Solid waste resource process product quality online monitoring method

Abstract

The present application provides a kind of solid waste resource process product quality on-line monitoring method, belong to solid waste resource technology field, the present application is by constructing parameter acquisition system simultaneously collecting multiple parameter matrix, application sampling rate determination function realizes the dynamic adjustment of sampling frequency, utilize solid waste component migration equation analysis material conversion law, adopt spectrum fusion optimization function to process multiple spectral data, use FWM model to realize product quality multidimensional evaluation, combined with electrochemical sensor network monitoring heavy metal content, finally form comprehensive quality evaluation result, and through FWM model output updates sampling rate determination function weight coefficient, form on-line monitoring closed loop control, realizes the real-time, accurate, comprehensive monitoring of solid waste resource product quality, effectively solves the missing data problem of key process transition point caused by fixed frequency sampling, provides reliable technical support for solid waste resource process optimization.

Description

Technical Field

[0001] This invention belongs to the field of solid waste resource utilization technology, and more specifically, relates to an online monitoring method for product quality in the solid waste resource utilization process. Background Technology

[0002] Solid waste resource utilization is an important way to achieve solid waste reduction and resource recycling, and it is of great significance to promoting green development and ecological civilization. In the traditional solid waste resource utilization process, product quality monitoring mainly relies on periodic manual sampling and analysis. With the development of automated control technology, some systems have adopted online monitoring equipment based on fixed-frequency sampling, such as spectrometers and gas chromatographs, to achieve real-time acquisition of product quality parameters. However, most existing solid waste resource utilization product quality monitoring systems adopt a constant sampling frequency design, that is, data is collected according to a pre-set fixed time interval. This monitoring method can meet the needs when dealing with industrial solid waste with stable components and smooth reaction processes, but it shows serious shortcomings when dealing with solid waste with complex components and high heterogeneity, such as municipal solid waste and hazardous waste. Especially in the solid waste resource utilization process, due to factors such as fluctuations in raw material composition, changes in reaction conditions, and adjustments in process parameters, product quality often exhibits a dynamic trend. Fixed-frequency sampling strategies cannot flexibly adjust sampling intervals according to the dynamic characteristics of the processing. This leads to redundant sampling and wasted resources during the stable phase of the process. At critical transition points, such as sudden temperature changes, catalyst deactivation, or drastic changes in the composition of input materials, untimely sampling may result in missing important data, failing to accurately capture quality fluctuations and affecting the accuracy and timeliness of product quality assessment. In other words, existing technologies suffer from a problem where fixed sampling frequencies in solid waste resource recovery monitoring lead to data gaps at critical process transition points, thus affecting the accuracy of product quality monitoring. Summary of the Invention

[0003] In view of this, the present invention provides an online monitoring method for product quality in the solid waste resource utilization process, which can solve the technical problem in the prior art where the sampling frequency for monitoring product quality in the solid waste resource utilization process is fixed, which may lead to data loss at key process transition points.

[0004] This invention is implemented as follows: It provides an online monitoring method for product quality in the solid waste resource utilization process, comprising: constructing a parameter acquisition system to collect a multi-parameter matrix; applying a sampling rate determination function to calculate the real-time sampling rate, wherein the sampling rate determination function is jointly determined based on the variable components of the multi-parameter matrix, processing temperature, material flow rate, solid waste component heterogeneity index, and FWM model output results; using solid waste component migration equations to analyze the material transformation law; using a spectral fusion optimization function to process spectral data; using an FWM model to achieve product quality assessment; monitoring heavy metal content through an electrochemical sensor network; and updating the weight coefficients in the sampling rate determination function according to the FWM model output results, forming an online monitoring closed-loop control.

[0005] Specifically, the step of constructing the parameter acquisition system to acquire a multi-parameter matrix involves initializing the sampling rate to 5 times per minute, synchronously acquiring spectral data and physicochemical parameters in the solid waste treatment process, and forming a multi-parameter matrix.

[0006] The multi-parameter matrix refers to a data set that includes real-time monitoring parameters such as temperature, pressure, humidity, oxidation-reduction potential, pH value, and flow rate during the solid waste treatment process.

[0007] Specifically, the sampling rate determination function calculates weighting coefficients based on the rate of change and standard deviation of each parameter in the multi-parameter matrix. When the weighting coefficients exceed a preset threshold, the sampling frequency is increased; otherwise, the sampling frequency is decreased. The real-time sampling rate is dynamically adjusted in conjunction with the output results of the FWM model.

[0008] When the FWM model has not yet output results, the sampling rate determination function sets the weight coefficients of the FWM model output results to 0.

[0009] The solid waste component heterogeneity index is a quantitative indicator that characterizes the heterogeneity of solid waste, calculated by analyzing the fluctuation range, uneven distribution, and interaction strength between solid waste components.

[0010] The solid waste component migration equation specifically expresses the physicochemical behaviors of pollutants during the solid waste resource utilization process, such as phase transformation, adsorption and desorption, oxidation-reduction and complexation precipitation. The concentration distribution of each component in the final product is obtained by solving the solid waste component migration equation set.

[0011] The input parameters of the solid waste component migration equation include temperature field distribution, redox potential, reactant concentration, catalyst activity, and pH value in a multi-parameter matrix, and the output is the predicted value of the target component concentration.

[0012] The spectral fusion optimization function is used to perform noise reduction, baseline correction, and feature extraction on Raman spectroscopy, near-infrared spectroscopy, and X-ray fluorescence spectroscopy data, and achieves optimized fusion of different spectral information through weight allocation.

[0013] The input parameters of the spectral fusion optimization function include wavelength range, signal-to-noise ratio threshold, characteristic peak position, peak intensity ratio, and baseline offset, and the output is a spectral feature vector.

[0014] The FWM model features a multi-layer encoder-decoder architecture, comprising a self-attention mechanism layer, a residual connection layer, and a normalization layer. The encoder extracts spectral feature vectors, predicted target component concentrations, and deep features from the multi-parameter matrix. The decoder reconstructs product quality evaluation indicators, while the intermediate connection layer uses a cross-attention mechanism to facilitate data exchange between different modalities. The input parameters of the FWM model include the spectral feature vectors output by the spectral fusion optimization function, the multi-parameter matrix, and the predicted target component concentrations output by the solid waste component migration equation. The self-attention mechanism is a multi-head attention structure, with weight allocation determined by a dynamic scaling function that considers the current sampling rate, parameter change trends, data stability index, and spectral feature similarity. The electrochemical sensor network refers to a system that uses selective electrodes at key nodes in the processing flow to measure changes in redox current for real-time monitoring of heavy metal ion concentrations. The comprehensive quality evaluation result is achieved by standardizing the heavy metal ion content data measured by the electrochemical sensor network and the spectral data processed by the spectral fusion optimization function, establishing a hierarchical weighted model to calculate the comprehensive product quality index, thus realizing dynamic evaluation of the quality of resource-based products.

[0015] This invention constructs a sampling rate determination function to achieve real-time adjustment of the sampling frequency based on parameters such as the changing components of a multi-parameter matrix, processing temperature, material flow rate, and the heterogeneity index of solid waste components. The sampling frequency is appropriately reduced during stable process phases and automatically increased at parameter fluctuations or process transition points. Combining solid waste component migration equations and spectral fusion optimization technology, this invention establishes a Free-Waste Model (FWM) as the core algorithm. This model, through a multi-layer encoder-decoder architecture and a self-attention mechanism, can deeply explore the intrinsic correlations between multiple parameters, achieving effective processing and fusion of data at different sampling frequencies. In particular, the dynamic scaling function and cross-attention mechanism enable the model to automatically adjust attention weights based on data stability and parameter change trends, ensuring the acquisition of key information even during low-frequency sampling phases. Through intelligent dynamic adjustment of the sampling rate and deep data fusion of the FWM model, this invention effectively solves the problem of missing data at key process transition points caused by fixed-frequency sampling, achieving accurate monitoring of the quality of solid waste resource recovery products throughout the entire process. This improves monitoring efficiency and reduces system resource consumption, providing reliable technical support for quality control and process optimization in solid waste resource recovery. Attached Figure Description

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

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0018] like Figure 1 The diagram shown is a flowchart of an online monitoring method for product quality in the solid waste resource utilization process provided by the present invention. This method includes the following steps:

[0019] S01. Construct a parameter acquisition system with an initial sampling rate of 5 times per minute. Simultaneously collect spectral data and physicochemical parameters in the solid waste treatment process to form a multi-parameter matrix.

