Multi-component blended material optimization method based on near-infrared spectroscopy and deep learning

Abstract

This application relates to the field of textile resource recycling and automation control technology, and discloses a method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning. The method includes the following steps: collecting the original spectral sequence and physical timestamp of the circulating material, processing it through a one-dimensional convolutional neural network to output a benchmark proportion vector of chemical components; generating a benchmark ideal spectral sequence by combining a dictionary matrix; extracting the difference to obtain the residual vector, and using an orthogonal projection matrix to separate and calculate the mechanical degradation index of independent fibers; calculating the target licker-in roller speed accordingly, inputting it into a surrogate model to obtain the original fiber compensation ratio vector; calculating the action lag time by combining linear velocity and station distance, adjusting the opening equipment speed when the trigger condition is met, and controlling the mixing device to perform supplementary feeding. This invention uses spectral extraction of component proportions and orthogonal projection to separate physical residuals and calculate the independent fiber degradation index, achieving the technical effect of online quantitative analysis of the aging state of waste materials.

Description

Technical Field

[0001] This invention relates to the field of textile resource recycling and automated control technology, specifically to an optimization method for multi-component blended materials based on near-infrared spectroscopy and deep learning. Background Technology

[0002] Recycling and remanufacturing waste textiles is a crucial step in achieving the circular utilization of textile resources. Daily recycling involves a wide variety of multi-component blended garments, whose internal fiber structures age to varying degrees after prolonged wear and washing. Industrialized recycling processes require first understanding the true composition and mechanical state of the raw materials, then performing opening and carding, and finally adding high-quality virgin fibers as needed for respinning. The precision of this processing system directly determines the final mechanical strength of the recycled yarn. How to accurately adjust subsequent processing equipment and ingredient ratios based on the real-time state of the waste materials is a core technological challenge for the textile recycling industry in reducing losses and ensuring the quality of finished products.

[0003] Current textile recycling production lines generally incorporate basic spectral scanning and automated control technologies. Using near-infrared spectroscopy equipment, the system can quickly compare preset characteristic absorption peaks to achieve online qualitative classification of common chemical components such as cotton and polyester, meeting the efficiency requirements of large-scale assembly line sorting. The subsequent opening stage typically employs standardized high-speed licker-in roller equipment, which can quickly separate large pieces of fabric, ensuring the throughput of the entire production line. Regarding the formulation of supplementary virgin fibers, factories often rely on preliminary sampling experiments to establish a fixed blending ratio. This static formulation mechanism is simple to implement, requires very low hardware computing power from the underlying control system, and is very easy to deploy and scale up on traditional textile production lines.

[0004] Conventional spectral analysis is limited to extracting chemical absorption features for component classification. The accompanying physical scattering signals in the material spectrum are not separated. The system cannot quantitatively determine the true mechanical degradation degree of blended fibers. Processing control lacks an accurate aging state benchmark. Under these circumstances, the opening roller maintains a fixed high-speed operation for extended periods. The fragile aging components cannot withstand the empirically set strong tearing force. Secondary physical damage occurs immediately, with a large number of effective fibers forcibly broken into short fibers. To compensate for the strength loss caused by this indiscriminate opening, subsequent processes are forced to introduce virgin material compensation. Existing control mechanisms mostly issue commands in real time, detached from the actual physical scale of the production line. The detection probe is tens of meters away from the mixing device. There is an objective physical time difference in conveyor belt operation. The batching action lacks spatial lag calculation based on linear velocity. Compensating new material is often mis-fed to unrelated waste batches, directly causing drastic fluctuations in the breaking strength index of the final recycled yarn. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning. This method solves the technical problems in existing technologies, such as the inability to quantitatively analyze the degree of mechanical degradation of waste fibers, which leads to secondary damage to fibers due to the fixed rotation speed of the opening equipment, and the lack of physical spatial temporal matching in the feeding of virgin fibers, which causes fluctuations in the final yarn breaking strength.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning, comprising the following steps: The original spectral sequence and physical timestamp of the material flowing on the waste material recycling conveyor belt are collected using a near-infrared spectral detector. The original spectral sequence is then input into a one-dimensional convolutional neural network for processing, and the basic chemical composition reference ratio vector of the flowing material is output. The pre-stored pure component dictionary matrix is ​​called, and a mapping calculation is performed in combination with the basic chemical component reference ratio vector to generate a reference ideal spectral sequence; The difference between the original spectral sequence and the reference ideal spectral sequence is calculated, the overall physical spectral residual vector is extracted, and the overall physical spectral residual vector is separated using an orthogonal projection matrix to calculate the independent fiber mechanical degradation index of the chemical components. The target licker-in speed of the opening equipment is calculated based on the independent fiber mechanical degradation index, and the basic chemical component reference ratio vector, the independent fiber mechanical degradation index, and the target licker-in speed are input into the hybrid proxy model for calculation to obtain the virgin fiber compensation ratio vector. The linear speed of the waste material recycling conveyor belt is obtained using a speed measuring device. The action lag time is calculated in combination with the physical station distance parameter. When the system clock meets the triggering condition determined based on the physical timestamp and the action lag time, an execution command is output to adjust the rotation speed of the opening equipment to the target licker roller speed. The dynamic mixing device is then controlled to perform supplementary feeding action according to the original fiber compensation ratio vector.

[0007] Preferably, a reference ideal spectral sequence is generated by mapping the reference proportion vector of the basic chemical components, including: The ideal absorption-scattering ratio sequence is obtained by weighted summation of the single-component row vectors in the pure component dictionary matrix based on the baseline proportion vector of the basic chemical components. The ideal absorption-scattering ratio sequence is mapped inversely to the reflectance space based on the inverse transform function to generate the reference ideal spectral sequence.

[0008] Preferably, the pure component dictionary matrix is ​​obtained through the following steps: Obtain the historical spectral matrix and the known ratio matrix corresponding to the historical spectral matrix, and convert the historical spectral matrix into a historical absorption-scattering ratio matrix; An objective function is constructed using reconstruction error and a sparse penalty term. Under non-negative constraints, a non-negative matrix factorization algorithm is used to decompose and solve the historical absorption-scattering ratio matrix and the known proportion matrix to extract the pure component dictionary matrix.

[0009] Preferably, the overall physical spectral residual vector is separated using an orthogonal projection matrix, and the independent fiber mechanical degradation index of the chemical components is calculated, including: Multiply the overall physical spectral residual vector with the orthogonal projection matrix of the sensitive band corresponding to the chemical component to obtain the independent residual feature vector of the chemical component; Calculate the norm of the independent residual eigenvectors; When the relative mass fraction of the chemical component in the baseline proportion vector of the basic chemical components is greater than the existence determination threshold, the norm is normalized using the relative mass fraction, and the independent fiber mechanical degradation index is obtained by combining the preset sensitivity weight coefficient.

