A leather solid waste resource utilization process management system based on big data analysis

Abstract

The application discloses a leather solid waste resourceization process management system based on big data analysis, comprising: a data acquisition module for collecting and preprocessing, constructing a basic data sequence; a feature extraction module for performing kernel regression operation according to the basic data sequence to form a solid waste feature sequence; a transformation prediction module for performing inference prediction operation on the solid waste feature sequence through a Fourier neural network; a path selection module for multi-objective sorting of the prediction result and screening of a target path; a parameter matching module for constructing a process parameter set according to the target path; a resource processing module for mapping the process parameter set into an instruction execution and recording feedback data; and a feedback updating module for updating parameters of the Fourier neural network. The application combines technologies such as feedforward neural network and Fourier neural network, has the advantages of high resource transformation efficiency, strong intelligent selection capability of processing path and sustainable optimization.

Description

Technical Field

[0001] This invention relates to the field of solid waste resource utilization technology, and in particular to a process management system for leather solid waste resource utilization based on big data analysis. Background Technology

[0002] The leather manufacturing process generates a large volume of complex solid waste, including chromium-containing scraps, leather dust, and oil residue. These solid wastes are characterized by significant compositional variations, high pollution loads, and unstable resource recovery value. Improper handling can cause serious environmental pollution. Therefore, the resource utilization of leather solid waste can not only reduce environmental risks but also increase the economic added value of industrial by-products, becoming an important direction for the green development of the leather industry.

[0003] In existing technologies, the resource recovery process for leather solid waste generally relies on manual experience, with treatment methods typically limited to a few fixed pathways such as incineration, composting, pyrolysis, and chemical recycling. Due to significant differences in the physical and chemical properties of different batches of solid waste, fixed pathways often fail to adapt to the dynamic characteristics of the waste, resulting in low resource recovery rates, high energy consumption, and difficulty in controlling residual emissions. Furthermore, the lack of systematic data analysis methods for modeling and optimizing the treatment process renders resource recovery decisions lacking in scientific basis and feedback mechanisms.

[0004] Furthermore, most existing systems employ traditional rule engines or linear regression models, which lack sufficient ability to predict the pathways of leather solid waste, struggle to cope with the impact of high-dimensional nonlinear characteristics, and make it difficult to dynamically adjust processing paths. While some systems possess basic data collection and display functions, they have not formed a closed-loop management system from "data-modeling-decision-execution-feedback," thus failing to support the needs of continuous optimization and intelligent scheduling.

[0005] Therefore, how to provide a process management system for the resource utilization of leather solid waste based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a process management system for the resource utilization of leather solid waste based on big data analysis. This invention fully integrates technologies such as feedforward neural networks, Fourier neural networks, Gumbel-Sinkhorn sorting algorithm, and Hungarian optimization algorithm, and constructs a closed-loop management process in detail from data acquisition, feature extraction, path prediction, target selection, parameter matching to resource processing and feedback updates. It has the advantages of high resource conversion efficiency, strong intelligent path selection capability, strong adaptability to the differences of different batches of solid waste, and continuous optimization of the processing process.

[0007] A leather solid waste resource utilization process management system based on big data analysis according to an embodiment of the present invention includes:

[0008] The data acquisition module is used to collect and preprocess basic data on leather solid waste and construct a basic data sequence.

[0009] The feature extraction module is used to construct a neural tangent kernel matrix by calculating gradient vectors of the basic data sequence through a feedforward neural network, and to perform kernel regression operation to extract high-dimensional implicit features at each time step to form a solid waste feature sequence.

[0010] The conversion prediction module is used to construct a set of candidate paths based on existing historical processing records, and to perform inference prediction operations on the solid waste feature sequence on each candidate path through a Fourier neural network to obtain the prediction results of each candidate path.

[0011] The path selection module is used to sort the prediction results by multiple objectives using the Gumbel-Sinkhorn network, construct the path permutation matrix and calculate the priority, and select the candidate path with the highest priority as the target path.

[0012] The parameter matching module is used to construct a process control sequence based on the target path, and to perform a minimum cost matching operation with the solid waste feature sequence using the Hungarian algorithm, extracting the control process parameters corresponding to the matching results to form a process parameter set;

[0013] The resource processing module is used to map the set of process parameters into execution instructions according to the preset action mapping table, perform resource recovery processing on leather solid waste according to the execution instructions, and record feedback data.

[0014] The feedback update module is used to calculate the difference between the prediction result and the feedback data, and update the parameters of the Fourier neural network based on the difference result.

[0015] Optionally, the basic data includes the batch number, source process, quality, moisture content, heavy metal content, and organic matter mass fraction of the leather solid waste; the preprocessing includes missing data filling, anomaly removal, numerical standardization, category coding, time alignment, and format unification operations; the prediction results include the predicted values ​​of resource conversion rate, processing energy consumption, processing cost, residue quality, and processing cycle corresponding to each candidate path; and the feedback data includes the actual resource conversion rate, actual processing energy consumption, actual processing cost, actual residue quality, and actual processing cycle.

[0016] Optionally, the feature extraction module includes:

[0017] The gradient direction calculation operation is performed on the basic data at each time step in the basic data sequence, and the gradient vector set corresponding to each time step is extracted through a feedforward neural network, specifically including:

[0018] The basic data at each time step in the basic data sequence are numerically arranged, and a data vector is constructed in a fixed order.

[0019] The data vector is fed into the feedforward neural network. In the hidden layer of the feedforward neural network, a weighted summation operation is performed according to the connection weights, and the hidden feature values ​​of the corresponding time step are obtained through nonlinear activation calculation.

[0020] Under the condition that the connection weights remain unchanged, partial derivative calculation is performed on each dimension of the data vector based on the hidden feature values ​​to obtain the rate of change of the hidden feature values ​​in each dimension, and the gradient change value corresponding to each dimension is formed.

[0021] The gradient change values ​​corresponding to each dimension within the same time step are numerically normalized and combined according to the order of the data vectors to form the gradient vector for that time step.

[0022] Arrange all gradient vectors in chronological order to construct a gradient vector set;

[0023] The gradient vector set at each time step is mapped by performing a kernel function mapping operation on the neural tangent kernel function to obtain the kernel similarity value between each time step. The neural tangent kernel matrix is ​​then constructed based on the kernel similarity value. The elements in the neural tangent kernel matrix represent the degree of similarity between the gradient vectors corresponding to any two time steps.