[0020] S02. The sampling rate determination function is used to calculate the real-time sampling rate. The sampling rate determination function is determined based on the multi-parameter matrix variation components, processing temperature, material flow rate, solid waste component heterogeneity index and FWM model output results. It includes the initial sampling rate and the real-time variation sampling rate. When the FWM model has not yet output results, the sampling rate determination function sets the weight coefficient of the FWM model output results to 0.

[0021] S03. Analyze the transformation law of substances in the resource utilization process through the solid waste component migration equation, and predict the content of key components in the product. The input parameters of the solid waste component migration equation include the temperature field distribution, redox potential, reactant concentration, catalyst activity and pH value in the multi-parameter matrix.

[0022] S04. Raman spectroscopy, near-infrared spectroscopy and X-ray fluorescence spectroscopy data are processed using a spectral fusion optimization function. The input parameters of the spectral fusion optimization function include wavelength range, signal-to-noise ratio threshold, characteristic peak position, peak intensity ratio and baseline offset.

[0023] S05. Using the FWM model to achieve multidimensional evaluation of product quality, the input parameters of the FWM model include the spectral feature vector output by the spectral fusion optimization function, the multi-parameter matrix, and the target component concentration prediction value output by the solid waste component migration equation, wherein the dimension of the attention mechanism parameter matrix is ​​jointly determined by the dimension of the multi-parameter matrix, the sampling rate, and the solid waste component heterogeneity index.

[0024] S06. Real-time monitoring of heavy metal ion content in solid waste resource products based on electrochemical sensor network, and fusion with the spectral data processed by the spectral fusion optimization function to form a comprehensive quality evaluation result;

[0025] S07. Update the weight coefficients in the sampling rate determination function according to the output of the FWM model, return to step S02 to recalculate the real-time sampling rate, and form an online monitoring closed-loop control.

[0026] The multi-parameter matrix refers to a data set that includes real-time monitoring parameters such as temperature, pressure, humidity, oxidation-reduction potential, pH value, and flow rate during the solid waste treatment process.

[0027] Specifically, the sampling rate determination function calculates weight coefficients based on the rate of change and standard deviation of each parameter in the multi-parameter matrix. When the weight coefficients exceed a preset threshold, the sampling frequency is increased; otherwise, the sampling frequency is decreased. The real-time sampling rate is dynamically adjusted in conjunction with the output results of the FWM model.

[0028] The solid waste component heterogeneity index is a quantitative indicator that characterizes the heterogeneity of solid waste, calculated by analyzing the fluctuation range, uneven distribution, and interaction strength between solid waste components.

[0029] The solid waste component migration equation specifically expresses the physicochemical behaviors of pollutants in the solid waste resource utilization process, such as phase transformation, adsorption and desorption, redox and complexation precipitation. By solving the solid waste component migration equation set, the concentration distribution of each component in the final product is obtained, providing a theoretical basis for quality assessment. The input of the solid waste component migration equation includes reactant concentration, temperature field distribution, redox potential, catalyst activity and pH value in the multi-parameter matrix. The output is the predicted concentration value of the target component. The predicted concentration value of the target component is used as one of the input parameters of the FWM model and a component of the comprehensive quality evaluation result.

[0030] The spectral fusion optimization function is used to perform noise reduction, baseline correction, and feature extraction on the raw data obtained from various spectral analysis techniques. It achieves optimized fusion of different spectral information through weight allocation, thereby improving detection accuracy. The input of the spectral fusion optimization function includes the wavelength range, signal-to-noise ratio threshold, characteristic peak position, peak intensity ratio, and baseline offset of each spectrum. The output is the spectral feature vector, which serves as one of the input parameters of the FWM model.

[0031] The FWM model has a multi-layer encoder-decoder architecture, including a self-attention mechanism layer, a residual connection layer, and a normalization layer. The encoder is responsible for extracting the spectral feature vector, the predicted concentration of the target component, and the deep features in the multi-parameter matrix. The decoder is responsible for reconstructing the product quality evaluation index. The intermediate connection part uses a cross-attention mechanism to realize the information interaction of data from different modalities. The output of the FWM model is used as one of the input parameters of the sampling rate determination function.

[0032] The steps for establishing the training dataset of the FWM model specifically include collecting historical monitoring data on the resource utilization process of solid waste from different sources, labeling the corresponding product quality indicators for each set of data, using data augmentation technology to simulate parameter fluctuations under different working conditions, constructing standard data pairs containing input parameter matrices and corresponding quality evaluation results, and forming training sets, validation sets and test sets through random partitioning.

[0033] The specific steps of training the FWM model include first performing unsupervised training on large-scale historical solid waste treatment data to learn the parameter distribution pattern and the inherent structure of the data; then performing supervised fine-tuning training for solid waste types; optimizing model parameters using adaptive learning rate and weight reduction strategies; evaluating the model's generalization ability through cross-validation; and finally selecting the model parameters with the best performance on the validation set.

[0034] The electrochemical sensor network refers to a system that monitors the concentration of heavy metal ions in real time by arranging selective electrodes at key nodes in the processing flow and measuring changes in redox current. The measurement results of the electrochemical sensor network are part of the comprehensive quality evaluation results.

[0035] The comprehensive quality evaluation result is achieved by standardizing the spectral data after processing the heavy metal ion content data measured by the electrochemical sensor network and the spectral data after processing by the spectral fusion optimization function, and then establishing a hierarchical weight model to calculate the comprehensive product quality index, thereby realizing the dynamic evaluation of the quality of resource-based products. The comprehensive product quality index and the predicted concentration value of the target component together constitute the comprehensive quality evaluation result.

[0036] The self-attention mechanism is a multi-head attention structure, and its weight allocation is determined by a dynamic scaling function. The dynamic scaling function takes into account the current sampling rate, parameter change trend, data stability index, and spectral feature similarity. The final attention weight distribution is obtained by calculating the dot product of the query matrix, key matrix, and value matrix and performing softmax normalization.

[0037] The specific implementation methods of the above steps are described in detail below. Step S01 involves constructing a parameter acquisition system that synchronously acquires various spectral data and physicochemical parameters from the solid waste treatment process at an initial frequency of 5 times per minute. Specifically, spectral probes and physicochemical parameter sensors, including Raman spectroscopy probes, near-infrared spectroscopy probes, X-ray fluorescence spectroscopy probes, as well as temperature sensors, pressure sensors, humidity sensors, redox potential sensors, pH sensors, and flow rate sensors, are first installed at key nodes of the solid waste treatment production line. These sensors include Raman spectroscopy probes, near-infrared spectroscopy probes, X-ray fluorescence spectroscopy probes, as well as temperature sensors, pressure sensors, humidity sensors, redox potential sensors, pH sensors, and flow rate sensors. Then, an industrial-grade data acquisition network is established, employing time-division multiplexing technology to ensure synchronous data acquisition from all sensors, while utilizing industrial-grade fieldbus technology for real-time data transmission. Next, a real-time database is used to store the acquired information, forming a multi-parameter matrix containing parameters such as temperature, pressure, humidity, redox potential, pH, and flow rate. Finally, a data preprocessing unit filters and denoises the raw data, eliminating electromagnetic interference and outliers, achieving data standardization. The purpose of this step is to ensure that the monitoring system can capture changes in key parameters during the solid waste treatment process in a timely and accurate manner, providing basic data support for subsequent analysis.

[0038] The specific implementation of step S02 involves calculating the real-time sampling rate using a sampling rate determination function. This function is determined based on the variation components of a multi-parameter matrix, processing temperature, material flow rate, solid waste component heterogeneity index, and the output results of the FWM model. Specifically, firstly, the rate of change and standard deviation of each parameter in the multi-parameter matrix within a specified time window are calculated. When the parameter's rate of change exceeds 12% or the standard deviation is greater than a set threshold (temperature parameter threshold: ±3℃, pH parameter threshold: ±0.4, redox potential parameter threshold: ±25mV), the weighting coefficient of that parameter is increased by 0.2; otherwise, it is decreased by 0.1. Next, the temperature influence factor is calculated based on the deviation between the processing temperature and the theoretical optimal reaction temperature. When the temperature deviation exceeds ±15℃, the sampling rate is appropriately increased. Then, the flow rate influence factor is determined based on the ratio of the material flow rate to the standard flow rate. When the flow rate exceeds 120% of the standard flow rate or is lower than 80%, the sampling rate is increased by at least 30%. The basic sampling rate is then adjusted based on the solid waste component heterogeneity index. When the heterogeneity index is greater than 0.75, the sampling rate is increased to at least 8 times per minute. Finally, considering the quality fluctuation prediction values ​​output by the FWM model, the sampling rate is further increased when the prediction values ​​indicate an increased risk of quality fluctuation. The final real-time sampling rate is calculated by multiplying the initial sampling rate by the weighted sum of all influencing factors. When the FWM model has not yet output results, its weight coefficient is set to 0. The purpose of this step is to achieve intelligent allocation of monitoring resources, increasing the monitoring frequency to ensure monitoring accuracy when there are large process fluctuations or increased product quality risks, and reducing the monitoring frequency to save system resources when the process is stable.