[0010] Preferably, calculating the target licker-in speed of the opening equipment based on the independent fiber mechanical degradation index includes: Determine whether the independent fiber mechanical degradation index of the brittle components in the preset brittle component set exceeds the aging tolerance threshold. When the aging tolerance threshold is exceeded, the difference between the independent fiber mechanical degradation index and the aging tolerance threshold is calculated, and the difference is multiplied by the rotational speed decay tuning coefficient to obtain the rotational speed decay of the brittle component. The reduced speed is obtained by subtracting the sum of the speed decay of all brittle components in the brittle component set from the base speed parameters; The reduced rotational speed is compared with a preset minimum safe rotational speed limit, and the maximum value is taken as the target licker-roll speed.

[0011] Preferably, the hybrid proxy model includes a parallel combination of a physical mechanism model and a Gaussian process error model; the process of obtaining the native fiber compensation ratio vector includes: The physical mechanism model performs a weighted superposition calculation based on the standard inherent strength of the chemical components, the baseline proportion vector of the basic chemical components, the mechanical degradation index of the independent fibers, the target licker-in roller speed, and the compensation ratio vector of the virgin fibers to be solved, and outputs a predicted value of the mechanism strength. The Gaussian process error model uses the basic chemical component reference ratio vector, the independent fiber mechanical degradation index, the target licker-in roller speed, and the original fiber compensation ratio vector to be solved as input features to output error compensation values. The estimated value of yarn strength is obtained by adding the estimated value of the mechanism strength to the estimated value of the error compensation.

[0012] Preferably, obtaining the native fiber compensation ratio vector further includes: The objective function is to minimize the sum of all supplementary mass ratios in the original fiber compensation ratio vector, and the constraint condition is that the estimated yarn strength is greater than the target yarn breaking strength. An optimization task is then constructed. The sequential quadratic programming algorithm is invoked to iteratively calculate the optimization task, and the minimum point at the convergence time is used as the original fiber compensation ratio vector.

[0013] Preferably, the motion lag time is calculated by combining the physical workstation distance parameter, including: Divide the first path length from the station where the near-infrared spectral detector is located to the station where the loosening equipment is located by the running linear velocity to calculate the lag time of the preceding action. Divide the second path length from the station where the near-infrared spectral detector is located to the station where the dynamic mixing device is located by the running linear velocity to calculate the lag time of the subsequent action.

[0014] Preferably, the output execution command adjusts the rotation speed of the opening device to the target licker-in roller speed, and controls the dynamic mixing device to perform supplementary feeding according to the virgin fiber compensation ratio vector, including: The physical timestamp, the target licker-in roller speed, and the original fiber compensation ratio vector are encapsulated into a control data frame and stored in the first-in-first-out data register queue. When the global clock is equal to the sum of the physical timestamp and the preceding action lag time, minus the preceding electrical response delay time, an execution command to adjust the target licker roller speed is output. When the global clock is equal to the sum of the physical timestamp and the lag time of the subsequent action, minus the delay time of the subsequent electrical response, an execution command is output to perform the supplementary feeding action according to the original fiber compensation ratio vector.

[0015] Preferably, the output execution instruction for performing the supplementary feeding action according to the original fiber compensation ratio vector includes: Read the pre-calibrated baseline linear density parameters of the waste material, multiply the baseline linear density parameters by the operating linear velocity, and calculate the absolute mass flow rate; The absolute mass flow rate is multiplied by the original fiber compensation ratio vector to calculate the absolute compensation mass feed rate; The execution command for controlling the dynamic mixing device to perform supplementary feeding action is output according to the absolute compensation mass feeding rate.

[0016] This invention provides a method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning. It offers the following advantages: 1. This invention employs near-infrared spectroscopy combined with a one-dimensional convolutional neural network to extract the proportions of basic components, and utilizes an orthogonal projection matrix to separate the physical spectral residuals to calculate the independent fiber mechanical degradation index. This achieves the technical effect of online quantitative analysis of the aging state of circulating waste blended materials. Compared with existing technologies that rely on offline sampling or simply on spectral absorption peaks for qualitative judgment, this invention overcomes the shortcomings of existing technologies that cannot accurately identify the physical damage characteristics of fibers, leading to ambiguity in subsequent processing standards.

[0017] 2. This invention employs a technical solution that dynamically calculates the speed reduction based on the mechanical degradation index assessment results of specific brittle components and adjusts the target licker roller speed of the opening equipment in real time. This achieves the technical effect of adaptively matching material strength and reducing secondary mechanical damage. Compared to the existing technology where opening machinery operates blindly with fixed empirical speed parameters for a long time, this invention solves the problem of easily causing excessive breakage of fragile fibers and severe loss of effective fiber length when dealing with variable recycled materials.

[0018] 3. This invention employs a parallel solution combining physical mechanisms and a Gaussian process error model to determine the optimal compensation ratio of virgin fibers. It also incorporates a technical solution that uses the physical workstation path length calculation to determine the lag time for material feeding triggering. This achieves precise material proportioning and tight synchronization of on-site spatial flow. Compared to existing technologies that rely on static formulas for rough material feeding and lack time-series tracking and verification, this invention overcomes the shortcomings of existing technologies where material delivery time differences cause misalignment in proportioning actions, leading to insufficient final yarn breaking strength. Attached Figure Description

[0019] Figure 1 This is a system architecture diagram of an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 This is a logic diagram of spectral data acquisition and component prediction in an embodiment of the present invention; Figure 4 This is a logical diagram of the ideal spectral reconstruction operation in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the decoupling of physical degradation features according to an embodiment of the present invention. Figure 6This is a logic diagram for the joint solution of material and machine parameters according to an embodiment of the present invention. Figure 7 This is a diagram illustrating the spatiotemporal synchronization control logic of an embodiment of the present invention. Figure 8 The data curves for the spectral analysis and optimization process of blended materials in an embodiment of the present invention are shown in (a) and (b) respectively. (a) is a comparison curve of actual spectrum, ideal spectrum and residual vector, and (b) is an iterative optimization convergence diagram of the compensation ratio of native fibers.

