A Deep Learning-Based Data Processing Method for Paper Additives Production

By constructing a multi-source heterogeneous process state information model based on deep learning, the problem of reaction endpoint prediction deviation in papermaking additive production was solved, achieving higher control accuracy and stability, and generating endpoint risk warning signals.

CN123113741APending Publication Date: 2026-11-17CANGZHOU TIANYUE CHEMICAL CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611318648.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the production process of papermaking additives, especially in the reaction preparation of cationic or acrylamide polymers, existing technologies are difficult to effectively monitor the reaction endpoint, resulting in problems such as endpoint prediction bias and insufficient consistency, leading to insufficient control precision and stability.

Method used

A deep learning-based approach is used to construct a multi-source heterogeneous process state information model by combining online viscosity, temperature and conductivity data. The state is predicted by iterative estimation model through deep learning of physical information, and an aggregated state residual constraint structure is constructed to generate an endpoint risk warning signal.

Benefits of technology

It improves the accuracy of polymerization reaction state perception, enhances the predictability and reliability of reaction endpoint determination, reduces the lag and misjudgment risk of traditional methods, and enhances the control accuracy and stability of the production process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113741A_ABST
    Figure CN123113741A_ABST
Patent Text Reader

Abstract

This invention relates to the field of deep learning and discloses a deep learning-based method for processing production data of papermaking additives. The method includes collecting online viscosity, temperature, and conductivity data during the polymerization reaction of papermaking additives to form multi-source heterogeneous process state information; extracting the temporal correlation between online viscosity, temperature, conductivity, and offline viscosity evolution under historical stable polymerization conditions to construct an initial state constraint space for the polymerization process; establishing an online-offline deviation correlation model based on the stage deviation between online observation signals and offline viscosity detection results to predict the polymerization state and generate a dynamic reference polymerization baseline state; mapping the dynamic reference polymerization baseline state to the online measurement space to construct a polymerization state residual constraint structure; and calculating the proximity of the reaction endpoint to generate an online-offline consistency index. This invention has the advantage of improving the accuracy of polymerization state perception.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of deep learning, specifically to a method for processing papermaking additive production data based on deep learning. Background Technology

[0002] In the production of papermaking additives, especially in the reaction preparation of cationic or acrylamide polymers, process parameters such as online viscosity, temperature and conductivity are usually used to monitor the polymerization reaction in real time. Data-driven models are then constructed by combining offline viscosity detection results for the prediction and control of the reaction endpoint.

[0003] However, in industrial continuous or semi-continuous polymerization processes, the system simultaneously exhibits shear-thinning characteristics and rapid formation of local micro-crosslinks, causing the reaction to show a significantly non-uniform response under different stirring and mixing conditions. Online signals only reflect local flow states and are difficult to characterize the overall degree of polymerization. Furthermore, offline viscosity detection is based on laboratory analysis after sampling, while post-polymerization reactions still occur during sample transfer and detection, resulting in time drift of the detection results compared to the instantaneous state at sampling. This also exhibits non-linear amplification characteristics near the endpoint, causing systematic bias in the label data. In addition, during the rapid transition phase at the reaction endpoint, the system is highly sensitive to temperature, initiator residue, and local shear conditions, making the relationship between online observation and offline labeling highly state-dependent and non-stationary. Traditional models based on fixed time alignment or static supervision assumptions are difficult to maintain consistency between input and label, and are prone to endpoint prediction bias. More importantly, in the rapid polymerization phase near the endpoint, the micro-crosslink network forms rapidly and is accompanied by rheological abrupt changes, making it difficult for both online and offline methods to stably characterize the same true state. This results in insufficient consistency of the supervision signal in this range, making it difficult for existing methods based on error minimization or time smoothing to effectively correct this type of structural bias.

[0004] Therefore, existing data-driven modeling methods with a single data source or fixed time alignment are insufficient to characterize the unstable mapping relationship between the actual state and the observed signal that changes dynamically with the evolution of the process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for processing papermaking additive production data, which has the advantage of improving the accuracy of polymerization state perception and solves the problems mentioned in the background technology.