[0024] A sliding window with a fixed step size is constructed for the neural tangent kernel matrix. Kernel regression is performed within each sliding window to extract high-dimensional implicit features at the corresponding time step, and the feature extraction results are continuously corrected.

[0025] Arrange all the corrected high-dimensional implicit features in chronological order to construct a solid waste feature sequence.

[0026] Optionally, the construction process of the neural tangent kernel matrix includes:

[0027] Perform numerical centering on each gradient vector in the gradient vector set according to the vector dimension, calculate the mean of each dimension over all time steps, and adjust the corresponding dimension values ​​at each time step based on the mean.

[0028] The adjusted gradient vector is length normalized by summing the squares of the values ​​of each dimension in each gradient vector and taking the square root to obtain the vector length. The values ​​of each dimension are then scaled proportionally according to the vector length.

[0029] Perform element-wise multiplication on the normalized gradient vectors of any two time steps according to the dimension, and sum them to obtain the inner product value between the corresponding two time steps. The inner product value represents the degree of consistency of the gradient change direction between the two time steps.

[0030] The inner product values ​​are mapped using a neural tangent kernel function. The inner product values ​​are used as inputs, and the kernel similarity values ​​are obtained through nonlinear function transformation.

[0031] The kernel similarity values ​​between all time steps are calculated and filled into a two-dimensional matrix structure in chronological order to construct a symmetric neural tangent kernel matrix.

[0032] Optionally, the conversion prediction module includes:

[0033] The existing historical processing records are classified and organized, the path identifiers and resource utilization result data corresponding to each historical batch are extracted, a candidate path set is constructed according to the path identifiers, and a corresponding path feature index is established for each candidate path.

[0034] The solid waste feature sequence is input into the Fourier neural network in chronological order. In the frequency domain mapping layer of the Fourier neural network, the solid waste feature sequence is subjected to spectral decomposition operation to generate the frequency domain representation sequence of the corresponding time step.

[0035] In the frequency domain weight layer of the Fourier neural network, the frequency domain representation sequence is mapped to the path feature indices in the candidate path set. For each candidate path, a corresponding path response sequence and weight coefficients are generated, specifically including:

[0036] According to the path feature index order in the candidate path set, each path feature index is numerically encoded to construct a path index sequence with the same dimension as the frequency domain representation sequence.

[0037] In the frequency domain weighting layer, a weight parameter matrix with the same dimension as the path index sequence is set, and the corresponding weight vector is extracted according to the index position of the path feature index in the weight parameter matrix.

[0038] Perform element-wise multiplication between the frequency domain representation vector of each time step in the frequency domain representation sequence and the corresponding weight vector to generate the weighted frequency domain vector of each time step;

[0039] The weighted frequency domain vectors at each time step under the same candidate path are summed in chronological order to construct the path response sequence of the candidate path;

[0040] The numerical values ​​in the path response sequence are normalized using the L2 norm, and the corresponding weight coefficients are extracted according to the path feature index order.

[0041] Based on the weighting coefficients, the response sequences of each path are multiplied element by element and summed to form the path prediction vector of each candidate path.

[0042] Numerical mapping processing is performed on the prediction vectors of each path according to the fixed mapping template, generating the predicted values ​​of resource conversion rate, processing energy consumption, processing cost, residue quality and processing cycle for each candidate path, and combining them to form the prediction results.

[0043] Optionally, the spectral decomposition operation specifically includes:

[0044] For each time step in the solid waste feature sequence, construct a fixed-length vector according to the feature dimension order of the high-dimensional implicit features, and form a feature matrix according to the time order.

[0045] In the frequency domain mapping layer, the fixed-length vector of each time step in the feature matrix is ​​used as input. The discrete Fourier transform results of the fixed-length vector of each time step on different preset frequency components are calculated and decomposed into real part values ​​and imaginary part values ​​to form a frequency component matrix.

[0046] The real and imaginary values ​​of each frequency component in the frequency component matrix are squared and summed. The summation result is then squared to obtain the amplitude sequence of the corresponding frequency component. Simultaneously, the arctangent operation is performed with the ratio of the imaginary to the real values ​​as input to obtain the phase sequence of the corresponding frequency component.

[0047] Numerical normalization is performed on each amplitude sequence to ensure that the amplitude sequences maintain a consistent scale within a preset range, and a unified reference angle calibration operation is performed on the phase sequence to ensure that the phases between different time steps have a consistent reference.

[0048] The normalized amplitude sequence and the calibrated phase sequence are combined in chronological order to form a frequency domain representation sequence.

[0049] Optionally, the path selection module includes:

[0050] The five predicted values ​​corresponding to each candidate path in the prediction results are arranged numerically according to a fixed index order to construct a path index set. The values ​​of each dimension in the path index set are then normalized to generate a path index matrix.

[0051] The path index matrix is ​​input into the Gumbel-Sinkhorn network, random perturbation values ​​are added to the path index corresponding to each candidate path to generate a perturbation path matrix, and a row-by-row exponential mapping operation is performed on the perturbation path matrix to form a sorting weight matrix.

[0052] Perform a fixed number of row and column normalization operations on the sorting weight matrix to ensure that the sum of the values ​​in each row and each column converges to a preset constant interval, forming a path permutation matrix. Specifically, this includes:

[0053] Read the values ​​of each row in the sorted weight matrix, and calculate the sum of the row values ​​of all values ​​in each row in row order;

[0054] The sorting weight matrix is ​​normalized based on the sum of the values ​​in each row, and all values ​​in each row are scaled proportionally so that the sum of the row values ​​in each row is within a preset constant range.

[0055] Calculate the sum of the column values ​​of all values ​​in each column of the row normalization result in column order;

[0056] The column normalization operation is performed on the row normalization result based on the sum of the values ​​in each column, and the scale is applied to all the values ​​in each column so that the sum of the column values ​​is within a preset constant range.

[0057] The column normalization result is used as the new sorting weight matrix. Row normalization and column normalization operations are repeated until the preset number of times are reached to generate the path permutation matrix.

[0058] The path priority is calculated based on the weight distribution corresponding to each row in the path permutation matrix, and the candidate path with the highest path priority is selected as the target path. The weight distribution represents the set of weights formed by arranging the values ​​of each row corresponding to the candidate path in the path permutation matrix in column order.