[0020] Among them, 101, near-infrared spectral detector; 102, velocity measurement device; 103, computing server; 10, data analysis and prediction module; 20, ideal spectrum reconstruction module; 30, residual feature decoupling module; 40, machine and material parameter joint solution module; 50, spatiotemporal alignment execution module; 104, programmable logic controller; 105, frequency conversion drive device; 106, opening equipment; 107, dynamic mixing device. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Reference Figure 1 The present invention provides a multi-component blended material optimization system based on near-infrared spectroscopy and deep learning, which may include: a near-infrared spectral detector 101, a velocity measurement device 102, a computing server 103, and a programmable logic controller 104.

[0023] A near-infrared spectrometer 101 is positioned directly above the waste material recycling conveyor belt, facing the conveyor belt surface, to collect the diffuse reflectance spectral sequence of the flowing material. A speed measuring device 102 is installed at the end of the conveyor belt drive shaft to measure the linear speed of the conveyor belt.

[0024] Near-infrared spectral detector 101 and velocity measuring device 102 establish data communication connections with computing server 103. Computing server 103 is communicatively connected to programmable logic controller 104. The output terminals of programmable logic controller 104 are electrically connected to frequency converter drive device 105 and dynamic mixing device 107, respectively.

[0025] The variable frequency drive device 105 is used to receive frequency adjustment signals and control the rotational speed of the licker-in roller of the opening equipment 106. The dynamic mixing device 107 is arranged in the downstream material flow channel of the opening equipment 106 and is used to replenish the material flow channel with a certain amount of virgin fiber material.

[0026] The computing server 103 is configured with various processing modules that execute logic programs. The computing server 103 has built-in data parsing and prediction module 10, ideal spectrum reconstruction module 20, residual feature decoupling module 30, joint solution module for machine and material parameters 40, and spatiotemporal alignment execution module 50.

[0027] The data parsing and prediction module 10 is configured to receive the spectral sequence and timestamp collected by the near-infrared spectral detector 101, process the input spectral data using a deep learning network, and output the baseline proportion vector of various chemical components in the target material.

[0028] The ideal spectrum reconstruction module 20 is communicatively connected to the data parsing and prediction module 10. It is configured to call the pre-stored pure component dictionary matrix and perform nonlinear inverse mapping calculation in the absorption and scattering ratio space in combination with the basic chemical component reference ratio vector to generate the reference ideal spectrum sequence.

[0029] The residual feature decoupling module 30 is communicatively connected to the data parsing and prediction module 10 and the ideal spectrum reconstruction module 20, respectively. It is configured to extract the difference between the actual spectrum and the reference ideal spectrum, and use the basic chemical component reference ratio vector as the constraint parameter. It uses the orthogonal projection matrix to separate the physical spectrum residuals and calculate the independent degradation index of each component.

[0030] The joint solution module 40 for machine and material parameters is connected to the residual feature decoupling module 30. It is configured to evaluate the mechanical speed parameters required by the opening equipment 106 based on the independent degradation index, and calculate the original fiber compensation ratio vector by combining the internally constructed hybrid proxy model and the target yarn strength constraint.

[0031] The spatiotemporal alignment execution module 50 is communicatively connected to the machine and material parameter joint solution module 40 and the speed measurement device 102, respectively. It is configured to calculate the action lag time using the physical distance of the workstation and the running linear speed, and send instructions to the programmable logic controller 104 when the system clock triggers the lag time condition.

[0032] Reference Figure 2 This invention provides a method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning, comprising the following steps: S100 uses the near-infrared spectral detector 101 to collect the original spectral sequence and timestamp, and calculates and outputs the basic chemical component benchmark ratio vector through the data analysis and prediction module 10. S200: The ideal spectral reconstruction module 20 is used to obtain the basic chemical component reference ratio vector, and nonlinear inverse mapping calculation is performed in the absorption scattering ratio space to generate the ideal spectral sequence. S300 uses the residual feature decoupling module 30 to extract the physical spectrum residual vector and calculates and outputs the independent fiber mechanical degradation index of various components to form a multidimensional physical degradation vector. S400 uses the machine and material parameter joint solution module 40 to calculate the target licker-roll speed of the opening equipment 106 based on the multi-dimensional physical degradation vector, and calculates the virgin fiber compensation ratio vector through the hybrid proxy model. S500 uses the time-space alignment execution module 50 to calculate the lag time of the control node. When the time condition is met, the programmable logic controller 104 adjusts the frequency conversion drive device 105 and controls the dynamic mixing device 107 to perform the feeding action.

[0033] The following will provide a detailed explanation of each step involved in the method.

[0034] Reference Figure 3 In this embodiment, the specific implementation of step S100 can be further divided into the following sub-steps: S110 is equipped with a near-infrared spectral detector 101 to collect diffuse reflectance spectra, generate a one-dimensional original spectral sequence, and simultaneously record physical timestamps and running linear velocity parameters.

[0035] During the operation of the waste material recycling production line, the built-in light source of the near-infrared spectrometer 101 illuminates the surface of the mixed material flowing directly below it. From a physical perspective, because the chemical bonds of hydrogen-containing groups in different chemical fibers have specific harmonic and combination frequency absorption characteristics in the near-infrared band, the material surface will absorb and scatter near-infrared light of specific wavelengths. The unabsorbed light forms a diffuse reflection signal, which is captured by the receiving field of view of the near-infrared spectrometer 101.

[0036] The photoelectric sensing array inside the near-infrared spectrometer 101 converts continuous diffuse reflection light signals into corresponding analog voltage signals. After high-frequency sampling by the built-in analog-to-digital converter, the continuous analog voltage signals are discretized to generate a one-dimensional original spectral sequence corresponding to a specific wavelength channel. The acquired one-dimensional original spectral sequence of the mixture is set as follows: ,in This represents the total number of discrete wavelength points, and its specific value depends on the resolution parameters of the spectral detector used.

[0037] To ensure spatial accuracy of subsequent instruction execution, the system reads the internal clock register while the near-infrared spectrometer 101 completes a single scan cycle, thereby obtaining the physical timestamp of the current material batch. The system synchronously calculates the constant linear speed of the conveyor belt by reading the pulse feedback values ​​from the speed measuring device 102. The system will generate a one-dimensional original spectral sequence. Physical timestamp With running linear velocity The data is packaged and encapsulated into structured data frames with time and space markers, and stored in the buffer memory of the computing server.

[0038] S120 processes the one-dimensional original spectral sequence through a one-dimensional convolutional neural network of the data parsing and prediction module 10, and calculates and outputs the basic chemical component reference ratio vector.

[0039] The data parsing and prediction module 10 reads structured data frames from buffer memory and extracts the one-dimensional raw spectral sequences. This serves as the input feature vector for a pre-trained one-dimensional convolutional neural network. Before being fed into the network, the module performs normalization preprocessing on the input sequence through standard normal variable transformation to suppress optical path changes and baseline drift caused by uneven material stacking height, thus preventing the algorithm from entering the zero-value dead zone.