[0006] To achieve the aforementioned goal of improving the accuracy of polymerization state perception, this invention provides the following technical solution: a deep learning-based method for processing papermaking additive production data, comprising the following steps: Online viscosity, temperature and conductivity data were collected during the polymerization reaction of papermaking additives, and offline viscosity detection results of corresponding batches at different reaction times were obtained to form multi-source heterogeneous process state information. Under historically stable polymerization conditions, the temporal correlation between online viscosity, temperature, conductivity and offline viscosity evolution is extracted, and the coupling change characteristics of online viscosity and conductivity at different flow rates during the cyclic conveying process are combined to construct the initial state constraint space of the polymerization process. Based on the stage deviation between online observation signals and offline viscosity detection results, an online and offline deviation correlation model is established. Under the initial state constraint space constraint, a physical information deep learning iterative estimation model with embedded polymerization reaction kinetic equations and mass conservation constraints is used to predict the polymerization state and generate a dynamic reference polymerization baseline state. Based on the preset observation mapping relationship, the dynamic reference aggregation benchmark state is mapped to the online measurement space. Based on the parameter deviation and trend deviation between the mapped reference state and the real-time multi-source heterogeneous process state information, the aggregation state residual constraint structure is constructed, and the state offset trajectory of the aggregation process is determined. Based on the state offset trajectory and the residual convergence characteristics in the aggregated state residual constraint structure, the degree of proximity to the reaction endpoint is calculated. Combined with the stage consistency change characteristics between online observation signals and offline detection results, an online and offline consistency index is generated. When the online and offline consistency index is lower than a preset threshold, an endpoint risk warning signal is generated.

[0007] Preferably, the process for forming multi-source heterogeneous process state information is as follows: Viscosity, temperature and conductivity acquisition units are installed in the polymerization reactor, circulation pipeline and online detection nodes to collect real-time data of the entire polymerization process. To address the issue of sensor clogging due to high viscosity of papermaking additives, abnormal sampling segments are identified and eliminated based on the characteristics of abnormal conductivity fluctuations and stagnant viscosity changes within a continuous sampling period. Offline sampling and detection are performed according to the time nodes corresponding to the preset reaction stages to obtain offline viscosity data for each stage. By aligning online continuous monitoring data with offline stage detection data in terms of time and stage, multi-source heterogeneous process status information for corresponding batches is formed.

[0008] Preferably, the process of constructing the initial state constraint space for the aggregation process is as follows: Retrieve historical stable and qualified polymerization production data, and extract online viscosity growth rate, temperature change rate and conductivity change trend based on the reaction stage; Based on the differences in online viscosity changes and conductivity responses corresponding to different flow rates during the cyclic transport process, the shear-thinning behavior characteristics of the system are extracted; By combining the offline viscosity evolution results of each reaction stage, a stage correspondence between online monitoring characteristics and offline detection results is established; The allowable fluctuation range of characteristic parameters at each stage, the stage change gradient, and the stage transition boundary are combined to form a stage constraint matrix, and the initial state constraint space of the aggregation process is constructed based on the stage constraint matrix. The initial state constraint space of the aggregation process is used to limit the state search range of the physical information deep learning iterative estimation model.

[0009] Preferably, the process of establishing an online-offline deviation correlation model is as follows: Based on the stage phase shift, gradient difference, and numerical deviation between the online observation signal and the offline viscosity detection result, which are collected in real time by the online detection unit arranged in the papermaking additive polymerization reaction system, the online observation deviation coefficient is calculated. Calculate the offline detection deviation coefficient based on the degree of deviation of the offline viscosity test results from the historical stable operating condition reference value; By combining the changes in the online observation deviation coefficient and the offline detection deviation coefficient at different reaction stages, an online and offline deviation correlation model is established.

[0010] Preferably, the process of generating the dynamic reference aggregation baseline state is as follows: Based on the online and offline deviation correlation model, the stage deviation range between online monitoring data and offline detection data is determined; By combining the stage constraint matrix in the initial state constraint space of the aggregation process, the state range of the current aggregation state is limited; A physical information deep learning iterative estimation model with embedded polymerization reaction kinetics equations, mass conservation constraints, and energy balance constraints is used to iteratively estimate the polymerization chain growth state, system flow state, and stage transition state. The input tensor of the physical information deep learning iterative estimation model includes online viscosity, temperature, conductivity and offline viscosity stage data, and the output tensor includes aggregated state parameters and stage transition parameters. By combining the estimation results with the historical stable operating condition reference trajectory, the reference state trajectory is dynamically corrected to generate a dynamic reference aggregated baseline state.

