Industrial solid waste recycling purpose intelligent matching method

Abstract

The invention provides an intelligent matching method for recycling purposes of industrial solid wastes, and belongs to the technical field of industrial solid waste resources, and the method comprises the following steps: firstly, carrying out comprehensive feature extraction and fusion on solid waste samples by adopting multi-mode sensing technologies such as near infrared spectroscopy and X-ray fluorescence to generate solid waste multi-dimensional feature fingerprints; then, utilizing a solid waste physical and chemical reaction mechanism equation to analyze the recycling potential of the solid waste; meanwhile, a pre-trained WRIM model is adopted to classify the solid waste, and the model combines a convolutional neural network and a multi-head attention mechanism; then, quantifying the matching degree between the solid waste characteristics and the resource application demand by using a tensor kernel similarity calculation function; and finally, through a multi-level priority evaluation engine, the comprehensive similarity scoring matrix, the resource potential index and the basic attribute category, generating an optimal resource scheme recommendation, and realizing accurate matching of solid waste resource purposes.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial solid waste resources, and specifically relates to an intelligent matching method for the resource utilization of industrial solid waste. Background Art

[0002] The resource utilization of industrial solid waste is an important way to achieve a circular economy and sustainable development. Traditional solid waste resource utilization technologies mainly rely on manual experience and simple physical and chemical index tests, and classify and match solid waste with uses through single parameters or limited physical and chemical property indicators. These methods have been applied to a certain extent in the fields of cement production, building materials utilization, soil improvement, metal recovery, etc., but the matching process mostly relies on expert experience and limited experimental data, making it difficult to achieve large-scale and standardized accurate matching. However, traditional matching methods have obvious limitations: first, relying on single or limited detection methods, the solid waste characteristic data obtained is incomplete, resulting in a lack of accuracy in the matching results; second, the matching process is highly dependent on expert experience, lacks a systematic quantitative evaluation standard, is highly subjective, and is difficult to guarantee matching accuracy; third, existing methods have difficulty in handling solid waste with complex components. For complex industrial solid waste with multiple components and multiple phases, it is impossible to accurately identify its optimal resource utilization path. Due to the wide variety, complex composition and high variability of industrial solid waste, existing technologies find it difficult to achieve high-precision matching of solid waste resource uses. Especially when faced with solid waste with complex composition, traditional matching methods can often only roughly judge the possible use categories and it is difficult to accurately locate the most suitable specific resource utilization path, resulting in inefficiency and value loss in the resource utilization process. This has become a key technical bottleneck restricting the efficient recycling of industrial solid waste. Summary of the Invention

[0003] In view of this, the present invention provides an intelligent matching method for the resource utilization of industrial solid waste, which can solve the technical problem of low accuracy in matching the resource utilization of industrial solid waste in the prior art.

[0004] The present invention is implemented as follows: The present invention provides an intelligent matching method for resource utilization of industrial solid waste, including: using multimodal sensing technology to extract features of solid waste samples; using a data fusion algorithm to integrate the features to generate a multi-dimensional feature fingerprint of solid waste; applying a solid waste physicochemical reaction mechanism equation to analyze solid waste samples, and calculating a solid waste resource utilization potential index; using a pre-trained solid waste identification and classification network WRIM model to classify solid waste samples; using a tensor kernel similarity calculation function to quantify the degree of matching between the multi-dimensional feature fingerprint of solid waste and preset resource utilization requirements, and obtain a similarity scoring matrix; using a multi-level priority evaluation engine in combination with the similarity scoring matrix, the solid waste resource utilization potential index and basic attribute categories to generate resource utilization plan recommendations.

[0005] Among them, the multimodal sensing technology refers to a comprehensive detection method that combines multiple detection means such as near-infrared spectroscopy, X-ray fluorescence, thermogravimetric analysis, and electron microscopy to simultaneously obtain solid waste characteristic data; the feature extraction includes obtaining spectral characteristics, elemental composition characteristics, thermogravimetric characteristics, and morphological characteristics.

[0006] Among them, the multi-dimensional characteristic fingerprint of solid waste refers to the conversion of the physical and chemical property data of solid waste into a unique identifier in a high-dimensional space through mathematical modeling, which facilitates computer similarity comparison and classification.

[0007] Among them, the solid waste physical and chemical reaction mechanism equation is used to calculate the reaction activity and conversion efficiency of solid waste in different resource utilization pathways. The input includes key element content, chemical bond energy distribution, surface active site density, crystal phase structure parameters and catalyst activity index, and the output is the solid waste resource utilization potential index.

[0008] Among them, the key element content refers to the mass percentage of elements in solid waste that have a decisive influence on resource utilization, including the content of major metal elements, non-metallic elements, rare earth elements, precious metal elements and harmful elements in solid waste, which is obtained through X-ray fluorescence analysis.

[0009] Among them, the chemical bond energy distribution refers to the energy distribution of various chemical bonds in solid waste, which is obtained through spectral analysis; the surface active site density refers to the number of active centers per unit area on the surface of solid waste particles that can participate in chemical reactions.

[0010] Among them, the crystal phase structure parameters refer to characteristic parameters such as the lattice constant, space group and crystal orientation of the crystalline phase in solid waste; the catalyst activity index refers to a dimensionless indicator that measures the degree to which the catalytically active components in solid waste promote resource utilization reactions.

[0011] Among them, the solid waste resource utilization potential index refers to a comprehensive indicator for quantitatively evaluating the transformation value of solid waste in a certain resource utilization path.

[0012] Among them, the specific structure of the pre-trained solid waste identification and classification network WRIM model is a hybrid architecture combining a multi-layer perceptron and a convolutional neural network, which includes a feature extraction layer, a feature fusion layer, a multi-head attention mechanism layer and a classification output layer.

[0013] Among them, the feature extraction layer adopts a composite structure combining parallel and series connections, including a convolutional residual module, a sparse attention module, a dense connection module and an adaptive pooling module. The information flow is transmitted between the modules through jump connections.

[0014] The feature fusion layer adopts an adaptive weight mechanism to integrate different modal features, and the number of heads in the multi-head attention mechanism layer is dynamically adjusted according to three parameters: the dimension of the solid waste multi-dimensional feature fingerprint, the complexity of the solid waste sample, and the expected classification accuracy.

[0015] The steps for establishing the training dataset for the WRIM model specifically include selecting representative samples from a global industrial solid waste database, standardizing the samples, expanding the training set through data augmentation techniques, labeling and classifying the samples using expert knowledge, and using cross-validation to divide the training set into a test set. The steps for training the WRIM model specifically include initializing model parameters, setting the learning rate and optimizer, performing mini-batch gradient descent training, regularly evaluating model performance, adjusting hyperparameters based on validation set performance, applying an early stopping strategy to prevent overfitting, and saving the optimal model weights.

[0016] The tensor kernel similarity calculation function refers to a similarity calculation method based on high-order matrix decomposition, which expresses the multi-dimensional feature fingerprint of solid waste and the preset resource utilization demand as tensors, calculates their spectral norm ratio after tensor kernel transformation, and outputs a similarity scoring matrix. The similarity scoring matrix directly affects the feature weight allocation mechanism in the WRIM model. The features corresponding to high-similarity areas obtain higher weights in subsequent classifications, realizing dynamic adjustment of feature importance. The multi-level priority evaluation engine refers to an intelligent decision-making system that comprehensively evaluates the priority of solid waste resource utilization schemes based on the similarity scoring matrix, the solid waste resource utilization potential index and the basic attribute categories. It calculates the comprehensive score of each scheme through a hierarchical analysis method, and continuously optimizes the decision-making rules based on historical application effects.

[0017] The present invention uses multimodal sensing technology to comprehensively extract solid waste features, uses data fusion algorithms to generate multidimensional feature fingerprints of solid waste, and combines the physical and chemical reaction mechanism equations of solid waste to deeply analyze the characteristics of solid waste, thereby achieving accurate matching of solid waste resource utilization uses. This method overcomes the defects of low matching accuracy of traditional technologies: first, the spectral features, elemental composition features, thermogravimetric features and morphological features obtained by multimodal sensing technology comprehensively characterize the characteristics of solid waste, improving the integrity and representativeness of feature extraction; secondly, the WRIM model combined with tensor kernel similarity calculation achieves accurate matching between solid waste and resource utilization needs, greatly improving the accuracy of matching; thirdly, the physical and chemical reaction mechanism equations of solid waste are combined with a multi-level priority evaluation engine, so that the matching results have theoretical support and practical guidance significance. Through the synergistic effect of the above-mentioned technical means, the present invention significantly improves the accuracy of matching industrial solid waste resource utilization uses, and has made breakthrough progress in the identification and matching of the uses of complex industrial solid waste. It effectively solves the technical problem of low matching accuracy in the existing technology and provides scientific and reliable technical support for the efficient resource utilization and value maximization of industrial solid waste. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0020] like Figure 1 FIG. 1 is a flow chart of an intelligent matching method for resource utilization of industrial solid waste provided by the present invention, and the method comprises the following steps:

