to... Perform the above weighted reconstruction on all spatial locations m to obtain... fusion feature vectors .Will The fused feature vectors are rearranged according to the spatial positional relationships of the feature maps before the spatial feature vector sequence is unfolded, resulting in a high-resolution version. and width The feature maps are arranged in the form of feature maps, and all rearranged feature maps constitute a solid waste resource enrichment distribution map. The feature vector of each feature map in the solid waste resource enrichment distribution map at each spatial location simultaneously contains the spatial packing intensity information of the corresponding location and the spectral feature information of the solid waste material components associated with that location. The two are correlated through cross-modal attention weights to achieve feature fusion of spatial mode and spectral mode.

[0059] The parameters involved in cross-modal attention fusion processing include the spatial feature vector dimension. Component feature vector dimension D, spatial feature vector sequence length and component feature vector sequence length The parameters are automatically determined through forward propagation of the first and second feature extraction networks. The cross-modal attention fusion processing itself does not contain learnable parameters and is not trained separately; it is embedded as a fixed computation module in the overall evaluation and management process.

[0060] See Figure 5 The figure shows different spatial locations in the solid waste spatial accumulation feature set during the cross-modal attention fusion processing step in Example 3. Attention weight distribution of spatial feature vectors and component feature vectors in the solid waste material component feature matrix With component eigenvector index The changing trend. The horizontal axis represents the component feature vector index. The range is from 0 to 500, and the vertical axis represents the corresponding attention weight value. The value ranges from 0 to 0.025.

[0061] As can be seen from the curves in the figure, the attention weight distribution at the five different spatial locations all exhibit a clear unimodal shape, with each curve concentrated within a specific interval of the feature index. This indicates that the cross-modal attention mechanism focuses primarily on a finite subset of the corresponding component feature vectors for each spatial location. Specifically:

[0062] Spatial location The attention weights are concentrated in the component feature index range of approximately 30 to 130, with a peak value of approximately 0.023, indicating that the stacking features at this spatial location are mainly associated with the component features of solid waste materials within this band range.

[0063] Spatial location The peak weight appears in the component feature index range of approximately 350 to 450, with a peak value of approximately 0.023, compared to... The positions are clearly separated, indicating that the characteristic spectral segments of material components corresponding to different spatial locations have strong distinguishability.

[0064] Spatial location The attention peak is located in the range of approximately 180 to 280 of the component feature index, with a peak value of approximately 0.023 and a relatively symmetrical peak shape, reflecting a strong response of this position to the component features of this band.

[0065] Spatial location The weight peaks are concentrated in the component feature index range of approximately 270 to 370, with a peak value close to 0.023. The peak position is similar to... The locations overlap to some extent, which may correspond to some common material composition characteristics.

[0066] Spatial location The weighted peak value is located in the range of approximately 380 to 480 of the component feature index, with a peak value of approximately 0.023. The peak shape is relatively concentrated, indicating that the solid waste accumulation characteristics at this spatial location are highly correlated with the components in the spectral feature range of this component.

[0067] Overall, the attention weight distribution of the five curves in the figure is clear, with peak positions that do not overlap and are all within a small index range, demonstrating the effective correlation and fine focusing capability of the cross-modal attention fusion mechanism between spatial packing features and component spectral features. The fused feature vectors at each spatial location are weighted and summed to obtain the corresponding component feature vectors, achieving precise fusion of spatial information and material composition information, providing rich fusion representations for the subsequent generation of solid waste resource enrichment distribution maps.

[0068] Example 4:

[0069] In practice, the steps for inputting the solid waste resource enrichment distribution map into the recycling value assessment network and outputting the comprehensive recycling potential index of the area to be assessed within a preset time window are as follows.

[0070] The recycling value assessment network comprises a first convolutional block, a second convolutional block, and a fully connected regression layer. The first convolutional block extracts the spatial location and enrichment density of recyclable component enrichment areas from the solid waste resource enrichment distribution map. The first convolutional block consists of two concatenated convolutional layers, each followed by a batch normalization layer and a linear unit activation function with leakage correction. The first convolutional layer of the first convolutional block uses 64 3×3 kernels with a stride of 1 and padding of 1; the second convolutional layer uses 128 3×3 kernels with a stride of 2 and padding of 1. The solid waste resource enrichment distribution map is input to the first convolutional layer of the first convolutional block as a multi-channel feature map. After convolution, batch normalization, and activation function processing in the first convolutional layer, a 64-channel feature map is output. This is then processed by convolution, batch normalization, and activation function processing in the second convolutional layer, resulting in a 128-channel feature map. The stride of 2 in the second convolutional layer reduces the spatial resolution of the feature map by half. In the 128-channel feature map output by the first convolutional block, the activation value of each channel at the spatial location (u,v) represents the response intensity of the location belonging to the recyclable component enrichment region. The higher the activation value, the higher the enrichment density of recyclable components. Multiple channels jointly encode the spatial location information of different recyclable component categories.