[0059] Optionally, the process of forming the sorting weight matrix includes:

[0060] In the Gumbel-Sinkhorn network, the path index values ​​corresponding to each candidate path in the path index matrix are read, and the number of index dimensions for each candidate path is counted.

[0061] For each candidate path, generate a random perturbation value that is consistent with the number of corresponding indicator dimensions, and add the random perturbation value to the path indicator value of the corresponding candidate path element by element according to the dimension to form the perturbation indicator value of the candidate path.

[0062] Arrange the perturbation index values ​​of all candidate paths in the original candidate path order to construct a perturbation path matrix;

[0063] Perform an exponential mapping operation on each element of each row of values ​​in the perturbation path matrix, so that the values ​​in each row are transformed into positive values ​​after the exponential mapping.

[0064] Summing operation is performed on all exponential mapping results in the same row to obtain the exponential sum of the row. Then, based on the exponential sum, a scaling operation is performed on each exponential mapping result in the row to make the sum of all values ​​in the row meet the preset requirements.

[0065] The values ​​in each row after scaling are arranged in the order of the candidate paths to form a sorting weight matrix.

[0066] Optionally, the parameter matching module includes:

[0067] Read the flow control sequence corresponding to the target path from the preset flow mapping table;

[0068] The Hungarian algorithm is used to extract the control flow parameters corresponding to each processing stage according to the order of each processing stage in the flow control sequence, and the number of processing stages is counted.

[0069] High-dimensional implicit features corresponding to all time steps of the current batch of leather solid waste are extracted from the solid waste feature sequence. The feature arrangement sequence is constructed in chronological order and then divided into equal segments according to the number of processing stages to obtain multiple feature subsequences.

[0070] For each processing stage, the control flow parameters are traversed, and each high-dimensional implicit feature in the corresponding feature subsequence is dimensionally aligned with the control flow parameters. The numerical difference between each corresponding dimension is calculated, and the absolute value summation operation is performed on all dimension differences to obtain the matching cost between the processing stage and the high-dimensional implicit feature.

[0071] Arrange all matching costs in the same processing stage in chronological order to form a stage cost sequence;

[0072] All stage cost sequences are combined in the order of processing stages to form a matching weight matrix, where rows represent processing stages and columns represent high-dimensional implicit features.

[0073] Perform row standardization on the matching weight matrix, record the minimum matching cost in each row, and subtract the corresponding minimum matching cost from each value in that row;

[0074] Perform column standardization on the row standardization results, record the minimum matching cost in each column, and subtract the corresponding minimum matching cost from each value in that column;

[0075] Perform zero-value search and marking operations on the column standardization results to achieve unique assignment while ensuring that each column is marked only once, and generate an assignment result matrix;

[0076] Based on the allocation result matrix, extract the high-dimensional implicit features corresponding to each processing stage, and map and assign the values ​​of each dimension of the high-dimensional implicit features according to the parameter order in the control flow parameters to generate a set of process parameters consistent with the order of the process control sequence.

[0077] Optionally, the resource processing module includes:

[0078] Read the process parameter set, divide the process parameters in the process parameter set into stages according to the processing stage order in the process control sequence, and generate a stage parameter subset;

[0079] Extract the processing action type corresponding to the parameter subset of each stage from the preset action mapping table, and perform mapping calculation on the processing action type according to the order in the parameter subset of the stage to generate a set of execution instructions including processing temperature, processing time, material feeding ratio and equipment operating frequency.

[0080] The leather solid waste is processed for resource recovery according to the set of execution instructions, and the start and end times of each processing stage are recorded to form a stage record sequence.

[0081] During the resource recovery process, the resource conversion rate, energy consumption, cost, residue quality and cycle of each processing stage are collected in real time and arranged in chronological order according to the stage record sequence to construct a stage feedback sequence.

[0082] Perform a phase summary calculation on the phase feedback sequence to generate feedback data.

[0083] The beneficial effects of this invention are:

[0084] First, this invention significantly improves the intelligence, precision, and efficiency of leather solid waste treatment by constructing a leather solid waste resource utilization process management system based on big data analysis. At the same time, it uses a feedforward neural network combined with a neural tangent kernel function to extract high-dimensional features from the original basic data, enabling the system to identify and model the differences in physicochemical properties of different batches of leather solid waste, thereby providing more detailed data support for subsequent treatment paths.

[0085] Secondly, by using a Fourier neural network to perform frequency domain modeling and response prediction for multiple candidate resource recovery paths, and combining this with a Gumbel-Sinkhorn network to rank and select multiple objective indicators, dynamic optimization and refined decision-making of resource recovery processing paths are achieved, significantly reducing processing energy consumption and resource waste.

[0086] Finally, the system introduces the Hungarian algorithm to optimally match path parameters with solid waste characteristics, and feeds back the resource recovery results to the prediction model through the feedback update module. This realizes closed-loop optimization of the processing flow and adaptive iteration of the model, effectively improving resource conversion rate and processing stability. This invention not only solves the problems of path fixation, slow response, and lack of feedback in the prior art, but also constructs a sustainable evolutionary intelligent resource recovery decision framework, which has significant engineering application value and promotion prospects. Attached Figure Description

[0087] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0088] Figure 1 This is a module structure diagram of a leather solid waste resource utilization process management system based on big data analysis proposed in this invention.

[0089] Figure 2 This invention presents an overall resource utilization management flowchart for a leather solid waste resource utilization process management system based on big data analysis.

[0090] Figure 3 This invention presents a flowchart illustrating parameter matching and resource processing feedback for a leather solid waste resource utilization process management system based on big data analysis. Detailed Implementation

[0091] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0092] refer to Figures 1-3 A process management system for the resource utilization of leather solid waste based on big data analysis, comprising:

[0093] The data acquisition module is used to collect and preprocess basic data on leather solid waste and construct a basic data sequence.

[0094] The feature extraction module is used to construct a neural tangent kernel matrix by calculating gradient vectors of the basic data sequence through a feedforward neural network, and to perform kernel regression operation to extract high-dimensional implicit features at each time step to form a solid waste feature sequence.

[0095] The conversion prediction module is used to construct a set of candidate paths based on existing historical processing records, and to perform inference prediction operations on the solid waste feature sequence on each candidate path through a Fourier neural network to obtain the prediction results of each candidate path.