[0040] In a preferred embodiment, the topology of this one-dimensional convolutional neural network sequentially includes an input layer, multiple alternating cascaded one-dimensional convolutional layers and max-pooling layers, a fully connected layer, and a normalized output layer. The one-dimensional convolutional layers employ one-dimensional convolutional kernels with sizes ranging from 3 to 7, along the one-dimensional original spectral sequence. A sliding window is used to calculate the absorption peak features in adjacent wavelength bands in a locally perceptive manner. The fully connected layer is responsible for concatenating and mapping the local features obtained after pooling to the global component space.

[0041] For the training process of this network model, the system needs to pre-acquire a large number of historical blended material spectra covering different component ratios to construct a training set. The input of the training samples is the historical spectral sequence, and the corresponding sample labels are the actual chemical component mass fractions calibrated offline using high-precision methods such as chemical dissolution. During model training, loss functions such as mean squared error are used, and the weights and bias parameters in the network are continuously updated through the backpropagation algorithm until the loss value on the validation set converges to the preset error tolerance range.

[0042] Let the forward mapping function of a one-dimensional convolutional neural network be... ,in These are the weights and bias parameters that have been trained and solidified in the network using a large number of historical samples. The data parsing and prediction module 10 outputs the baseline chemical composition proportion vector of the material through forward propagation. The mathematical expression of this calculation process is as follows: ; in, The total number of target chemical components set for the system is usually set to... Elements within a vector The network prediction yields the first... The relative mass fraction of a chemical component in the current mixture.

[0043] To conform to the mass conservation principle of the physical system, the output layer of a one-dimensional convolutional neural network is configured with Softmax activation nodes, aiming to constrain the output baseline chemical component proportion vector. The dimensional constraint condition that the sum of the mass fractions of all components equals 1 is satisfied. This constraint condition is expressed as: ; Under this calculation mechanism, the basic chemical component reference ratio vector It only reflects the macroscopic chemical composition of the material and characterizes the theoretical chemical composition ratio of the material under the condition of no physical wear. It does not include information on structural damage such as fiber breakage caused by mechanical impact or thermal oxidation during the actual recycling process.

[0044] Reference Figure 4 In this embodiment, the specific implementation of step S200 can be further divided into the following sub-steps: S210 uses a non-negative matrix factorization algorithm with matching constraints to process historical production data and extract a dictionary matrix of pure components.

[0045] Obtaining absolutely pure, un-physically aged single-component materials prepared in the laboratory requires significant costs. To address the challenge of physical calibration of single-component dictionaries, this embodiment employs data-driven adaptive extraction based on historical production line data.

[0046] The ideal spectrum reconstruction module 20 reads historical regular production batch data accumulated in the system database and constructs a historical spectrum matrix. and the corresponding known proportion matrix .in, This represents the total number of historical samples retrieved. The number of discrete wavelength points. This represents the total number of chemical components. To avoid meaningless solutions from subsequent matrix decomposition, the ideal spectral reconstruction module 20 performs rank verification before constructing the matrix to ensure that the selected historical samples have sufficient linearly independent diversity in component proportions.

[0047] Since the near-infrared diffuse reflectance spectrum exhibits a nonlinear relationship with the substance concentration in the original reflectance space, the ideal spectral reconstruction module 20 will reconstruct the historical spectral matrix. The reflectance data in the model is converted into absorption-scattering ratio data in Kubelka-Munk space to generate a historical absorption-scattering ratio matrix. In physical optics theory, the ratio of a substance's absorption coefficient to its scattering coefficient is directly proportional to the concentration of its chemical components. Let the matrix... The specific reflectance at a certain wavelength point is The specific mathematical expression for this conversion process is: ; In the actual calculation of this formula, in order to prevent the measured reflectivity from approaching zero due to the dark current noise of the device and causing a division-to-zero anomaly, the module will impose a lower limit truncation restriction on the input reflectivity.

[0048] After completing the spatial transformation, the ideal spectral reconstruction module 20 constructs a nonnegative matrix decomposition optimization problem with known stoichiometric constraints to calculate the dictionary matrix of pure components. The objective function expression for this optimization problem is: ; The objective function is subject to nonnegativity constraints: .

[0049] in, The Frobenius norm is used to measure reconstruction error. Let L1 be the norm of the dictionary matrix of pure components; This is the coefficient for the set sparsity penalty term. As a preferred approach, this coefficient... The value of is set between 0.01 and 0.1, and its specific value can be determined by performing grid search and cross-validation on the validation set spectrum. Introducing a sparsity penalty term makes the extracted spectral features purer and eliminates broadband baseline drift interference. The non-negativity constraint conforms to physical reality, meaning that the absorption-to-scattering ratio of a substance cannot be negative.

[0050] By employing numerical calculation methods such as the multiplier update rule or alternating least squares method to solve the above optimization problem, the ideal spectrum reconstruction module 20 extracts a feature dictionary, i.e., a pure component dictionary matrix, from a large historical sample containing complex noise and nonlinear physical interference. Each row vector in this matrix represents the ideal optical response of a single pure component in Kubelka-Munk space.

[0051] S220 combines the baseline proportion vector of basic chemical components with the dictionary matrix of pure components to generate a baseline ideal spectral sequence through nonlinear inverse mapping.

[0052] According to the additivity principle of the Kubelka-Munk theory, the total absorption-scattering ratio of a mixture is equivalent to the linear superposition of the absorption-scattering ratios of each individual component according to their mass proportions. The ideal spectrum reconstruction module 20 acquires the basic chemical component reference proportion vector output by the data analysis and prediction module 10. Within the Kubelka-Munk space, the ideal spectral reconstruction module 20 calculates the ideal absorption-scattering ratio sequence of the current mixture. The specific calculation formula is as follows: ; in, Basic chemical component reference ratio vector The Middle The relative mass fraction of the chemical components; Dictionary matrix representing pure components The corresponding number in the middle Row vectors of chemical components.

[0053] Since subsequent difference calculations need to be consistent with the dimensions of the actual acquisition equipment, the ideal spectral reconstruction module 20 applies the Kubelka-Munk inverse transform formula to convert the ideal absorption-scattering ratio sequence... By inversely mapping to reflectance space, a reconstructed reference ideal spectral sequence is generated. Based on the root-finding rule for bivariate equations and combined with the physical boundary constraints of reflectivity between 0 and 1, the mathematical expression for this inverse transform is: ; The reference ideal spectral sequence obtained through the above inverse mapping calculation This sequence characterizes only the absorption properties of pure chemical components at the corresponding formulation. It excludes physical scattering interference introduced during material processing and recycling, serving as a reference surface for subsequent feature decoupling calculations.

