On-line detection system for material composition of production line based on spectral analysis

Abstract

The application discloses a production line material component on-line detection system based on spectrum analysis and relates to the technical field of material detection. The system comprises an on-line acquisition module, a formula benchmark construction module, a wavelet scattering module, a residual scattering generation module, a component sensitive reconstruction module, a component state output module and a feedback control module. The on-line acquisition module acquires a target material spectrum and generates a standardized spectrum. The formula benchmark construction module generates a qualified formula benchmark state. The wavelet scattering module forms measured, benchmark and perturbation scattering characteristics. The residual scattering generation module obtains component deviation residuals by deducting the perturbation. The component sensitive reconstruction module filters and calibrates to generate sensitive residuals. The component state output module generates a component state. The feedback control module outputs a detection result and a control instruction. The application realizes high-precision on-line detection of production line material components.

Description

Technical Field

[0001] This invention relates to the field of material testing technology, and in particular to an online detection system for material composition in production lines based on spectral analysis. Background Technology

[0002] Current production line material composition detection technologies mostly employ offline laboratory testing methods, obtaining composition results through chemical or instrumental analysis after sampling. This method has a long testing cycle and cannot reflect real-time changes in material composition under continuous conveying conditions, making it difficult to detect fluctuations in proportions and uneven mixing during production. Some industrial scenarios use near-infrared or visible light spectroscopy for online detection, achieving rapid prediction by establishing statistical models between spectra and composition. While these methods can achieve certain detection results under stable operating conditions, in continuous conveying environments, changes in material thickness, fluctuations in conveying speed, and ambient light interference can introduce significant spectral drift, causing a shift in the correspondence between spectral characteristics and actual composition.

[0003] Existing spectral analysis methods are mostly based on single spectral feature extraction or traditional regression models, making it difficult to simultaneously distinguish between component changes and non-component perturbations. This is especially true in situations with dust obstruction, uneven surface spreading, and batch-to-batch variations, resulting in poor stability of test results and a high risk of misjudgment. Furthermore, existing methods typically lack a formulation benchmark reference mechanism, failing to perform deviation analysis based on a pre-set formulation state for the current production status. Consequently, test results only reflect changes in component values, lacking a comprehensive ability to characterize the degree of formulation deviation, local uniformity, and quality risks.

[0004] Traditional spectral models suffer from significant drift issues when applied across batches. Differences in spectral distribution between different production batches are difficult to model uniformly, requiring frequent model recalibration and impacting the continuity of online detection. Furthermore, the lack of structured modeling of the relationship between production disturbances and spectral response makes it difficult for existing technologies to achieve stable and accurate online detection and quality control of material composition in complex industrial environments.

[0005] Therefore, how to provide an online detection system for material composition in production lines based on spectral 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 an online material composition detection system for production lines based on spectral analysis. This invention combines spectral analysis, wavelet scattering feature extraction, and formula benchmark modeling methods to achieve disturbance stripping and residual analysis of continuously conveyed material composition. This can improve the accuracy and stability of composition detection in complex production environments and enhance cross-batch detection consistency and quality control capabilities.

[0007] According to an embodiment of the present invention, an online material composition detection system for production lines based on spectral analysis includes: The online acquisition module is used to collect online spectral data of target materials in the production line and preprocess it to generate standardized material spectral data; The formulation baseline construction module is used to construct the formulation baseline spectral state of the target material under qualified formulation conditions based on the preset formulation information, calibration spectral samples and retest component data of the target material. The wavelet scattering module is used to input standardized material spectral data, production disturbance state data and formulation reference spectral state into the improved continuous wavelet scattering network to establish the measured scattering branch, the reference scattering branch and the disturbance estimation branch; The residual scattering generation module is used to map the measured scattering characteristics, the formulation reference scattering characteristics, and the perturbation estimated scattering characteristics to the same scattering path space to generate the composition deviation residual scattering characteristics; The component-sensitive reconstruction module is used to perform path filtering and batch drift calibration on the component deviation residual scattering characteristics based on the calibrated spectral samples and re-examined component data, and generate component-sensitive residual scattering characteristics. The composition status output module is used to generate production line material composition status information based on the composition-sensitive residual scattering characteristics. The production line material composition status information includes target composition parameters, formula deviation parameters, local uniformity parameters, and quality risk parameters. The feedback control module is used to generate component deviation detection results based on the material composition status information and preset quality thresholds in the production line, and output production feedback control commands based on the component deviation detection results.

[0008] Optionally, the target material is a mixture continuously conveyed by the target production line according to a preset formula; the preset formula information includes target component items, target component parameter benchmark values, and component deviation judgment thresholds; the calibration spectral sample is generated by binding the sample spectral data of the target material in different component states with the corresponding retest component data, and is called by the formula benchmark construction module and the component sensitive reconstruction module; the retest component data includes the measured parameters of the target component corresponding to the sample spectral data.

[0009] Optionally, the preprocessing includes removing background interference, correcting spectral bias, removing outlier data, and normalization to obtain standardized material spectral data.

[0010] Optionally, the process of constructing the formulation reference spectral state in the formulation reference construction module includes: Extract the measured parameters of the target component corresponding to the target component from the re-examination component data, compare the deviation between the measured parameters of the target component and the benchmark value of the target component parameter, and obtain the sample component deviation value. The sample component deviation value is compared with the component deviation judgment threshold, and the calibrated spectral samples whose sample component deviation value meets the component deviation judgment threshold are selected to generate a qualified formula sample set. Extract the spectral data of the qualified formula sample set, perform spectral response alignment and spectral intensity unification processing on the sample spectral data, and generate a qualified formula spectral set. The qualified formulation spectral set is categorized according to the target component items, and the baseline spectral response characteristics corresponding to each target component item are extracted. Based on the reference spectral response characteristics of each target component, the reference spectral state of the target material under qualified formulation conditions is generated.

[0011] Optionally, the improved continuous wavelet scattering network in the wavelet scattering module adds a state coding layer, a path constraint layer, and a branch scattering layer to the traditional continuous wavelet scattering structure. Specifically, it includes a state coding layer, a path constraint layer, a branch scattering layer, a feature fusion layer, and a scattering output layer connected in sequence. The state coding layer is used to receive standardized material spectral data, production disturbance state data, and formulation reference spectral state, and to perform unified state characterization on data from different sources, generating measured spectral state vectors, reference spectral state vectors, and disturbance state vectors. The path constraint layer is used to establish scattering path constraint rules based on the formulation reference spectral state, and to constrain the scale selection, frequency band propagation and scattering level in the continuous wavelet scattering network, generating measured path constraints, reference path constraints and perturbation path constraints respectively. The branch scattering layer is used to establish the measured scattering branch, the reference scattering branch, and the disturbance estimation branch based on the corresponding path constraints. The standardized material spectral data, the formulation reference spectral state, and the production disturbance state data are subjected to continuous wavelet scattering transformation along independent scattering paths, respectively. The feature fusion layer is used to correlate and fuse the measured scattering features output by the measured scattering branch, the reference scattering features output by the reference scattering branch, and the perturbation scattering features output by the perturbation estimation branch to construct a joint scattering feature that includes formulation difference information and perturbation compensation information. The scattering output layer is used to stabilize and map the joint scattering features and output the features, generating multi-scale scattering feature results.