[0021] S01. Use multimodal sensing technology to extract features of solid waste samples to obtain spectral features, elemental composition features, thermogravimetric features, and morphological features;

[0022] S02. Using a data fusion algorithm to integrate the spectral characteristics, the elemental composition characteristics, the thermogravimetric characteristics, and the morphological characteristics to generate a multi-dimensional feature fingerprint of the solid waste;

[0023] S03. Performing an in-depth analysis of the solid waste sample using a solid waste physicochemical reaction mechanism equation, wherein the input parameters of the solid waste physicochemical reaction mechanism equation include the key element content, chemical bond energy distribution, surface active site density, crystal phase structure parameters, and catalyst activity index extracted from the multi-dimensional characteristic fingerprint of the solid waste, and calculating a solid waste resource potential index;

[0024] S04. Classify the solid waste sample using a pre-trained solid waste identification and classification network WRIM model, where the input of the WRIM model is the multi-dimensional feature fingerprint of the solid waste, and the output determines the basic attribute category of the solid waste sample;

[0025] S05. quantifying the degree of match between the multi-dimensional feature fingerprint of the solid waste and the preset resource utilization requirements using a tensor core similarity calculation function to obtain a similarity scoring matrix, wherein the similarity scoring matrix is ​​used to adjust the feature weights in the WRIM model;

[0026] S06. Generate resource recovery plan recommendations using a multi-level priority evaluation engine in combination with the similarity scoring matrix, the solid waste resource recovery potential index, and the basic attribute categories.

[0027] Among them, multimodal sensing technology refers to a comprehensive detection method that combines multiple detection methods such as near-infrared spectroscopy, X-ray fluorescence, thermogravimetric analysis, and electron microscopy to simultaneously obtain solid waste characteristic data.

[0028] Among them, the multi-dimensional characteristic fingerprint of solid waste refers to the conversion of the physical and chemical property data of solid waste into a unique identifier in a high-dimensional space through mathematical modeling, which facilitates computer similarity comparison and classification.

[0029] Among them, the solid waste physical and chemical reaction mechanism equation is used to calculate the reaction activity and conversion efficiency of solid waste in different resource utilization pathways. The input includes key element content, chemical bond energy distribution, surface active site density, crystal phase structure parameters and catalyst activity index, and the output is the solid waste resource utilization potential index.

[0030] Among them, the key element content refers to the mass percentage of elements in solid waste that have a decisive influence on resource utilization, including the content of major metal elements, non-metallic elements, rare earth elements, precious metal elements and harmful elements in solid waste, which is obtained through X-ray fluorescence analysis.

[0031] Among them, chemical bond energy distribution refers to the energy distribution of various chemical bonds in solid waste, which is obtained through spectral analysis.

[0032] Among them, the surface active site density refers to the number of active centers per unit area on the surface of solid waste particles that can participate in chemical reactions.

[0033] Among them, the crystal phase structure parameters refer to the characteristic parameters such as the lattice constant, space group and crystal orientation of the crystalline phase in solid waste.

[0034] Among them, the catalyst activity index refers to a dimensionless indicator that measures the degree to which the catalytically active components in solid waste promote resource utilization reactions.

[0035] Among them, the solid waste resource utilization potential index refers to a comprehensive indicator for quantitatively evaluating the transformation value of solid waste in a certain resource utilization path.

[0036] Among them, the specific structure of the pre-trained solid waste identification and classification network WRIM model is a hybrid architecture combining a multi-layer perceptron and a convolutional neural network, which includes a feature extraction layer, a feature fusion layer, a multi-head attention mechanism layer and a classification output layer. The feature extraction layer adopts a composite structure combining parallel and series connections, including a convolutional residual module, a sparse attention module, a dense connection module and an adaptive pooling module. The modules achieve efficient transmission of information flow through jump connections. The feature fusion layer adopts an adaptive weight mechanism to integrate different modal features. The number of heads in the multi-head attention mechanism layer is dynamically adjusted according to three parameters: the dimension of the solid waste multi-dimensional feature fingerprint, the complexity of the solid waste sample and the expected classification accuracy.

[0037] Among them, the convolutional residual module refers to a neural network structural unit that adds residual connections on the basis of the traditional convolutional layer, which is used to extract local features of solid waste samples while alleviating the gradient vanishing problem.

[0038] Among them, the sparse attention module refers to an attention mechanism network unit that introduces sparse constraints, which is used to capture key information in solid waste features and suppress redundant information.

[0039] Among them, the densely connected module refers to a neural network structural unit in which each layer is directly connected to all previous layers, which is used to maximize information flow and promote feature reuse.

[0040] Among them, the adaptive pooling module refers to a pooling layer that can automatically adjust the size of the pooling area, which is used to extract multi-scale information of solid waste characteristics.

[0041] Among them, skip connection refers to the connection method that directly transmits the output of the front layer in the neural network to the back layer, which is used to alleviate the difficulty of deep network training and retain the original feature information.

[0042] Among them, the adaptive weight mechanism refers to a calculation method that automatically assigns fusion weights according to the importance of input features.

[0043] The steps for establishing the training data set for the WRIM model include screening typical samples from the global industrial solid waste database, standardizing the samples, expanding the size of the training set through data augmentation technology, labeling and classifying the samples using expert knowledge, and dividing the training set and test set using the cross-validation method.

[0044] The steps of WRIM model training include initializing model parameters, setting learning rate and optimizer, performing small-batch gradient descent training, regularly evaluating model performance, adjusting hyperparameters based on validation set performance, applying early stopping strategy to prevent overfitting, and saving the optimal model weights.

[0045] Among them, the tensor kernel similarity calculation function refers to a similarity calculation method based on high-order matrix decomposition. The multi-dimensional feature fingerprint of the solid waste and the preset resource utilization demand are represented as tensors respectively, and their spectral norm ratios are calculated after tensor kernel transformation. The similarity score matrix is ​​output with a value range of 0 to 1. The similarity score matrix directly affects the feature weight allocation mechanism in the WRIM model. The features corresponding to the high-similarity areas obtain higher weights in subsequent classifications, thereby realizing dynamic adjustment of feature importance.

[0046] The spectral norm ratio refers to the ratio of the maximum singular values ​​of two matrices and is used to quantify the similarity of the matrices.

[0047] Among them, the multi-level priority evaluation engine refers to an intelligent decision-making system that comprehensively evaluates the priority of solid waste resource utilization plans based on the similarity scoring matrix, the solid waste resource utilization potential index and the basic attribute categories. It calculates the comprehensive score of each plan through a hierarchical analysis method and continuously optimizes the decision-making rules based on historical application results.

[0048] Among them, the hierarchical analysis method refers to a decision-making method that decomposes complex decision-making problems into multiple levels and compares the weights layer by layer.

[0049] The specific implementation of the above steps is described in detail below. The specific implementation of step S01 is to use a variety of sensor devices to perform comprehensive feature extraction on the solid waste sample. First, a near-infrared spectrometer is used to scan the solid waste sample to obtain spectral data in the wavelength range of 800 to 2500 nm. The Savitzky-Golay smoothing method is used to eliminate spectral noise, with a sampling interval of 2 nm to form a spectral feature vector. Then, an X-ray fluorescence spectrometer is used to perform elemental composition analysis to measure the content of each element in the solid waste. The excitation source power is set to 60 kV, the detection time is 120 s, and the detection limit is 0.001%. The element content distribution vector is generated. Then, a thermogravimetric analyzer is used to heat from room temperature to 900 ° C at a heating rate of 10 ° C / min under a nitrogen atmosphere. The sample mass change curve and its first-order derivative are recorded to obtain thermogravimetric feature data. Finally, the solid waste surface is imaged by a scanning electron microscope with a magnification of 500 to 5000 times to obtain the morphological characteristics of the solid waste particles. The morphological parameters such as particle size distribution, surface roughness, and porosity are extracted through image processing algorithms. The purpose of this step is to obtain comprehensive characteristic information of solid waste samples through multimodal sensing technology, providing a data basis for subsequent intelligent matching.

[0050] The specific implementation method of step S02 is to perform data fusion processing on the different dimensional features obtained by multimodal sensing technology. First, the data standardization method is used to convert the feature data of different dimensions into a unified scale. The spectral features are converted by standard normal variables, the element composition features are normalized by maximum and minimum values, the thermogravimetric features are processed by mean centering, and the morphological features are transformed by logarithm. Then, a multi-level data fusion architecture is used to integrate the standardized features. The low-level fusion uses the feature cascade method to directly splice the different modal data. The middle-level fusion uses principal component analysis to reduce the dimension and retain the principal components with a cumulative contribution rate of 95%. The high-level fusion uses a deep autoencoder to extract the potential feature representation. The encoder structure of the autoencoder is [1024, 512, 256, 128], and the decoder structure is

[0051] The multi-layer fusion results were integrated into a multidimensional feature fingerprint of solid waste using tensor decomposition technology. The tensor decomposition used was a third-order Tucker decomposition with a core tensor size of [32, 32, 32] and a decomposition error threshold of 0.01. This step aims to integrate the dispersed solid waste feature data into a unified digital representation, facilitating subsequent processing by intelligent algorithms.