[0054] Reference Figure 5 In this embodiment, the specific implementation of step S300 can be further divided into the following sub-steps: S310 utilizes the residual feature decoupling module 30 to perform differential stripping operation on the physical spectral residual vector.

[0055] The residual feature decoupling module 30 reads the actual acquired one-dimensional raw spectral sequences from the system cache. and the reference ideal spectral sequence generated by the ideal spectral reconstruction module 20. Under the same data dimension space, the residual feature decoupling module 30 performs a difference calculation on the vectors of the two sequences mentioned above to extract the overall physical spectrum residual vector. The mathematical expression for this difference operation is: ; From a physical optics perspective, the scattering coefficient of photons propagating in a medium is positively correlated with the material's internal porosity and surface roughness. Stripping away the intrinsic absorption matrix, dominated by macroscopic chemical concentration, from the spectral signals of mixtures helps to highlight changes in the microscopic physical structure. This is due to the acquisition of a benchmark ideal spectral sequence. The overall physical spectrum residual vector remaining after subtraction operation, representing only the pure and undamaged chemical absorption response. It can independently reveal the nonlinear physical scattering characteristics caused by fiber surface wear, internal fracture, and increased porosity.

[0056] S320 utilizes the residual feature decoupling module 30 to perform directional orthogonal projection based on prior component constraints, and calculates the independent fiber mechanical degradation index of various components.

[0057] Because the physical aging optical characteristics of different fibers in blended materials exhibit significant band overlap, directly using the overall physical residual assessment can easily introduce coupling errors. To reduce such interference, the residual characteristic decoupling module 30 calls the orthogonal projection matrix of the sensitive band preset for each component. (Setting the...) The orthogonal projection matrix of the sensitive bands of the components is as follows ,in This represents the original number of wavelength points. This represents the feature dimension of the subspace after dimensionality reduction. As a preferred approach, the orthogonal projection matrix of this sensitive band... The orthogonal basis can be determined by pre-collecting the sample spectra of a single component at different aging stages and performing principal component analysis or singular value decomposition on its residual matrix to extract its characteristic orthogonal basis.

[0058] The residual feature decoupling module 30 will decouple the overall physical spectrum residual vector Orthogonal projection matrix of sensitive band Multiply them, map them to the feature subspace of a specific component, and calculate the first... Independent residual eigenvectors of each component : ; After obtaining the independent residual feature vectors of each component, the residual feature decoupling module 30 combines the basic chemical component benchmark proportion vector output by the data analysis and prediction module 10. Normalization calculations are performed to eliminate the influence of differences in the absolute content of materials on scattering intensity.

[0059] In real-world industrial environments, when a batch of material contains absolutely no certain component, the underlying noise of the sensor can cause a small positive value in the L2 norm of the corresponding independent residual eigenvector. Directly performing division can easily amplify the noise and output false high-degradation characteristic parameters. Therefore, the module introduces a truncation discrimination mechanism before calculation. The residual feature decoupling module 30 calculates the... Independent fiber mechanical degradation index of each component The specific calculation formula is as follows: ; in, Independent residual eigenvectors The L2 norm is used to quantify the scattering energy intensity within this subspace; Basic chemical component reference ratio vector The Middle The relative mass fraction of the chemical components; The sensitivity weighting coefficient for the corresponding component can be determined by linear regression fitting between destructive mechanical tensile test data and optical residual data of historical samples. The system sets a threshold value for the components to be judged, which is usually set between 0.01 and 0.05.

[0060] After calculating the degradation index of each monomer component, the residual feature decoupling module 30 aggregates and splices the values ​​to form a multidimensional physical degradation vector that characterizes the overall structural damage state of the mixture. : ; This multidimensional physical degradation vector serves as a feedforward parameter for subsequent machine-material collaborative optimization control, providing a quantitative physical basis for the system to dynamically adjust the mechanical operating state and material compensation ratio.

[0061] Reference Figure 6 In this embodiment, the specific implementation of step S400 can be further divided into the following sub-steps: S410, using the joint solution module 40 for machine and material parameters, calculates the target licker roller speed of the opening equipment 106 based on the degradation index of a specific brittle component in the multidimensional physical degradation vector.

[0062] Different materials in blended fabrics exhibit varying degrees of resistance to mechanical impact. Brittle fibers such as regenerated cellulose experience a decrease in structural strength after physical aging during the recycling cycle, making them prone to secondary cutting damage if processed using standard mechanical parameters. To suppress this excessive tearing, the machine-material parameter joint solution module 40 uses a multi-dimensional physical degradation vector decoupled from the front end. Adaptive degradation control is implemented on the upstream mechanical parameters.

[0063] The joint solution module 40 for machine and material parameters predefines a set of indices of brittle components susceptible to mechanical damage within the system. For each component within this set, the system combines its corresponding independent fiber mechanical degradation index. The system is conditionally matched against a preset aging tolerance threshold. To prevent an excessively high degradation index from causing a negative calculated speed and triggering a frequency converter malfunction, the machine and material parameter joint solution module 40 calculates the optimal target licker-roll speed for the downstream opening equipment 106. The specific calculation equation is as follows: ; in, Basic rotational speed parameters set for industrial control systems for standard non-destructive materials, unit: rpm; To maintain the minimum safe operating speed limit for the equipment's basic throughput; For the corresponding number The rotational speed attenuation tuning factor for a brittle component is also in rpm, and its specific value is determined based on the slope of the fracture strength test curve of the component. This represents the tolerable aging index threshold for this component; internal This is the function for finding the maximum value.

[0064] In this operational structure, the following is introduced This forms a nonlinear dead zone control mechanism. When the degradation index of a specific component falls below the tolerance threshold... When the difference is negative, the internal... When the function output is zero, the system maintains the base speed to ensure opening efficiency; only when the degree of physical damage exceeds the safety threshold will the system trigger speed decay calculation, thereby establishing a physical balance between material protection and processing efficiency.

[0065] S420 utilizes the joint solution module 40 for machine and material parameters to construct a hybrid proxy model guided by physical mechanisms, and performs numerical optimization calculation of the original fiber compensation ratio vector under constraints.

[0066] After determining the speed control parameters on the upstream mechanical side, the system needs to simultaneously calculate the material compensation scheme on the downstream chemical side. For recycled materials containing physical defects, incorporating high-quality virgin fibers can compensate for the mechanical shortcomings of the final yarn. The required virgin fiber compensation ratio vector to be added to the current material flow is set as follows: ,in The supplementary number The ratio of the mass of the virgin fiber to the total mass of the current original batch is a non-negative real number.