[0071] The spatial location and enrichment density of the recyclable component enrichment region extracted by the first convolutional block (i.e., the 128-channel feature map output by the first convolutional block) are input into the second convolutional block of the recycling value assessment network. The second convolutional block performs multi-scale feature extraction on the recyclable component enrichment region using different receptive fields, generating multi-scale enriched features. The second convolutional block adopts an initial module structure consisting of three parallel convolutional branches and one pooling branch. The first parallel branch contains a 1×1 convolutional layer, outputting a 32-channel feature map with a receptive field of 1×1. The second parallel branch contains a 1×1 convolutional layer followed by a 3×3 convolutional layer. The 1×1 convolutional layer compresses the number of channels to 48, and the 3×3 convolutional layer outputs a 64-channel feature map with a receptive field of 3×3. The third parallel branch contains a 1×1 convolutional layer followed by two stacked 3×3 convolutional layers. The 1×1 convolutional layer compresses the number of channels to 48, and both 3×3 convolutional layers output 64-channel feature maps, with an equivalent receptive field of 5×5. The pooling branch contains a 2×2 max-pooling layer with a stride of 1 and padding of 1, followed by a 1×1 convolutional layer, outputting a 32-channel feature map. The output feature maps from the four branches are concatenated along the channel dimension to form the output feature map of the second convolutional block, with a total of 32+64+64+32=192 channels. The output feature map of the second convolutional block is the multi-scale enriched feature, which spatially integrates enriched information from different receptive fields, ranging from local details to the surrounding environment.

[0072] The multi-scale enriched features are input into the fully connected regression layer of the recycling value assessment network. The fully connected regression layer consists of a global average pooling layer, a first fully connected layer, a modified linear unit activation function layer, a dropout layer, and a second fully connected layer. The global average pooling layer calculates the spatial average of each channel of the multi-scale enriched features, compressing the 192-channel multi-scale enriched features into a 192-dimensional feature vector. The first fully connected layer has 128 neurons, mapping the 192-dimensional feature vector to a 128-dimensional intermediate feature vector. The modified linear unit activation function layer provides a non-linear transformation. The dropout probability of the dropout layer is set to 0.5, randomly dropping 50% of the neurons during training to prevent overfitting, and automatically shutting down during inference. The second fully connected layer has a number of neurons equal to the total number of recycling categories R, the value of which is determined based on the actual list of recyclable material categories in the area to be evaluated. The second fully connected layer maps the 128-dimensional intermediate feature vector to an R-dimensional output vector. The r-th dimension of the output vector is the potential score of the area to be evaluated in the r-th recycling category, denoted as . .

[0073] The potential scores of all recycling categories output by the fully connected regression layer are used for a weighted summation. The weight coefficients of each recycling category used in the weighted summation are correlated with the historical recycling price series of the region to be evaluated. The historical recycling price series of each recycling category in the region to be evaluated within a preset time window are obtained. For the r-th recycling category, the arithmetic mean of the daily market recycling price of that category within the preset time window is calculated to obtain the average market price of the r-th recycling category, denoted as [value missing]. The average market price of all R recycling categories is normalized to obtain the weight coefficient of the r-th recycling category. The normalization process uses sum-based normalization, calculated by taking the average market price of the r-th recycling category. Divide by the sum of the average market prices of all recycling categories, i.e.:

[0074]

[0075] in, This represents the weight coefficient of the r-th recycling category, with a value range of (0,1), and the sum of the weight coefficients of all recycling categories is 1; This represents the average market price of the r-th recycling category within a preset time window, expressed in monetary units per kilogram. This represents the average market price of the u-th recycling category within a preset time window, where u is the summation index variable, and the process iterates through all R recycling categories; R is the total number of recycling categories.

[0076] Potential scores for each recycling category With the corresponding weighting coefficients Multiply and sum to obtain the comprehensive recycling potential index. The expression is ,in The comprehensive potential index for recycling, The potential score for the r-th recycling category is directly output by the fully connected regression layer. This represents the weighting coefficient for the r-th recycling category. The comprehensive recycling potential index. This is a scalar value; the larger the value, the higher the overall value of recyclable resources in that area within a preset time window.