[0012] Optionally, the process of establishing the measured scattering branch, the reference scattering branch, and the perturbation estimation branch in the branched scattering layer includes: Based on the measured path constraints and the reference path constraints, corresponding path state segments are extracted from the measured spectral state vector and the reference spectral state vector respectively to generate the measured path state segment and the reference path state segment. According to the scale number, frequency band number and level number in the measured path constraint and the reference path constraint, continuous wavelet scattering transform is performed on the measured path state segment and the reference path state segment respectively to generate the measured path scattering response and the reference path scattering response. According to the path numbers in the measured path constraints and the reference path constraints, the measured path scattering response and the reference path scattering response are collected respectively to generate the measured scattering characteristics and the formulation reference scattering characteristics. Based on the perturbation path constraints, perturbation path state segments are extracted from the perturbation state vector, and the perturbation path state segments are mapped to the scattering path space corresponding to the perturbation path constraints to generate the perturbation path scattering response; The scattering responses of the perturbation paths are aggregated according to the path numbers in the perturbation path constraints to generate the estimated scattering features of the perturbation.

[0013] Optionally, the perturbation estimation branch is also used to perform correction operations, specifically including: Extract the disturbance intensity state and disturbance change state, determine the corresponding scattering path number, and generate disturbance path weights; The perturbation path scattering response under the corresponding scattering path number is weighted and corrected using the perturbation path weight to generate a weighted perturbation path scattering response; The weighted perturbation path scattering response is written into the perturbation estimation scattering features according to the scattering path number to obtain the corrected perturbation estimation scattering features.

[0014] Optionally, the process of generating component-biased residual scattering features in the residual scattering generation module includes: The measured scattering characteristics, the formulation-based scattering characteristics, and the corrected perturbation-estimated scattering characteristics are matched for path correspondence. The measured scattering characteristics are subtracted from the perturbation-estimated scattering characteristics based on the corrected perturbation estimation scattering characteristics, and the deviation is mapped with the formulation reference scattering characteristics to obtain the residual scattering characteristics; The residual scattering features corresponding to each path are combined to generate component-biased residual scattering features.

[0015] Optionally, the component-sensitive reconstruction module specifically includes: Read the component deviation residual scattering characteristics, calibration spectral samples, and re-examination component data, and convert the calibration spectral samples into calibration residual scattering characteristics; The calibration residual scattering characteristics are bound to the measured parameters of the target components in the re-examination component data to generate residual scattering calibration data. Based on the residual scattering calibration data, the correlation strength between the residual scattering changes corresponding to each scattering path and the measured parameter changes of the target component is calculated; Retain scattering paths whose correlation strength meets the preset conditions, generate a set of component-sensitive paths, and extract corresponding path features from the component deviation residual scattering features based on the set of component-sensitive paths. Based on the differences in path characteristics corresponding to different acquisition batches in the calibrated residual scattering characteristics, batch drift calibration is performed on the corresponding path characteristics to generate component-sensitive residual scattering characteristics.

[0016] Optionally, the operation of generating component deviation detection results in the feedback control module includes: Obtain the target component parameters, formulation deviation parameters, local homogeneity parameters, and quality risk parameters. Compare the target component parameters with the corresponding benchmark values ​​to obtain the target component deviation values. The target component deviation value, formulation deviation parameter, and local uniformity parameter are compared with the preset quality threshold to generate corresponding anomaly markers; Based on the anomaly markers, quality risk parameters, and the degree of deviation of each parameter from the quality threshold, the component deviation type and component deviation level are determined, and the component deviation detection results are generated.

[0017] The beneficial effects of this invention are: (1) This invention constructs a reference spectral state of the formulation and introduces a comparison mechanism between measured scattering characteristics and reference scattering characteristics to realize a direct mapping relationship between the current composition state of the target material and the qualified formulation state, thereby structurally improving the reference consistency of component identification and reducing the identification deviation problem caused by the lack of a reference system in traditional spectral models; (2) By introducing production disturbance state data and establishing a disturbance estimation scattering branch, this invention separates and subtracts non-component spectral changes caused by factors such as conveying speed fluctuations, uneven material distribution, and environmental interference, so that the residual scattering characteristics retain only information related to component changes, thereby improving the anti-interference capability and stability of component detection under complex industrial site conditions. (3) This invention constructs a component-sensitive path reconstruction mechanism based on component-sensitive residual scattering characteristics, and combines batch drift calibration and local uniformity analysis to achieve unified processing of cross-batch data distribution differences, so that the detection results can simultaneously express component content, characterize formula deviation and determine quality risk, thereby improving the comprehensive usability and quality control capability of online detection results. Attached Figure Description

[0018] 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: Figure 1 This is a module connection diagram of the online material composition detection system for production lines based on spectral analysis proposed in this invention. Figure 2The structure diagram of the improved continuous wavelet scattering network for the online detection system of material composition in a production line based on spectral analysis proposed in this invention is shown below. Figure 3 The present invention proposes a flowchart of the three-branch wavelet scattering coordination and output of an online material composition detection system for production lines based on spectral analysis; Figure 4 This is a diagram illustrating the process of generating component deviation residual scattering features in Embodiment 2 of the present invention. Figure 5 This is a distribution diagram of scattering path characteristics in Embodiment 2 of the present invention; Figure 6 This is the sensitive path scoring map in Embodiment 2 of the present invention; Figure 7 This is a diagram showing the results of sensitive path filtering in Embodiment 2 of the present invention; Figure 8 This is a distribution diagram of the batch before drift calibration in Embodiment 2 of the present invention; Figure 9 This is a distribution diagram after batch drift calibration in Embodiment 2 of the present invention. Detailed Implementation

[0019] Example 1: Reference Figures 1-3 An online material composition detection system for production lines based on spectral analysis includes: The online acquisition module is used to collect online spectral data of target materials in the production line and preprocess it to generate standardized material spectral data; The formulation baseline construction module is used to construct the formulation baseline spectral state of the target material under qualified formulation conditions based on the preset formulation information, calibration spectral samples and retest component data of the target material. The wavelet scattering module is used to input standardized material spectral data, production disturbance state data and formulation reference spectral state into the improved continuous wavelet scattering network to establish the measured scattering branch, the reference scattering branch and the disturbance estimation branch; The measured scattering branch is used to generate measured scattering characteristics based on standardized material spectral data, the reference scattering branch is used to generate reference scattering characteristics based on the reference spectral state of the formulation, and the disturbance estimation branch is used to generate disturbance estimation scattering characteristics based on production disturbance state data. The residual scattering generation module is used to map the measured scattering characteristics, the formulation reference scattering characteristics, and the perturbation estimated scattering characteristics to the same scattering path space, and generate the component deviation residual scattering characteristics that deduct non-component perturbations and deviate from the formulation reference. The component-sensitive reconstruction module is used to perform component-sensitive path screening and batch drift calibration on the component deviation residual scattering characteristics based on the calibrated spectral samples and re-examined component data, and generate component-sensitive residual scattering characteristics. The composition status output module is used to generate production line material composition status information based on the composition-sensitive residual scattering characteristics. The production line material composition status information includes target composition parameters, formula deviation parameters, local uniformity parameters, and quality risk parameters. The feedback control module is used to generate component deviation detection results based on the material composition status information and preset quality thresholds in the production line, and output production feedback control commands based on the component deviation detection results.