[0096] The path selection module is used to sort the prediction results by multiple objectives using the Gumbel-Sinkhorn network, construct the path permutation matrix and calculate the priority, and select the candidate path with the highest priority as the target path.

[0097] The parameter matching module is used to construct a process control sequence based on the target path, and to perform a minimum cost matching operation with the solid waste feature sequence using the Hungarian algorithm, extracting the control process parameters corresponding to the matching results to form a process parameter set;

[0098] The resource processing module is used to map the set of process parameters into execution instructions according to the preset action mapping table, perform resource recovery processing on leather solid waste according to the execution instructions, and record feedback data.

[0099] The feedback update module is used to calculate the difference between the prediction result and the feedback data, and update the parameters of the Fourier neural network based on the difference result.

[0100] In this embodiment, the basic data includes the batch number, source process, quality, moisture content, heavy metal content, and organic matter mass fraction of the leather solid waste. The preprocessing includes missing data filling, anomaly removal, numerical standardization, category coding, time alignment, and format unification. The prediction results include the predicted values ​​of resource conversion rate, processing energy consumption, processing cost, residue quality, and processing cycle for each candidate path. The feedback data includes the actual resource conversion rate, actual processing energy consumption, actual processing cost, actual residue quality, and actual processing cycle.

[0101] In this embodiment, the feature extraction module includes:

[0102] The gradient direction calculation operation is performed on the basic data at each time step in the basic data sequence, and the gradient vector set corresponding to each time step is extracted through a feedforward neural network, specifically including:

[0103] The basic data at each time step in the basic data sequence are numerically arranged, and a data vector is constructed in a fixed order.

[0104] The data vector is fed into the feedforward neural network. In the hidden layer of the feedforward neural network, a weighted summation operation is performed according to the connection weights, and the hidden feature values ​​of the corresponding time step are obtained through nonlinear activation calculation.

[0105] Under the condition that the connection weights remain unchanged, partial derivative calculation is performed on each dimension of the data vector based on the hidden feature values ​​to obtain the rate of change of the hidden feature values ​​in each dimension, and the gradient change value corresponding to each dimension is formed.

[0106] The gradient change values ​​corresponding to each dimension within the same time step are numerically normalized and combined according to the order of the data vectors to form the gradient vector for that time step.

[0107] Arrange all gradient vectors in chronological order to construct a gradient vector set;

[0108] The kernel function mapping operation is performed on the gradient vector set of each time step by the neural tangent kernel function to obtain the kernel similarity value between each time step, and the neural tangent kernel matrix is ​​constructed based on the kernel similarity value. The elements in the neural tangent kernel matrix represent the degree of similarity between the gradient vectors corresponding to any two time steps.

[0109] A sliding window with a fixed step size is constructed for the neural tangent kernel matrix. Kernel regression is performed within each sliding window to extract high-dimensional implicit features at the corresponding time step, and the feature extraction results are continuously corrected.

[0110] Arrange all the corrected high-dimensional implicit features in chronological order to construct a solid waste feature sequence.

[0111] In this embodiment, the process of constructing the neural tangent kernel matrix includes:

[0112] Perform numerical centering on each gradient vector in the gradient vector set according to the vector dimension, calculate the mean of each dimension over all time steps, and adjust the corresponding dimension values ​​at each time step based on the mean.

[0113] The adjusted gradient vector is length normalized by summing the squares of the values ​​of each dimension in each gradient vector and taking the square root to obtain the vector length. The values ​​of each dimension are then scaled proportionally according to the vector length.

[0114] Perform element-wise multiplication on the normalized gradient vectors of any two time steps according to the dimension, and sum them to obtain the inner product value between the corresponding two time steps. The inner product value represents the degree of consistency of the gradient change direction between the two time steps.

[0115] The inner product values ​​are mapped using a neural tangent kernel function. The inner product values ​​are used as inputs, and the kernel similarity values ​​are obtained through nonlinear function transformation.

[0116] The kernel similarity values ​​between all time steps are calculated and filled into a two-dimensional matrix structure in chronological order to construct a symmetric neural tangent kernel matrix.

[0117] In this embodiment, the kernel regression operation specifically includes:

[0118] In the neural tangent kernel matrix, a sliding window with a fixed step size is constructed with the current time step as the center. The kernel similarity values ​​between the time steps within the coverage area of ​​the sliding window are extracted to form the local kernel matrix of the current time step.

[0119] The kernel similarity values ​​in the local kernel matrix are used as weights to perform a weighted summation operation on the gradient vectors corresponding to each time step within the sliding window.

[0120] Perform a scaling operation on the weighted summation result, dividing the value of each dimension in the weighted summation result by the sum of all kernel similarity values ​​within the sliding window to obtain the high-dimensional implicit feature of the current time step;

[0121] The sliding window is moved sequentially according to time steps, and the weighted accumulation and scaling operations are repeatedly performed to generate high-dimensional implicit features corresponding to all time steps.

[0122] In this embodiment, the conversion prediction module includes:

[0123] The existing historical processing records are classified and organized, the path identifiers and resource utilization result data corresponding to each historical batch are extracted, a candidate path set is constructed according to the path identifiers, and a corresponding path feature index is established for each candidate path.

[0124] The solid waste feature sequence is input into the Fourier neural network in chronological order. In the frequency domain mapping layer of the Fourier neural network, the solid waste feature sequence is subjected to spectral decomposition operation to generate the frequency domain representation sequence of the corresponding time step.

[0125] In the frequency domain weight layer of the Fourier neural network, the frequency domain representation sequence is mapped to the path feature indices in the candidate path set. For each candidate path, a corresponding path response sequence and weight coefficients are generated, specifically including:

[0126] According to the path feature index order in the candidate path set, each path feature index is numerically encoded to construct a path index sequence with the same dimension as the frequency domain representation sequence.

[0127] In the frequency domain weighting layer, a weight parameter matrix with the same dimension as the path index sequence is set, and the corresponding weight vector is extracted according to the index position of the path feature index in the weight parameter matrix.

[0128] Perform element-wise multiplication between the frequency domain representation vector of each time step in the frequency domain representation sequence and the corresponding weight vector to generate the weighted frequency domain vector of each time step;

[0129] The weighted frequency domain vectors at each time step under the same candidate path are summed in chronological order to construct the path response sequence of the candidate path;

[0130] The numerical values ​​in the path response sequence are normalized using the L2 norm, and the corresponding weight coefficients are extracted according to the path feature index order.