[0011] Preferably, the process of constructing the aggregated state residual constraint structure is as follows: Extract the reference state trajectory corresponding to the dynamic reference aggregation baseline state; Based on the preset observation mapping relationship, the reference state trajectory is mapped to the online measurement space; Calculate the numerical deviation, gradient deviation, and stage deviation between the real-time online state and the reference state; The direction of the deviation is determined based on the sign of the deviation, and the intensity of the deviation is determined based on the magnitude of the absolute value of the deviation. The deviation data of each stage are combined to form the stage residual matrix, deviation direction matrix and deviation intensity matrix, and an aggregated state residual constraint structure is constructed.

[0012] Preferably, the process of determining the state offset trajectory of the aggregation process is as follows: Calculate the phase offset between the online observation signal and the offline detection result at the offline sampling time; By combining the characteristics of viscosity, temperature and conductivity changes at the corresponding stages, the correlation between stage phase shift and process parameters is established. The trend of polymerization state shift is calculated based on the phase shift change between adjacent reaction stages; Based on the historical stable operating condition reference trajectory, the Kalman filter method is used to correct the state offset; By combining the corrected stage phase offset changes, the state offset trajectory of the aggregation process is obtained through fitting.

[0013] Preferably, the process for calculating the proximity of the reaction endpoint is as follows: Extract the convergence rate of state migration at each stage based on the state migration trajectory; The degree of state convergence is calculated by combining the residual reduction magnitude and residual stability duration in the aggregated state residual constraint structure. Calculate the endpoint reference proximity based on the normalized distance between the current stage residual and the historical endpoint reference residual; The convergence degree of the state and the endpoint reference proximity are weighted and fused to quantify the endpoint proximity of the current polymerization reaction.

[0014] Preferably, the process of generating online and offline consistency indicators is as follows: The phase consistency attenuation degree is calculated based on the phase offset between the online observation signal and the offline detection result. The cumulative phase shift and duration of anomalies during the continuous reaction phase were statistically analyzed. By combining the stage deviation values ​​in the aggregated state residual constraint structure, and weighting them according to the phase consistency attenuation degree, anomaly duration, and stage deviation values, an online and offline consistency index is generated.

[0015] Preferably, the process of generating the endpoint risk warning signal is as follows: Based on the consistency index between online and offline processes and the stage deviation changes in the residual constraint structure of the aggregation state, the degree of instability at the endpoint of the aggregation reaction is calculated. When the consistency index between online and offline processes falls below a preset threshold and the stage deviation continues to expand, it is determined that the current aggregation process has a risk of endpoint instability. Risk levels are classified based on the degree of instability at the endpoint and the duration of the abnormality. An endpoint risk warning signal is generated according to the corresponding risk level, and control instructions such as delaying material discharge, increasing the frequency of offline re-inspection, or reducing the reaction temperature are triggered.

[0016] Compared with existing technologies, this invention provides a deep learning-based method for processing papermaking additive production data, which has the following beneficial effects: This invention constructs a multi-source heterogeneous process state information fusion mechanism to collaboratively characterize real-time online viscosity, temperature, and conductivity monitoring data during the polymerization reaction of papermaking additives with corresponding batch offline viscosity detection results. Based on historical stable operating conditions, it introduces temporal correlation patterns and coupled variation characteristics under circulating conveying flow rate conditions, thereby forming an initial state constraint space with process constraint significance. This enhances the adaptability and stability of reaction process state modeling. Simultaneously, by establishing a stage deviation correlation model between online observation signals and offline detection results, and combining it with a deep learning iterative estimation mechanism embedding reaction kinetic equations and mass conservation constraints, it achieves dynamic correction and continuous prediction of the polymerization reaction state, effectively improving... The system improves the characterization accuracy of intermediate reaction states that cannot be directly observed. Furthermore, by mapping the dynamic reference polymerization baseline state to the online measurement space and introducing parameter deviations and trend deviations to construct a residual constraint structure, it achieves continuous tracking and quantitative expression of the reaction process deviation trajectory, enabling the system to structurally identify the evolution trend of the polymerization process. Finally, based on the residual convergence characteristics and the stage consistency change law, the system comprehensively calculates the proximity of the reaction endpoint and generates a consistency index. When the consistency index is lower than a preset threshold, it provides an endpoint risk warning, thereby significantly improving the advance and reliability of polymerization reaction endpoint judgment, reducing the lag and misjudgment risk caused by traditional reliance on offline detection, and overall improving the control accuracy, stability, and industrial application safety of the papermaking additive production process. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method of the present invention; Figure 2 This is a schematic diagram of the process framework of the present invention. Detailed Implementation