[0020] In this embodiment, the target material is a mixture continuously conveyed by the target production line according to a preset formula; the preset formula information includes target component items, target component parameter benchmark values, and component deviation judgment thresholds; the calibration spectral sample is generated by binding the sample spectral data of the target material in different component states with the corresponding retest component data, which is called by the formula benchmark construction module and the component sensitive reconstruction module; the retest component data includes the measured parameters of the target component corresponding to the sample spectral data.

[0021] Specifically, the target material refers to the material that has completed batching, mixing, and conveying in the target production line and needs to be tested online. The target material has a predetermined material composition, production process requirements and quality control requirements. The mixture can be a powder mixture, a granular mixture, a flake mixture or a crushed mixture. The target material is not an arbitrary unknown material, but a material to be tested that has established preset formula information, calibrated spectral samples and retested component data. The target component item refers to the component item to be detected that is predetermined according to the quality control requirements of the target material. It is used to limit the specific component objects to be identified by online detection. The target component item can be determined as one or more of the main component item, auxiliary component item, impurity component item, moisture component item and active ingredient item according to the actual production needs of the target production line. After the target component item is determined, the subsequent formula benchmark construction module, improved continuous wavelet scattering network, component sensitive reconstruction module and component status output module all perform data processing around the target component item. Different component states refer to the sample states formed when the target material is in a qualified formula state, a state with high component levels, a state with low component levels, or a state of uneven mixing; the binding generation refers to associating the sample spectral data and the retest component data of the same target material sample according to the sample number, forming a calibration relationship in which the spectral data and the real component data correspond one-to-one. The measured parameters of the target component refer to the actual detected values ​​of the target component obtained after retesting the same target material sample. They are used to determine the true component state corresponding to the sample spectral data and provide calibration basis for the construction of the formulation reference spectral state and the screening of component sensitive pathways.

[0022] In this embodiment, the preprocessing includes removing background interference, correcting spectral deviations, eliminating abnormal data, and normalization processing to obtain standardized material spectral data.

[0023] Among them, background interference removal refers to acquiring background spectral data when the target material has not entered the detection window or the detection window is in an empty state, and performing difference processing on the online spectral data and background spectral data to reduce the influence of ambient light, detection window reflection and background response of the conveyor bearing surface on the online spectral data. Correcting spectral bias refers to correcting the overall intensity drift, baseline shift, and wavelength position shift in online spectral data based on a preset reference spectrum or a standard reference sample spectrum, so that the online spectral data obtained from different acquisition cycles are under a uniform spectral response scale. Removing abnormal data refers to identifying abnormal spectral segments in online spectral data caused by insufficient obstruction of target materials, instantaneous accumulation, strong reflection, dust obstruction, or loss of acquisition, and deleting or replacing the abnormal spectral segments from the online spectral data with correction values ​​of adjacent spectral segments of the same acquisition time sequence. Normalization refers to the process of unifying the intensity scale of online spectral data after background interference removal, spectral deviation correction and outlier data removal, so that the spectral intensity distribution corresponding to each acquisition cycle is within the same numerical range, thus obtaining standardized material spectral data. Standardized material spectral data and synchronously acquired production disturbance status data maintain a time-series correspondence and are used as input data for the wavelet scattering module.

[0024] In this embodiment, the process of constructing the spectral state of the formulation reference in the formulation reference construction module includes: Read the target ingredient items, target ingredient parameter baseline values, and ingredient deviation judgment thresholds from the preset formula information, and read the calibration spectral sample and the retest ingredient data bound to the calibration spectral sample; Extract the measured parameters of the target component corresponding to the target component from the re-examination component data, compare the deviation between the measured parameters of the target component and the benchmark value of the target component parameter, and obtain the sample component deviation value. Deviation comparison refers to calculating the difference or proportional difference between the measured parameters of the target component under the same target component item and the benchmark value of the target component parameter, and obtaining the sample component deviation value used to characterize the degree of deviation of the calibration spectral sample from the qualified formula state; the sample component deviation value corresponds one-to-one with the calibration spectral sample. The sample component deviation value is compared with the component deviation judgment threshold, and the calibrated spectral samples whose sample component deviation value meets the component deviation judgment threshold are selected to generate a qualified formula sample set. Meeting the component deviation judgment threshold means that the sample component deviation value is within the allowable deviation range defined by the component deviation judgment threshold; the component state of the target material corresponding to the selected calibration spectrum sample conforms to the preset formula information and serves as the sample basis for constructing the formula reference spectrum state; Extract the spectral data of the qualified formula sample set, perform spectral response alignment and spectral intensity unification processing on the sample spectral data, and generate a qualified formula spectral set. Spectral response alignment refers to adjusting the spectral data of different calibrated spectral samples to the same wavelength sampling position, so that the spectral responses of different samples can be compared and aggregated at the same wavelength dimension; spectral intensity unification refers to adjusting the spectral intensity of different calibrated spectral samples to the same intensity scale, reducing the impact of differences in acquisition conditions on the extraction of reference spectral response features. The qualified formulation spectral set is categorized according to the target component items, and the baseline spectral response characteristics corresponding to each target component item are extracted. Component item aggregation refers to classifying sample spectral data related to the same target component item into the same component item set based on the correspondence between the measured parameters of the target component in the re-inspection component data and the target component item; reference spectral response characteristics refer to characteristic data extracted from the same component item set that can characterize the spectral response law of the target component item under qualified formulation conditions.

[0025] Based on the reference spectral response characteristics of each target component, the reference spectral state of the target material under qualified formulation conditions is generated.