[0052] The specific implementation of step S03 is to evaluate the resource potential of solid waste based on the solid waste physicochemical reaction mechanism equation. First, key parameters are extracted from the multi-dimensional feature fingerprint of solid waste, including the content of main metal elements {M i}(such as iron, aluminum, copper, etc.), non-metallic element content {N j}(such as silicon, phosphorus, sulfur, etc.), rare earth element content {R k}、Precious metal element content {P m} and harmful element content {H n}; Extract chemical bond energy distribution information {B p}, surface active site density D a , crystal structure parameters {C q} and catalyst activity index I c These parameters are then substituted into the solid waste physicochemical reaction mechanism equation to calculate the solid waste resource potential index (PI). This equation takes into account the stability, reactivity, and energy efficiency of solid waste under different thermodynamic and chemical conditions, and comprehensively evaluates the conversion value of solid waste in various resource recovery pathways, such as building materials utilization, metal recovery, and chemical synthesis. The calculation assumes the reaction conditions are standard atmospheric pressure, a temperature range of 25 to 1200°C, an oxidation-reduction potential range of -2.0 to +2.0V, and a reaction time of 0 to 24 hours. The purpose of this step is to quantitatively evaluate the potential value of solid waste in different resource recovery pathways through theoretical calculations, providing a scientific basis for subsequent solution recommendations.

[0053] The specific implementation of step S04 is to classify solid waste samples using a pre-trained solid waste identification and classification network WRIM model. First, the multi-dimensional feature fingerprint of the solid waste is passed into the WRIM model as an input vector. The feature extraction layer of the WRIM model performs preliminary processing on the input features. The convolution residual module uses a 3×3 convolution kernel to extract local features. The sparse attention module uses L1 regularization to achieve feature selection. The number of output channels per layer in the dense connection module is set to 64, and the output feature map size of the adaptive pooling module is unified to 8×8. The feature fusion layer uses an adaptive weight mechanism to integrate different modal features, and the weight coefficient is dynamically adjusted through the backpropagation algorithm. The number of heads in the multi-head attention mechanism layer is automatically selected according to the current sample complexity, ranging from 4 to 16. The WRIM model finally outputs the basic attribute category of the solid waste sample. The category classification includes 12 basic types, including metal slag, non-metallic slag, combustion residue, chemical waste slag, and construction waste. The classification probability threshold is set to 0.75. Samples below the threshold are marked as uncertain categories and undergo secondary analysis. The purpose of this step is to determine the basic attribute categories of solid waste samples and provide a classification basis for subsequent resource utilization plan matching.

[0054] The specific implementation of step S05 is to use the tensor kernel similarity calculation function to quantify the matching degree between solid waste and resource utilization. First, a preset resource utilization demand database is constructed, which includes 20 typical resource utilization directions such as building material use, metal recovery use, catalyst use, adsorption material use, and environmental remediation use. Each use is represented by a characteristic demand vector. Then, the multi-dimensional characteristic fingerprint of solid waste and the preset resource utilization demand are respectively represented as a third-order tensor T f and T u ; Then calculate the tensor kernel transformation K(T f , T u ), use the high-order singular value decomposition algorithm to decompose the tensor and calculate the spectral norm ratio of the two tensors The similarity score is obtained, and the value range is 0 to 1. Finally, the similarity score matrix S is generated, and the matrix dimension is n×m, where n is the solid waste feature dimension, m is the number of resource uses, and the score threshold is set to 0.7. Feature items above the threshold are marked as key matching features. The similarity score matrix S is directly used to adjust the feature weights in the WRIM model. According to the formula Update the weights, where α is the weight adjustment coefficient, which is set to 0.5. The purpose of this step is to quantitatively assess the degree of match between solid waste characteristics and the needs of various resource utilization applications, and to provide a similarity basis for recommending resource utilization plans.

[0055] The specific implementation of step S06 is to generate resource recovery scheme recommendations using a multi-level priority evaluation engine. First, a three-level evaluation structure is established, with the bottom layer being the basic attribute category evaluation, the middle layer being the solid waste resource recovery potential index evaluation, and the top layer being the similarity score matrix evaluation. A hierarchical analysis method is then used to determine the weights of each layer. By constructing a judgment matrix and calculating the eigenvectors, the weight coefficients of each layer are obtained, with the bottom layer weight being 0.2, the middle layer weight being 0.35, and the top layer weight being 0.45. The comprehensive score of each resource recovery scheme is then calculated as CS = 0.2 × BC + 0.35 × PI + 0.45 × SM, where BC is the basic attribute category matching score, PI is the solid waste resource recovery potential index, and SM is the maximum eigenvalue of the similarity score matrix. A list of resource recovery scheme recommendations is generated based on the comprehensive score ranking, with a score threshold set at 0.65. Schemes above the threshold are recommended. Finally, the decision rules are adjusted through feedback from historical application data, and the evaluation parameters of the multi-level priority evaluation engine are optimized using a reinforcement learning algorithm, with a learning rate set at 0.01, a discount factor set at 0.9, and 1000 iterations. The purpose of this step is to comprehensively consider the characteristics of solid waste, resource potential and matching degree, generate the optimal resource plan recommendation, and achieve efficient utilization of solid waste resources.

[0056] The detailed structure of the WRIM model and the specific implementation of its training dataset are as follows. The WRIM model adopts a hybrid architecture combining a multi-layer perceptron and a convolutional neural network, consisting of a feature extraction layer, a feature fusion layer, a multi-head attention mechanism layer, and a classification output layer. The feature extraction layer adopts a composite structure combining parallel and serial connections and includes four functional modules: the convolutional residual module uses depthwise separable convolution to reduce the number of parameters. The convolution kernel size is 3×3, the number of channels is 64, and there are three stacked layers. Each layer is followed by a batch normalization layer and a Li Kai activation function, and the residual connection uses an identity mapping. The sparse attention module implements feature selection through a gating mechanism and L1 regularization. The attention weights are normalized using a softmax function, and the sparsity parameter is set to 0.1. The dense connection module has 64 output channels per layer, with a total of four layers, a growth rate of 16, and a transition layer compression factor of 0.5. The adaptive pooling module uses a spatial pyramid pooling structure with pooling scales of 1×1, 2×2, and 4×4. The modules are connected through jump connections to achieve efficient information flow transmission. Jump connections use additive merging, with one jump connection set for every two layers. The feature fusion layer uses an adaptive weight mechanism to integrate features of different modalities, and dynamically allocates weights through learnable gating units. The gating network structure is a two-layer perceptron with a hidden layer dimension of 128, an output layer dimension equal to the number of modalities, and an activation function of the sigmoid function. The multi-head attention mechanism layer adopts a dynamic head number selection strategy, with the basic head number set to 8. According to the three parameters of input feature dimension d, sample complexity c, and expected classification accuracy a, the attention layer is selected according to the formula Dynamically adjusted, the dimension of each attention head is 64, and the attention score is calculated using scaled dot products. The classification output layer uses a fully connected network structure with hidden layer dimensions of [512, 256, 128]. The number of neurons in the output layer is equal to the number of solid waste categories, and the activation function is the softmax function.

[0057] The establishment of the WRIM model training dataset specifically includes the following steps: first, typical samples were selected from the global industrial solid waste database, covering solid waste types from different regions and industries, with a total sample size of 50,000; all samples were standardized, including crushing to a size of less than 100 mesh, drying to constant weight, and storing in an inert atmosphere to prevent oxidation and deterioration; the training set size was expanded through data augmentation techniques, including adding Gaussian noise (variance of 0.01), random rotation (angle range of -15° to +15°), random scaling (ratio range of 0.8 to 1.2), and feature masking (masking ratio of 0.1), to expand the sample size to 200,000; expert knowledge was used to annotate and classify the samples, and experts in materials science, chemical engineering, environmental engineering and other fields were invited to form an annotation team. The Delphi method was used to reach a classification consensus, and the classification system included 12 main categories and 42 subcategories; the stratified five-fold cross-validation method was used to divide the training set into training and test sets, with the training set accounting for 80%, the validation set accounting for 10%, and the test set accounting for 10%, to ensure that the distribution ratio of samples of each category was consistent in each dataset. The specific steps of WRIM model training include: initializing model parameters using He Kaiming initialization method, initializing weights with mean 0 and standard deviation The model is trained with a normal distribution of n, where n is the number of input units in this layer. The initial learning rate is set to 0.001, and a cosine annealing learning rate scheduling strategy is adopted with a period of 30 rounds and a minimum learning rate of 0.01 of the initial value. The Adam algorithm is selected as the optimizer with parameters β1 of 0.9, β2 of 0.999, and a weight decay coefficient of 0.0001. Mini-batch gradient descent training is performed with a batch size of 64 and a total of 100 rounds of training. The model performance is evaluated on the validation set every 5 rounds, and the evaluation indicators include accuracy, precision, recall rate, and F1 score. The hyperparameters are dynamically adjusted according to the performance of the validation set to complete the training.