[0131] Based on the weighting coefficients, the response sequences of each path are multiplied element by element and summed to form the path prediction vector of each candidate path.

[0132] Numerical mapping processing is performed on the prediction vectors of each path according to the fixed mapping template, generating the predicted values ​​of resource conversion rate, processing energy consumption, processing cost, residue quality and processing cycle for each candidate path, and combining them to form the prediction results.

[0133] In this embodiment, the spectrum decomposition operation specifically includes:

[0134] For each time step in the solid waste feature sequence, construct a fixed-length vector according to the feature dimension order of the high-dimensional implicit features, and form a feature matrix according to the time order.

[0135] In the frequency domain mapping layer, the fixed-length vector of each time step in the feature matrix is ​​used as input. The discrete Fourier transform results of the fixed-length vector of each time step on different preset frequency components are calculated and decomposed into real part values ​​and imaginary part values ​​to form a frequency component matrix.

[0136] The real and imaginary values ​​of each frequency component in the frequency component matrix are squared and summed. The summation result is then squared to obtain the amplitude sequence of the corresponding frequency component. Simultaneously, the arctangent operation is performed with the ratio of the imaginary to the real values ​​as input to obtain the phase sequence of the corresponding frequency component.

[0137] Numerical normalization is performed on each amplitude sequence to ensure that the amplitude sequences maintain a consistent scale within a preset range, and a unified reference angle calibration operation is performed on the phase sequence to ensure that the phases between different time steps have a consistent reference.

[0138] The normalized amplitude sequence and the calibrated phase sequence are combined in chronological order to form a frequency domain representation sequence.

[0139] In this embodiment, the path selection module includes:

[0140] The five predicted values ​​corresponding to each candidate path in the prediction results are arranged numerically according to a fixed index order to construct a path index set. The values ​​of each dimension in the path index set are then normalized to generate a path index matrix.

[0141] The path index matrix is ​​input into the Gumbel-Sinkhorn network, random perturbation values ​​are added to the path index corresponding to each candidate path to generate a perturbation path matrix, and a row-by-row exponential mapping operation is performed on the perturbation path matrix to form a sorting weight matrix.

[0142] Perform a fixed number of row and column normalization operations on the sorting weight matrix to ensure that the sum of the values ​​in each row and each column converges to a preset constant interval, forming a path permutation matrix. Specifically, this includes:

[0143] Read the values ​​of each row in the sorted weight matrix, and calculate the sum of the row values ​​of all values ​​in each row in row order;

[0144] The sorting weight matrix is ​​normalized based on the sum of the values ​​in each row, and all values ​​in each row are scaled proportionally so that the sum of the row values ​​in each row is within a preset constant range.

[0145] Calculate the sum of the column values ​​of all values ​​in each column of the row normalization result in column order;

[0146] The column normalization operation is performed on the row normalization result based on the sum of the values ​​in each column, and the scale is applied to all the values ​​in each column so that the sum of the column values ​​is within a preset constant range.

[0147] The column normalization result is used as the new sorting weight matrix. Row normalization and column normalization operations are repeated until the preset number of times are reached to generate the path permutation matrix.

[0148] The path priority is calculated based on the weight distribution corresponding to each row in the path permutation matrix, and the candidate path with the highest path priority is selected as the target path. Here, the weight distribution represents the set of weights formed by arranging the values ​​of each row corresponding to the candidate path in the path permutation matrix in column order.

[0149] In this embodiment, the process of forming the sorting weight matrix includes:

[0150] In the Gumbel-Sinkhorn network, the path index values ​​corresponding to each candidate path in the path index matrix are read, and the number of index dimensions for each candidate path is counted.

[0151] For each candidate path, generate a random perturbation value that is consistent with the number of corresponding indicator dimensions, and add the random perturbation value to the path indicator value of the corresponding candidate path element by element according to the dimension to form the perturbation indicator value of the candidate path.

[0152] Arrange the perturbation index values ​​of all candidate paths in the original candidate path order to construct a perturbation path matrix;

[0153] Perform an exponential mapping operation on each element of each row of values ​​in the perturbation path matrix, so that the values ​​in each row are transformed into positive values ​​after the exponential mapping.

[0154] Summing operation is performed on all exponential mapping results in the same row to obtain the exponential sum of the row. Then, based on the exponential sum, a scaling operation is performed on each exponential mapping result in the row to make the sum of all values ​​in the row meet the preset requirements.

[0155] The values ​​in each row after scaling are arranged in the order of the candidate paths to form a sorting weight matrix.

[0156] In this embodiment, the parameter matching module includes:

[0157] Read the flow control sequence corresponding to the target path from the preset flow mapping table;

[0158] The Hungarian algorithm is used to extract the control flow parameters corresponding to each processing stage according to the order of each processing stage in the flow control sequence, and the number of processing stages is counted.

[0159] High-dimensional implicit features corresponding to all time steps of the current batch of leather solid waste are extracted from the solid waste feature sequence. The feature arrangement sequence is constructed in chronological order and then divided into equal segments according to the number of processing stages to obtain multiple feature subsequences.

[0160] For each processing stage, the control flow parameters are traversed, and each high-dimensional implicit feature in the corresponding feature subsequence is dimensionally aligned with the control flow parameters. The numerical difference between each corresponding dimension is calculated, and the absolute value summation operation is performed on all dimension differences to obtain the matching cost between the processing stage and the high-dimensional implicit feature.

[0161] Arrange all matching costs in the same processing stage in chronological order to form a stage cost sequence;

[0162] All stage cost sequences are combined in the order of processing stages to form a matching weight matrix, where rows in the matching weight matrix represent processing stages and columns represent high-dimensional implicit features.

[0163] Perform row standardization on the matching weight matrix, record the minimum matching cost in each row, and subtract the corresponding minimum matching cost from each value in that row;

[0164] Perform column standardization on the row standardization results, record the minimum matching cost in each column, and subtract the corresponding minimum matching cost from each value in that column;

[0165] Perform zero-value search and marking operations on the column standardization results to achieve unique assignment while ensuring that each column is marked only once, and generate an assignment result matrix;

[0166] Based on the allocation result matrix, extract the high-dimensional implicit features corresponding to each processing stage, and map and assign the values ​​of each dimension of the high-dimensional implicit features according to the parameter order in the control flow parameters to generate a set of process parameters consistent with the order of the process control sequence.