[0026] In this embodiment, the improved continuous wavelet scattering network in the wavelet scattering module adds a state coding layer, a path constraint layer, and a branch scattering layer to the traditional continuous wavelet scattering structure. Specifically, it includes a state coding layer, a path constraint layer, a branch scattering layer, a feature fusion layer, and a scattering output layer connected in sequence. The state coding layer receives standardized material spectral data, production disturbance state data, and formulation baseline spectral state, performs unified state characterization on data from different sources, and generates measured spectral state vectors, baseline spectral state vectors, and disturbance state vectors; specifically: The state encoding layer reads standardized material spectral data and production disturbance state data according to the same acquisition sequence, and reads the reference spectral response corresponding to the target component item in the formula reference spectral state. It extracts the spectral response sequence, local peak shape change sequence, and scattering intensity change sequence from the standardized material spectral data, arranges the extraction results in wavelength order, and performs vectorization to generate a measured spectral state vector. It extracts the reference spectral response sequence, reference peak shape state, and reference scattering response range under the qualified formula state from the formula reference spectral state, and performs vectorization according to the same wavelength order as the measured spectral state vector to generate a reference spectral state vector. It performs time-series alignment and scale normalization on the production disturbance state data, extracts the disturbance intensity state and disturbance change state corresponding to the current acquisition sequence, and maps the disturbance intensity state and disturbance change state to the feature dimension corresponding to the measured spectral state vector to generate a disturbance state vector. Through the above processing, the measured spectral state vector, the reference spectral state vector, and the disturbance state vector have the same time sequence identifier and corresponding feature dimensions. The path constraint layer is used to establish scattering path constraint rules based on the formulation's baseline spectral state, constraining scale selection, bandwidth propagation, and scattering hierarchy in the continuous wavelet scattering network, generating measured path constraints, baseline path constraints, and perturbation path constraints respectively; specifically: The path constraint layer extracts the reference response band, reference response intensity range, and reference response variation range corresponding to the target component from the reference spectral state of the formulation; it determines the scale selection range of the continuous wavelet scattering network based on the reference response band, so that the wavelet scale covers the spectral response width corresponding to the target component; it determines the frequency band propagation range based on the reference response intensity range, so that the scattering propagation between adjacent frequency bands retains the spectral response variation related to the target component; it determines the scattering level based on the reference response variation range, so that the continuous wavelet scattering network unfolds within the level that can characterize the variation of the target component; the path constraint layer combines the scale selection range, frequency band propagation range, and scattering level to form scattering path constraint rules, and generates measured path constraints, reference path constraints, and perturbation path constraints according to the measured scattering branch, the reference scattering branch, and the perturbation estimation branch, respectively; Among them, scale selection refers to determining the scale range of wavelet analysis based on the width of the reference response band corresponding to the target component item; frequency band propagation refers to limiting the transmission relationship of spectral response between different frequency bands during continuous wavelet scattering; scattering level refers to limiting the expansion depth of continuous wavelet scattering transform; by constraining scale selection, frequency band propagation and scattering level, the measured scattering branch, the reference scattering branch and the perturbation estimation branch can generate scattering features in the corresponding scattering path space.

[0027] The branch scattering layer is used to establish the measured scattering branch, the reference scattering branch, and the disturbance estimation branch based on the corresponding path constraints. The standardized material spectral data, the formulation reference spectral state, and the production disturbance state data are subjected to continuous wavelet scattering transformation along independent scattering paths, respectively. The feature fusion layer is used to correlate and fuse the measured scattering features from the measured scattering branch output, the reference scattering features from the reference scattering branch output, and the perturbation scattering features from the perturbation estimation branch output, constructing a joint scattering feature that includes formulation difference information and perturbation compensation information; specifically: The feature fusion layer reads the path number, scale number, frequency band number, and level number corresponding to the measured scattering feature, the reference scattering feature, and the perturbation estimated scattering feature, respectively. Based on the path number, scale number, frequency band number, and level number, the three types of scattering features are path aligned, forming a set of associated features for measured, reference, and perturbation estimated scattering features belonging to the same scattering path location. In each set of associated features, the formula difference relationship between the measured and reference scattering features is preserved, as well as the perturbation compensation relationship between the perturbation estimated scattering feature and the measured scattering feature. The path-aligned formula difference relationship and perturbation compensation relationship are combined according to the scattering path order to generate joint scattering features. The scattering output layer is used to stabilize and map the joint scattering features and output the features, generating multi-scale scattering feature results.

[0028] In this invention, the improved continuous wavelet scattering network is a network structure formed by introducing production state constraints and formula reference constraints on the basis of a continuous wavelet scattering structure. A continuous wavelet scattering structure refers to a structure that performs wavelet scattering transforms on the input spectral signal at different scales, frequency bands, and levels to extract changes in absorption peaks, peak shapes, local fluctuations, and scattering responses from the spectral curve.

[0029] The state coding layer is used to provide a unified state representation for data from different sources. Unified state representation refers to converting standardized material spectral data, production disturbance state data, and formulation reference spectral states into a data expression form that can be processed by the same network, enabling the three types of data to establish a correspondence in the subsequent path constraint layer. The measured spectral state vector is used to represent the online spectral state of the current target material, the reference spectral state vector is used to represent the spectral reference state of the target material under qualified formulation conditions, and the disturbance state vector is used to represent the impact of the current production disturbance state on the spectral acquisition process.

[0030] The path constraint layer is used to establish scattering path constraint rules based on the formulation reference spectral state. The scattering path constraint rules refer to the limitation of scale selection, frequency band propagation and scattering level in the continuous wavelet scattering network based on the reference spectral response position and response range of the target component item in the formulation reference spectral state, so that the scattering process preferentially retains the spectral change path related to the target component item. The scale selection is used to determine the resolution range of wavelet scattering analysis, the frequency band propagation is used to determine the transmission relationship of spectral response between adjacent frequency bands, and the scattering level is used to determine the expansion depth of wavelet scattering transform.

[0031] Measured path constraints, baseline path constraints, and perturbation path constraints are all generated by scattering path constraint rules. Measured path constraints define the path for the measured scattering branch to perform scattering analysis on standardized material spectral data; baseline path constraints define the path for the baseline scattering branch to perform scattering analysis on the formulation baseline spectral state; and perturbation path constraints define the path for the perturbation estimation branch to map production perturbation state data to the scattering feature space. These three types of path constraints have corresponding path numbers and hierarchical positions, enabling the residual scattering generation module to map the three types of scattering features to the same scattering path space.

[0032] The branch scattering layer is used to establish the measured scattering branch, the reference scattering branch, and the perturbation estimation branch. The measured scattering branch performs a continuous wavelet scattering transform on the measured spectral state vector to generate measured scattering features. The reference scattering branch performs a continuous wavelet scattering transform on the reference spectral state vector to generate the formula reference scattering features. The perturbation estimation branch generates perturbation estimation scattering features based on the perturbation state vector and perturbation path constraints. The scattering features are retained and used as inputs to the residual scattering generation module.

[0033] The feature fusion layer is used to construct joint scattering features. The joint scattering features are intermediate features formed by associating and labeling the measured scattering features, the formulation reference scattering features, and the perturbation estimated scattering features. They are used to characterize the correspondence between the current measured spectral state, the qualified formulation spectral state, and the production perturbation state. The feature fusion layer does not replace the independent output of the three types of scattering features. The three types of scattering features are still sent to the residual scattering generation module separately. The scattering output layer is used for stabilization mapping and feature output. Stabilization mapping refers to unifying the scale, path numbering, and feature dimension of the joint scattering features and the three independent scattering features, so that the output features can be directly called by the residual scattering generation module. The multi-scale scattering feature results output by the scattering output layer include measured scattering features, formulated baseline scattering features, perturbation estimated scattering features, and joint scattering features.