[0058] It should be noted that in order to address the technical issue of low matching accuracy for industrial solid waste resource utilization, the following three key technical features are of significant significance in improving matching accuracy:

[0059] Multimodal sensing technology and the construction of multidimensional feature fingerprints of solid waste are the fundamental innovations of this method. Traditional methods often rely on only a single or limited detection method and are unable to fully capture the complex characteristics of solid waste. This method simultaneously obtains solid waste characteristic data through multiple detection methods such as near-infrared spectroscopy, X-ray fluorescence, thermogravimetric analysis and electron microscopy, and uses data fusion algorithms to integrate these heterogeneous data into high-dimensional feature fingerprints. This comprehensive characterization method greatly enriches the description of solid waste characteristics and provides a comprehensive information basis for subsequent accurate matching. As the unique identifier of solid waste in high-dimensional space, the multidimensional feature fingerprint enables computers to compare the degree of fit between solid waste and potential resource utilization from multiple dimensions, significantly improving the accuracy and comprehensiveness of matching, and overcoming the matching bias problem caused by insufficient information in traditional methods.

[0060] The solid waste physicochemical reaction mechanism equation and resource utilization potential index have achieved a deep understanding of the solid waste resource utilization process. Traditional matching methods are mostly based on empirical rules or simple data comparisons, and lack in-depth analysis of the reaction mechanism of solid waste in the resource utilization process. This method directly links microscopic properties such as key element content, chemical bond energy distribution, and surface active site density with resource conversion efficiency by establishing a physicochemical reaction mechanism equation, thereby realizing a theoretical mapping from the essential properties of materials to resource value. This mechanism-based analysis method can predict the reaction activity and conversion efficiency of solid waste in different resource utilization pathways, and quantitatively evaluate it as a solid waste resource utilization potential index, providing a scientific basis for matching decisions and avoiding the waste of resources and environmental risks caused by blind attempts in traditional methods.

[0061] Tensor kernel similarity calculation and multi-level priority evaluation engine provide high-precision matching algorithms and comprehensive decision support. Traditional matching methods usually use simple similarity calculations or manual experience judgments, which are difficult to handle complex relationships in high-dimensional feature spaces. This method uses a tensor kernel similarity calculation function based on high-order matrix decomposition. By expressing the multi-dimensional feature fingerprint of solid waste and the preset resource utilization requirements as tensors, the spectral norm ratio is calculated after kernel transformation to accurately quantify the degree of matching between the two. This advanced mathematical tool can capture nonlinear relationships and potential patterns between features, significantly improving matching accuracy. The multi-level priority evaluation engine comprehensively considers the similarity scoring matrix, solid waste resource utilization potential index and basic attribute categories, and generates optimal resource utilization plan recommendations through hierarchical analysis, realizing the intelligent transformation from data to decision. The engine can also continuously optimize decision rules based on historical application results, forming an adaptive learning mechanism to continuously improve matching results.

[0062] The organic combination of the above three key technical features has improved the accuracy of matching industrial solid waste resource uses from feature extraction, mechanism analysis to decision-making evaluation, and realized the methodological transformation from experience-driven to data-driven and mechanism-driven, providing a scientific path for efficient and accurate resource utilization of industrial solid waste.

[0063] Specifically, the principle of the present invention is:

[0064] The intelligent matching method for resource utilization of industrial solid waste of the present invention is based on the technical principles of combining multimodal data fusion, deep learning and mechanism model, and realizes accurate matching of resource utilization of solid waste through systematic feature extraction, analysis and matching process.

[0065] At the feature extraction level, this invention uses multimodal sensing technology to simultaneously acquire the spectral, elemental, thermogravimetric, and morphological characteristics of solid waste samples. Unlike traditional methods that rely on a single or limited detection method, multimodal sensing technology can comprehensively capture the physical and chemical properties of solid waste from multiple dimensions. A data fusion algorithm integrates these various features to generate a multidimensional feature fingerprint of the solid waste. This fingerprint serves as a unique identifier of the solid waste in a high-dimensional feature space, providing a comprehensive and accurate data foundation for subsequent precise matching.

[0066] At the analytical and evaluation level, this invention combines mechanistic models with data-driven methods. It applies the solid waste physicochemical reaction mechanism equation for in-depth analysis, calculating the theoretical suitability of solid waste for various resource utilization pathways by considering key parameters such as key element content, chemical bond energy distribution, and surface active site density. Simultaneously, it uses a pre-trained WRIM model for solid waste classification. This model combines the advantages of a multi-layer perceptron and a convolutional neural network. Through a composite feature extraction layer, a feature fusion layer with an adaptive weighting mechanism, and a multi-head attention mechanism layer, it can effectively process high-dimensional heterogeneous features, identify the basic attribute categories of solid waste, and lay the foundation for accurate matching.

[0067] At the matching decision level, the innovation of this invention lies in the introduction of a tensor core similarity calculation function. This function uses high-order matrix decomposition technology to compare the multidimensional feature fingerprint of solid waste with the preset resource utilization requirements in tensor space, quantify the degree of match, and generate a similarity score matrix. This matrix is ​​not only used to adjust the feature weights in the WRIM model to achieve dynamic feature importance adjustment, but also serves as the core input of the multi-level priority evaluation engine. Through hierarchical analysis, it comprehensively evaluates the priority of each resource utilization solution and generates the optimal matching recommendation.

[0068] The technical principle of the present invention is logical. Through comprehensive feature extraction, in-depth mechanism analysis, intelligent model classification and precise similarity calculation, a complete intelligent matching system for industrial solid waste resource utilization is formed. It can effectively solve the technical problem of low matching accuracy in the existing technology and achieve high-precision matching of solid waste resource utilization.

[0069] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0070] The specific implementation method of step S01 is to use a variety of sensor equipment to perform comprehensive feature extraction on the solid waste sample. First, a near-infrared spectrometer is used to scan the solid waste sample to obtain spectral data in the wavelength range of 800 to 2500 nm. The Savitzky-Golay smoothing method is used to eliminate spectral noise, and the sampling interval is 2 nm to form a spectral feature vector; then, an X-ray fluorescence spectrometer is used to perform elemental composition analysis to measure the content of each element in the solid waste. The excitation source power is set to 60 kV, the detection time is 120 s, and the detection limit is 0.001% to generate an element content distribution vector; then, a thermogravimetric analyzer is used to heat from room temperature to 900 ° C at a heating rate of 10 ° C / min in a nitrogen atmosphere, and the sample mass change curve and its first-order derivative are recorded to obtain thermogravimetric feature data; finally, the solid waste surface is imaged by a scanning electron microscope, and the magnification is set to 500 to 5000 times to obtain the morphological characteristics of the solid waste particles, and the morphological parameters such as particle size distribution, surface roughness, and porosity are extracted through image processing algorithms. The purpose of this step is to obtain comprehensive characteristic information of solid waste samples through multimodal sensing technology, providing a data basis for subsequent intelligent matching. The mathematical expression of the Savitsky-Golay smoothing method is:

[0071]

[0072] Where y j is the smoothed spectral data point; x j+i is the original spectrum data point; c i is the smoothing coefficient; N is the normalization factor, m is the half-width of the smoothing window, which is 5. The spectral feature vector is recorded as S NIR =[s1, s2, ..., s n ], where n is the number of spectral data points, which is determined by the wavelength range and sampling interval, and the calculation formula is The element content distribution vector is recorded as E = [e1, e2, ..., e p ], where p is the number of elements detected, usually 40 to 60 elements. Thermogravimetric characteristic data include mass change curve TG(T) = [tg1, tg2, ..., tg q] and the first-order derivative curve DTG(T)=[dtg1,dtg2,...,dtg q ], where T is the temperature variable and q is the number of temperature sampling points. The morphology parameters include the particle size distribution function P(d) = [p1, p2, ..., p r ], surface roughness index R a , porosity φ, etc.

[0073] The specific implementation method of step S02 is to perform data fusion processing on the different dimensional features obtained by multimodal sensing technology. First, the data standardization method is used to convert the feature data of different dimensions into a unified scale. The spectral features are converted by standard normal variables, the element composition features are normalized by maximum and minimum values, the thermogravimetric features are processed by mean centering, and the morphological features are transformed by logarithm. Then, a multi-level data fusion architecture is used to integrate the standardized features. The low-level fusion uses the feature cascade method to directly splice the different modal data. The middle-level fusion uses principal component analysis to reduce the dimension and retain the principal components with a cumulative contribution rate of 95%. The high-level fusion uses a deep autoencoder to extract the potential feature representation. The encoder structure of the autoencoder is [1024, 512, 256, 128], and the decoder structure is

[0074] The multi-layer fusion results were integrated into a multidimensional feature fingerprint of solid waste using tensor decomposition technology. The tensor decomposition used was a third-order Tucker decomposition, with the core tensor size set to [32, 32, 32] and the decomposition error threshold set to 0.01. The purpose of this step was to integrate the dispersed solid waste feature data into a unified digital representation to facilitate subsequent intelligent algorithm processing. The mathematical expression of the data standardization method is as follows:

[0075] Spectral characteristics standard normal variable transformation:

[0076] Where, is the normalized spectral feature vector; μ NIR is the mean of the spectral data; σ NIR is the standard deviation of the spectral data.