[0167] In this embodiment, the resource processing module includes:

[0168] Read the process parameter set, divide the process parameters in the process parameter set into stages according to the processing stage order in the process control sequence, and generate a stage parameter subset;

[0169] Extract the processing action type corresponding to the parameter subset of each stage from the preset action mapping table, and perform mapping calculation on the processing action type according to the order in the parameter subset of the stage to generate a set of execution instructions including processing temperature, processing time, material feeding ratio and equipment operating frequency.

[0170] The leather solid waste is processed for resource recovery according to the set of execution instructions, and the start and end times of each processing stage are recorded to form a stage record sequence.

[0171] During the resource recovery process, the resource conversion rate, energy consumption, cost, residue quality and cycle of each processing stage are collected in real time and arranged in chronological order according to the stage record sequence to construct a stage feedback sequence.

[0172] Perform a phase summary calculation on the phase feedback sequence to generate feedback data.

[0173] In this embodiment, the feedback update module includes:

[0174] Match the prediction results with the feedback data to construct the prediction data sequence and the feedback data sequence;

[0175] Perform term-by-term difference calculations on the predicted data sequence and the feedback data sequence to generate an error sequence;

[0176] The error values ​​in the error sequence are arranged in a fixed index order to construct an error matrix, and the error matrix is ​​then normalized to form a standard error matrix.

[0177] The standard error matrix is ​​input into the loss calculation unit of the Fourier neural network, and the backpropagation operation is performed on the frequency domain weight parameters to obtain the gradient update values ​​corresponding to each frequency domain weight parameter.

[0178] The frequency domain weight parameters of the Fourier neural network are corrected based on the gradient update values.

[0179] Example 1: To verify the feasibility of this invention in practice, it was applied to the solid waste treatment process of a typical leather manufacturing production line. The system focuses on leather solid waste generated in actual production, covering typical industrial solid waste types such as leather scraps, chromium-containing leather shavings, and oily sludge from four different processes: leveling, cutting, spraying, and dyeing / finishing. Addressing key challenges such as large batch-to-batch variations in solid waste indicators, poor adaptability of existing fixed treatment pathways, and low resource conversion rates, an intelligent resource-based process management system was deployed and applied.

[0180] In this scenario, the system integrates daily automatically collected solid waste data. The collected data includes six categories of indicators: batch number, source process, total mass, moisture content, organic matter mass fraction, and heavy metal content. This data is automatically organized by time dimension to construct a structured basic data sequence. Through the system's data acquisition module and preprocessing algorithms, missing values ​​are linearly interpolated, outliers are removed by three times the standard deviation, and Z-Score standardization and one-hot encoding are uniformly applied to data of different dimensions. Finally, standard input data for downstream network processing is constructed.

[0181] The system calls a feedforward neural network to extract gradient vectors from the basic data, constructs a high-dimensional kernel similarity matrix by combining it with the neural tangent kernel, and performs kernel regression fitting through a sliding window to obtain implicit expressions at different time steps, forming a solid waste feature sequence. This sequence serves as the core input, fed into a Fourier neural network to predict the treatment effects of multiple candidate paths. The candidate paths include five different treatment modes: chemical extraction, pyrolysis gasification, bio-fermentation, acid-base neutralization, and centrifugal drying. Each path is associated with historical treatment records and recovery result indicators in the system. The Fourier neural network transforms the solid waste feature sequence into the frequency domain space, combines it with path feature vectors to perform multi-path prediction, and outputs the predicted resource conversion rate, energy consumption, treatment cycle, treatment cost, and residue ratio for each candidate path in the current batch.

[0182] Next, the system utilizes a Gumbel-Sinkhorn network for path ranking optimization. The prediction results are mapped to a path index matrix, and a path permutation matrix is ​​generated through perturbation and Sinkhorn normalization operations. The system selects the highest priority path as the target path and proceeds to the parameter matching stage. By invoking the Hungarian algorithm, a minimum matching graph is constructed between the feature sequence and the path parameters, and the parameter values ​​are extracted and mapped to the process control sequence. The control parameters include key process parameters such as processing temperature, material dosage ratio, reaction pH, stirring frequency, and processing time.

[0183] During the resource processing stage, the system drives the on-site processing equipment to automatically perform resource recovery operations according to the mapping instructions, and records the temperature, power consumption, feed amount, reaction time, and residual product ratio at each stage in real time during the processing. After each batch of processing is completed, the system automatically generates feedback data and calculates the difference with the previous prediction results to update the weights of the Fourier neural network and complete closed-loop optimization.

[0184] To verify the system performance, the present invention was compared with the traditional method of "fixed process + manual experience to set parameters" on five different batches of solid waste data. The test data are shown in the table below:

[0185] Table 1. Comparison of the effects of intelligent resource recovery systems and traditional methods.

[0186]

[0187] As shown in Table 1, the system proposed in this invention outperforms traditional treatment methods in several key dimensions: the average resource conversion rate is increased by 7.8 percentage points, the residue ratio is reduced by approximately 9 percentage points, the average energy consumption is reduced by more than 30 kWh / t, the treatment cost is reduced by approximately 60 yuan / t, and the average treatment cycle is shortened by more than 25 minutes. These data fully demonstrate that the system can intelligently select the optimal treatment path based on the specific properties of leather solid waste, avoiding resource waste and energy redundancy under traditional empirical approaches. It also possesses sustainable optimization and model iteration capabilities, significantly improving the precision and overall operational efficiency of leather solid waste resource utilization.

[0188] In summary, this invention not only achieves a leap from static rules to dynamic intelligent prediction in algorithm structure, but also fully verifies its significant advantages in improving resource recycling efficiency, reducing cost and energy consumption, and enhancing processing adaptability in practical applications, thus possessing broad promotional value and industrialization potential.