[0077] The training process of the recycling value assessment network is as follows: A training dataset is constructed, in which each sample contains a historical solid waste resource enrichment distribution map and a corresponding true comprehensive recycling potential index label. The true comprehensive recycling potential index label is calculated using historical actual recycling revenue data: for a historical time period and region, the actual recycled weight of each recycling category is calculated, multiplied by the market price of each recycling category during the same period to obtain the revenue for each recycling category, and the total revenue is divided by the area of ​​the region to obtain the true comprehensive recycling potential index. During training, the historical solid waste resource enrichment distribution map is input into the recycling value assessment network, and forward propagation yields the predicted comprehensive recycling potential index. , with real labels For comparison, the mean squared error loss function is used to calculate the loss value. The Adam optimizer is used, with an initial learning rate of 0.0001, a batch size of 16, and 200 training epochs. Backpropagation is used to update all learnable parameters of the first convolutional block, the second convolutional block, and the fully connected regression layer until the loss function converges. During training, the weight coefficients... It does not participate in training, but only calculates and participates in weighted summation in real time based on the recycling price data within a preset time window during forward inference; during training, it uses the training set to construct the weight coefficients at that time to match the label construction method of the corresponding period.

[0078] See Figure 6 In the graph, the vertical axis represents the mean squared error loss, ranging from 0 to approximately 0.5, and the horizontal axis represents the number of training epochs, ranging from 0 to 200. The solid curve in the graph represents the training loss, and the dashed curve represents the validation loss. Overall, both the training loss and the validation loss show a significant decreasing trend, indicating that the recovery value assessment network is gradually optimized during training, and the prediction accuracy is continuously improved.

[0079] In the early stages of training (from 0 to approximately 20 epochs), both the training loss and validation loss rapidly decreased from approximately 0.3 to approximately 0.1, with the validation loss slightly higher than the training loss, indicating a certain risk of overfitting. As the number of training epochs increased (from approximately 20 to 100 epochs), both curves continued to decline steadily, and the validation loss gradually approached the training loss, indicating that the model's generalization ability improved and the training process was effectively stabilized.

[0080] In the later stages of training (approximately 100 to 200 epochs), the training loss and validation loss tend to stabilize with small fluctuations, eventually converging to a low level close to 0.01, indicating that the network has achieved a good fit. This loss decrease curve reflects the training process of the recovery value evaluation network in the above embodiment, which uses the mean squared error loss function to iteratively update parameters through the Adam optimizer, verifying the effectiveness of the network structure and training scheme.

[0081] Example 5:

[0082] In practice, the steps for obtaining the historical recycling efficiency sequence and logistics cost sequence of the area to be evaluated within a preset time window, and inputting the comprehensive recycling potential index, the historical recycling efficiency sequence, and the logistics cost sequence into the management strategy generation network to output the differentiated recycling scheduling strategy for the area to be evaluated are as follows.

[0083] The comprehensive recycling potential index is spatially back-projected onto each monitoring station within a preset time window to generate a station potential distribution vector. There are T monitoring stations within the preset time window, and the total number of monitoring stations is T. The comprehensive recycling potential index V is a scalar value representing the overall recycling potential of the entire area to be evaluated. Spatial back-projection is achieved through a pre-established regional spatial division weight matrix, denoted as E, with a shape of T rows multiplied by 1. List, The total length of the spatial feature vector sequence during the generation of the solid waste resource enrichment distribution map, and the element in the t-th row and m-th column of E. This represents the area overlap ratio between the spatial coverage area of ​​the t-th monitoring station and the spatial location corresponding to the m-th spatial feature vector. The value range is [0,1], and the sum of all elements in the same row is 1. The site potential distribution vector is denoted as... The site potential score of the t-th monitoring station. The comprehensive recycling potential index V is determined as follows: It is allocated to each monitoring station according to spatial location correspondence. Specifically, for the t-th monitoring station, the sum of the activation values ​​of the fused feature vectors corresponding to all spatial locations within the spatial coverage area of ​​the t-th monitoring station is extracted from the solid waste resource enrichment distribution map. After normalization, this sum is multiplied by the comprehensive recycling potential index V to obtain the station's potential score. Site potential distribution vector The dimension is T, which represents the distribution of the recovery value of each monitoring station within the current preset time window.