[0039] The specific implementation of step S03 involves analyzing the transformation patterns of substances during resource recovery through solid waste component migration equations to predict the content of key components in the product. Specifically, a set of migration equations is first constructed to describe the physicochemical behaviors of pollutants during solid waste resource recovery, including phase transitions, adsorption / desorption, redox reactions, and complexation / precipitation. This set of equations covers three aspects: mass conservation, momentum conservation, and energy conservation. Next, using the temperature field distribution, redox potential, reactant concentration, catalyst activity, and pH value from the multi-parameter matrix as input parameters, the solid waste component migration equations are discretized and solved using the finite element method. Then, an adaptive step-size Runge-Kutta integral algorithm is used for numerical solution to calculate the concentration distribution of pollutants under different operating conditions. Subsequently, combining the thermodynamic equilibrium constant and reaction kinetic parameters, the speciation of heavy metal elements under different pH and redox conditions is predicted. Finally, the predicted concentration values ​​of the target components are output as one of the input parameters of the FWM model and a component of the comprehensive quality evaluation results. The predicted concentrations of typical metal ions such as lead, cadmium, mercury, and arsenic are controlled within ±7% error, while the predicted errors for organic pollutants are controlled within ±12%. The purpose of this step is to accurately predict the transformation behavior of each component during solid waste resource recovery and the concentration distribution in the final product using physicochemical theoretical models, providing a theoretical basis for quality assessment.

[0040] The specific implementation of step S04 involves processing Raman, near-infrared, and X-ray fluorescence spectral data using a spectral fusion optimization function. Specifically, the spectral data are first preprocessed, including denoising, baseline correction, and normalization. For Raman spectroscopy, wavelet transform is used for denoising, with a signal-to-noise ratio threshold set to 4.5. For near-infrared spectroscopy, standard normal transformation is used for baseline drift correction. For X-ray fluorescence spectroscopy, Compton scattering peak normalization is employed. Next, feature extraction is performed, focusing on the 400–1800 cm⁻¹ region for Raman spectroscopy. -1 Characteristic peaks within the wavenumber range were selected. For near-infrared spectroscopy, absorption peaks in the wavelength range of 1100–2500 nm were chosen, while for X-ray fluorescence spectroscopy, the characteristic peak positions and intensity ratios of typical elements were considered. Principal component analysis (PCA) was then used for dimensionality reduction, retaining principal components that explained up to 95% of the variance. A multispectral fusion model was then constructed, employing a feature-level fusion algorithm based on a convolutional neural network to optimize the fusion of different spectral information. Finally, the spectral feature vector was output as one of the input parameters of the FWM model. The inputs to this spectral fusion optimization function included the wavelength range of each spectrum, the signal-to-noise ratio threshold, the characteristic peak position, the peak intensity ratio, and the baseline offset. The purpose of this step is to fully utilize the advantages of different spectral techniques, extract complementary information, improve detection accuracy, and provide multidimensional spectral features for product quality assessment.

[0041] The specific implementation of step S05 involves using the FWM model to achieve multidimensional product quality assessment. Specifically, the spectral feature vector output from the spectral fusion optimization function, the multi-parameter matrix, and the predicted target component concentration output from the solid waste component migration equation are first used as input parameters to the FWM model. Next, the encoder part of the FWM model employs a multilayer perceptron and a self-attention mechanism to extract deep feature representations of the input data. The self-attention mechanism uses a multi-head attention structure with 8 heads and a hidden layer dimension of 512. Then, in the intermediate connection part, a cross-attention mechanism is used to achieve information interaction between different modalities. The attention weights are determined by a dynamic scaling function that comprehensively considers the current sampling rate, parameter change trends, data stability index, and spectral feature similarity. Afterwards, the decoder part of the FWM model reconstructs product quality evaluation indicators using residual connections and layer normalization techniques. The decoder has 6 layers, each containing a self-attention sublayer, a cross-attention sublayer, and a feedforward neural network sublayer. Finally, the multidimensional quality assessment results of the product are output, including purity, safety, stability, and applicability indicators. The purpose of this step is to integrate multi-source heterogeneous data through deep learning models to achieve a comprehensive and accurate assessment of the quality of solid waste resource recovery products, and to provide a basis for decision-making in the production process.

[0042] The specific implementation of step S06 is based on real-time monitoring of heavy metal ion content in solid waste resource products using an electrochemical sensor network. This data is then fused with the spectral data processed by a spectral fusion optimization function to form a comprehensive quality evaluation result. Specifically, a selective electrode array, including ion-selective electrodes, modified electrodes, and reference electrodes, is first deployed at key nodes in the processing flow to form an electrochemical sensor network. Next, the redox current changes of each electrode are measured, and the concentration of heavy metal ions is quantitatively analyzed using voltammetry and potentiometry, focusing primarily on toxic heavy metals such as lead, cadmium, mercury, arsenic, and chromium, with detection limits reaching the 1 μg / L level. Then, a multivariate correction algorithm is used to eliminate cross-interference between ions, improving measurement accuracy; the measurement accuracy of typical heavy metal ions is controlled within ±5%. Afterward, the heavy metal ion content data measured by the electrochemical sensor network and the spectral data processed by the spectral fusion optimization function are standardized to eliminate dimensional differences. Finally, a hierarchical weighted model is established to calculate the comprehensive product quality index. This model assigns weights to each indicator based on its importance in different application scenarios, forming a comprehensive quality evaluation result. The purpose of this step is to achieve accurate monitoring of heavy metal pollutants in products by combining electrochemical and spectroscopic methods, thereby ensuring the environmental safety of resource-based products.

[0043] The specific implementation of step S07 involves updating the weight coefficients in the sampling rate determination function based on the output of the FWM model, returning to step S02 to recalculate the real-time sampling rate, and forming an online monitoring closed-loop control. Specifically, the product quality assessment results output by the FWM model are first analyzed, and the deviations of each quality indicator from the target value are calculated. Then, based on the magnitude and trend of the deviations, the weight coefficients of each parameter in the sampling rate determination function are adjusted. When the deviation of the quality indicator increases, the weight coefficients of the relevant parameters are increased; when the quality indicator stabilizes within the target range, the weight coefficients are appropriately decreased. Then, the adjustment step size of the weight coefficients is set, generally between 0.05 and 0.15, dynamically adjusted according to the quality fluctuation range. Afterwards, the updated weight coefficients are applied to recalculate the real-time sampling rate, and the new sampling rate is applied to the parameter acquisition system. Finally, a closed-loop feedback mechanism for quality monitoring and sampling control is formed, realizing the intelligent allocation and optimized utilization of monitoring resources. The purpose of this step is to enable the monitoring system to adaptively adjust the monitoring strategy according to the product quality status through a closed-loop control strategy, thereby improving monitoring efficiency and ensuring monitoring quality.

[0044] The detailed structure of the FWM model adopts a multi-layer encoder-decoder architecture, including a self-attention mechanism layer, a residual connection layer, and a normalization layer. The specific structural design is as follows: The encoder consists of six stacked encoder layers, each containing a multi-head self-attention sub-layer and a feedforward neural network sub-layer. Each sub-layer employs residual connections and layer normalization techniques. The multi-head self-attention mechanism has 8 heads and a hidden layer dimension of 512. The feedforward neural network uses a two-layer fully connected network with an intermediate layer dimension of 2048 and uses the GELU activation function. The decoder also consists of six stacked decoder layers, each containing a self-attention sub-layer, a cross-attention sub-layer, and a feedforward neural network sub-layer. The cross-attention sub-layer is responsible for associating the encoder output with the decoder's current state, enabling information interaction between different modalities. The intermediate connection layer uses a cross-attention mechanism, calculating the dot product of the query matrix, key matrix, and value matrix and performing softmax normalization to obtain the attention weight distribution, achieving effective fusion of different modalities. The model output layer is a fully connected layer that maps the decoder output to the product quality evaluation index space. The self-attention mechanism adopts a multi-head attention structure, and its weight allocation is determined by a dynamic scaling function. This function takes into account the current sampling rate, parameter change trend, data stability index and spectral feature similarity. The final attention weight distribution is obtained by calculating the dot product of the query matrix, key matrix and value matrix and performing softmax normalization.