[0034] In this embodiment, the process of establishing the measured scattering branch, the reference scattering branch, and the perturbation estimation branch in the branched scattering layer includes: Based on the measured path constraints and the reference path constraints, corresponding path state segments are extracted from the measured spectral state vector and the reference spectral state vector respectively to generate the measured path state segment and the reference path state segment. Specifically, the branch scattering layer reads the path number, scale number, frequency band number, and level number from the measured path constraint, the reference path constraint, and the perturbation path constraint. Based on the spectral response position or perturbation response position corresponding to the path number, it extracts the corresponding data segment from the measured spectral state vector, the reference spectral state vector, and the perturbation state vector. The extracted data segment retains the original path number, scale number, frequency band number, and level number to form the measured path state segment, the reference path state segment, and the perturbation path state segment. According to the scale number, frequency band number and level number in the measured path constraint and the reference path constraint, continuous wavelet scattering transform is performed on the measured path state segment and the reference path state segment respectively to generate the measured path scattering response and the reference path scattering response. Performing continuous wavelet scattering transform refers to applying wavelet filtering, modulus-based processing, and local smoothing to the measured path state segments and the reference path state segments according to the scale number, frequency band number, and level number in the path constraints, extracting the local spectral variation features under the corresponding path layer by layer; for the measured path state segments, continuous wavelet scattering transform yields the measured path scattering response; for the reference path state segments, continuous wavelet scattering transform yields the reference path scattering response; the measured path scattering response is used to characterize the scattering change of the target material under the current spectral state, and the reference path scattering response is used to characterize the scattering reference change of the target material under the qualified formulation state; According to the path numbers in the measured path constraints and the reference path constraints, the measured path scattering response and the reference path scattering response are collected respectively to generate the measured scattering characteristics and the formulation reference scattering characteristics. Specifically, measured path scattering responses with the same path number are grouped into the same measured path set, and reference path scattering responses with the same path number are grouped into the same reference path set. Within the same set, the scattering responses are arranged according to scale number, frequency band number, and hierarchical number to form a path feature sequence corresponding to each scattering path. The path feature sequences corresponding to each measured path set are combined to generate measured scattering features, and the path feature sequences corresponding to each reference path set are combined to generate formulated reference scattering features. Based on the perturbation path constraints, perturbation path state segments are extracted from the perturbation state vector, and the perturbation path state segments are mapped to the scattering path space corresponding to the perturbation path constraints to generate the perturbation path scattering response; Mapping to the scattering path space corresponding to the perturbation path constraint refers to converting the perturbation path state segment into a perturbation response representation with the same path identifier as the measured path scattering response and the reference path scattering response, based on the path number, scale number, frequency band number, and level number in the perturbation path constraint. During the conversion, the perturbation path state segment is matched with the corresponding path according to the acquisition sequence, and the position of the perturbation path state segment acting on the corresponding scattering path is determined according to the perturbation path constraint, generating the perturbation path scattering response. The perturbation path scattering response is used to characterize the scattering offset caused by non-component factors on the corresponding scattering path. The scattering responses of perturbation paths are aggregated according to the path numbers in the perturbation path constraints to generate estimated scattering features. Specifically, scattering responses of perturbation paths with the same path number are grouped into the same perturbation path set, and then arranged according to scale number, frequency band number, and hierarchical number to form a perturbation feature sequence corresponding to each perturbation path. The perturbation feature sequences corresponding to each perturbation path set are combined to generate the estimated scattering features. The estimated scattering features have corresponding path numbers, scale numbers, frequency band numbers, and hierarchical numbers as the measured scattering features and the formulation reference scattering features.

[0035] In this embodiment, the perturbation estimation branch is also used to perform correction operations, specifically including: Extract the disturbance intensity state and disturbance change state, and determine the scattering path number corresponding to the disturbance intensity state and disturbance change state based on the disturbance path constraint; Based on the disturbance intensity state and disturbance change state under the same scattering path number, generate disturbance path weights; The perturbation path scattering response under the corresponding scattering path number is weighted and corrected using the perturbation path weight to generate a weighted perturbation path scattering response; The weighted perturbation path scattering response is written into the perturbation estimation scattering features according to the scattering path number to obtain the corrected perturbation estimation scattering features.

[0036] In this embodiment, weighted correction refers to adjusting the amplitude and correcting the degree of influence of the scattering response of the perturbation path under the same scattering path number by using the perturbation path weight, so that the corrected perturbation path scattering response can characterize the actual degree of influence of the production perturbation state data on the corresponding scattering path. Specifically, the disturbance estimation branch reads the disturbance intensity state, disturbance change state, and disturbance path scattering response under the same scattering path number; it determines the strength of the influence of production disturbance on the scattering path based on the disturbance intensity state, and determines the changing trend of the influence of production disturbance on the scattering path based on the disturbance change state; then, it combines the disturbance intensity state and the disturbance change state into a disturbance path weight, and multiplies the disturbance path weight with the disturbance path scattering response under the corresponding scattering path number to obtain the weighted disturbance path scattering response; When the weight of the perturbation path increases, the scattering response of the corresponding scattering path number is enhanced, indicating that the scattering path is significantly affected by non-component factors. Conversely, when the weight of the perturbation path decreases, the scattering response of the corresponding scattering path number is weakened, indicating that the scattering path is less affected by non-component factors. The weighted perturbation path scattering response is written into the perturbation estimation scattering features according to the scattering path number, so that the perturbation estimation scattering features can be used as the basis for non-component perturbation subtraction in the residual scattering generation module.

[0037] In this embodiment, the process of generating component-biased residual scattering features in the residual scattering generation module includes: The measured scattering characteristics, the formulation-based scattering characteristics, and the corrected perturbation-estimated scattering characteristics are matched for path correspondence. Path correspondence matching refers to the residual scattering generation module reading the path identifiers carried in the measured scattering features, the formulation reference scattering features, and the corrected perturbation estimated scattering features, and establishing a correspondence according to the same scattering path position; the path identifier is used to characterize at least the scattering path number, scale position, frequency band position, and scattering level, so that the measured scattering features, formulation reference scattering features, and the corrected perturbation estimated scattering features under the same scattering path can be classified into the same corresponding path; The measured scattering characteristics are subtracted from the perturbation-estimated scattering characteristics based on the corrected perturbation estimation scattering characteristics, and the deviation is mapped with the formulation reference scattering characteristics to obtain the residual scattering characteristics; Perturbation subtraction of measured scattering characteristics refers to using the corrected perturbation-estimated scattering characteristics as a non-component perturbation in the same corresponding path to subtract the scattering offset caused by production disturbance in the measured scattering characteristics, thus obtaining the measured scattering response after perturbation subtraction. This processing is used to weaken the influence of non-component factors on the measured scattering characteristics, making the retained scattering response closer to the spectral scattering change corresponding to the real composition change of the target material. Deviation mapping refers to the difference transformation between the measured scattering response after disturbance subtraction and the formulation reference scattering characteristics in the same corresponding path, to obtain the residual scattering characteristics relative to the qualified formulation state under the corresponding path; the residual scattering characteristics are used to characterize the component-related deviation between the current detection state of the target material and the qualified formulation state; The residual scattering features corresponding to each path are combined to generate composition deviation residual scattering features. Specifically, the residual scattering features obtained under different corresponding paths are arranged and summarized according to the path order in the scattering path space to form composition deviation residual scattering features that can characterize the overall composition deviation state of the target material. The composition deviation residual scattering features serve as input to the composition-sensitive reconstruction module for subsequent composition-sensitive path screening and batch drift calibration.