[0084] The historical recycling efficiency sequence is broken down by monitoring station and type of recycled material, generating a historical efficiency matrix for each monitoring station across all recycling categories. The historical recycling efficiency sequence contains recycling data from each monitoring station for Q consecutive time units prior to a preset time window. The granularity of the time unit is set to days, and the value of Q is set to 30 days based on data availability and periodicity. The historical recycling efficiency sequence records the estimated amount of materials collected and the amount that should be collected for each recycling category at each monitoring station within each time unit. For the t-th monitoring station and the r-th recycling category, the actual collection amount of materials in the r-th recycling category within each of the most recent Q time units is extracted from the historical recycling efficiency sequence and divided by the estimated amount that should be collected to obtain the recycling efficiency ratio. The average of the Q recycling efficiency ratios is then calculated as the historical efficiency value of the t-th monitoring station in the r-th recycling category, denoted as [missing value]. The historical efficiency values ​​of all T monitoring sites and R recycling categories together constitute a historical efficiency matrix, with a shape of T rows multiplied by R columns, denoted as . The element in the t-th row and r-th column is , The value range is [0,1], where 1 indicates that the historical recycling efficiency has reached the full amount that should be collected.

[0085] The logistics cost sequence is broken down according to the transportation routes between each monitoring station, generating a path cost matrix between each monitoring station. The logistics cost sequence includes historical transportation cost records prior to a preset time window, recording transportation cost data between any two monitoring stations or between a monitoring station and a recycling center. Transportation routes include paths from monitoring station to monitoring station and paths from a monitoring station to a recycling center, with the recycling center as the (T+1)th node in the path matrix. The path cost matrix is ​​denoted as... The shape is T rows multiplied by (T+1) columns, where the first... Line number Column elements Indicates from the first The monitoring station to the The cost of a single transport to each monitoring station; when =T+1, Indicates from the first The cost of a single transport from a monitoring station to the recycling center. The transport cost is measured in monetary units and is calculated based on the distance of the transport route, road condition parameters, and fuel consumption rate, or obtained by averaging from historical logistics cost series.

[0086] The site potential distribution vector, historical efficiency matrix, and path cost matrix are concatenated to generate a joint input feature vector. (Site potential distribution vector) The historical efficiency matrix has a length of T. Expanding by rows yields a one-dimensional vector of length T multiplied by R; the path cost matrix. Expanding by rows, we obtain a one-dimensional vector of length T multiplied by (T+1). Concatenating the three vectors sequentially, end to end, forms a joint input feature vector, denoted as... . Total dimensions Calculated as .

[0087] Joint input feature vector The input is fed into the graph attention layer of the management policy generation network. The core of the management policy generation network is the graph attention layer, which models each monitoring station as a graph node. The total number of monitoring stations is T, and the set of graph nodes is denoted as Ti. The negative correlation values ​​of the path cost matrix between each monitoring station are used as the initial weights of the edges between graph nodes. For the The monitoring station and the first For each monitoring station, if the path cost matrix includes... Existence and Then in the first The graph node and the first An edge is established between each node in the graph, and the initial weight of the edge is denoted as . The calculation method is as follows ,in Let be the numerical stability constant, set to . If there is no direct transport path between the two monitoring stations, then Initial weights of edges This indicates that the lower the path cost, the stronger the initial association.

[0088] The site potential distribution vector is used as the initial node feature for each graph node, and the historical efficiency matrix is ​​used as the auxiliary node feature for each graph node. For the t-th graph node, the initial node feature is denoted as... , It consists of two parts: site potential score. As the main feature, and all elements in the t-th row of the historical efficiency matrix As an auxiliary feature, i.e. The initial node feature matrix is ​​formed by the initial node features of all T graph nodes, with dimensions 1+R. The shape is T rows multiplied by (1+R) columns.

[0089] The graph attention layer iteratively updates the scheduling priorities among monitoring stations. The number of iterations is denoted as K, and the value of K is positively correlated with the sparsity of the path cost matrix among the monitoring stations. The sparsity of the path cost matrix is ​​denoted as... sparsity Defined as path cost matrix The ratio of the number of non-zero elements to the total number of elements. The specific method for determining the number of iterations K is as follows: set the base number of iterations. Set the maximum number of iterations. Sparsity threshold When sparsity Less than the sparsity threshold When K is set to When sparsity Greater than or equal to the sparsity threshold At that time, Increase based on the existing basis, the increase amount is At the same time, ensure that the K value does not exceed Higher sparsity indicates a more complex transportation connectivity between monitoring stations, requiring more iterations to ensure sufficient information dissemination.