[0067] Because collecting a large number of training samples labeled "yarn breaking strength" in the online environment of long-process spinning is time-consuming, purely data-driven neural network models are prone to overfitting due to insufficient samples. The machine and material parameter joint solution module 40 constructs a hybrid surrogate model that combines a physical analytical model and a data-driven model in parallel. .

[0068] This hybrid surrogate model is used to calculate the estimated yarn strength under the current condition. Its mathematical structure is defined as follows: ; in, This is an analytical mechanism sub-model based on the physical mixing principle. According to the system's physical topology, the virgin fiber is mixed in downstream after the opening and beating process; therefore, it does not possess the physical degradation properties of the front end and is not subject to secondary mechanical damage from the rotational speed of the licker-in roller. Based on the classical mechanics theory of composite materials, the macroscopic mechanical strength of a multi-component system depends on the weighted sum of the mass fraction of each component and its inherent strength. The component strength of the recycled material is affected by the nonlinear reduction caused by physical aging and machining parameters, while the virgin fiber retains its standard inherent strength. As a preferred approach, the specific calculation formula for this mechanism sub-model is as follows: ; in, For the first Standard inherent strength of the fiber components; This is the empirical reduction constant for system fitting. The formula independently maps and weights the reduced strength of recycled materials to the standard strength of virgin fibers, and uses the denominator to normalize the quality after adding compensating materials, ensuring strict consistency between the underlying mathematical calculation logic of the algorithm and the overall hardware layout and execution timing of the system.

[0069] This is a sub-model for error compensation in a Gaussian process, and its input feature vector is... Integrated , , as well as This Gaussian process model uses nonparametric Bayesian inference to capture nonlinear cross-feature coupling effects and system residuals that are difficult to describe by the mechanistic sub-model. In this embodiment, the Gaussian process uses radial basis functions (RBFs) as the covariance kernel function.

[0070] For the error compensation sub-model of this Gaussian process For training and consolidation, during the trial operation phase, the system used offline sampling to place the actually spun yarn on an electronic single yarn strength tester for tensile and breakage tests, obtaining the actual physical strength values ​​as label data. The actual strength values ​​and the mechanism sub-model were then used as the basis for the label data. The residuals of the output values ​​are used as the training target, and the length scale and variance hyperparameter of the RBF kernel function are optimized using the marginal log-likelihood loss function.

[0071] After the model is built, the machine and material parameter joint solution module 40 constructs an optimization constraint search task. Let the minimum target yarn breaking strength required for industrial production be... The module takes minimizing the total consumption of native fibers as its objective function: ; The objective function is subject to the boundary physical constraints of yarn strength and the non-negative mass constraint: ; The joint solution module for machine and material parameters 40 calls its internal Sequential Quadratic Programming (SQP) algorithm as a nonlinear solver to iteratively optimize the aforementioned constrained problem. In each iteration, this algorithm approximates the optimal solution by constructing a quadratic approximation subproblem of the Lagrangian function. The minimum point obtained through numerical solution represents the original fiber compensation ratio vector that satisfies the mechanical performance requirements and minimizes cost. This vector, along with the target roller rotation speed, is output as a control command to the execution logic layer.

[0072] Reference Figure 7 In this embodiment, the specific implementation of step S500 can be further divided into the following sub-steps: S510 uses the spatiotemporal alignment execution module 50 to calculate the station logistics delay time parameters based on the equipment's physical topology.

[0073] In a continuous waste material recycling production line, the computation time of the algorithm model and the physical distance between different processing stations can cause spatial control misalignment. If hardware actions are executed immediately after the machine-material parameter joint solution module 40 outputs parameters, the adjusted control state may be applied to the wrong batch of materials. To reduce this spatial misalignment error, the spatiotemporal alignment execution module 50 performs delayed alignment calculations based on the equipment topology layout.

[0074] The spatiotemporal alignment execution module 50 acquires the structural distance parameters from the spectral acquisition station where the near-infrared spectral detector 101 is located to different downstream physical stations. Let the physical flow path length from the spectral acquisition station to the downstream opening equipment 106 licker-in roller station be... The physical flow path length to the downstream dynamic mixing device 107 multi-bin weighing hopper station is... Combined with the constant linear speed of the conveyor belt obtained at the system's data acquisition front end. The spatiotemporal alignment execution module 50 calculates the corresponding upstream and downstream logistics transmission lag time parameters based on kinematic equations. The specific physical calculation formula is as follows: ; ; in, Characterizes the physical time period required for a specific batch of material to be transported from the start of spectral scanning to the station where the opening equipment 106 is located; This represents the time period required for the same batch of materials to arrive at the dynamic mixing device 107. The above calculation logic establishes a mapping relationship between the control system's time base and the physical progress of logistics transportation.

[0075] The S520 uses the space-time alignment execution module 50 to compare the global system clock and trigger an interrupt to issue a space-time synchronization instruction through a first-in-first-out queue buffer mechanism.

[0076] After acquiring the material control parameters and lag time, the calculation server 103 and the programmable logic controller 104 collaboratively perform hardware-level timing matching. The spatiotemporal alignment execution module 50 pushes the control tuple data frame calculated by the front end into the first-in-first-out (FIFO) data register queue allocated in the memory of the programmable logic controller 104. This control tuple contains the target licker roll speed obtained from the previous calculation. , original fiber compensation ratio vector And the physical timestamp recorded when the batch of materials was initially scanned. .

[0077] The First-In-First-Out (FIFO) queue design is intended to handle the continuous flow and accumulation of multiple batches of materials on a production line. The Programmable Logic Controller 104 (PLC104) uses a timer interrupt service routine at its lowest level to poll and monitor the system's global clock at a fixed frequency. And extract the tuple data from the head of the queue to perform time condition judgment.

[0078] As a preferred approach, considering the inherent response time of industrial fieldbus communication and the speed change of the inverter-driven motor, the electrical response delay time of the preceding action is set to [value missing]. When the system global clock meets the time alignment requirements When this occurs, it indicates that the corresponding batch of material is about to move into the mechanical impact range of the opening equipment 106. The programmable logic controller 104 then activates the previous execution interrupt, based on the extracted target licker roller speed. The frequency adjustment command is sent to the variable frequency drive unit 105 via the industrial communication bus. The variable frequency drive unit 105 changes the stator power supply frequency output to the three-phase AC motor, causing the speed of the licker roller of the opening device 106 to switch to a new speed at the instant of material contact. This enables adaptive mechanical protection for damaged materials. At this point, the controller only updates the preceding execution completion flag of the tuple and does not remove it from the queue.