[0038] In this embodiment, the component-sensitive reconstruction module specifically includes: Read the component deviation residual scattering characteristics, calibration spectral samples, and re-examination component data, and convert the calibration spectral samples into calibration residual scattering characteristics; Converting calibration spectral samples into calibration residual scattering features involves the component-sensitive reconstruction module calling the sample spectral data from the calibration spectral samples and generating standardized sample spectral data using the same spectral standardization processing method as the online spectral data. The standardized sample spectral data, the production disturbance state data corresponding to the sample acquisition, and the formula reference spectral state are then input into the wavelet scattering module to generate measured scattering features, sample formula reference scattering features, and sample disturbance estimated scattering features. The residual scattering generation module performs disturbance subtraction on the measured scattering features and performs deviation mapping relative to the sample formula reference scattering features to generate calibration residual scattering features corresponding to the calibration spectral samples. The calibration residual scattering features and the component deviation residual scattering features obtained from online detection are located in the same scattering path space. The calibration residual scattering characteristics are bound to the measured parameters of the target components in the re-examination component data to generate residual scattering calibration data. Specifically, the calibration residual scattering feature value corresponding to each scattering path is extracted according to the scattering path number, and the measured parameters of the target component in the re-examination component data under the same sample number are extracted; the variation amplitude of the calibration residual scattering feature value of different samples under the same scattering path is statistically analyzed, and the variation amplitude of the corresponding measured parameters of the target component is statistically analyzed; the consistency of the change direction, the correspondence of the change amplitude, and the intra-batch stability of the two are compared to obtain the correlation strength corresponding to the scattering path; the higher the correlation strength, the better the residual scattering change corresponding to the scattering path can reflect the change of the measured parameters of the target component; Based on the residual scattering calibration data, the correlation strength between the residual scattering changes corresponding to each scattering path and the measured parameter changes of the target component is calculated; Retain scattering paths whose correlation strength meets the preset conditions, generate a set of component-sensitive paths, and extract corresponding path features from the component deviation residual scattering features based on the set of component-sensitive paths. Specifically, the correlation strength meeting the preset conditions means that the correlation strength corresponding to the scattering path is not lower than the preset correlation strength threshold, and the correlation direction of the scattering path remains consistent in different collection batches. The preset correlation strength threshold can be preset according to the number of calibrated spectral samples, the detection accuracy requirements of the target component item, and the component deviation judgment threshold. For scattering paths with correlation strength lower than the preset correlation strength threshold, or scattering paths with inconsistent correlation directions in different collection batches, the component-sensitive reconstruction module judges them as non-component-sensitive paths and removes them.

[0039] Based on the differences in path characteristics corresponding to different acquisition batches in the calibrated residual scattering characteristics, batch drift calibration is performed on the corresponding path characteristics to generate component-sensitive residual scattering characteristics.

[0040] Specifically, in the calibration residual scattering characteristics, a sampling batch under the qualified formulation state is selected as the benchmark sampling batch, and the benchmark path feature center of each component sensitive path in the benchmark sampling batch is calculated; the batch path feature center of the same component sensitive path in the remaining sampling batches is calculated respectively; the difference between the batch path feature center and the benchmark path feature center is determined as the path offset of the component sensitive path of the sampling batch; the path offset of the corresponding sampling batch is subtracted from the current corresponding path feature to obtain the path feature after batch drift calibration, and the path feature after batch drift calibration is combined to generate the component sensitive residual scattering characteristics.

[0041] In this embodiment, the operation of generating component deviation detection results in the feedback control module includes: Obtain the target component parameters, formulation deviation parameters, local homogeneity parameters, and quality risk parameters. Compare the target component parameters with the corresponding benchmark values ​​to obtain the target component deviation values. Specifically, the target component parameter refers to the actual characterization value of the component to be detected in the target material obtained by the component status output module based on the component-sensitive residual scattering characteristics, which reflects the content status of the target component; the formula deviation parameter refers to the degree of deviation of the target component parameter from the corresponding benchmark value in the preset formula information, which reflects whether the target material deviates from the qualified formula state; the local uniformity parameter refers to the degree of fluctuation of the target component parameter within the continuous detection window, which reflects the local mixing consistency of the target material during the transportation process; the quality risk parameter refers to the risk characterization value obtained by combining the target component parameter, formula deviation parameter, and local uniformity parameter, which is used to participate in the generation of component deviation detection results. The target component deviation value, formulation deviation parameter, and local uniformity parameter are compared with the preset quality threshold to generate corresponding anomaly markers; The preset quality thresholds are set for at least the target component deviation value, the formulation deviation parameter, and the local homogeneity parameter, so that different types of component anomalies can be identified separately. The feedback control module compares the target component deviation value, the formulation deviation parameter, and the local homogeneity parameter with the corresponding preset quality thresholds. When the corresponding parameter exceeds the allowable range, a corresponding anomaly mark is generated; when the corresponding parameter does not exceed the allowable range, a normal mark is generated. Based on the anomaly markers, quality risk parameters, and the degree of deviation of each parameter from the quality threshold, the component deviation type and component deviation level are determined, and the component deviation detection results are generated. The feedback control module first determines the type of component deviation based on the target component deviation marker, the formula deviation marker, and the local uniformity anomaly marker. When the target component deviation marker is abnormal, it is determined that there is a deviation in the content of the target component. When the formula deviation marker is abnormal, it is determined that there is a formula deviation. When the local uniformity anomaly marker is abnormal, it is determined that there is local mixing unevenness. The feedback control module then determines the component deviation level based on the degree of deviation of each parameter from the preset quality threshold and in combination with the quality risk parameter. The greater the degree of deviation and the higher the quality risk parameter, the higher the component deviation level. Finally, the component deviation type and component deviation level are combined to generate the component deviation detection result.

[0042] Example 2: Reference Figures 4-9 To verify the feasibility of this invention in practice, it was applied to an online component detection scenario in a granular mixture production line. This production line employs continuous batching and mixing, where multiple basic components are mixed according to a preset formula to form the target material. During production, fluctuations in upstream feed, changes in conveyor belt speed, variations in material layer thickness, and environmental factors can easily cause drift in the online spectral signals, making it difficult for traditional spectral detection methods to accurately distinguish between changes in actual composition and spectral changes caused by production disturbances. When the material composition deviates slightly, traditional models often produce false alarms or missed alarms, resulting in delayed quality control.

[0043] In Example 2, the online acquisition module uses a near-infrared spectral acquisition device installed above the conveyor belt to continuously scan the target material, acquiring online spectral data in the wavelength range of 400nm to 2400nm. The acquired spectral data is then processed through abnormal spectrum removal, noise filtering, spectral intensity normalization, and wavelength alignment to generate standardized material spectral data.