[0045] The detailed steps for establishing the training dataset for the FWM model are as follows: First, historical monitoring data on the resource utilization process of solid waste from different sources are collected, including treatment data for different types of solid waste such as municipal solid waste, industrial solid waste, and agricultural waste, with the data covering at least 100 complete production cycles. Next, each set of data is cleaned and labeled. The cleaning process includes outlier detection and handling, missing value imputation, and data consistency verification. The labeling process includes labeling corresponding product quality indicators based on laboratory analysis results, such as purity, hazardous substance content, and stability. Then, data augmentation techniques are used to simulate parameter fluctuations under different operating conditions, including adding Gaussian noise to simulate sensor errors, random scaling to simulate parameter fluctuations, and time shifting to simulate process delays, expanding the dataset size to 3-5 times the original data. Afterward, standard data pairs containing input parameter matrices and corresponding quality evaluation results are constructed. The dimension of the input parameter matrix is ​​time step × number of parameters, and the dimension of the quality evaluation results is the number of indicators. Finally, the dataset is divided into training, validation, and test sets using stratified random sampling, with proportions of 70%, 15%, and 15%, respectively, ensuring that the distribution of different types of solid waste data in each subset remains consistent. The FWM model training process first involves unsupervised pre-training on large-scale historical solid waste treatment data, employing a masked autoencoder structure to learn the parameter distribution patterns and the inherent structure of the data. Then, supervised fine-tuning training is performed specifically for the current type of solid waste, with the loss function comprehensively considering mean squared error and weighted cross-entropy. An adaptive learning rate strategy is used during training, with an initial learning rate set to 1×10⁻⁶. -4 When the validation set loss shows no improvement for five consecutive epochs, the learning rate is reduced to 0.8 times its original value; simultaneously, a weight-decreasing regularization technique is introduced, with a decrease coefficient set to 5 × 10⁻⁶. -5 To prevent overfitting, the model's generalization ability is evaluated through 5-fold cross-validation, and the model parameters with the best performance on the validation set are finally selected as the final model.

[0046] It should be noted that the sampling rate determination function incorporates the FWM model output as a parameter, enabling dynamic adjustment of the sampling strategy based on product quality assessment results, thus achieving a two-way feedback mechanism between data collection and quality assessment. When the system detects fluctuations in product quality, the FWM model outputs a higher coefficient of variation, and the sampling rate determination function correspondingly increases the sampling frequency to ensure sufficiently dense data acquisition during critical quality transition periods. This adaptive sampling mechanism based on quality assessment results overcomes the limitations of traditional fixed-interval sampling, enabling the monitoring system to possess quality-oriented intelligent decision-making capabilities.

[0047] Setting the weight coefficients of the FWM model to 0 before it outputs results ensures the system's rationality and smooth transition during the startup phase. In the initial monitoring stage, the system relies on fundamental parameters such as the varying components of the multi-parameter matrix and processing temperature to determine the sampling frequency, preventing the sampling strategy from failing due to a lack of model output. As the FWM model begins to generate evaluation results, the system gradually increases its weight influence, achieving a smooth transition from purely parameter-driven to quality assessment-driven approaches. This progressive weight adjustment mechanism ensures the stability and continuity of system operation, adapting to the parameter learning and model optimization stages in the solid waste resource recovery process.

[0048] The multi-head attention mechanism employs a dynamic scaling function to determine weight allocation, enabling the FWM model to automatically adjust the attention given to different features based on real-time conditions. The dynamic scaling function comprehensively considers the current sampling rate, parameter change trends, data stability index, and spectral feature similarity. At high sampling rates, it emphasizes short-term change patterns, while at low sampling rates, it strengthens the grasp of long-term trends. This adaptive attention allocation mechanism solves the problem of uneven information density under different sampling frequencies, improving the model's ability to handle sparsely sampled data.

[0049] The three key technologies described above form a complete closed-loop adaptive system. The sampling rate determination function receives the evaluation results from the FWM model and adjusts the sampling strategy, while the FWM model adapts to the data characteristics of different sampling frequencies through a dynamically scaled multi-head attention mechanism. The two technologies mutually promote and optimize each other. In the initial stage of system operation, a zero-weight coefficient mechanism ensures a smooth start-up. As the operating time increases, the system gradually establishes a closed-loop control of sampling-analysis-evaluation-feedback, optimizing system resource utilization while ensuring monitoring quality. This collaborative mechanism enables the monitoring system to automatically evolve according to the dynamic characteristics of the solid waste treatment process, forming an optimal monitoring strategy adapted to the current process, and solving the problem that traditional fixed-frequency sampling cannot cope with process fluctuations.

[0050] Specifically, the principle of this invention is as follows: The core technical principle of this invention lies in the organic combination of an adaptive sampling frequency adjustment mechanism and deep fusion of multimodal data. First, the sampling rate determination function is the key to achieving adaptive sampling. This function determines the real-time sampling rate by calculating the weighting coefficients of the rate of change and standard deviation of each parameter in the multi-parameter matrix in real time, and comprehensively considering factors such as processing temperature, material flow rate, and solid waste component heterogeneity index. When the parameter rate of change is detected to exceed a preset threshold, the system automatically increases the sampling frequency to ensure that sufficiently dense data is obtained at process transition points; while during the parameter stabilization phase, the sampling frequency is appropriately reduced to save computing resources and storage space.

[0051] Secondly, the solid waste component migration equation provides a theoretical basis for adaptive sampling. This equation describes the physicochemical behaviors of pollutants during solid waste resource utilization, including phase transitions, adsorption / desorption, redox reactions, and complexation / precipitation. By analyzing the influence of parameters such as temperature field distribution, redox potential, and reactant concentration on component migration, the potential impact of process parameter changes on product quality can be predicted, providing a scientific basis for adjusting the sampling frequency. When the migration equation predicts a significant change in component concentration, the system will increase the sampling density in advance to prevent data loss.

[0052] The spectral fusion optimization function achieves optimized integration of different spectral information by performing noise reduction, baseline correction, and feature extraction on Raman, near-infrared, and X-ray fluorescence spectral data. This multispectral fusion method can improve the information density of a single sample by leveraging spectral complementarity while reducing the sampling frequency, thereby reducing the overall number of samplings required without compromising monitoring quality.

[0053] The FWM model, the core of this invention's algorithm, employs a multi-layer encoder-decoder architecture and utilizes a self-attention mechanism to process data acquired at different sampling frequencies. The model's unique feature lies in its self-attention mechanism, which uses a dynamic scaling function to determine weight allocation, considering factors such as the current sampling rate, parameter trends, and data stability index. This allows the model to automatically adjust its attention allocation strategy based on changes in data acquisition frequency. In low-frequency sampling phases, the model focuses more on historical data and trends; while in high-frequency sampling phases, it emphasizes capturing instantaneous correlations between parameters.

[0054] Finally, the weight coefficients in the sampling rate determination function are updated using the FWM model output, forming a complete closed-loop control system. This feedback mechanism enables the sampling strategy to continuously optimize itself as the monitoring system operates, gradually forming a sampling mode best suited to the current solid waste treatment process, fundamentally solving the problem that fixed-frequency sampling cannot cope with dynamic changes in the process.

[0055] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0056] The specific implementation of step S01 involves constructing a parameter acquisition system with an initial sampling rate of 5 times per minute. This system synchronously acquires spectral data and physicochemical parameters from the solid waste treatment process, forming a multi-parameter matrix. First, spectral probes and physicochemical parameter sensors, including Raman spectroscopy probes, near-infrared spectroscopy probes, X-ray fluorescence spectroscopy probes, as well as temperature sensors, pressure sensors, humidity sensors, redox potential sensors, pH sensors, and flow rate sensors, are installed at key nodes of the solid waste treatment production line. Then, an industrial-grade data acquisition network is established, employing time-division multiplexing technology to ensure synchronous data acquisition from each sensor. Simultaneously, industrial-grade fieldbus technology is used to achieve real-time data transmission, with the acquisition frequency initialized to 5 times per minute. Next, a real-time database is used to store the acquired information, forming a multi-parameter matrix M, which can be represented as:

[0057]

[0058] In the formula, T ij P represents the temperature value at the j-th time step of the i-th sampling point, in °C; ij H represents the pressure value at the j-th time step of the i-th sampling point, in kPa; ij E represents the humidity value at the j-th time step of the i-th sampling point, in percentage (%). ij The pH value represents the redox potential at the i-th sampling point and the j-th time step, in mV. ij V represents the pH value at the j-th time step of the i-th sampling point; ij S represents the flow velocity value at the j-th time step of the i-th sampling point, in m / s; ij，k Let represent the k-th spectral value at the j-th time step of the i-th sampling point; m represents the number of sampling points; and n represents the dimension of the spectral data. Finally, the raw data is filtered and denoised by a data preprocessing unit to eliminate electromagnetic interference and outliers, thus achieving data standardization. A wavelet transform denoising algorithm is used to preprocess the spectral data. The expression for wavelet transform denoising is:

[0059]

[0060] In the formula, The denoised spectral data is represented by W; the wavelet transform operator is represented by w. -1 Denotes the inverse wavelet transform operator; δ λ The threshold function is represented by λ, which is set to 4.5 times the noise standard deviation. The noise standard deviation is estimated using the first-order difference method. The purpose of this step is to ensure that the monitoring system can capture changes in key parameters during the solid waste treatment process in a timely and accurate manner, providing basic data support for subsequent analysis.