[0090] The single-iteration update process of the graph attention layer is as follows. In the k-th iteration, k ranges from 1 to K, and for each graph node... Computation graph nodes With all adjacent graph nodes The attention coefficient between the t-th graph node and the j-th graph node in the k-th iteration is denoted as . Attention coefficients are based on graph nodes Node features in the (k-1)th iteration With Adjacency Graph Nodes Node features in the (k-1)th iteration The similarity between them is determined. Similarity calculation uses a learnable linear transformation and a dot product attention mechanism. The graph attention layer of the management policy generation network contains a learnable weight matrix. The shape is (1+R) rows multiplied by List, The attention hidden dimension is set to 32, along with a learnable attention vector. , dimension First, the graph node features are processed through... Perform a linear transformation: , ,in and All 3D vector. The transformed features are concatenated and the attention coefficients are calculated to obtain the original score.

[0091]

[0092] in, This represents the transpose of the learnable attention vector in the k-th iteration, with a dimension of 1 row multiplied by 1. List; Indicates will and Concatenate along the vector dimension to obtain 3D vector; the negative slope parameter of LeakyReLU is set to 0.2. This represents the raw score of the attention coefficient between the t-th graph node and the j-th graph node in the k-th iteration. For graph nodes... The set of all adjacency graph nodes The attention coefficients are obtained by normalizing using the Softmax function:

[0093]

[0094] in, Represents graph nodes There exists a set of indices of all adjacent graph nodes connected to an edge, including the graph nodes. The original score of itself and its loop. It is obtained by calculating its own characteristics, and the calculation method is the same as that of adjacent nodes.

[0095] Based on the attention coefficient, weighted aggregation is performed on the node features of each adjacent graph node in the (k-1)th iteration to generate graph nodes. After the k-th iteration, the node features are updated. :

[0096]

[0097] in, To represent a nonlinear activation function, the exponential linear unit activation function is used; Let be the weight matrix for the linear transformation described above. After updating all T graph nodes, the node feature matrix for the k-th iteration is obtained. .

[0098] After K iterations, the final updated node features are obtained. Based on the differences between the updated node characteristics and the initial node characteristics of each graph node, the scheduling priority of each monitoring station is determined. The difference value of the t-th graph node is denoted as... , ,in The L2 norm of a vector. Difference value. The magnitude of the difference value reflects the degree of change in the node characteristics of the t-th monitoring station after integrating information on the global station potential distribution, historical recovery efficiency, and transportation costs. A larger difference value indicates a greater degree of feature reconstruction after information interaction at that monitoring station, and a more urgent corresponding scheduling need. All monitoring stations are sorted from largest to smallest difference value, with the monitoring station with the largest difference value receiving the highest scheduling priority and the one with the smallest difference value receiving the lowest scheduling priority, generating a scheduling priority ranking sequence. Based on the scheduling priority ranking sequence, and combined with the material collection time window constraints of each monitoring station and the material transfer volume constraints of each transportation route, a differentiated recovery scheduling strategy is generated. The differentiated recovery scheduling strategy includes the suggested material collection frequency for each monitoring station, and the material transfer batch and transfer volume allocation scheme for each transportation route.

[0099] The training process of the management strategy generation network is as follows: A training dataset is constructed, where each training sample contains a joint input feature vector corresponding to a historical period, and a manually defined optimized retrieval scheduling strategy as a label. The labels are given in the form of the scheduling priority order of each monitoring node and the supervision signal. The learnable parameters of the graph attention layer are also described. and The scheduling priority sequence is shared across each iteration, and training employs either a policy gradient method or a ranking loss to supervise the learning of the scheduling priority sequence. The loss function is a list-level ranking loss function, using the difference between the normalized depreciation cumulative gain and the labeled scheduling priority sequence as the loss value. The optimizer is the Adam optimizer, with an initial learning rate of 0.0005, a batch size of 4, and 150 training epochs. The weight matrix of the graph attention layer is updated across all iterations via backpropagation. and attention vector Until the losses are mitigated.