[0044] The formulation benchmark construction module reads the preset formulation information corresponding to the target material and retrieves historical calibration spectral samples and corresponding retest component data. The retest component data is obtained through standard laboratory testing methods. The module then selects data sets from historical samples that meet the requirements for qualified formulations to construct the formulation benchmark spectral status. After selection, a total of 1260 valid calibration samples were obtained as the benchmark sample library.

[0045] The wavelet scattering module simultaneously inputs standardized material spectral data, production disturbance state data, and formulation baseline spectral state into an improved counterfactual state-coupled continuous wavelet scattering network. The production disturbance state data includes state information such as conveyor speed, material layer thickness, ambient temperature, and dust concentration. The improved continuous wavelet scattering network establishes measured scattering branches, a baseline scattering branch, and a disturbance estimation branch. The network uses continuous Morlet wavelets as the scattering basis function, completing frequency band decomposition and feature propagation at different scales in two scattering layers. The measured scattering branch generates measured scattering features, the baseline scattering branch generates formulation baseline scattering features, and the disturbance estimation branch generates disturbance estimated scattering features.

[0046] The residual scattering generation module maps three types of scattering features to a unified scattering path space and performs perturbation subtraction on the measured scattering features. Before processing, some scattering paths exhibit a strong response to transport velocity fluctuations, resulting in spectral variations of up to 11.8%. After perturbation subtraction, the non-component fluctuations of the same paths are reduced to 3.1%. Subsequently, deviation mapping is performed using the formulation-based scattering features to generate composition-deviation residual scattering features. The generation results show that the residual response corresponding to changes in the target composition is significantly enhanced, and composition-sensitive information is effectively separated from background interference.

[0047] The component-sensitive reconstruction module establishes a path correlation analysis model using calibrated spectral samples and re-examined component data. After correlation analysis, 86 highly sensitive scattering paths were extracted, accounting for 21.4% of all scattering paths. Subsequently, batch drift calibration was performed to correct path offsets caused by different production batches. Before calibration, the average feature offset between different batches reached 7.6%; after calibration, it decreased to 1.9%. This yields the component-sensitive residual scattering characteristics.

[0048] The composition status output module generates material composition status information for the production line based on the composition-sensitive residual scattering characteristics. Detection results show that the average absolute error between the target composition parameters and the laboratory retest results is 0.37%, the average error of the formula deviation parameter is 0.42%, the error of the local uniformity parameter is 0.03%, and the error of the quality risk parameter is 0.04. The feedback control module generates composition deviation detection results based on the material composition status information for the production line. When the detected formula deviation exceeds a preset quality threshold, it automatically outputs a production feedback control command to correct the feeding ratio.

[0049] To further verify the effectiveness of this invention, it was compared with traditional partial least squares regression (PLS) models and traditional convolutional neural network (CNN) models. A total of 500 test samples were used, including normal samples, slightly deviated samples, and significantly deviated samples. The results show that the traditional PLS model achieved an accuracy of 89.4% under complex perturbation environments, the traditional CNN model achieved an accuracy of 93.1%, and this invention achieved an accuracy of 98.2%. Regarding component prediction error, the average error of the traditional PLS model was 1.84%, the average error of the traditional CNN model was 1.09%, and the average error of this invention was reduced to 0.37%. Under batch drift conditions, the error growth rate of the traditional PLS model reached 31.6%, the error growth rate of the traditional CNN model was 18.9%, and the error growth rate of this invention was only 5.4%.

[0050] Test results show that this invention utilizes an improved counterfactual state-coupled continuous wavelet scattering network to establish measured scattering branches, reference scattering branches, and disturbance estimation branches. By constructing component deviation residual scattering characteristics, it effectively decouples production disturbance factors from actual component changes. Through component-sensitive path screening and batch drift calibration, it improves cross-batch detection consistency. Through component state output and feedback control mechanisms, it achieves real-time monitoring and quality risk warning of material components in the production line, meeting the high-precision online detection requirements in complex industrial production environments.

[0051]

[0052] As shown in Table 1, the present invention exhibits higher detection stability and recognition accuracy in online material composition detection. Compared with traditional partial least squares regression models and traditional convolutional neural network models, the component recognition accuracy of the present invention reaches 98.2%, and the average component prediction error is reduced to 0.37%. This indicates that by jointly processing the formulation baseline spectral state, perturbation estimation scattering characteristics, and component deviation residual scattering characteristics, spectral information related to the actual component changes can be extracted more accurately.

[0053] In terms of formula deviation, local uniformity, and quality risk identification, the present invention achieves accuracy rates of 97.8%, 97.1%, and 97.6%, respectively, all significantly higher than the comparative model. This indicates that the present invention can not only detect the content of target components, but also make comprehensive judgments on formula deviation, local mixing unevenness, and quality abnormalities in the continuous conveying process of the production line.

[0054] In terms of disturbance resistance and cross-batch stability, the batch feature offset of this invention is reduced to 1.9%, the batch drift error growth rate is only 5.4%, and the non-component disturbance residual rate is reduced to 3.1%. These results indicate that the disturbance estimation branch, residual scattering generation module, and batch drift calibration process can effectively mitigate the impact of production disturbances and batch differences on spectral detection results.

[0055] In terms of detection efficiency, the single-detection response time of this invention is 72ms, which is lower than the 182ms of the traditional partial least squares regression model and the 136ms of the traditional convolutional neural network model. This indicates that while improving detection accuracy, this invention can still meet the real-time requirements of online detection on the production line. Overall, the data in Table 1 demonstrates that this invention has significant advantages in detection accuracy, anti-interference ability, cross-batch consistency, and real-time feedback capability.

[0056] 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. An online material composition detection system for production lines based on spectral analysis, characterized in that, include: The online acquisition module is used to collect online spectral data of target materials in the production line and preprocess it to generate standardized material spectral data; The formulation baseline construction module is used to construct the formulation baseline spectral state of the target material under qualified formulation conditions based on the preset formulation information, calibration spectral samples and retest component data of the target material. The wavelet scattering module is used to input standardized material spectral data, production disturbance state data and formulation reference spectral state into the improved continuous wavelet scattering network to establish the measured scattering branch, the reference scattering branch and the disturbance estimation branch. The residual scattering generation module is used to map the measured scattering characteristics, the formulation reference scattering characteristics, and the perturbation estimated scattering characteristics to the same scattering path space to generate the composition deviation residual scattering characteristics; The component-sensitive reconstruction module is used to perform path filtering and batch drift calibration on the component deviation residual scattering characteristics based on the calibrated spectral samples and re-examined component data, and generate component-sensitive residual scattering characteristics. The composition status output module is used to generate production line material composition status information based on the composition-sensitive residual scattering characteristics. The production line material composition status information includes target composition parameters, formula deviation parameters, local uniformity parameters, and quality risk parameters. The feedback control module is used to generate component deviation detection results based on the material composition status information and preset quality thresholds in the production line, and output production feedback control commands based on the component deviation detection results.