[0079] As the batch of materials continues to flow backward in the channel, the electrical response delay time for the subsequent actions is set to... When the system global clock further satisfies the back-end time alignment condition. At this time, the programmable logic controller 104 activates the subsequent execution interrupt. At this point, the programmable logic controller 104, based on the parsed native fiber compensation ratio vector... The system sends an action timing command to the dynamic mixing device 107. The multi-compartment stepper motor assembly inside the dynamic mixing device 107 starts. This is due to the target virgin fiber proportioning vector... As a relative mass ratio parameter, to ensure accurate feeding during continuous operation, in this embodiment, the system calculation server 103 pre-calibrates and stores the basic linear density parameter of the waste material. This basic linear density is ensured by a conventional pre-processing homogenization web laying technique. The control system uses real-time readings of the operating linear speed. With this constant linear density The system performs dynamic multiplication calculations to obtain the real-time dynamic absolute mass flow rate of the current logistics flow. The programmable logic controller 104 then compares this dynamic absolute mass flow rate with the parsed ratio vector. Multiplying these values ​​yields the absolute compensation mass feeding rate, which is then used to control the opening of the corresponding feeding valve and the motor speed in real time, precisely and quantitatively mixing the corresponding virgin fibers into the current material flow.

[0080] After confirming that the actions of both the preceding and following workstations corresponding to the tuple data have been completed, the programmable logic controller 104 pops the tuple data from the head of the first-in-first-out queue and clears it, releasing memory space and waiting for the synchronous interrupt trigger of the next batch of materials. This hardware-software decoupling and queue step-by-step dwell mechanism helps maintain the orderliness of closed-loop compensation under continuous production cycle.

[0081] To further clarify the collaborative working process of the technical solution described in this invention, a specific working scenario example will be used for illustration below.

[0082] The waste blended materials processed by the current recycling line are set to mainly consist of two components: cotton fiber and polyester fiber.

[0083] Reference Figure 8 When the material passes through the detection area, the near-infrared spectrometer 101 collects the diffuse reflectance spectrum of the material surface. Simultaneously, the speed measuring device 102 measures the current linear speed of the conveyor belt as 2.0 m / s, and the system records the current physical timestamp. The data analysis and prediction module 10 receives the spectral data and inputs it into a one-dimensional convolutional neural network, calculating and outputting a baseline proportion vector of basic chemical components. In this working scenario, the model predicts that the relative mass fraction of cotton fiber in the current batch of material is 0.65, and the relative mass fraction of polyester fiber is 0.35.

[0084] The ideal spectrum reconstruction module 20 obtains the reference scale vector, calls the system's pre-stored pure cotton fiber dictionary and pure polyester fiber dictionary, linearly superimposes them in the Kubelka-Munk space according to weights of 0.65 and 0.35, and generates the reference ideal spectrum sequence through inverse transformation. For example... Figure 8As shown in (a), the dashed line represents the actual one-dimensional original spectrum collected, and the solid line represents the reconstructed reference ideal spectrum. There is a significant deviation in the shape between the two.

[0085] The residual feature decoupling module 30 performs vector subtraction between the actual spectrum and the reference ideal spectrum to extract features such as... Figure 8 (a) shows the overall physical spectral residual vector as indicated by the dashed line. Subsequently, the module calls the orthogonal projection matrices of the sensitive bands for cotton and polyester fibers respectively to perform dimensionality reduction projection on the residual vectors, and calculates the independent degradation index in conjunction with the baseline scaling vector. The calculation results show that the independent fiber mechanical degradation index for cotton fiber is 0.42, and the degradation index for polyester fiber is 0.12.

[0086] The joint solution module 40 for machine and material parameters obtains a multidimensional physical degradation vector. The system presupposes that cotton fiber is a brittle component and sets its aging tolerance threshold to 0.20. The current degradation index of the cotton fiber, 0.42, exceeds this threshold, triggering the mechanical speed reduction control logic. The basic rotational speed parameter of the opening equipment 106 is set to 800 rpm. Based on the formula calculation, the target licker-in roller speed is adjusted to 660 rpm to reduce secondary mechanical impact on the damaged cotton fiber.

[0087] After the rotational speed parameters are determined, the machine and material parameter joint solution module 40 initiates the virgin fiber compensation calculation. The system sets the target yarn breaking strength requirement to 15.0 cN / tex. The module inputs the baseline ratio, degradation index, and adjusted rotational speed of 660 rpm into the hybrid surrogate model composed of the mechanistic sub-model and the Gaussian error model. Figure 8 As shown in (b), the sequential quadratic programming algorithm begins its iteration with the objective of minimizing the cost of the compensated material. After multiple iterations, the objective function converges. The solution results show that, to meet the strength requirements, the virgin fiber compensation ratio vector needs to be set to add an additional 0.14% mass of virgin cotton fiber to the current material, without the need to add polyester fiber.

[0088] The spatiotemporal alignment execution module 50 calculates the lag time based on the physical distances between the equipment. The physical distance from the spectral acquisition station to the opening device 106 is set to 5.0m, and the physical distance to the dynamic mixing device 107 is set to 12.0m. Combined with a linear operating velocity of 2.0m / s, the lag time for the preceding action is calculated to be 2.5s, and the lag time for the following action is 6.0s. The module stores the target rotational speed of 660rpm, the compensation ratio of 0.14, and the corresponding timestamps in a first-in, first-out queue.