[0100] In practical implementation, the method also includes inputting the differentiated recycling scheduling strategy into a scheduling simulation environment for feasibility verification. The digital twin map of the scheduling simulation environment is constructed based on the geographic information data of the area to be evaluated, including digital elevation models, road network vector data, and geographic coordinates of monitoring stations. The digital twin map contains the location of each monitoring station, the distance of each transportation route, and road condition parameters, including road grade, speed limit, and congestion coefficient. The differentiated recycling scheduling strategy is converted into a simulation scheduling instruction sequence, which is represented in the form of discrete events. Each event includes execution time, execution station, operation type, and material transfer volume. Operation types include two main categories: collection initiation and transfer initiation. The collection initiation instruction specifies the start time and duration of the material collection time window for the corresponding monitoring station, while the transfer initiation instruction specifies the start and end points of the transportation route and the material transfer volume. The simulation environment executes a sequence of simulation scheduling instructions, progressing in discrete time steps of minutes. At each time step, the material backlog at each monitoring station is updated. The material backlog is calculated cumulatively based on the difference between the material generation rate and collection rate at that station. Simultaneously, the cumulative transport time of vehicles on each transport route is updated. The time series of material backlog at each monitoring station and the cumulative transport time of each transport route are recorded for all time steps in the simulation process. It is determined whether the maximum material backlog at each monitoring station exceeds a preset backlog threshold throughout the entire simulation period. The preset backlog threshold is set based on a certain percentage of the monitoring station's storage capacity, such as 90% of the storage capacity. Simultaneously, it is determined whether the cumulative transport time of each transport route exceeds a preset time threshold, which is determined comprehensively based on the maximum allowable working time of the transport route and the driver's working time limit. If the maximum material backlog at all monitoring stations does not exceed the preset backlog threshold and the cumulative transport time of all transport routes does not exceed the preset time threshold, then the differentiated recycling scheduling strategy is deemed to have passed the feasibility verification. If the maximum backlog of materials at any monitoring station exceeds a preset backlog threshold, or the cumulative transportation time of any transportation route exceeds a preset time threshold, the corresponding scheduling parameters in the differentiated recycling scheduling strategy will be corrected. The correction process is as follows: for monitoring stations with excessive backlog, increase their collection frequency parameter by 30%; for transportation routes with excessive cumulative transportation time, reduce the material transfer volume allocation for that route by 20%, and allocate the reduced transfer volume to nearby alternative transportation routes. Then, the corrected differentiated recycling scheduling strategy will be regenerated and its feasibility will be verified again until it passes the verification.

[0101] In its implementation, the method also includes time-series tracking correction of the comprehensive recycling potential index. It acquires a series of thermal maps and spectral sequences of tracked solid waste components for multiple consecutive tracking time windows following a preset time window for the area to be evaluated. The method for setting the tracking time windows is consistent with the preset time window, and the number of tracking time windows is set to three. For each tracking time window, the processing flow of the preset time window is executed sequentially: the thermal map sequence of tracked solid waste is input into the first feature extraction network to obtain a spatial accumulation feature map of tracked solid waste; the spectral sequence of tracked solid waste components is input into the second feature extraction network to obtain a feature matrix of material components of tracked solid waste; and cross-modal attention fusion processing is performed to generate a resource enrichment distribution map of tracked solid waste for that tracking time window. The resource enrichment distribution maps of tracked solid waste for each tracking time window are sequentially input into the recycling value assessment network to generate the comprehensive recycling potential index for each tracking time window, denoted as... Calculate the trend change between the comprehensive recycling potential index V for the preset time window and the comprehensive recycling potential index for each tracking time window. The trend change is calculated using a first-order difference operator. The first trend change is... The second trend change is The third trend change is The arithmetic mean of the three trend changes is taken as the average trend change. The comprehensive recycling potential index for a preset time window is adjusted by a moving average based on the average trend change, resulting in a corrected comprehensive recycling potential index. The correction method is as follows ,in For correction factor, Setting it to 0.5 means that the original assessment value is corrected only by half of the trend momentum to avoid overcorrection caused by fluctuations in single-window data.