[0018] The technical solutions of 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.

[0019] Example 1: Please refer to Figure 1 As shown in the embodiment of the present invention, a method for processing papermaking auxiliary agent production data based on deep learning includes the following steps: S1: Collect online viscosity data, temperature data, and conductivity data during the polymerization reaction of papermaking additives, and obtain offline viscosity detection results of corresponding batches at different reaction times to form multi-source heterogeneous process status information.

[0020] The process of forming multi-source heterogeneous process state information in S1 is as follows: Viscosity, temperature, and conductivity acquisition units are installed in the polymerization reactor, circulation pipeline, and online detection nodes to collect real-time data throughout the polymerization process. Viscosity sensors, temperature sensors, and conductivity sensors are installed in the papermaking additive polymerization reactor, circulation pipeline, and key online detection nodes. The sensors are connected to the data acquisition and control system via industrial fieldbus or Ethernet, and data is continuously collected at a uniform sampling period of 1 second or a preset sampling period. Each data point is timestamped and synchronously written to the database to form real-time monitoring data for the entire polymerization process. During the acquisition process, to ensure the measurement accuracy under the flow state of high-viscosity materials, rotary or vibratory online viscosity sensors can be used, and flow rate monitoring units are arranged in the circulation pipeline to record material flow conditions. To address the issue of sensor clogging under high viscosity conditions of papermaking additives, abnormal sampling segments are identified and removed based on the characteristics of abnormal conductivity fluctuations and stagnant viscosity changes within continuous sampling periods. Furthermore, to address the problem of sensor clogging under high viscosity conditions, a sliding time window is set to segment and analyze online data. Within each window, the amplitude of viscosity and conductivity changes is calculated. When a sudden change in conductivity and a stagnant viscosity change occur, that time period is identified as an abnormal sampling segment. Data from abnormal segments can be repaired using linear interpolation or nearest-neighbor averaging, and this time period is marked for subsequent reference, used for model input and stage analysis, ensuring data quality and usability. Offline sampling and testing are performed according to the time nodes corresponding to the preset reaction stages to obtain offline viscosity data for each stage. Based on the preset key stage time nodes of the polymerization reaction process, such as the initiation reaction, chain growth period, cross-linking period and endpoint period, samples are extracted from the reactor at the corresponding time points, and the viscosity is immediately measured in the laboratory using a rotational viscometer. Each sampling is conducted under constant temperature and constant shear rate conditions, and the sampling time is recorded. The obtained offline viscosity data, along with the sample batch number, reaction time, temperature conditions, etc., are archived together to form staged offline data. By aligning and associating online continuous monitoring data with offline stage detection data in terms of time, multi-source heterogeneous process state information for corresponding batches is formed. Using the reaction start time as a unified zero point, the offline sampling time is mapped to the online data time axis. At each offline sampling time point, online viscosity, temperature, and conductivity data within a fixed time window (e.g., ±5 seconds) are extracted, and the average value is calculated as the corresponding online state for that offline detection value. Each offline detection value is paired with its corresponding online average state, and the data is grouped and stored according to the reaction stage. This results in batch-level multi-source heterogeneous process state information. This information includes both the changing trend of online continuous data and retains the accurate labels of offline detection results, providing a reliable foundation for aggregated state modeling.

[0021] S2: Under historical stable polymerization conditions, extract the temporal correlation between online viscosity, temperature, conductivity and offline viscosity evolution, and combine the coupling change characteristics of online viscosity and conductivity at different flow rates during the cyclic conveying process to construct the initial state constraint space of the polymerization process.