[0061] The specific implementation of step S02 involves calculating the real-time sampling rate using a sampling rate determination function. This function is determined based on the variation components of the multi-parameter matrix, processing temperature, material flow rate, solid waste component heterogeneity index, and the output results of the FWM model. The specific expression for the sampling rate determination function is as follows:

[0062]

[0063] In the formula, R real R represents the real-time sampling rate, expressed in times per minute. init This represents the initial sampling rate, with a value of 5 samples per minute; w i Let w represent the weight coefficient of the i-th influence factor. The initial values ​​are empirically set as w1 = 0.3, w2 = 0.25, w3 = 0.2, w4 = 0.15, and w5 = 0.1, and satisfy the following condition: f i Let f represent the i-th influencing factor, including parameter variation factor f1, temperature deviation factor f2, flow rate deviation factor f3, component heterogeneity factor f4, and FWM model output factor f5. The calculation methods for each influencing factor are as follows:

[0064] The formula for calculating the parameter variation factor f1 is:

[0065]

[0066] In the formula, ΔP represents the rate of change vector of each parameter in the multi-parameter matrix within a specified time window; σ P σ represents the standard deviation vector of each parameter in a multi-parameter matrix within a specified time window; thresh This represents the standard deviation threshold vector, corresponding to a temperature threshold of 3℃, a pH threshold of 0.4, and a redox potential threshold of 25mV. The formula for calculating the temperature deviation factor f2 is:

[0067]

[0068] In the formula, T represents the current processing temperature, in °C; T opt This represents the theoretical optimal reaction temperature, expressed in °C. The temperature is determined based on different solid waste types and treatment processes, with typical values ​​ranging from 400 to 800 °C. The formula for calculating the flow rate deviation factor f3 is:

[0069]

[0070] In the formula, V represents the current material flow rate, in m / s; V std This represents the standard flow rate, expressed in m / s, and is determined according to process requirements; typical values ​​range from 0.5 to 2 m / s. The formula for calculating the component heterogeneity factor f4 is:

[0071]

[0072] In the formula, I het The solid waste component heterogeneity index, ranging from 0 to 1, is calculated by analyzing the fluctuation range, degree of non-uniform distribution, and strength of interactions between solid waste components.

[0073]

[0074] In the formula, C max，i C represents the maximum concentration of the i-th component; min，i This represents the minimum concentration of the i-th component; Indicates average concentration; σ C Indicates the standard deviation of concentration; r ij Let f represent the correlation coefficient between the i-th component and the j-th component; m represents the number of components. The formula for calculating the FWM model output factor f5 is:

[0075]

[0076] In the formula, Q risk This represents the quality fluctuation risk value output by the FWM model, ranging from 0 to 1. The purpose of this step is to achieve intelligent allocation of monitoring resources, increasing the monitoring frequency to ensure monitoring accuracy when there are large process fluctuations or increased product quality risks, and reducing the monitoring frequency to save system resources when the process is stable.

[0077] The specific implementation of step S03 involves analyzing the transformation patterns of substances during resource recovery through solid waste component migration equations to predict the content of key components in the product. The solid waste component migration equations are constructed based on three aspects: mass conservation, momentum conservation, and energy conservation, as specifically expressed below:

[0078] mass conservation equation:

[0079]

[0080] In the formula, C i The concentration of the i-th component is expressed in mg / L or mg / kg; t represents time in seconds. This represents a velocity vector, with units of m / s; D represents the gradient operator; i This represents the diffusion coefficient of the i-th component, expressed in m. 2 / s;r ij This represents the rate constant for the conversion of the j-th component into the i-th component, expressed in seconds. -1 S i Let represent the source and sink terms of the i-th component, in units of mg / (L·s) or mg / (kg·s). Momentum conservation equation:

[0081]

[0082] In the formula, ρ represents density, with units of kg / m³. 3 p represents pressure, in Pa; τ represents stress tensor. This represents the vector of gravitational acceleration, with units of m / s². 2 Energy conservation equation:

[0083]

[0084] In the formula, c p represents specific heat capacity, in J / (kg·K); T represents temperature, in K; k represents thermal conductivity, in W / (m·K); r j The reaction rate of the j-th reaction is expressed in mol / (L·s) or mol / (kg·s); ΔH j The heat of reaction for the j-th reaction is expressed in J / mol. The reaction rate r... j Described by the modified Arrhenius equation:

[0085]

[0086] In the formula, A j E represents the pre-exponential factor of the j-th reaction, in units related to the reaction order; j α represents the activation energy of the j-th reaction, in J / mol; R represents the gas constant, with a value of 8.314 J / (mol·K); α ij f(pH, E) represents the reaction order of the i-th component in the j-th reaction; h The expression for the effect of pH and redox potential on the reaction rate is:

[0087]

[0088] In the formula, k1 and k2 represent the influence coefficients of pH value and redox potential, respectively, and their typical values ​​are 0.5 and 0.01, determined experimentally. opt This indicates the optimal pH value, determined according to different reaction types, with typical values ​​ranging from 5 to 9; E h，optThe optimal redox potential, expressed in mV, is determined based on different reaction types, with typical values ​​ranging from -400 to 400 mV. The finite element method (FEM) is used to discretize and solve the solid waste component migration equations. Spatial discretization employs the Galerkin weighted residual method, while temporal discretization uses an adaptive step-size Runge-Kutta integral algorithm to obtain predicted concentrations of target components. By substituting the temperature field distribution, redox potential, reactant concentration, catalyst activity, and pH value from the multi-parameter matrix into the solution, the speciation of heavy metals under different pH and redox conditions is predicted. The predicted concentration errors for typical metal ions such as lead, cadmium, mercury, and arsenic are controlled within ±7%, and the prediction errors for organic pollutants are controlled within ±12%. The purpose of this step is to accurately predict the transformation behavior of each component during solid waste resource recovery and the concentration distribution in the final product using a physicochemical theoretical model, providing a theoretical basis for quality assessment.

[0089] The specific implementation of step S04 involves processing Raman, near-infrared, and X-ray fluorescence spectral data using a spectral fusion optimization function. The spectral fusion optimization function first preprocesses each spectral data, including denoising, baseline correction, and normalization. Wavelet transform is used for denoising Raman spectra, standard normal transformation is used for baseline drift correction in near-infrared spectra, and Compton scattering peak normalization is used in X-ray fluorescence spectra. Then, feature extraction is performed, focusing on the 400–1800 cm⁻¹ region for Raman spectra. -1 Characteristic peaks within the wavenumber range are selected for near-infrared spectroscopy, focusing on absorption peaks in the 1100–2500 nm wavelength range, while X-ray fluorescence spectroscopy focuses on the position and intensity ratio of characteristic peaks of typical elements. The mathematical expression for the spectral fusion optimization function is:

[0090] F(S) = W R ·F R (S R )+W N ·F N (S N )+W X ·F X (S X );

[0091] In the formula, F(S) represents the fused spectral eigenvector; S R S N and S X F represents the raw data for Raman spectroscopy, near-infrared spectroscopy, and X-ray fluorescence spectroscopy, respectively; R F N and F X W represents the feature extraction functions for Raman spectroscopy, near-infrared spectroscopy, and X-ray fluorescence spectroscopy, respectively; R W N and W XLet W represent the weighting coefficients for Raman spectroscopy, near-infrared spectroscopy, and X-ray fluorescence spectroscopy, respectively, satisfying W R +W N +W X =1, initial value set to W R =0.4, W N =0.3, W X =0.3, and will be dynamically adjusted based on feature relevance. The feature extraction function uses principal component analysis, and its expression is:

[0092]

[0093] In the formula, i represents the spectral type index, which takes the value of R, N, or X; P i Represents the principal component loading matrix; μ i This represents the mean vector of the spectral data. The dynamic adjustment method for the weighting coefficients is as follows:

[0094]

[0095] In the formula, r i This represents the correlation coefficient between the i-th spectral feature and the concentration of the target component; β represents the adjustment parameter, with a value of 5. The purpose of this step is to fully utilize the advantages of different spectroscopic techniques, extract complementary information, improve detection accuracy, and provide multidimensional spectral features for product quality assessment.