[0077] Normalization of the maximum and minimum values ​​of elemental composition characteristics:

[0078] Where, E norm is the normalized element content distribution vector; E min is the minimum value of element content; E max is the maximum value of the element content.

[0079] Thermogravimetric characteristic mean centering: TG norm(T) = TG (T) - μ TG ;DTG norm (T)=DTG(T)-μ DTG ;

[0080] Where, TG norm (T) and DTG norm (T) are the thermogravimetric curve and first-order derivative curve after centralization treatment; μ TG and μ DTG are the means of the thermogravimetric curve and the first-order derivative curve, respectively.

[0081] Logarithmic transformation of morphological features: P norm (d) = log(1 + P(d)); φ norm =log(1+φ);

[0082] Where, P norm (d) and φ norm are the logarithmically transformed particle size distribution function, surface roughness index, and porosity, respectively.

[0083] The mathematical expression of principal component analysis dimensionality reduction is: Y = W T X;

[0084] Where Y is the feature matrix after dimensionality reduction; X is the feature matrix after standardization; W is the eigenvector matrix, and the covariance matrix is ​​decomposed by eigenvalue It is obtained that ∑W=WΛ, where Λ is the eigenvalue diagonal matrix, and the eigenvectors corresponding to the eigenvalues ​​whose cumulative contribution rate reaches 95% are selected to form W.

[0085] The mathematical expression of deep autoencoder is:

[0086] Coding process: h i =f(W i h i-1 +b i ), i∈[1,l];

[0087] Decoding process:

[0088] Where h i is the output of the i-th layer encoder; is the j-th layer decoder output; W i and b i are the weight matrix and bias vector of the encoder layer i respectively; W j ′ and b j′ are the weight matrix and bias vector of the jth layer of the decoder respectively; f(·) is the Li Kai activation function, f(x) = max(0.01x, x); l is the number of encoder layers, here l = 4; h0 is the input feature; Reconstructed output.

[0089] The mathematical expression of the third-order tensor Tucker decomposition is:

[0090] Where, is the original data tensor; is a kernel tensor of size [32, 32, 32]; A, B, and C are factor matrices on the three modes respectively; × k represents the tensor product along the kth mode; the decomposition error is measured by the Frobenius norm, The final multi-dimensional feature fingerprint of solid waste is expressed as

[0091] The specific implementation of step S03 is to evaluate the resource potential of solid waste based on the solid waste physicochemical reaction mechanism equation. First, key parameters are extracted from the multi-dimensional feature fingerprint of solid waste, including the content of main metal elements {M i}(such as iron, aluminum, copper, etc.), non-metallic element content {N j}(such as silicon, phosphorus, sulfur, etc.), rare earth element content {R k}、Precious metal element content {P m} and harmful element content {H n}; Extract chemical bond energy distribution information {B p}, surface active site density D a , crystal structure parameters {C q} and catalyst activity index I c ; These parameters are then substituted into the solid waste physicochemical reaction mechanism equation to calculate the solid waste resource potential index PI. This equation takes into account the stability, reactivity and energy efficiency of solid waste under different thermodynamic conditions and chemical environments, and comprehensively evaluates the conversion value of solid waste in various resource utilization pathways such as building materials utilization, metal recycling, and chemical synthesis; the reaction conditions in the calculation are set to standard atmospheric pressure, the temperature range is 25 to 1200 ° C, the redox potential range is -2.0 to +2.0 V, and the reaction time is 0 to 24 h. The purpose of this step is to quantitatively evaluate the potential value of solid waste in different resource utilization pathways through theoretical calculations, and provide a scientific basis for subsequent scheme recommendations. The mathematical expression of the solid waste physicochemical reaction mechanism equation is:

[0092] PI=α·V e +β·R a +γ·E e -δ·P h ;

[0093] Where, PI is the solid waste resource potential index, ranging from 0 to 1; V e is the comprehensive index of value elements; R a is the reactivity index; E e is the energy efficiency index; P h is the environmental hazard index; α, β, γ, and δ are weight coefficients, and their values ​​vary according to different resource utilization pathways. Typical values ​​are α = 0.4, β = 0.3, γ = 0.2, and δ = 0.1.

[0094] Value element comprehensive index V e The calculation formula is:

[0095]

[0096] Where n M 、n N 、n R 、n P are the amounts of major metal elements, non-metal elements, rare earth elements and precious metal elements respectively; They are the economic value weight coefficients of various elements, which are adjusted dynamically according to market prices.

[0097] Reactivity index R a The calculation formula is:

[0098] R a =f B ({B p})·f D (D a )·f C ({C q})·f I (I c );

[0099] Where, f B ({B p}) is the chemical bond energy distribution function, which calculates the breaking energy of chemical bonds in solid waste and its contribution to the reaction activity; f D (D a ) is the surface active site density function, which characterizes the number of active centers that can participate in the reaction per unit area; f C ({C q}) is the crystal structure function, which describes the effect of the crystal structure of solid waste on the reaction activity; f I (I c ) is the catalytic activity function, which measures the promotion effect of catalytic active components in solid waste on resource recovery reactions. The specific form of these functions is:

[0100]

[0101] Where n B is the number of chemical bond types; k p is the reaction rate constant; B p is the content of the p-type chemical bond; E p is the activation energy of the p-th type chemical bond, in kJ / mol; R is the gas constant, 8.314 J / (mol·K); T is the reaction temperature, in K.

[0102] f D (D a )=1-exp(-λ·D a );

[0103] Where λ is the proportionality coefficient, which ranges from 0.5 to 2.0, depending on the type of solid waste.

[0104]

[0105] Where n C is the number of crystal phase structure parameters; ω q is the weight coefficient of the qth crystal phase structure parameter.

[0106]

[0107] Where, I c，max It is the maximum reference value of the catalyst activity index, which is 100.

[0108] Energy efficiency index E e The calculation formula is:

[0109]

[0110] Where ΔG out is the Gibbs free energy of the resource product, in kJ / mol; ΔG in The input energy for the resource recovery process is in kJ / mol; ∈ is a small positive number to prevent the denominator from being zero, and its value is 0.001.

[0111] Environmental Hazard Index P h The calculation formula is:

[0112]

[0113] Where n H is the amount of harmful elements; μ n is the hazard weight coefficient of the nth harmful element; H n is the content of the nth harmful element; η n It is the removal efficiency of the nth harmful element in the resource recovery process, and its value range is 0 to 1.

[0114] The specific implementation of step S04 is to classify solid waste samples using a pre-trained solid waste identification and classification network WRIM model. First, the multi-dimensional feature fingerprint of the solid waste is passed into the WRIM model as an input vector. The feature extraction layer of the WRIM model performs preliminary processing on the input features. The convolution residual module uses a 3×3 convolution kernel to extract local features. The sparse attention module uses L1 regularization to achieve feature selection. The number of output channels per layer in the dense connection module is set to 64, and the output feature map size of the adaptive pooling module is unified to 8×8. The feature fusion layer uses an adaptive weight mechanism to integrate different modal features, and the weight coefficient is dynamically adjusted through the backpropagation algorithm. The number of heads in the multi-head attention mechanism layer is automatically selected according to the current sample complexity, ranging from 4 to 16. The WRIM model finally outputs the basic attribute category of the solid waste sample. The category classification includes 12 basic types, including metal slag, non-metallic slag, combustion residue, chemical waste slag, and construction waste. The classification probability threshold is set to 0.75. Samples below the threshold are marked as uncertain categories and undergo secondary analysis. The purpose of this step is to determine the basic attribute categories of solid waste samples and provide a classification basis for subsequent resource recovery scheme matching. The mathematical expression of the WRIM model is as follows:

[0115] Classification output: C = softmax(MLP(ATT(F(X))));

[0116] Where C is the classification probability vector, and the dimension is the number of categories; X is the multi-dimensional feature fingerprint of solid waste; F(·) is the feature extraction layer function; ATT(·) is the multi-head attention mechanism layer function; MLP(·) is the multi-layer perceptron function; softmax is the normalized exponential function,

[0117] The feature extraction layer function F(·) contains the composite operation of four functional modules:

[0118] F(X)=POOL(DENSE(SPARSE(CONV(X))));

[0119] Where CONV(·) is the convolution residual module function; SPARSE(·) is the sparse attention module function; DENSE(·) is the dense connection module function; POOL(·) is the adaptive pooling module function.

[0120] The mathematical expression of the convolution residual module function is:

[0121]

[0122] Where, is the residual function, which consists of three layers of depth-wise separable convolution. Wi is the convolution kernel parameter of the i-th layer; σ(·) is the Li Kai activation function; BN(·) is the batch normalization function.