[0022] The process of constructing the initial state constraint space for the aggregation process in S2 is as follows: Historical stable and qualified polymerization production data are retrieved, and online viscosity growth rate, temperature change rate, and conductivity change trend are extracted according to the reaction stage. Complete process data records, including online viscosity, temperature, conductivity, and corresponding reaction time series, are retrieved from historically verified stable and qualified polymerization batch data. The entire polymerization process is divided into several stage intervals according to preset reaction stage division rules. Within each stage, time series fitting is performed on the online viscosity data to calculate the viscosity growth rate per unit time. Simultaneously, the temperature rise and fall rate per unit time is calculated, and the direction and slope of the conductivity change trend within the stage are calculated. These three types of parameters are summarized by stage to form stage characteristics under historical stable operating conditions, used to characterize the stable evolution law of different reaction stages. Based on the differences in online viscosity and conductivity response corresponding to different flow rates during the circulating transport process, the shear-thinning behavior characteristics of the system are extracted. In the circulating transport pipeline, the polymerization reaction system is circulated under different flow rate conditions by adjusting the circulating pump or flow control device, ensuring that the flow rate covers the three typical ranges allowed by the process: low, medium, and high. For each flow rate condition, the reactor temperature, feed concentration, and initiator dosage are kept constant to avoid interference from temperature fluctuations or changes in material composition. Online viscosity and conductivity sensors are simultaneously deployed to continuously collect data on the system throughout the entire circulating transport process. The sampling frequency is adjusted according to the reaction rate. The flow rate is set to ensure that the dynamic changes in viscosity and conductivity can be captured precisely. In the data processing stage, the online viscosity curves and conductivity response curves collected at different flow rates are compared and analyzed. The viscosity decrease, recovery rate and conductivity change synchronicity under each flow rate condition are calculated. The quantitative influence relationship of flow rate change on the viscosity response of the system is extracted. Combined with the difference in conductivity response, the microstructure adjustment or local depolymerization characteristics of the system under different shear conditions are identified, thereby forming quantifiable shear thinning behavior characteristic parameters. Finally, the parameters are classified and stored according to the flow rate range as key inputs when constructing the initial state constraint space of the polymerization process. By combining the offline viscosity evolution results of each reaction stage, a stage correspondence between online monitoring characteristics and offline detection results is established. According to the stage time nodes set by the reaction process, samples are extracted from the reaction system at the end of each stage and standard laboratory viscosity tests are performed to obtain the offline viscosity results of the corresponding stage. The offline detection values ​​of each stage are matched with the online monitoring data within the corresponding time interval of that stage. The viscosity growth rate, temperature change rate, and conductivity change trend parameters within the corresponding time window are extracted from the online data and bound to the offline viscosity results one by one. By summarizing multiple historical batches of data, a corresponding mapping relationship between online characteristic parameters and offline detection results under different reaction stages is established and stored according to stage classification. The allowable fluctuation range of characteristic parameters at each stage, the stage change gradient, and the stage transition boundary are combined to form a stage constraint matrix. Based on the stage constraint matrix, an initial state constraint space for the polymerization process is constructed. Based on historical stable operating data, the numerical distribution range of online viscosity growth rate, temperature change rate, conductivity change trend, and shear thinning behavior characteristics are statistically analyzed for each reaction stage. The maximum, minimum, and average fluctuation ranges are calculated. At the same time, based on the stage changes corresponding to the offline viscosity detection results, the state switching intervals and characteristic abrupt change locations between each stage are determined. The characteristic fluctuation intervals, change gradient ranges, and stage transition boundaries of each stage are arranged in a structured manner to form a parameter constraint matrix at the stage level. This matrix is ​​used to describe the stable operating boundary conditions of different reaction stages. Using the stage constraint matrix as the basic input, the constraint parameters of all reaction stages are uniformly spatialized and integrated to form an initial state constraint space covering the entire polymerization reaction process. This space is used to restrict the value range of online monitoring characteristics, inter-stage migration conditions, and state evolution paths, ensuring that online state evolution must meet the statistical distribution boundaries under historical stable operating conditions. Meanwhile, in the subsequent deep learning iterative estimation of physical information, this constraint space is used as the boundary condition input for state update, so that the model only performs state search and convergence within a reasonable process range, thereby improving the stability and consistency of state estimation. The initial state constraint space of the aggregation process is used to limit the state search range of the physical information deep learning iterative estimation model. When estimating the state of the aggregation process, the constructed initial state constraint space is introduced into the physical information deep learning iterative estimation model as an input constraint. In each iteration update of the model, candidate state solutions are generated based on the current online viscosity, temperature, and conductivity input data. The candidate state solutions are compared with the allowable parameter range in the initial state constraint space. When the candidate state solution exceeds the allowable fluctuation range or gradient range of the corresponding stage, the state update result is truncated or corrected to return it to the boundary of the constraint space. In this way, the state update process of the model is always limited to the reasonable state distribution range corresponding to the historical stable operating conditions, thereby avoiding invalid state estimates that exceed the physical laws of the process and improving the convergence stability and prediction consistency of the iterative estimation process.