[0096] The specific implementation of step S05 involves using the FWM model to achieve multi-dimensional product quality assessment. The FWM model employs a multi-layer encoder-decoder architecture, including a self-attention mechanism layer, a residual connection layer, and a normalization layer. The model's input parameters include the spectral feature vector output by the spectral fusion optimization function, the multi-parameter matrix, and the predicted target component concentration output by the solid waste component migration equation. The self-attention mechanism uses a multi-head attention structure, and its weight allocation is determined by a dynamic scaling function, the expression of which is:

[0097]

[0098] In the formula, α ij The s represents the attention weight of the i-th query vector on the j-th key vector; ij The similarity score is calculated using the following formula:

[0099]

[0100] In the formula, Q i K represents the i-th query vector; j d represents the j-th key vector; k Represents the dimension of the key vector; g(R) real ,ΔP,I stab Ssim () represents the dynamic scaling function, considering the current sampling rate R. real Parameter variation trend ΔP, data stability index I stab Spectral feature similarity S sim The expression is:

[0101] g(R real ,ΔP,I stab S sim )=a·log(1+R real )+b·‖ΔP‖2+c·(1-I stab )+d·S sim ;

[0102] In the formula, a, b, c, and d are weighting coefficients that satisfy a + b + c + d = 1, with initial values ​​set as a = 0.2, b = 0.3, c = 0.25, and d = 0.25; ||ΔP||² represents the L2 norm of the parameter change trend vector; I stab This represents the data stability index, with a value range of 0 to 1, and is calculated using the following formula:

[0103]

[0104] In the formula, σ i μ represents the standard deviation of the i-th parameter. i S represents the mean of the i-th parameter; m represents the number of parameters. sim The similarity of spectral features is represented by a value ranging from 0 to 1, and the calculation formula is as follows:

[0105]

[0106] In the formula, F(S) represents the current spectral eigenvector; F(S) ref This represents the reference spectral feature vector. The output layer of the FWM model is a fully connected layer that maps the decoder output to the product quality evaluation index space, outputting multi-dimensional quality assessment results for the product, including purity, safety, stability, and applicability indices. The purpose of this step is to integrate multi-source heterogeneous data through a deep learning model to achieve a comprehensive and accurate assessment of the quality of solid waste resource recovery products, providing a basis for decision-making in the production process.

[0107] The specific implementation of step S06 is based on real-time monitoring of heavy metal ion content in solid waste resource products using an electrochemical sensor network, and the spectral data processed by a spectral fusion optimization function is fused to form a comprehensive quality evaluation result. First, selective electrode arrays, including ion-selective electrodes, modified electrodes, and reference electrodes, are deployed at key nodes in the processing flow to form an electrochemical sensor network. Then, the changes in redox current of each electrode are measured, and the concentration of heavy metal ions is quantitatively analyzed using voltammetry and potentiometry. The formula for calculating the concentration of heavy metal ions is:

[0108] C M =k·(I peak -I baseline )·f corr ;

[0109] In the formula, C M Indicates the concentration of heavy metal ions, in mg / L; I peak I represents the peak current, in μA. baseline The baseline current is represented in μA; k represents the correction factor in mg / (L·μA), determined using the standard curve method; f corr The interference correction factor is expressed by the following formula:

[0110]

[0111] In the formula, k i C represents the interference coefficient of the i-th interfering ion; i C represents the concentration of the i-th interfering ion, in mg / L; i，ref This represents the reference concentration of the i-th interfering ion, in mg / L. A multivariate correction algorithm is then used to eliminate cross-interference between ions, improving measurement accuracy; the measurement accuracy for typical heavy metal ions is controlled within ±5%. Next, the heavy metal ion content data measured by the electrochemical sensor network and the spectral data processed by the spectral fusion optimization function are standardized to eliminate dimensional differences. The standardization method is as follows:

[0112]

[0113] In the formula, X norm The data represents the standardized data; X represents the original data; μ X σ represents the mean; X Let represent the standard deviation. Finally, a hierarchical weighted model is established to calculate the comprehensive product quality index. The expression for this model is:

[0114]

[0115] In the formula, Q total This represents the comprehensive product quality index, with a value range of 0 to 1; wi Let the weight coefficient of the i-th type of indicator satisfy the following condition: w ij This represents the weight coefficient of the j-th sub-indicator under the i-th category of indicators, satisfying... q ij This represents the normalized score of the j-th sub-indicator under the i-th index category, with a value range of 0 to 1, where m represents the number of index categories; n i This indicates the number of sub-indicators included in the i-th category of indicators. The purpose of this step is to achieve accurate monitoring of heavy metal pollutants in products through the complementary combination of electrochemical and spectroscopic methods, thereby ensuring the environmental safety of resource-based products.

[0116] The specific implementation of step S07 involves updating the weight coefficients in the sampling rate determination function based on the output of the FWM model, returning to step S02 to recalculate the real-time sampling rate, thus forming an online monitoring closed-loop control. First, the product quality assessment results output by the FWM model are analyzed, and the deviations of each quality indicator from the target value are calculated:

[0117] ΔQ i =|Q i -Q i，target |;

[0118] In the formula, ΔQ i Q represents the deviation of the i-th quality indicator; i Q represents the actual value of the i-th quality indicator; i，target This represents the target value of the i-th quality indicator. Then, based on the magnitude and trend of the deviation, the sampling rate is adjusted to determine the weighting coefficients of each parameter in the function.

[0119]

[0120] In the formula, This represents the updated weight coefficients; This represents the weight coefficients before the update; Δw i The adjustment amount for the weighting coefficient is expressed by the following formula:

[0121]

[0122] In the formula, η represents the learning rate, which ranges from 0.01 to 0.1; sign represents the sign function; Δw represents the partial derivative of the quality index deviation with respect to the weighting coefficient. max This represents the maximum step size for weight adjustment, ranging from 0.05 to 0.15. After weight adjustment, normalization is required to ensure...

[0123]

[0124] In the formula, This represents the normalized weighting coefficients. The updated weighting coefficients are then applied to recalculate the real-time sampling rate, and the new sampling rate is applied to the parameter acquisition system.

[0125]

[0126] In the formula, This represents the updated real-time sampling rate, expressed in times per minute. Finally, a closed-loop feedback mechanism for quality monitoring and sampling control is established, enabling intelligent allocation and optimized utilization of monitoring resources. The purpose of this step is to enable the monitoring system to adaptively adjust its monitoring strategy based on product quality status through a closed-loop control strategy, thereby improving monitoring efficiency and ensuring monitoring quality.

[0127] The detailed structure of the FWM model employs a multi-layer encoder-decoder architecture, including a self-attention mechanism layer, residual connection layers, and normalization layers. The encoder part consists of six stacked encoder layers, each containing a multi-head self-attention sub-layer and a feedforward neural network sub-layer. Each sub-layer uses residual connections and layer normalization. The mathematical expression for the multi-head self-attention mechanism is:

[0128] MultiHead(Q,K,V)=Concat(head1,head2,...,head h W O ;

[0129] In the formula, MultiHead(Q, K, V) represents the output of multi-head attention; head i The output of the i-th attention head is represented by the following formula:

[0130]

[0131] In the formula, d represents the learnable parameter matrix; model This represents the model dimension, with a value of 512; d k and d v The dimension representing the keys and values ​​is 64; h represents the number of heads, with a value of 8. The formula for calculating the attention function Attention(Q, K, V) is:

[0132]

[0133] In the formula, Q represents the query matrix; K represents the key matrix; V represents the value matrix; d k The dimension of the key is represented by `<key>`; `softmax` represents the softmax function. The formulas for calculating residual connections and layer normalization are:

[0134] LayerNorm(x+Sublayer(x));

[0135] In the formula, x represents the input of the sublayer; Sublayer(x) represents the function of the sublayer; LayerNorm represents the layer normalization function, and its calculation formula is as follows:

[0136]

[0137] In the formula, γ and β represent the learnable scale and offset parameters; μ and σ represent the mean and standard deviation, respectively; ∈ represents a small constant to prevent division by zero errors, with a value of 1×10. -6 The calculation formula for the sublayer of the feedforward neural network is as follows:

[0138] FFN(x)=max(0,xW1+b1)W2+b2;

[0139] In the formula, b1 and b2 represent learnable parameters; d ff This represents the hidden layer dimension of the feedforward network, with a value of 2048. The decoder part also consists of 6 stacked decoder layers, each containing a self-attention sub-layer, a cross-attention sub-layer, and a feedforward neural network sub-layer. The calculation formula for the cross-attention sub-layer is the same as for multi-head attention, but the query matrix comes from the decoder input, and the key and value matrices come from the encoder output. The model output layer is a fully connected layer, mapping the decoder output to the product quality evaluation metric space:

[0140] Y = XW + b;

[0141] In the formula, Y represents the output vector; X represents the output of the decoder; W and b represent the learnable weight matrix and bias vector, respectively.