[0123] The mathematical expression of the sparse attention module function is:

[0124] SPARSE(X)=X⊙GATE(X)+λ||GATE(X)||1;

[0125] Where ⊙ is element-by-element multiplication; GATE(X) is the gating function, GATE(X) = softmax(W g ·X);W g is the gating weight matrix; λ is the L1 regularization coefficient, which is 0.1; ||·||1 is the L1 norm.

[0126] The mathematical expression of the densely connected module function is:

[0127] DENSE(X)=TRANS([X0,X1,X2,X3,X4]);

[0128] Where, X0=X is the initial input; X i =H i ([X0, X1, ..., X i-1 ]) is the output of the i-th layer, i∈[1,4]; H i is a composite function, including normalization, activation and convolution operations; [·] represents the feature concatenation operation; TRANS(·) is the transition layer function used to reduce the number of feature channels, TRANS(Y)=CONV 1×1 (BN(Y)), where CONV 1×1 It is a 1×1 convolution operation.

[0129] The mathematical expression of the adaptive pooling module function is:

[0130] POOL(X)=CONCAT(POOL1(X), POOL2(X), POOL4(X));

[0131] Where, POOL k Indicates pooling the input feature map into a feature map of size k×k; CONCAT represents the feature concatenation operation.

[0132] The mathematical expression of the multi-head attention mechanism layer function ATT(·) is:

[0133] ATT(X)=CONCAT(head1, head2,..., head H )W O ;

[0134] Where, is the i-th attention head; are the query, key, and value weight matrices of the i-th attention head respectively; W O is the output weight matrix; d k is the dimension of the key vector; H is the number of attention heads, which is determined by the three parameters of input feature dimension d, sample complexity c and expected classification accuracy a according to the formula Dynamic adjustment, H∈[4,16].

[0135] The specific implementation of step S05 is to use the tensor kernel similarity calculation function to quantify the matching degree between solid waste and resource utilization. First, a preset resource utilization demand database is constructed, which includes 20 typical resource utilization directions such as building material use, metal recovery use, catalyst use, adsorption material use, and environmental remediation use. Each use is represented by a characteristic demand vector. Then, the multi-dimensional characteristic fingerprint of solid waste and the preset resource utilization demand are respectively represented as a third-order tensor T f and T u ; Then calculate the tensor kernel transformation K(T f , T u ), use the high-order singular value decomposition algorithm to decompose the tensor and calculate the spectral norm ratio of the two tensors The similarity score is obtained, and the value range is 0 to 1. Finally, the similarity score matrix S is generated, and the matrix dimension is n×m, where n is the solid waste feature dimension, m is the number of resource uses, and the score threshold is set to 0.7. Feature items above the threshold are marked as key matching features. The similarity score matrix S is directly used to adjust the feature weights in the WRIM model. According to the formula Update the weights, where α is the weight adjustment coefficient, which is set to 0.5. The purpose of this step is to quantitatively evaluate the degree of match between solid waste characteristics and various resource utilization requirements, and provide a similarity basis for resource utilization scheme recommendations. f , T u ) is expressed as:

[0136] K(T f , T u )=T f ×1M1×2M2×3M3×T u ;

[0137] Where, × i represents the tensor product along the i-th mode; M1, M2, and M3 are the kernel mapping matrices of the tensor kernel transformation, which are used to enhance the expressive power of feature representation. The spectral norm of the tensor is defined as:

[0138]

[0139] Where x1, x2, and x3 are unit vectors; sup represents the supremum. The calculation formula of the similarity score matrix S is:

[0140]

[0141] Where S i，j Score the similarity between solid waste feature i and resource use j; is the i-th feature of the solid waste multidimensional feature fingerprint; It is the jth resource utilization demand.

[0142] The specific implementation method of step S06 is to use a multi-level priority evaluation engine to generate resource recovery scheme recommendations. First, a three-level evaluation structure is established, with the bottom layer being the basic attribute category evaluation, the middle layer being the solid waste resource recovery potential index evaluation, and the top layer being the similarity score matrix evaluation; then, a hierarchical analysis method is used to determine the weights of each layer, and the weight coefficients of each layer are obtained by constructing a judgment matrix and calculating the eigenvectors. The bottom layer weight is 0.2, the middle layer weight is 0.35, and the top layer weight is 0.45; then, the comprehensive score of each resource recovery scheme is calculated as CS = 0.2 × BC + 0.35 × PI + 0.45 × SM, where BC is the basic attribute category matching score, PI is the solid waste resource recovery potential index, and SM is the maximum eigenvalue of the similarity score matrix; a resource recovery scheme recommendation list is generated based on the comprehensive score sorting, and the score threshold is set to 0.65. Schemes above the threshold are recommended; finally, the decision rules are adjusted through historical application data feedback, and a reinforcement learning algorithm is used to optimize the parameters of the multi-level priority evaluation engine. The Q learning method is used, and the Q value update formula is:

[0143] Q(s t , a t )=Q(s t , a t )+α[r t +γmax a Q(s t+1 ,a)-Q(s t , a t )];

[0144] Where, Q(s t , a t ) is state s t Next take action a t The value function of r t is the immediate reward; α is the learning rate, which is 0.01; γ is the discount factor, which is 0.9; max a Q(s t+1 , a) is the next state s t+1 The maximum Q value of all possible actions under state s tDefined as the combination vector of solid waste characteristics and recommended solutions, action a t Defined as the operation of adjusting the evaluation parameters, the reward r t Calculated based on the actual application effect of the recommended solution.

[0145] The detailed structure of the WRIM model and the specific implementation of its training dataset are as follows. The WRIM model adopts a hybrid architecture combining a multi-layer perceptron and a convolutional neural network, consisting of a feature extraction layer, a feature fusion layer, a multi-head attention mechanism layer, and a classification output layer. The feature extraction layer adopts a composite structure combining parallel and serial connections and includes four functional modules: the convolutional residual module uses depthwise separable convolution to reduce the number of parameters. The convolution kernel size is 3×3, the number of channels is 64, and there are three stacked layers. Each layer is followed by a batch normalization layer and a Li Kai activation function, and the residual connection uses an identity mapping. The sparse attention module implements feature selection through a gating mechanism and L1 regularization. The attention weights are normalized using a softmax function, and the sparsity parameter is set to 0.1. The dense connection module has 64 output channels per layer, with a total of four layers, a growth rate of 16, and a transition layer compression factor of 0.5. The adaptive pooling module uses a spatial pyramid pooling structure with pooling scales of 1×1, 2×2, and 4×4. The modules are connected through jump connections to achieve efficient information flow transmission. Jump connections use additive merging, with one jump connection set for every two layers. The feature fusion layer uses an adaptive weight mechanism to integrate features of different modalities, and dynamically allocates weights through learnable gating units. The gating network structure is a two-layer perceptron with a hidden layer dimension of 128, an output layer dimension equal to the number of modalities, and an activation function of the sigmoid function. The multi-head attention mechanism layer adopts a dynamic head number selection strategy, with the basic head number set to 8. According to the three parameters of input feature dimension d, sample complexity c, and expected classification accuracy a, the attention layer is selected according to the formula Dynamically adjusted, the dimension of each attention head is 64, and the attention score is calculated using scaled dot products. The classification output layer uses a fully connected network structure with hidden layer dimensions of [512, 256, 128]. The number of neurons in the output layer is equal to the number of solid waste categories, and the activation function is the softmax function.

[0146] The establishment of the WRIM model training dataset specifically includes the following steps: first, typical samples were selected from the global industrial solid waste database, covering solid waste types from different regions and industries, with a total sample size of 50,000; all samples were standardized, including crushing to a size of less than 100 mesh, drying to constant weight, and storing in an inert atmosphere to prevent oxidation and deterioration; the training set size was expanded through data augmentation techniques, including adding Gaussian noise (variance of 0.01), random rotation (angle range of -15° to +15°), random scaling (ratio range of 0.8 to 1.2), and feature masking (masking ratio of 0.1), to expand the sample size to 200,000; expert knowledge was used to annotate and classify the samples, and experts in materials science, chemical engineering, environmental engineering and other fields were invited to form an annotation team. The Delphi method was used to reach a classification consensus, and the classification system included 12 main categories and 42 subcategories; the stratified five-fold cross-validation method was used to divide the training set into training and test sets, with the training set accounting for 80%, the validation set accounting for 10%, and the test set accounting for 10%, to ensure that the distribution ratio of samples of each category was consistent in each dataset.

[0147] The specific steps of WRIM model training include: initializing model parameters using He Kaiming initialization method, initializing weights with mean 0 and standard deviation Normal distribution, where n is the number of input units in this layer; set the initial learning rate to 0.001, adopt the cosine annealing learning rate scheduling strategy, the cycle is 30 rounds, and the minimum learning rate is 0.01 of the initial value; select the Adam algorithm as the optimizer, with parameters β1 of 0.9, β2 of 0.999, and the weight decay coefficient of 0.0001; perform small batch gradient descent training with a batch size of 64 and a total of 100 rounds of training; evaluate the model performance on the validation set every 5 rounds of training, and the evaluation indicators include accuracy, precision, recall rate and F1 score; dynamically adjust the hyperparameters according to the performance of the validation set to complete the training. Among them, the mathematical expression of He Kaiming's initialization method is:

[0148]

[0149] Where w is the weight parameter; represents a normal distribution with mean μ and standard deviation σ; n is the number of input units in this layer.