2. The online material composition detection system for production lines based on spectral analysis according to claim 1, characterized in that, The target material is a mixture continuously conveyed by the target production line according to a preset formula; the preset formula information includes target component items, target component parameter benchmark values, and component deviation judgment thresholds; The calibration spectral sample is generated by binding the sample spectral data of the target material in different composition states with the corresponding retest component data, and is called by the formulation benchmark construction module and the component sensitivity reconstruction module; the retest component data includes the measured parameters of the target component corresponding to the sample spectral data.

3. The online material composition detection system for production lines based on spectral analysis according to claim 2, characterized in that, The preprocessing includes removing background interference, correcting spectral bias, removing outlier data, and normalization to obtain standardized material spectral data.

4. The online material composition detection system for production lines based on spectral analysis according to claim 3, characterized in that, The process of constructing the spectral state of the formulation reference in the formulation reference construction module includes: Extract the measured parameters of the target component corresponding to the target component from the re-examination component data, compare the deviation between the measured parameters of the target component and the benchmark value of the target component parameter, and obtain the sample component deviation value. The sample component deviation value is compared with the component deviation judgment threshold, and the calibrated spectral samples whose sample component deviation value meets the component deviation judgment threshold are selected to generate a qualified formula sample set. Extract the spectral data of the qualified formula sample set, perform spectral response alignment and spectral intensity unification processing on the sample spectral data, and generate a qualified formula spectral set; The qualified formulation spectral set is categorized according to the target component items, and the baseline spectral response characteristics corresponding to each target component item are extracted. Based on the reference spectral response characteristics of each target component, the reference spectral state of the target material under qualified formulation conditions is generated.

5. The online material composition detection system for production lines based on spectral analysis according to claim 4, characterized in that, The improved continuous wavelet scattering network in the wavelet scattering module adds a state coding layer, a path constraint layer, and a branch scattering layer to the traditional continuous wavelet scattering structure. Specifically, it includes a state coding layer, a path constraint layer, a branch scattering layer, a feature fusion layer, and a scattering output layer connected in sequence. The state coding layer is used to receive standardized material spectral data, production disturbance state data, and formulation reference spectral state, and to perform unified state characterization on data from different sources, generating measured spectral state vectors, reference spectral state vectors, and disturbance state vectors. The path constraint layer is used to establish scattering path constraint rules based on the formulation reference spectral state, and to constrain the scale selection, frequency band propagation and scattering level in the continuous wavelet scattering network, generating measured path constraints, reference path constraints and perturbation path constraints respectively. The branch scattering layer is used to establish the measured scattering branch, the reference scattering branch, and the disturbance estimation branch based on the corresponding path constraints. The standardized material spectral data, the formulation reference spectral state, and the production disturbance state data are subjected to continuous wavelet scattering transformation along independent scattering paths, respectively. The feature fusion layer is used to correlate and fuse the measured scattering features output by the measured scattering branch, the reference scattering features output by the reference scattering branch, and the perturbation scattering features output by the perturbation estimation branch to construct a joint scattering feature that includes formulation difference information and perturbation compensation information. The scattering output layer is used to stabilize and map the joint scattering features and output the features, generating multi-scale scattering feature results.

6. The online material composition detection system for production lines based on spectral analysis according to claim 5, characterized in that, The process of establishing the measured scattering branch, the reference scattering branch, and the perturbation estimation branch in the branched scattering layer includes: Based on the measured path constraints and the reference path constraints, corresponding path state segments are extracted from the measured spectral state vector and the reference spectral state vector respectively to generate the measured path state segment and the reference path state segment. According to the scale number, frequency band number and level number in the measured path constraint and the reference path constraint, continuous wavelet scattering transform is performed on the measured path state segment and the reference path state segment respectively to generate the measured path scattering response and the reference path scattering response. According to the path numbers in the measured path constraints and the reference path constraints, the measured path scattering response and the reference path scattering response are collected respectively to generate the measured scattering characteristics and the formulation reference scattering characteristics. Based on the perturbation path constraints, perturbation path state segments are extracted from the perturbation state vector, and the perturbation path state segments are mapped to the scattering path space corresponding to the perturbation path constraints to generate the perturbation path scattering response; The scattering responses of the perturbation paths are aggregated according to the path numbers in the perturbation path constraints to generate the estimated scattering features of the perturbation.

7. The online material composition detection system for production lines based on spectral analysis according to claim 6, characterized in that, The perturbation estimation branch is also used to perform correction operations, specifically including: Extract the disturbance intensity state and disturbance change state, determine the corresponding scattering path number, and generate disturbance path weights; The perturbation path scattering response under the corresponding scattering path number is weighted and corrected using the perturbation path weight to generate a weighted perturbation path scattering response; The weighted perturbation path scattering response is written into the perturbation estimation scattering features according to the scattering path number to obtain the corrected perturbation estimation scattering features.

8. The online material composition detection system for production lines based on spectral analysis according to claim 7, characterized in that, The process of generating compositional deviation residual scattering features in the residual scattering generation module includes: The measured scattering characteristics, the formulation-based scattering characteristics, and the corrected perturbation-estimated scattering characteristics are matched for path correspondence. The measured scattering characteristics are subtracted from the perturbation-estimated scattering characteristics based on the corrected perturbation estimation scattering characteristics, and the deviation is mapped with the formulation reference scattering characteristics to obtain the residual scattering characteristics; The residual scattering features corresponding to each path are combined to generate component-biased residual scattering features.

9. The online material composition detection system for production lines based on spectral analysis according to claim 8, characterized in that, The component-sensitive reconstruction module specifically includes: Read the component deviation residual scattering characteristics, calibration spectral samples, and re-examination component data, and convert the calibration spectral samples into calibration residual scattering characteristics; The calibration residual scattering characteristics are bound to the measured parameters of the target components in the re-examination component data to generate residual scattering calibration data. Based on the residual scattering calibration data, the correlation strength between the residual scattering changes corresponding to each scattering path and the measured parameter changes of the target component is calculated; Retain scattering paths whose correlation strength meets the preset conditions, generate a set of component-sensitive paths, and extract corresponding path features from the component deviation residual scattering features based on the set of component-sensitive paths. Based on the differences in path characteristics corresponding to different acquisition batches in the calibrated residual scattering characteristics, batch drift calibration is performed on the corresponding path characteristics to generate component-sensitive residual scattering characteristics.

10. The online detection system for production line material composition based on spectral analysis according to claim 9, characterized in that, The operation of generating component deviation detection results in the feedback control module includes: Obtain the target component parameters, formulation deviation parameters, local homogeneity parameters, and quality risk parameters. Compare the target component parameters with the corresponding benchmark values ​​to obtain the target component deviation values. The target component deviation value, formulation deviation parameter, and local uniformity parameter are compared with the preset quality threshold to generate corresponding anomaly markers; Based on the anomaly markers, quality risk parameters, and the degree of deviation of each parameter from the quality threshold, the component deviation type and component deviation level are determined, and the component deviation detection results are generated.

Images