[0023] S3: Based on the stage deviation between online observation signals and offline viscosity detection results, an online and offline deviation correlation model is established. Under the initial state constraint space constraint, a physical information deep learning iterative estimation model with embedded polymerization reaction kinetic equations and mass conservation constraints is used to predict the polymerization state and generate a dynamic reference polymerization baseline state.

[0024] The process of establishing the online and offline deviation correlation model in S3 is as follows: Based on the phase shift, gradient difference, and numerical deviation between the online observation signals acquired in real time by online detection units arranged in the papermaking additive polymerization reaction system and the offline viscosity detection results, the online observation deviation coefficient is calculated. Online viscosity, temperature, and conductivity acquisition units are arranged in the papermaking additive polymerization reactor and circulation pipeline to acquire data in real time throughout the polymerization process. The acquired online observation signals are divided according to the reaction stages. For each stage, the gradient of the online observation signal relative to the previous stage is calculated, and the gradient of each stage is compared with the reference rate of change under historical stable conditions. At the same time, the numerical deviation and phase shift of the stage are calculated. By weighting the phase shift, gradient difference, and numerical deviation, the online observation deviation coefficient of each stage is obtained as an indicator reflecting the degree of deviation of the current online observation state from the historical stable conditions. Based on the degree of deviation of the offline viscosity test results from the historical stable operating condition reference value, the offline test deviation coefficient is calculated. According to the preset reaction stage time nodes, the polymer sample is sampled and tested offline to obtain the offline viscosity data corresponding to each stage. The offline viscosity value of each stage is compared with the reference value of the same stage under the historical stable operating condition. The deviation amplitude is calculated and the correction of the results by possible time drift in offline test is taken into account. The offline test deviation coefficient is obtained by standardizing the deviation amplitude. This coefficient can quantify the degree of deviation of the offline measurement results from the stable operating condition and is used to reflect the changing trend of the true polymerization state of the sample. By combining the changing relationship between online observation deviation coefficient and offline detection deviation coefficient at different reaction stages, an online-offline deviation correlation model is established. The online observation deviation coefficient and offline detection deviation coefficient are matched according to the reaction stages, and the relationship between the two changes with the stages throughout the aggregation process is analyzed to identify the systematic deviation pattern between online signals and offline labels. Regression analysis or multivariate correlation calculation methods are used to generate the online-offline deviation correlation model, enabling the model to predict the possible deviation range of offline viscosity based on online observation data. This model also considers the phase difference, gradient difference, and numerical difference at different stages and integrates them into a stage mapping relationship for the generation of dynamic reference aggregation baseline state and online prediction correction.