[0102] In practical implementation, the method also includes triggering incremental category learning updates based on the revised comprehensive recycling potential index. The revised comprehensive recycling potential index... With preset potential threshold Compare and preset potential thresholds. The target value is set at a certain percentage based on the highest historical comprehensive recycling potential index of the area to be assessed, with the percentage set at 75%. , This represents the maximum value of the comprehensive recycling potential index in historical records. When the adjusted comprehensive recycling potential index... Exceeding the preset potential threshold At this time, the current time window is marked as the category incremental learning trigger window. Within the category incremental learning trigger window, component spectral samples and spatial accumulation samples of the newly added solid waste category are acquired. The newly added solid waste category refers to the solid waste material category not included in the training phase of the second feature extraction network and the recycling value assessment network. Component spectral samples of the newly added category are acquired by deploying supplementary spectral acquisition equipment at monitoring sites, and spatial accumulation samples of the newly added category are acquired by high spatial resolution remote sensing or on-site image acquisition. The component spectral samples of the newly added solid waste category are input into the second feature extraction network. After forward propagation through the pre-trained multi-branch convolutional layer and feature splicing layer in the second feature extraction network, the material component feature vector of the newly added category is generated. After generating material component feature vectors for multiple spectral samples of the newly added category, the average value is calculated to obtain the prototype feature vector of the newly added category. Add new category prototype feature vectors As new category prototypes, they are embedded into the category prototype library of the fully connected regression layer of the recycling value assessment network. The category prototype library of the fully connected regression layer maintains a prototype matrix, where each trained recycling category corresponds to a prototype feature vector, and the newly added category prototype feature vector... Append to the prototype matrix to update the total number of recovery categories from R to R+1. Freeze all parameters of the first feature extraction network, freeze all parameters of the multi-branch convolutional layers and feature concatenation layers of the second feature extraction network, and only perform incremental updates on the classification weights of the fully connected regression layer of the recovery value evaluation network. The original weight matrix of the second fully connected layer of the fully connected regression layer has a shape of 128 rows by R columns. Expand the weight matrix by one column, using the prototype feature vectors of the newly added categories in the expanded column. The mapping values ​​in the feature space preceding the second fully connected layer are initialized. These mapping values ​​are obtained through forward propagation of the first few layers of the frozen fully connected regression layer. Then, a small number of newly added category samples are used to fine-tune the weights of the expanded fully connected regression layer with a small learning rate of 0.00001 and a fine-tuning period of 10, generating the updated recycling value assessment network. The updated recycling value assessment network can simultaneously evaluate the potential scores of both existing and newly added recycling categories, outputting a potential score vector with dimension R+1, and calculating the comprehensive recycling potential index including the newly added categories based on this vector.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An assessment and management method based on solid waste classification, recycling, and utilization, characterized in that, The method includes: Obtain solid waste thermogram sequences, solid waste component spectral sequences, and recycling flow records from multiple monitoring stations in the area to be evaluated within a preset time window; The solid waste heat map sequence is input into the first feature extraction network to generate a solid waste spatial accumulation feature map set of the area to be evaluated. The solid waste component spectral sequence is input into the second feature extraction network to generate the solid waste material component feature matrix of the region to be evaluated. The solid waste spatial accumulation feature atlas and the solid waste material composition feature matrix are subjected to cross-modal attention fusion processing to generate a solid waste resource enrichment distribution map of the area to be evaluated. The solid waste resource enrichment distribution map is input into the recycling value assessment network, and the comprehensive recycling potential index of the area to be assessed is output within the preset time window. The historical recycling efficiency sequence and logistics cost sequence of the area to be evaluated within the preset time window are obtained. The comprehensive recycling potential index, the historical recycling efficiency sequence, and the logistics cost sequence are jointly input into the management strategy generation network to output the differentiated recycling scheduling strategy of the area to be evaluated.

2. The assessment and management method based on solid waste classification and recycling according to claim 1, characterized in that, The step of inputting the solid waste heat map sequence into the first feature extraction network to generate a solid waste spatial accumulation feature map set for the area to be evaluated includes: Spatial resolution normalization is performed on each frame of the solid waste heat map sequence to unify the pixel size of all solid waste heat maps to a preset pixel size. Each frame of solid waste heat map after spatial resolution normalization is input into the spatiotemporal convolutional layer of the first feature extraction network in chronological order. The spatiotemporal convolutional layer extracts the local spatial accumulation intensity features and the accumulation diffusion trend features between adjacent frames of solid waste heat map. The local spatial accumulation intensity features and the accumulation diffusion trend features are input into the spatial attention layer of the first feature extraction network. The spatial attention layer performs weighted enhancement processing on the accumulation intensity at different spatial locations in the solid waste thermogram to generate an enhanced spatial accumulation feature map. The enhanced spatial stacking feature map is input into the feature pyramid structure of the first feature extraction network. The feature pyramid structure is used to perform multi-level fusion of spatial stacking features at different scales, and the solid waste spatial stacking feature map set is output.

3. The assessment and management method based on solid waste classification and recycling according to claim 1, characterized in that, The step of inputting the solid waste component spectral sequence into the second feature extraction network to generate the solid waste material component feature matrix of the region to be evaluated includes: Baseline correction and noise filtering are performed on each spectral data in the spectral sequence of the solid waste components to generate a corrected spectral data sequence. The corrected spectral data sequence is divided into band subsets according to the known characteristic bands of the solid waste material category, generating sub-spectral sequences corresponding to multiple band subsets; Each of the sub-spectral sequences is input in parallel into the multi-branch convolutional layer of the second feature extraction network, and the material type response features corresponding to each band subset are extracted by each branch convolutional layer. The response features of all material types output from each branch convolutional layer are input into the feature splicing layer of the second feature extraction network. The feature splicing layer splices all material type response features along the feature dimension to output the solid waste material component feature matrix.

4. The assessment and management method based on solid waste classification and recycling according to claim 3, characterized in that, The kernel size of each branch of the multi-branch convolutional layer is inversely proportional to the characteristic spectral bandwidth of the solid waste material category corresponding to each of the band subsets.