[0142] The steps for establishing the training dataset for the FWM model are as follows: First, historical monitoring data on the resource utilization process of solid waste from different sources are collected, including treatment data for different types of solid waste such as municipal solid waste, industrial solid waste, and agricultural waste. The data volume should cover at least 100 complete production cycles. Then, each set of data is cleaned and labeled. The cleaning process includes outlier detection and handling, missing value imputation, and data consistency verification. Outlier detection uses an improved local outlier factor algorithm, calculated using the following formula:

[0143]

[0144] In the formula, LOF k (p) represents the local outlier factor at point p; N k (p) represents the set of k nearest neighbors of point p; lrd k (p) represents the local reachability density of point p, calculated using the following formula:

[0145]

[0146] In the formula, reach-dist k (p, o) represents the reachable distance from point p to point o, calculated using the following formula:

[0147] reach-dist k (p, o) = max {k-distance (o), d (p, o)};

[0148] In the formula, k-distance(o) represents the distance from point o to its k-th nearest neighbor; d(p, o) represents the Euclidean distance from point p to point o. Missing value imputation uses a multiple linear regression method, calculated as follows:

[0149]

[0150] In the formula, β0 represents the missing value to be imputed; β0 represents the intercept term; β j Represents the regression coefficient; x j Let represent known variables; ε represent the random error term; and m represent the number of known variables. The annotation process includes annotating corresponding product quality indicators based on laboratory analysis results, such as purity, content of harmful substances, and stability. Then, data augmentation techniques are used to simulate parameter fluctuations under different operating conditions, including adding Gaussian noise to simulate sensor errors, random scaling to simulate parameter fluctuations, and time shifting to simulate process delays, expanding the dataset size to 3-5 times the original data. The noise addition formula is:

[0151]

[0152] In the formula, x noisy This represents the data after adding noise; x represents the original data; α represents the noise intensity, ranging from 0.01 to 0.1. This indicates that the mean is 0 and the variance is σ. 2 The Gaussian noise, σ, is determined based on the standard deviation of the original data, typically taken as 5% to 10% of the original data standard deviation. The random scaling formula is:

[0153]

[0154] In the formula, x scaled This represents the scaled data; β represents the scaling strength, with a value ranging from 0.05 to 0.2. This represents a uniform distribution on the interval [-1, 1]. The time shift formula is:

[0155]

[0156] In the formula, The shifted time series data is represented by Δt, which represents the time offset, ranging from -5 to 5 time steps. Then, standard data pairs containing the input parameter matrix and corresponding quality evaluation results are constructed. The input parameter matrix dimension is time step size × number of parameters, and the quality evaluation results dimension is the number of indicators. Finally, the dataset is divided into training, validation, and test sets using stratified random sampling, with proportions of 70%, 15%, and 15%, respectively, ensuring that the distribution of different types of solid waste data remains consistent across each subset.

[0157] The FWM model training process first involves unsupervised pre-training on large-scale historical solid waste treatment data, employing a masked autoencoder structure to learn the parameter distribution patterns and the inherent structure of the data. The loss function of the masked autoencoder is:

[0158]

[0159] In the formula, L MAE The loss function of the masked autoencoder is represented by ; N represents the number of samples; M represents the mask matrix, with elements of 0 or 1; ⊙ represents the Hadamard product; X represents the original data matrix. Represents the reconstruction of the data matrix; ||·|| F Let Frobenius norm be represented. Then, supervised fine-tuning training is performed for the current solid waste type, with the loss function comprehensively considering mean squared error and weighted cross-entropy:

[0160] L total =λ1L MSE +λ2L WCE ;

[0161] In the formula, L total L represents the total loss function; MSE L represents the mean squared error loss; WCE Let λ1 and λ2 represent the weighted cross-entropy loss; λ1 and λ2 represent the weight coefficients, satisfying λ1 + λ2 = 1, with initial values ​​set to λ1 = 0.7 and λ2 = 0.3. The mean squared error loss function is:

[0162]

[0163] In the formula, y i This represents the true label of the i-th sample; Let represent the predicted value of the i-th sample. The weighted cross-entropy loss function is:

[0164]

[0165] In the formula, C represents the number of categories; w j y represents the weight coefficient of the j-th class; ij This represents the true probability that the i-th sample belongs to the j-th class; This represents the predicted probability that the i-th sample belongs to the j-th class. An adaptive learning rate strategy is used during training, with the initial learning rate set to 1×10j. -4 When the validation set loss does not improve for 5 consecutive epochs, the learning rate is reduced to 0.8 times its original value.

[0166]

[0167] In the formula, lr new Indicates the updated learning rate; lr old n represents the learning rate before the update. no_improve This represents the number of epochs in which the validation set loss shows no continuous improvement. A weight-decreasing regularization technique is also introduced, with a decrease coefficient set to 5 × 10⁻⁶. -5 To prevent model overfitting:

[0168]

[0169] In the formula, L reg This represents the loss function after adding regularization; γ represents the decreasing coefficient, with a value of 5 × 10. -5 P represents the total number of parameters; W i This represents the value of the i-th parameter. The model's generalization ability is evaluated using 5-fold cross-validation, and the model parameters with the best performance on the validation set are ultimately selected as the final model.

[0170] To better understand and implement this invention, a specific application scenario, Example 2, is provided below: Researchers have developed an online monitoring system for the product quality of the fly ash resource recovery process from municipal solid waste incineration, which is applied to a production line for preparing building materials from fly ash. Based on the proposed online monitoring method for the product quality of solid waste resource recovery, this system achieves parameter acquisition, quality prediction, and evaluation throughout the entire fly ash resource recovery production process.

[0171] First, a parameter acquisition system was constructed, installing spectral probes and physicochemical parameter sensors in key processes such as pretreatment, stabilization, molding, and curing on the production line. The spectral probes included 3 Raman spectrometers, 2 near-infrared spectrometers, and 2 X-ray fluorescence spectrometers. The physicochemical parameter sensors included 12 temperature sensors, 8 pressure sensors, 6 humidity sensors, 5 redox potential sensors, 4 pH sensors, and 3 flow rate sensors. The initial sampling rate was set to 5 times per minute, and real-time data transmission was achieved via an industrial-grade fieldbus, forming a multi-parameter matrix containing 38 parameters.

[0172] The system dynamically adjusts the monitoring frequency based on a sampling rate determination function. In actual operation, when the temperature of the stabilization process fluctuated from the set value of 790℃ to 830℃, the temperature deviation factor f2 increased from 0.5 to 1.0, and the sampling rate increased from 5 times per minute to 7.5 times per minute. Simultaneously, the system monitored a decrease in material flow rate from the standard value of 1.2 m / s to 0.85 m / s, with a flow rate deviation factor f3 of 1.0, further supporting the increase in the sampling rate. Fly ash component analysis showed its heterogeneity index I... het The value was 0.68, corresponding to a component heterogeneity factor f4 of 1.0. Considering the influence of various factors, the real-time sampling rate was adjusted to 8.5 times per minute to ensure sufficient monitoring data is obtained during process fluctuations.

[0173] The solid waste component migration equations were applied to analyze the transformation of heavy metals in fly ash during the stabilization process. This example focuses on the migration behavior of four heavy metals: lead, cadmium, chromium, and zinc. Parameters such as temperature field distribution, redox potential, and pH value were substituted into the corresponding equations. By solving the component migration equations, the system predicted the heavy metal concentrations in the stabilized product, as shown in Table 1.

[0174] Table 1. Heavy metal concentrations predicted by the migration equations of solid waste components.