[0150] The mathematical expression of the cosine annealing learning rate scheduling strategy is:

[0151]

[0152] Where η t is the learning rate of the tth round; η min is the minimum learning rate, set to 0.00001; η maxis the maximum learning rate, set to 0.001; T is the cosine period, set to 30 rounds.

[0153] The update rule of the Adam optimization algorithm is:

[0154] m t =β1m t-1 +(1-β1)g t ;

[0155]

[0156] Where m t and v t are the first-order moment estimate and the second-order moment estimate respectively; g t is the current gradient; β1 and β2 are exponential decay rates, set to 0.9 and 0.999 respectively; and is the bias-corrected moment estimate; θ t is the model parameter; η t is the learning rate; ∈ is a small constant to prevent division by zero errors, set to 10 -8 .

[0157] Through the implementation of the above six steps, intelligent matching of industrial solid waste resource uses can be achieved, providing a scientific basis and decision-making support for solid waste resource utilization, and promoting the development of circular economy and efficient resource utilization. The core innovation of this method lies in: using multimodal sensing technology to comprehensively acquire solid waste characteristics; using advanced data fusion algorithms to generate multi-dimensional feature fingerprints of solid waste; evaluating the resource potential of solid waste based on physical and chemical reaction mechanism equations; using the deep learning model WRIM to achieve accurate classification; introducing the tensor kernel similarity calculation function to quantify the degree of matching; and constructing a multi-level priority evaluation engine to generate the optimal recommendation plan. This method can adapt to the resource needs of different types of solid waste and has broad application prospects. The algorithm optimizes the evaluation parameters of the multi-level priority evaluation engine, with the learning rate set to 0.01, the discount factor to 0.9, and the number of iterations to 1000. The purpose of this step is to comprehensively consider the characteristics of solid waste, resource potential and matching degree, generate the optimal resource plan recommendation, and achieve efficient utilization of solid waste resources. The construction method of the judgment matrix in the hierarchical analysis method is:

[0158]

[0159] Where a ij Indicates the importance of the i-th factor relative to the j-th factor, using a 1 to 9 scale. ji =1 / a ij , a ii = 1. The weight vector w is obtained by solving the eigenvalue equation Aw = λ max w is obtained, where λ maxis the maximum eigenvalue of the judgment matrix A.

[0160] The calculation formula for the basic attribute category matching score BC is:

[0161]

[0162] Where n c is the number of basic attribute categories of solid waste; v k is the matching value coefficient of the k-th attribute; p k is the probability that the solid waste sample belongs to the kth category, output by the WRIM model.

[0163] The calculation formula of the maximum eigenvalue SM of the similarity score matrix is:

[0164] SM=λ max (S)=max i λ i (S);

[0165] Where λ i (S) is the i-th eigenvalue of the similarity score matrix S.

[0166] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: A research team conducted research on intelligent matching of resource utilization of blast furnace slag solid waste generated by a steel plant. The team first used multimodal sensing technology to extract features from blast furnace slag samples. The SpectraScan-NIR3000 near-infrared spectrometer was used to scan the blast furnace slag samples to obtain spectral data with a wavelength range of 800 to 2500 nm. The Savitzky-Golay smoothing method was used to eliminate noise and obtain spectral feature vectors. The elemental composition of the sample was then analyzed using an X-MET8000 elemental analyzer, and the measured contents of the main elements are shown in Table 1.

[0167] Table 1 Analysis results of main element contents in blast furnace slag

[0168] element content(%) element content(%) element content(%) <![CDATA[SiO2]]> 34.28 <![CDATA[Al2O3]]> 13.76 CaO 38.65 MgO 8.23 <![CDATA[Fe2O3]]> 0.89 <![CDATA[TiO2]]> 0.52 <![CDATA[Na2O]]> 0.32 <![CDATA[K2O]]> 0.58 MnO 0.42 <![CDATA[P2O5]]> 0.21 S 0.65 other 1.49

[0169] Then, a STA449F3 thermogravimetric analyzer was used to heat the sample from room temperature to 900°C at a heating rate of 10°C / min under a nitrogen atmosphere, and the mass change curve and its first-order derivative were recorded. Finally, the surface morphology of the blast furnace slag was observed using a field emission scanning electron microscope, and parameters such as particle size distribution, surface roughness, and porosity were extracted. The particle size distribution ranged from 0.5 to 250 μm, with an average particle size of 42.6 μm and a surface roughness of R a The thickness is 5.87 μm and the porosity is 38.2%.

[0170] The research team used a data fusion algorithm to integrate the multimodal feature data. First, each feature was standardized and then integrated using a multi-level data fusion architecture. Low-level fusion used a feature concatenation method to directly concatenate data from different modalities. Mid-level fusion employed principal component analysis for dimensionality reduction, selecting the top 15 principal components with a cumulative contribution rate of 96.4%. High-level fusion employed a deep autoencoder to extract latent feature representations. The encoder structure was [1024, 512, 256, 128], the decoder structure was [128, 256, 512, 1024], the intermediate hidden layer dimension was 64, and the reconstruction error during training was 0.058. Finally, tensor decomposition technology was used to generate a multidimensional feature fingerprint of blast furnace slag, with a decomposition error of 0.008.

[0171] The resource potential of blast furnace slag samples was evaluated by applying the solid waste physicochemical reaction mechanism equation. Key parameters were extracted from the multidimensional feature fingerprint and the comprehensive index of value elements V was calculated. e is 0.726, and the reaction activity index R a The energy efficiency index E is 0.638. e is 0.815, and the environmental hazard index P h The resource utilization potential index PI calculated based on the solid waste physicochemical reaction mechanism equation is 0.752.

[0172] The pre-trained WRIM model was used to classify blast furnace slag samples. The model output determined that the blast furnace slag belonged to the metallic slag class with a classification probability of 0.943, far exceeding the set threshold of 0.75. The model achieved a classification accuracy of 97.2% on the validation set, with a precision of 96.8%, a recall of 95.7%, and an F1 score of 96.2%.

[0173] The results of tensor kernel similarity calculation show that the matching degree between the blast furnace slag and various resource utilization uses is shown in Table 2.

[0174] Table 2 Matching scores of blast furnace slag and different resource utilization uses

[0175] Resource utilization Similarity score Resource utilization Similarity score Cement production raw materials 0.953 Glass ceramic materials 0.892 concrete admixtures 0.917 Roadbed filling materials 0.875 Insulation materials 0.826 soil conditioner 0.764 Adsorption materials 0.683 Metal recycling 0.452 catalyst carrier 0.513 Environmental remediation materials 0.486

[0176] The multi-level priority assessment engine generated recommended recycling options for blast furnace slag, which was found to be the most suitable for cement production (overall score of 0.897), concrete admixture (overall score of 0.864), and glass-ceramic materials (overall score of 0.829). Based on these recommendations, the research team selected blast furnace slag for cement production. Actual application results have shown that using this blast furnace slag to replace some clay as a cement raw material can reduce energy consumption by approximately 15% and CO2 emissions by approximately 18%, while also meeting national performance standards.

[0177] Traditional matching of solid waste resource uses usually uses empirical judgment and simple physical property testing methods, which lacks systematicity and scientificity and easily leads to low resource efficiency. The present invention uses multimodal sensing technology to comprehensively acquire solid waste characteristics, generates multi-dimensional feature fingerprints of solid waste through data fusion, combines the solid waste physical and chemical reaction mechanism equation to evaluate resource potential, uses the WRIM deep learning model for accurate classification, and finally generates the optimal recommendation plan through a multi-level evaluation engine. Compared with traditional methods, the present invention has the advantages of more comprehensive feature extraction, more accurate matching algorithm, and more reliable recommendation results, which greatly improves the efficiency and value of solid waste resource utilization.

[0178] Example 3: A research team conducted intelligent resource matching research on phosphogypsum solid waste generated by a chemical plant. The team first used multimodal sensing technology to extract features from the phosphogypsum samples. An FTNIR-4500 near-infrared spectrometer was used to scan the phosphogypsum samples, acquiring spectral data in the wavelength range of 800 to 2500 nm with a sampling interval of 2 nm. Spectral feature vectors for 851 data points were obtained. Elemental composition analysis was performed using an Axios-mAX-4kW X-ray fluorescence spectrometer. The main element contents of the phosphogypsum are shown in Table 3.