5. The assessment and management method based on solid waste classification and recycling according to claim 1, characterized in that, The step of performing cross-modal attention fusion processing on the solid waste spatial accumulation feature atlas and the solid waste material composition feature matrix to generate a solid waste resource enrichment distribution map of the area to be evaluated includes: The solid waste spatial accumulation feature map is expanded into a spatial feature vector sequence along the channel dimension, and the solid waste material component feature matrix is ​​expanded into a component feature vector sequence along the row direction; Perform a dot product operation between each spatial feature vector in the spatial feature vector sequence and each component feature vector in the component feature vector sequence to generate a cross-modal correlation weight matrix; The cross-modal correlation weight matrix is ​​normalized along the direction of the component feature vector sequence to generate an attention weight distribution; Based on the attention weight distribution, the spatial accumulation features at each location in the solid waste spatial accumulation feature map are weighted and reconstructed to generate the solid waste resource enrichment distribution map that integrates spatial accumulation information and material composition information.

6. The assessment and management method based on solid waste classification and recycling according to claim 1, characterized in that, The step of inputting the solid waste resource enrichment distribution map into the recycling value assessment network and outputting the comprehensive recycling potential index of the area to be assessed within the preset time window includes: The solid waste resource enrichment distribution map is input into the first convolutional block of the recycling value assessment network, and the spatial location and enrichment density of the recyclable component enrichment area in the solid waste resource enrichment distribution map are extracted through the first convolutional block. The spatial location and enrichment density of the recyclable component enrichment region are input into the second convolutional block of the recycling value assessment network. The second convolutional block performs multi-scale feature extraction of different receptive fields on the recyclable component enrichment region to generate multi-scale enrichment features. The multi-scale enriched features are input into the fully connected regression layer of the recycling value assessment network, and the multi-scale enriched features are mapped to the potential scores of the region to be assessed in each recycling category through the fully connected regression layer. A weighted summation operation is performed on the potential scores of the area to be evaluated in each recycling category. The weight coefficient of each recycling category is associated with the historical recycling price series of the area to be evaluated. The weighted summation result is used as the comprehensive recycling potential index.

7. The assessment and management method based on solid waste classification and recycling according to claim 6, characterized in that, The weighting coefficient for each recycling category is the normalized value of the average market price of each recycling category of materials in the region to be evaluated within the preset time window.

8. The assessment and management method based on solid waste classification and recycling according to claim 1, characterized in that, The steps of obtaining the historical recycling efficiency sequence and logistics cost sequence of the area to be evaluated within the preset time window, and inputting the comprehensive recycling potential index together with the historical recycling efficiency sequence and logistics cost sequence into the management strategy generation network to output the differentiated recycling scheduling strategy for the area to be evaluated include: The comprehensive recycling potential index is spatially back-projected according to each monitoring station within the preset time window to generate a station potential distribution vector for each monitoring station. The historical recycling efficiency sequence is broken down according to each monitoring station and each type of recycled material to generate a historical efficiency matrix for each monitoring station in each recycling category; The logistics cost sequence is broken down according to the transportation paths between each monitoring station to generate a path cost matrix between each monitoring station. The site potential distribution vector, the historical efficiency matrix, and the path cost matrix are concatenated to generate a joint input feature vector. This joint input feature vector is then input into the graph attention layer of the management strategy generation network. The graph attention layer iteratively updates the scheduling priority among the monitoring sites and outputs the differentiated recycling scheduling strategy.

9. The assessment and management method based on solid waste classification and recycling according to claim 8, characterized in that, The number of iterations of the graph attention layer is positively correlated with the sparsity of the path cost matrix between each monitoring station.

10. The assessment and management method based on solid waste classification and recycling according to claim 8, characterized in that, The steps for iteratively updating the scheduling priority among monitoring stations through the graph attention layer include: Each monitoring station is used as a graph node in the graph attention layer, and the negative correlation value of the path cost matrix between each monitoring station is used as the initial weight of the edges between the graph nodes. The site potential distribution vector is used as the initial node feature of each graph node, and the historical efficiency matrix is ​​used as the auxiliary node feature of each graph node. For each graph node, the attention coefficient between the graph node and all its neighboring graph nodes is calculated. The attention coefficient is determined based on the similarity between the initial node features and auxiliary node features of the graph node and the initial node features and auxiliary node features of its neighboring graph nodes. Based on the attention coefficient, the initial node features of each adjacent graph node are weighted and aggregated to generate the updated node features of each graph node. The scheduling priority of each monitoring station is determined based on the difference between the updated node characteristics of each graph node and the initial node characteristics of each graph node.

Images