[0175]

[0176] The spectral fusion optimization function is used to process three types of spectral data. Raman spectroscopy focuses on the 400–1800 cm⁻¹ range. -1 Characteristic peaks within the wavenumber range were selected. For near-infrared spectroscopy, absorption peaks in the 1100–2500 nm wavelength range were chosen. For X-ray fluorescence spectroscopy, the position and intensity ratio of characteristic peaks for heavy metal elements were considered. By calculating the correlation coefficient between spectral characteristics and target component concentrations, the system dynamically adjusted the weighting coefficients for each spectrum: Raman spectroscopy weighted at 0.38, near-infrared spectroscopy weighted at 0.27, and X-ray fluorescence spectroscopy weighted at 0.35. The fused and optimized spectral data were transformed into a 128-dimensional feature vector, which served as input to the FWM model.

[0177] The FWM model is based on a multi-layer encoder-decoder architecture, and its self-attention mechanism employs an 8-head attention structure. The model input includes a 128-dimensional spectral feature vector, a 38-dimensional parameter matrix, and 4-dimensional predicted heavy metal concentrations. The model is trained on a training set containing 5000 historical datasets, with an initial learning rate of 1×10⁻⁶. -4 When the validation set loss showed no improvement for five consecutive epochs, the learning rate was reduced to 0.8 times its original value. After 100 epochs of training, the model achieved a mean squared error of 0.035 and an accuracy of 94.7% on the test set.

[0178] An electrochemical sensor network was used to monitor the content of heavy metal ions in the product in real time. Researchers installed six selective electrode arrays at the end of the production line, targeting heavy metal ions such as lead, cadmium, chromium, and zinc. Table 2 shows a comparison between the electrochemical monitoring results and laboratory analysis of a batch of products.

[0179] Table 2 Comparison of Electrochemical Sensor Network Monitoring Results and Laboratory Analysis

[0180] heavy metal elements Electrochemical sensor measurement (mg / L) Laboratory analytical value (mg / L) Relative error (%) Lead (Pb) 0.43 0.45 4.4 Cadmium (Cd) 0.07 0.075 6.7 Chromium (Cr) 0.36 0.39 7.7 Zinc (Zn) 1.59 1.67 4.8

[0181] The system integrates electrochemical sensor network measurements with spectral data to form a comprehensive quality evaluation result, including safety indicators, mechanical performance indicators, and durability indicators. Table 3 shows the comprehensive quality index of a certain batch of products calculated by the system.

[0182] Table 3. Evaluation Results of Comprehensive Product Quality Index

[0183] Quality Indicator Categories Weighting coefficient Evaluation score Quality status Safety indicators 0.4 0.86 good Mechanical performance indicators 0.35 0.92 excellent Durability index 0.25 0.78 qualified Overall Quality Index 1.0 0.86 good

[0184] Based on the quality fluctuation risk value Q output by the FWM model risk The system sets the FWM model output factor f5 to 0.5, and updates the weight coefficients in the sampling rate determination function accordingly, with the learning rate η set to 0.05. After the update, the values ​​of w1, w2, w3, w4, and w5 are adjusted from 0.1 to 0.18, w1 from 0.3 to 0.32, w2 from 0.25 to 0.27, w3 from 0.2 to 0.18, w4 from 0.15 to 0.14, and w5 from 0.1 to 0.09. Based on the updated weight coefficients, the real-time sampling rate is recalculated to 6.3 times per minute, forming a closed-loop monitoring control.

[0185] Traditional solid waste resource recovery product quality monitoring mainly relies on offline laboratory analysis, which suffers from problems such as monitoring lag, insufficient sample representativeness, and high monitoring costs. This embodiment employs an online monitoring method that solves these core problems: firstly, it achieves real-time monitoring of multi-dimensional parameters through multi-parameter matrix and spectral fusion technology, reducing monitoring latency from hours to minutes; secondly, a dynamic sampling rate adjustment mechanism improves sampling representativeness, especially by increasing sampling frequency during process fluctuations, enabling more accurate capture of quality fluctuations. This method not only improves monitoring efficiency and accuracy but also provides decision-making support for solid waste resource recovery product quality control, promoting the development of solid waste resource recovery technology.

[0186] It should be noted that the variables involved in this invention are explained in detail in Tables 4, 5, 6, and 7 below.

[0187] Table 4. Variable Explanation Table (Part 1)

[0188]

[0189] Table 5. Variable Explanation Table (Part Two)

[0190]

[0191]

[0192] Table 6. Variable Explanation Table (Part 3)

[0193]

[0194]

[0195] Table 7. Variable Explanation Table (Part Four)

[0196]

[0197] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention includes any person skilled in the art within the technical scope disclosed in the present invention.

Claims

1. A method for online monitoring of product quality in the solid waste resource utilization process, characterized in that, include: S01. Construct a parameter acquisition system with an initial sampling rate of 5 times per minute. Simultaneously collect spectral data and physicochemical parameters in the solid waste treatment process to form a multi-parameter matrix. S02. The sampling rate determination function is used to calculate the real-time sampling rate. The sampling rate determination function is determined based on the multi-parameter matrix variation components, processing temperature, material flow rate, solid waste component heterogeneity index and FWM model output results. It includes the initial sampling rate and the real-time variation sampling rate. When the FWM model has not yet output results, the sampling rate determination function sets the weight coefficient of the FWM model output results to 0. S03. Analyze the transformation law of substances in the resource utilization process through the solid waste component migration equation, and predict the content of key components in the product. The input parameters of the solid waste component migration equation include the temperature field distribution, redox potential, reactant concentration, catalyst activity and pH value in the multi-parameter matrix. S04. Raman spectroscopy, near-infrared spectroscopy and X-ray fluorescence spectroscopy data are processed using a spectral fusion optimization function. The input parameters of the spectral fusion optimization function include wavelength range, signal-to-noise ratio threshold, characteristic peak position, peak intensity ratio and baseline offset. S05. Using the FWM model to achieve multidimensional evaluation of product quality, the input parameters of the FWM model include the spectral feature vector output by the spectral fusion optimization function, the multi-parameter matrix, and the target component concentration prediction value output by the solid waste component migration equation, wherein the dimension of the attention mechanism parameter matrix is ​​jointly determined by the dimension of the multi-parameter matrix, the sampling rate, and the solid waste component heterogeneity index. S06. Real-time monitoring of heavy metal ion content in solid waste resource products based on electrochemical sensor network, and fusion with the spectral data processed by the spectral fusion optimization function to form a comprehensive quality evaluation result; S07. Update the weight coefficients in the sampling rate determination function according to the output of the FWM model, return to step S02 to recalculate the real-time sampling rate, and form an online monitoring closed-loop control. The multi-parameter matrix refers to a data set that includes real-time monitoring parameters such as temperature, pressure, humidity, oxidation-reduction potential, pH value, and flow rate during the solid waste treatment process. Specifically, the sampling rate determination function calculates weight coefficients based on the rate of change and standard deviation of each parameter in the multi-parameter matrix. When the weight coefficients exceed a preset threshold, the sampling frequency is increased; otherwise, the sampling frequency is decreased. The real-time sampling rate is dynamically adjusted in conjunction with the output results of the FWM model. The specific structure of the FWM model is a multi-layer encoder-decoder architecture, including a self-attention mechanism layer, a residual connection layer, and a normalization layer. The encoder is responsible for extracting the spectral feature vector, the predicted concentration of the target component, and the deep features in the multi-parameter matrix. The decoder is responsible for reconstructing the product quality evaluation indicators. The intermediate connection part uses a cross-attention mechanism to realize information interaction between different modal data. The output of the FWM model is used as one of the input parameters of the sampling rate determination function. Specifically, the inputs of the FWM model are: the spectral feature vector output by the spectral fusion optimization function, the multi-parameter matrix, and the predicted concentration of the target component output by the solid waste component migration equation. The output of the FWM model is: the multi-dimensional quality evaluation results of the product, including purity indicators, safety indicators, stability indicators, and applicability indicators. Specifically, step S07 involves: first, analyzing the product quality assessment results output by the FWM model and calculating the deviations of each quality indicator from the target value; then, adjusting the weight coefficients of each parameter in the sampling rate determination function based on the magnitude and trend of the deviations; increasing the weight coefficients of relevant parameters when the quality indicator deviations increase, and appropriately decreasing the weight coefficients when the quality indicators are stable within the target range; then setting the adjustment step size of the weight coefficients, which is generally set to 0.05 to 0.15, and dynamically adjusted according to the fluctuation range of quality; finally, recalculating the real-time sampling rate using the updated weight coefficients and applying the new sampling rate to the parameter acquisition system.

2. The method according to claim 1, characterized in that, The solid waste component heterogeneity index is a quantitative indicator that characterizes the heterogeneity of solid waste, calculated by analyzing the fluctuation range, uneven distribution, and interaction strength between solid waste components.

Images