[0189] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A process management system for the resource utilization of leather solid waste based on big data analysis, characterized in that, include: The data acquisition module is used to collect and preprocess basic data on leather solid waste and construct a basic data sequence. The feature extraction module is used to construct a neural tangent kernel matrix by calculating gradient vectors of the basic data sequence through a feedforward neural network, and to perform kernel regression operation to extract high-dimensional implicit features at each time step to form a solid waste feature sequence. The conversion prediction module is used to construct a set of candidate paths based on existing historical processing records, and to perform inference prediction operations on the solid waste feature sequence on each candidate path through a Fourier neural network to obtain the prediction results of each candidate path. The path selection module is used to sort the prediction results by multiple objectives using the Gumbel-Sinkhorn network, construct the path permutation matrix and calculate the priority, and select the candidate path with the highest priority as the target path. The parameter matching module is used to construct a process control sequence based on the target path, and to perform a minimum cost matching operation with the solid waste feature sequence using the Hungarian algorithm, extracting the control process parameters corresponding to the matching results to form a process parameter set; The resource processing module is used to map the set of process parameters into execution instructions according to the preset action mapping table, perform resource recovery processing on leather solid waste according to the execution instructions, and record feedback data. The feedback update module is used to calculate the difference between the prediction result and the feedback data, and update the parameters of the Fourier neural network based on the difference result.

2. The leather solid waste resource utilization process management system based on big data analysis according to claim 1, characterized in that, The basic data includes batch number, source process, quality, moisture content, heavy metal content, and organic matter mass fraction of leather solid waste. The preprocessing includes missing data filling, anomaly removal, numerical standardization, category coding, time alignment, and format unification. The prediction results include predicted resource conversion rate, energy consumption, cost, residue quality, and processing cycle for each candidate path. The feedback data includes actual resource conversion rate, actual energy consumption, actual cost, actual residue quality, and actual processing cycle.

3. The leather solid waste resource utilization process management system based on big data analysis according to claim 1, characterized in that, The feature extraction module includes: The gradient direction calculation operation is performed on the basic data at each time step in the basic data sequence, and the gradient vector set corresponding to each time step is extracted through a feedforward neural network, specifically including: The basic data at each time step in the basic data sequence are numerically arranged, and a data vector is constructed in a fixed order. The data vector is fed into the feedforward neural network. In the hidden layer of the feedforward neural network, a weighted summation operation is performed according to the connection weights, and the hidden feature values ​​of the corresponding time step are obtained through nonlinear activation calculation. Under the condition that the connection weights remain unchanged, partial derivative calculation is performed on each dimension of the data vector based on the hidden feature values ​​to obtain the rate of change of the hidden feature values ​​in each dimension, and the gradient change value corresponding to each dimension is formed. The gradient change values ​​corresponding to each dimension within the same time step are numerically normalized and combined according to the order of the data vectors to form the gradient vector for that time step. Arrange all gradient vectors in chronological order to construct a gradient vector set; The gradient vector set at each time step is mapped by performing a kernel function mapping operation on the neural tangent kernel function to obtain the kernel similarity value between each time step. The neural tangent kernel matrix is ​​then constructed based on the kernel similarity value. The elements in the neural tangent kernel matrix represent the degree of similarity between the gradient vectors corresponding to any two time steps. A sliding window with a fixed step size is constructed for the neural tangent kernel matrix. Kernel regression is performed within each sliding window to extract high-dimensional implicit features at the corresponding time step, and the feature extraction results are continuously corrected. Arrange all the corrected high-dimensional implicit features in chronological order to construct a solid waste feature sequence.

4. The leather solid waste resource utilization process management system based on big data analysis according to claim 3, characterized in that, The construction process of the neural tangent kernel matrix includes: Perform numerical centering on each gradient vector in the gradient vector set according to the vector dimension, calculate the mean of each dimension over all time steps, and adjust the corresponding dimension values ​​at each time step based on the mean. The adjusted gradient vector is length normalized by summing the squares of the values ​​of each dimension in each gradient vector and taking the square root to obtain the vector length. The values ​​of each dimension are then scaled proportionally according to the vector length. Perform element-wise multiplication on the normalized gradient vectors of any two time steps according to the dimension, and sum them to obtain the inner product value between the corresponding two time steps. The inner product value represents the degree of consistency of the gradient change direction between the two time steps. The inner product values ​​are mapped using a neural tangent kernel function. The inner product values ​​are used as inputs, and the kernel similarity values ​​are obtained through nonlinear function transformation. The kernel similarity values ​​between all time steps are calculated and filled into a two-dimensional matrix structure in chronological order to construct a symmetric neural tangent kernel matrix.

5. The leather solid waste resource utilization process management system based on big data analysis according to claim 1, characterized in that, The conversion prediction module includes: The existing historical processing records are classified and organized, the path identifiers and resource utilization result data corresponding to each historical batch are extracted, a candidate path set is constructed according to the path identifiers, and a corresponding path feature index is established for each candidate path. The solid waste feature sequence is input into the Fourier neural network in chronological order. In the frequency domain mapping layer of the Fourier neural network, the solid waste feature sequence is subjected to spectral decomposition operation to generate the frequency domain representation sequence of the corresponding time step. In the frequency domain weight layer of the Fourier neural network, the frequency domain representation sequence is mapped to the path feature indices in the candidate path set. For each candidate path, a corresponding path response sequence and weight coefficients are generated, specifically including: According to the path feature index order in the candidate path set, each path feature index is numerically encoded to construct a path index sequence with the same dimension as the frequency domain representation sequence. In the frequency domain weighting layer, a weight parameter matrix with the same dimension as the path index sequence is set, and the corresponding weight vector is extracted according to the index position of the path feature index in the weight parameter matrix. Perform element-wise multiplication between the frequency domain representation vector of each time step in the frequency domain representation sequence and the corresponding weight vector to generate the weighted frequency domain vector of each time step; The weighted frequency domain vectors at each time step under the same candidate path are summed in chronological order to construct the path response sequence of the candidate path; The numerical values ​​in the path response sequence are normalized using the L2 norm, and the corresponding weight coefficients are extracted according to the path feature index order. Based on the weighting coefficients, the response sequences of each path are multiplied element by element and summed to form the path prediction vector of each candidate path. Numerical mapping processing is performed on the prediction vectors of each path according to the fixed mapping template, generating the predicted values ​​of resource conversion rate, processing energy consumption, processing cost, residue quality and processing cycle for each candidate path, and combining them to form the prediction results.