[0089] After the system clock has elapsed for 2.5 seconds and the inverter's electrical response time has been deducted, the programmable logic controller 104 triggers a preceding interrupt, controlling the variable frequency drive 105 to switch the licker-in roller speed to 660 rpm. After the system clock has elapsed for 6.0 seconds and the motor response time has been deducted, the programmable logic controller 104 triggers a following interrupt. At this time, the controller calculates the absolute mass flow rate based on the current linear density and linear velocity, and controls the stepper motor and valves of the dynamic mixing device 107 according to the ratio parameter of 0.14, feeding a fixed amount of virgin cotton fiber into the material flow channel. After execution, the control element is cleared from the queue, and the system completes the closed-loop control of this batch of materials.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-component blend material optimization method based on near-infrared spectroscopy and deep learning, characterized in that, Includes the following steps: The original spectral sequence and physical timestamp of the material flowing on the waste material recycling conveyor belt are collected using a near-infrared spectral detector. The original spectral sequence is then input into a one-dimensional convolutional neural network for processing, and the basic chemical composition reference ratio vector of the flowing material is output. The pre-stored pure component dictionary matrix is ​​called, and a mapping calculation is performed in combination with the basic chemical component reference ratio vector to generate a reference ideal spectral sequence; The difference between the original spectral sequence and the reference ideal spectral sequence is calculated, the overall physical spectral residual vector is extracted, and the overall physical spectral residual vector is separated using an orthogonal projection matrix to calculate the independent fiber mechanical degradation index of the chemical components. The target licker-in speed of the opening equipment is calculated based on the independent fiber mechanical degradation index, and the basic chemical component reference ratio vector, the independent fiber mechanical degradation index, and the target licker-in speed are input into the hybrid proxy model for calculation to obtain the virgin fiber compensation ratio vector. The linear speed of the waste material recycling conveyor belt is obtained using a speed measuring device. The action lag time is calculated in combination with the physical station distance parameter. When the system clock meets the triggering condition determined based on the physical timestamp and the action lag time, an execution command is output to adjust the rotation speed of the opening equipment to the target licker roller speed. The dynamic mixing device is then controlled to perform supplementary feeding action according to the original fiber compensation ratio vector. 2.The multi-component blend optimization method based on near-infrared spectroscopy and deep learning according to claim 1, wherein, By combining the baseline chemical component reference ratio vector with mapping calculations, a reference ideal spectral sequence is generated, including: The ideal absorption-scattering ratio sequence is obtained by weighted summation of the single-component row vectors in the pure component dictionary matrix based on the baseline proportion vector of the basic chemical components. The ideal absorption-scattering ratio sequence is mapped inversely to the reflectance space based on the inverse transform function to generate the reference ideal spectral sequence.

3. The method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning according to claim 2, characterized in that, The pure component dictionary matrix is ​​obtained through the following steps: Obtain the historical spectral matrix and the known ratio matrix corresponding to the historical spectral matrix, and convert the historical spectral matrix into a historical absorption-scattering ratio matrix; An objective function is constructed using reconstruction error and a sparse penalty term. Under non-negative constraints, a non-negative matrix factorization algorithm is used to decompose and solve the historical absorption-scattering ratio matrix and the known proportion matrix to extract the pure component dictionary matrix.

4. The method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning according to claim 1, characterized in that, The overall physical spectral residual vector is separated using an orthogonal projection matrix, and the independent fiber mechanical degradation index of the chemical components is calculated, including: Multiply the overall physical spectral residual vector with the orthogonal projection matrix of the sensitive band corresponding to the chemical component to obtain the independent residual feature vector of the chemical component; Calculate the norm of the independent residual eigenvectors; When the relative mass fraction of the chemical component in the baseline proportion vector of the basic chemical components is greater than the existence determination threshold, the norm is normalized using the relative mass fraction, and the independent fiber mechanical degradation index is obtained by combining the preset sensitivity weight coefficient.

5. The method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning according to claim 1, characterized in that, The target licker-in speed of the opening equipment is calculated based on the independent fiber mechanical degradation index, including: Determine whether the independent fiber mechanical degradation index of the brittle components in the preset brittle component set exceeds the aging tolerance threshold. When the aging tolerance threshold is exceeded, the difference between the independent fiber mechanical degradation index and the aging tolerance threshold is calculated, and the difference is multiplied by the rotational speed decay tuning coefficient to obtain the rotational speed decay of the brittle component. The reduced speed is obtained by subtracting the sum of the speed decay of all brittle components in the brittle component set from the base speed parameters; The reduced rotational speed is compared with a preset minimum safe rotational speed limit, and the maximum value is taken as the target licker-roll speed.

6. The method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning according to claim 1, characterized in that, The hybrid proxy model includes a parallel combination of a physical mechanism model and a Gaussian process error model; The process of obtaining the native fiber compensation ratio vector includes: The physical mechanism model performs a weighted superposition calculation based on the standard inherent strength of the chemical components, the baseline proportion vector of the basic chemical components, the mechanical degradation index of the independent fibers, the target licker-in roller speed, and the compensation ratio vector of the virgin fibers to be solved, and outputs a predicted value of the mechanism strength. The Gaussian process error model uses the basic chemical component reference ratio vector, the independent fiber mechanical degradation index, the target licker-in roller speed, and the original fiber compensation ratio vector to be solved as input features to output error compensation values. The estimated value of yarn strength is obtained by adding the estimated value of the mechanism strength to the estimated value of the error compensation.

7. The method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning according to claim 6, characterized in that, Obtaining the native fiber compensation ratio vector also includes: The objective function is to minimize the sum of all supplementary mass ratios in the original fiber compensation ratio vector, and the constraint condition is that the estimated yarn strength is greater than the target yarn breaking strength. An optimization task is then constructed. The sequential quadratic programming algorithm is invoked to iteratively calculate the optimization task, and the minimum point at the convergence time is used as the original fiber compensation ratio vector.

8. The method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning according to claim 1, characterized in that, Calculate motion lag time by combining physical workstation distance parameters, including: Divide the first path length from the station where the near-infrared spectral detector is located to the station where the loosening equipment is located by the running linear velocity to calculate the lag time of the preceding action. Divide the second path length from the station where the near-infrared spectral detector is located to the station where the dynamic mixing device is located by the running linear velocity to calculate the lag time of the subsequent action.

9. The method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning according to claim 8, characterized in that, Output execution commands to adjust the rotation speed of the opening equipment to the target licker-in roller speed, and control the dynamic mixing device to perform supplementary feeding actions according to the virgin fiber compensation ratio vector, including: The physical timestamp, the target licker-in roller speed, and the original fiber compensation ratio vector are encapsulated into a control data frame and stored in the first-in-first-out data register queue. When the global clock is equal to the sum of the physical timestamp and the preceding action lag time, minus the preceding electrical response delay time, an execution command to adjust the target licker roller speed is output. When the global clock is equal to the sum of the physical timestamp and the lag time of the subsequent action, minus the delay time of the subsequent electrical response, an execution command is output to perform the supplementary feeding action according to the original fiber compensation ratio vector.

10. The method for optimizing multi-component blended materials based on near-infrared spectroscopy and deep learning according to claim 9, characterized in that, The output executes the supplementary feeding action according to the original fiber compensation ratio vector, including: Read the pre-calibrated baseline linear density parameters of the waste material, multiply the baseline linear density parameters by the operating linear velocity, and calculate the absolute mass flow rate; The absolute mass flow rate is multiplied by the original fiber compensation ratio vector to calculate the absolute compensation mass feed rate; The execution command for controlling the dynamic mixing device to perform supplementary feeding action is output according to the absolute compensation mass feeding rate.

Images