[0179] Table 3 Analysis results of main element contents in phosphogypsum

[0180]

[0181] The sample was heated from room temperature to 900°C at a heating rate of 10°C / min under a nitrogen atmosphere using a TGA / DSC 3+ thermogravimetric analyzer, and the mass change curve of the sample was recorded. Obvious mass loss was detected in the three temperature ranges of 105-180°C, 350-450°C, and 700-780°C, corresponding to the removal of crystal water, dehydration, and decomposition processes, respectively. The surface morphology of phosphogypsum was observed using a Nova NanoSEM 450 field emission scanning electron microscope, and the morphological parameters were extracted. The particles were a mixture of plate-like and needle-like shapes, with a particle size distribution in the range of 2-120μm, an average particle size of 35.7μm, and a surface roughness of R a The thickness is 2.34 μm and the porosity is 27.6%.

[0182] The research team used a data fusion algorithm to integrate multimodal feature data. First, each feature was standardized. Spectral features were transformed using standard normal variables, elemental composition features were normalized using maximum and minimum values, thermogravimetric features were mean-centered, and morphological features were logarithmically transformed. A multi-level data fusion architecture was then used for integration. The feature dimension after low-level fusion was 1426, and mid-level fusion was reduced to 43 principal components using principal component analysis, with a cumulative contribution rate of 95.8%. High-level fusion used a deep autoencoder to extract latent feature representations. The training process used a mini-batch gradient descent method with a batch size of 64, a total of 120 training rounds, and an initial learning rate of 0.001. The final reconstruction error obtained from training was 0.043. Finally, a third-order tensor Tucker decomposition was used to generate a multidimensional feature fingerprint of phosphogypsum. The kernel tensor size was [32, 32, 32], and the decomposition error was 0.009.

[0183] The solid waste physicochemical reaction mechanism equation was used to conduct an in-depth analysis of the phosphogypsum sample. The key parameters were extracted from the multi-dimensional feature fingerprint to calculate the comprehensive index of value elements V. e The reactivity index R is 0.437. a The energy efficiency index E is 0.582. e The environmental hazard index P is 0.742. h The total waste recovery potential index (PI) was 0.563. In the solid waste physicochemical reaction mechanism equation, the weight coefficients α, β, γ, and δ were 0.35, 0.30, 0.25, and 0.10, respectively.

[0184] The pre-trained WRIM model was used to classify a sample of phosphogypsum as chemical waste with a classification probability of 0.912, exceeding the threshold of 0.75. The WRIM model achieved 94.6% accuracy, 93.8% precision, 92.7% recall, and a 93.2% F1 score for this sample.

[0185] The results of tensor kernel similarity calculation show that the matching degree between phosphogypsum and different resource utilization uses is shown in Table 4.

[0186] Table 4. Matching scores of phosphogypsum and different resource utilization uses

[0187] Resource utilization Similarity score Resource utilization Similarity score Gypsum building materials 0.967 Cement retarder 0.921 soil conditioner 0.883 Ammonium sulfate fertilizer production 0.852 roadbed materials 0.795 Lightweight wall panels 0.775 Filling material 0.743 Calcium source recovery 0.657 water treatment chemicals 0.512 Insulation materials 0.487

[0188] The resource utilization plan recommendations generated by the multi-level priority evaluation engine show that the three most suitable resource utilization pathways for phosphogypsum are: gypsum building materials (comprehensive score 0.834), cement retarders (comprehensive score 0.793) and soil conditioners (comprehensive score 0.746). The research team verified the resource utilization pathways for gypsum building materials and used the phosphogypsum for the production of gypsum boards after purification, neutralization and drying. The prepared gypsum board product has a flexural strength of 5.8MPa and a water absorption rate of 26.4%, which meets the requirements of the national standard GB / T 9775-2008. Compared with natural gypsum, the cost of building materials produced using phosphogypsum is reduced by about 25%, and the environmental burden of phosphogypsum is reduced.

[0189] Traditional methods for matching solid waste resource uses mainly rely on chemical composition analysis and limited physical property testing, and lack systematic consideration of the multidimensional characteristics of solid waste, resulting in a single resource utilization plan with low efficiency. The present invention extracts comprehensive feature information of solid waste through multimodal sensing technology, adopts advanced data fusion algorithms to generate multidimensional feature fingerprints of solid waste, and combines deep learning and tensor computing technology to achieve high-precision resource use matching. Compared with traditional methods, the present invention has a more comprehensive description of solid waste characteristics, a more intelligent matching process, and a more diversified resource utilization plan, which greatly improves the scientific nature and economic value of solid waste resource utilization. In the embodiment of phosphogypsum, traditional methods are often limited to using it as a cement retarder or a simple filling material, while the method of the present invention can explore more high-value-added resource utilization paths and improve resource utilization efficiency.

[0190] It should be noted that the industrial solid waste resource classification system is shown in Table 5 below.

[0191] Table 5 Industrial solid waste resource classification system

[0192]

[0193]

[0194] This classification system was developed through a Delphi process, reaching consensus among experts in materials science, chemical engineering, environmental engineering, and resource recovery. It considers multiple factors, including solid waste sources, physical and chemical properties, and resource potential, providing a foundational framework for intelligently matching solid waste to resource utilization.

[0195] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 6 and 7 below.

[0196] Table 6 Variable Explanation Table (Part 1)

[0197]

[0198]

[0199] Table 7 Variable Explanation Table (Part 2)

[0200]

[0201]

[0202] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. An intelligent matching method for resource utilization of industrial solid waste, characterized in that: include: The multimodal sensing technology is used to extract features from solid waste samples; the features are integrated using a data fusion algorithm to generate a multi-dimensional feature fingerprint of the solid waste; Solid waste samples are analyzed using the solid waste physicochemical reaction mechanism equation, and the solid waste resource utilization potential index is calculated. The pre-trained solid waste identification and classification network WRIM model is used to classify the solid waste samples. The tensor kernel similarity calculation function is used to quantify the degree of match between the multi-dimensional feature fingerprint of the solid waste and the preset resource utilization requirements to obtain a similarity scoring matrix. A multi-level priority evaluation engine is used to combine the similarity scoring matrix, the solid waste resource utilization potential index and the basic attribute categories to generate resource utilization plan recommendations.

2. The intelligent matching method for resource utilization of industrial solid waste according to claim 1 is characterized in that: The multimodal sensing technology refers to a comprehensive detection method that combines multiple detection methods such as near-infrared spectroscopy, X-ray fluorescence, thermogravimetric analysis, and electron microscopy to simultaneously obtain solid waste characteristic data; the feature extraction includes obtaining spectral characteristics, elemental composition characteristics, thermogravimetric characteristics, and morphological characteristics.

3. The intelligent matching method for resource utilization of industrial solid waste according to claim 2, characterized in that: The multi-dimensional characteristic fingerprint of solid waste refers to converting the physical and chemical property data of solid waste into a unique identifier in a high-dimensional space through mathematical modeling, which facilitates computer similarity comparison and classification.

4. The intelligent matching method for resource utilization of industrial solid waste according to claim 3 is characterized in that: The solid waste physicochemical reaction mechanism equation is used to calculate the reaction activity and conversion efficiency of solid waste in different resource utilization pathways. The input includes key element content, chemical bond energy distribution, surface active site density, crystal phase structure parameters and catalyst activity index, and the output is the solid waste resource utilization potential index.

5. The intelligent matching method for resource utilization of industrial solid waste according to claim 4 is characterized in that: The key element content refers to the mass percentage of elements in solid waste that have a decisive influence on resource utilization, including the content of major metal elements, non-metallic elements, rare earth elements, precious metal elements and harmful elements in solid waste, which is obtained through X-ray fluorescence analysis.

6. The intelligent matching method for resource utilization of industrial solid waste according to claim 5, characterized in that: The chemical bond energy distribution refers to the energy distribution of various chemical bonds in solid waste, which is obtained through spectral analysis; the surface active site density refers to the number of active centers per unit area on the surface of solid waste particles that can participate in chemical reactions.

7. The intelligent matching method for resource utilization of industrial solid waste according to claim 6, characterized in that: The crystal phase structure parameters refer to characteristic parameters such as the lattice constant, space group and crystal orientation of the crystalline phase in solid waste; the catalyst activity index refers to a dimensionless indicator that measures the degree to which the catalytically active components in solid waste promote resource recovery reactions.

8. The intelligent matching method for resource utilization of industrial solid waste according to claim 7, characterized in that: The solid waste resource utilization potential index refers to a comprehensive indicator for quantitatively evaluating the conversion value of solid waste in a certain resource utilization path.

9. The intelligent matching method for resource utilization of industrial solid waste according to claim 8, characterized in that: The specific structure of the pre-trained solid waste identification and classification network WRIM model is a hybrid architecture combining a multi-layer perceptron and a convolutional neural network, which includes a feature extraction layer, a feature fusion layer, a multi-head attention mechanism layer and a classification output layer.

10. The intelligent matching method for resource utilization of industrial solid waste according to claim 9, characterized in that: The feature extraction layer adopts a composite structure combining parallel and series connections, including a convolutional residual module, a sparse attention module, a dense connection module and an adaptive pooling module. The information flow is transmitted between the modules through jump connections.

Images