6. A leather solid waste resource utilization process management system based on big data analysis according to claim 5, characterized in that, The spectral decomposition operation specifically includes: For each time step in the solid waste feature sequence, construct a fixed-length vector according to the feature dimension order of the high-dimensional implicit features, and form a feature matrix according to the time order. In the frequency domain mapping layer, the fixed-length vector of each time step in the feature matrix is ​​used as input. The discrete Fourier transform results of the fixed-length vector of each time step on different preset frequency components are calculated and decomposed into real part values ​​and imaginary part values ​​to form a frequency component matrix. The real and imaginary values ​​of each frequency component in the frequency component matrix are squared and summed. The summation result is then squared to obtain the amplitude sequence of the corresponding frequency component. Simultaneously, the arctangent operation is performed with the ratio of the imaginary to the real values ​​as input to obtain the phase sequence of the corresponding frequency component. Numerical normalization is performed on each amplitude sequence to ensure that the amplitude sequences maintain a consistent scale within a preset range, and a unified reference angle calibration operation is performed on the phase sequence to ensure that the phases between different time steps have a consistent reference. The normalized amplitude sequence and the calibrated phase sequence are combined in chronological order to form a frequency domain representation sequence.

7. A leather solid waste resource utilization process management system based on big data analysis according to claim 1, characterized in that, The path selection module includes: The five predicted values ​​corresponding to each candidate path in the prediction results are arranged numerically according to a fixed index order to construct a path index set. The values ​​of each dimension in the path index set are then normalized to generate a path index matrix. The path index matrix is ​​input into the Gumbel-Sinkhorn network, random perturbation values ​​are added to the path index corresponding to each candidate path to generate a perturbation path matrix, and a row-by-row exponential mapping operation is performed on the perturbation path matrix to form a sorting weight matrix. Perform a fixed number of row and column normalization operations on the sorting weight matrix to ensure that the sum of the values ​​in each row and each column converges to a preset constant interval, forming a path permutation matrix. Specifically, this includes: Read the values ​​of each row in the sorted weight matrix, and calculate the sum of the row values ​​of all values ​​in each row in row order; The sorting weight matrix is ​​normalized based on the sum of the values ​​in each row, and all values ​​in each row are scaled proportionally so that the sum of the row values ​​in each row is within a preset constant range. Calculate the sum of the column values ​​of all values ​​in each column of the row normalization result in column order; The column normalization operation is performed on the row normalization result based on the sum of the values ​​in each column, and the scale is applied to all the values ​​in each column so that the sum of the column values ​​is within a preset constant range. The column normalization result is used as the new sorting weight matrix. Row normalization and column normalization operations are repeated until the preset number of times are reached to generate the path permutation matrix. The path priority is calculated based on the weight distribution corresponding to each row in the path permutation matrix, and the candidate path with the highest path priority is selected as the target path. The weight distribution represents the set of weights formed by arranging the values ​​of each row corresponding to the candidate path in the path permutation matrix in column order.

8. A leather solid waste resource utilization process management system based on big data analysis according to claim 7, characterized in that, The process of forming the sorting weight matrix includes: In the Gumbel-Sinkhorn network, the path index values ​​corresponding to each candidate path in the path index matrix are read, and the number of index dimensions for each candidate path is counted. For each candidate path, generate a random perturbation value that is consistent with the number of corresponding indicator dimensions, and add the random perturbation value to the path indicator value of the corresponding candidate path element by element according to the dimension to form the perturbation indicator value of the candidate path. Arrange the perturbation index values ​​of all candidate paths in the original candidate path order to construct a perturbation path matrix; Perform an exponential mapping operation on each element of each row of values ​​in the perturbation path matrix, so that the values ​​in each row are transformed into positive values ​​after the exponential mapping. Summing operation is performed on all exponential mapping results in the same row to obtain the exponential sum of the row. Then, based on the exponential sum, a scaling operation is performed on each exponential mapping result in the row to make the sum of all values ​​in the row meet the preset requirements. The values ​​in each row after scaling are arranged in the order of the candidate paths to form a sorting weight matrix.

9. A leather solid waste resource utilization process management system based on big data analysis according to claim 1, characterized in that, The parameter matching module includes: Read the flow control sequence corresponding to the target path from the preset flow mapping table; The Hungarian algorithm is used to extract the control flow parameters corresponding to each processing stage according to the order of each processing stage in the flow control sequence, and the number of processing stages is counted. High-dimensional implicit features corresponding to all time steps of the current batch of leather solid waste are extracted from the solid waste feature sequence. The feature arrangement sequence is constructed in chronological order and then divided into equal segments according to the number of processing stages to obtain multiple feature subsequences. For each processing stage, the control flow parameters are traversed, and each high-dimensional implicit feature in the corresponding feature subsequence is dimensionally aligned with the control flow parameters. The numerical difference between each corresponding dimension is calculated, and the absolute value summation operation is performed on all dimension differences to obtain the matching cost between the processing stage and the high-dimensional implicit feature. Arrange all matching costs in the same processing stage in chronological order to form a stage cost sequence; All stage cost sequences are combined in the order of processing stages to form a matching weight matrix, where rows represent processing stages and columns represent high-dimensional implicit features. Perform row standardization on the matching weight matrix, record the minimum matching cost in each row, and subtract the corresponding minimum matching cost from each value in that row; Perform column standardization on the row standardization results, record the minimum matching cost in each column, and subtract the corresponding minimum matching cost from each value in that column; Perform zero-value search and marking operations on the column standardization results to achieve unique assignment while ensuring that each column is marked only once, and generate an assignment result matrix; Based on the allocation result matrix, extract the high-dimensional implicit features corresponding to each processing stage, and map and assign the values ​​of each dimension of the high-dimensional implicit features according to the parameter order in the control flow parameters to generate a set of process parameters consistent with the order of the process control sequence.

10. A leather solid waste resource utilization process management system based on big data analysis according to claim 1, characterized in that, The resource processing module includes: Read the process parameter set, divide the process parameters in the process parameter set into stages according to the processing stage order in the process control sequence, and generate a stage parameter subset; Extract the processing action type corresponding to the parameter subset of each stage from the preset action mapping table, and perform mapping calculation on the processing action type according to the order in the parameter subset of the stage to generate a set of execution instructions including processing temperature, processing time, material feeding ratio and equipment operating frequency. The leather solid waste is processed for resource recovery according to the set of execution instructions, and the start and end times of each processing stage are recorded to form a stage record sequence. During the resource recovery process, the resource conversion rate, energy consumption, cost, residue quality and cycle of each processing stage are collected in real time and arranged in chronological order according to the stage record sequence to construct a stage feedback sequence. Perform a phase summary calculation on the phase feedback sequence to generate feedback data.

Images