[0025] The process of generating the dynamic reference aggregation baseline state in S3 is as follows: Based on the online and offline deviation correlation model, the stage deviation interval between online monitoring data and offline detection data is determined. The online observation deviation coefficient and offline detection deviation coefficient output by the online and offline deviation correlation model are matched for stage correspondence. The aggregation process is divided into several continuous stage intervals according to the reaction time axis. In each stage, the difference between the online deviation coefficient and the offline deviation coefficient is calculated, and their distribution range under historical stable operating conditions is statistically analyzed, including the maximum deviation value, the minimum deviation value, and the mean offset, thereby determining the deviation interval corresponding to that stage. This deviation interval is used to constrain the degree of deviation between the current online monitoring data and offline detection data, which is used to limit the offset range of subsequent state estimation. By combining the stage constraint matrix in the initial state constraint space of the aggregation process, the current aggregation state is limited in terms of state range. Taking the stage constraint matrix in the initial state constraint space as input, the viscosity, temperature and conductivity data obtained from the current online monitoring are used to determine the stage affiliation. Based on the stage label corresponding to the current reaction time, the allowable fluctuation range and change gradient boundary of the corresponding stage are extracted from the stage constraint matrix, and the current online state is verified in terms of range. When the online data exceeds the allowable fluctuation range, the boundary truncation or correction method is used to limit it to the stage constraint range, thereby obtaining a standardized state input that meets the historical stable operating condition distribution conditions. A physical information deep learning iterative estimation model, embedding polymerization reaction kinetics equations, mass conservation constraints, and energy balance constraints, is used to iteratively estimate the polymerization chain growth state, system flow state, and stage transition state. The online viscosity, temperature, and conductivity data of the current stage, along with the corresponding offline viscosity data, are used as input vectors. In each iteration, the chain growth trend is corrected based on the polymerization reaction kinetics, while the mass conservation condition constrains the material changes in the system, and the temperature evolution trend is corrected based on the energy balance condition. Through multiple rounds of iterative calculations, the state estimation results gradually converge to a stable state that satisfies the physical constraints, thus obtaining the polymerization chain growth state, system flow state, and stage transition state of the current stage. The input tensor of the physical information deep learning iterative estimation model includes online viscosity, temperature, conductivity, and offline viscosity stage data, and the output tensor includes aggregation state parameters and stage transition parameters. In the physical information deep learning iterative estimation model, the online viscosity data, temperature data, conductivity data, and offline viscosity stage data are aligned in chronological order and a unified input tensor is constructed. Each time step corresponds to a multivariate state vector. After iterative calculation, the model outputs aggregation state parameters and stage transition parameters. The aggregation state parameters are used to characterize the change in the degree of aggregation of the current system, and the stage transition parameters are used to determine the switching trend between reaction stages. The output results are synchronized with the input data in time and are used for subsequent dynamic trajectory correction. By combining the estimation results with historical stable operating condition reference trajectories, the reference state trajectory is dynamically corrected to generate a dynamic reference aggregation baseline state. The estimation results are compared with the historical stable operating condition reference trajectory point by point to calculate the offset between the current estimated state and the historical reference trajectory. The reference trajectory is then corrected point by point according to the magnitude of the offset. When the offset is small, a smooth correction method is used to maintain the original trajectory trend. When the offset continues to accumulate, the trajectory of the corresponding stage is locally reconstructed and corrected, thereby generating a dynamic reference aggregation baseline state that can reflect the evolution trend of the current production state, which is used for subsequent residual calculation and endpoint proximity assessment.

[0026] S4: Based on the preset observation mapping relationship, the dynamic reference aggregation benchmark state is mapped to the online measurement space. Based on the parameter deviation and trend deviation between the mapped reference state and the real-time multi-source heterogeneous process state information, the aggregation state residual constraint structure is constructed, and the state offset trajectory of the aggregation process is determined.

[0027] The process of constructing the aggregated state residual constraint structure in S4 is as follows: Extract the reference state trajectory corresponding to the dynamic reference aggregation baseline state; extract the reference state trajectory data arranged in chronological order from the dynamic reference aggregation baseline state, the reference state trajectory includes aggregation chain growth state parameters, system flow state parameters and stage transition state parameters, perform time axis standardization processing on the reference state trajectory, map different batches of data to the same reaction time baseline, and use a sliding time window to smooth the trajectory to eliminate the influence of local fluctuations, thereby obtaining a standard reference state trajectory with continuity and stability for mapping and comparison calculations; Based on a preset observation mapping relationship, the reference state trajectory is mapped to the online measurement space. According to the preset observation mapping relationship, each state parameter in the reference state trajectory is converted into a variable space consistent with the online measurement data. The mapping relationship is established based on the correspondence between online viscosity, temperature, conductivity and aggregate state parameters under historical stable operating conditions. The aggregate state parameters in the reference trajectory are converted into equivalent online viscosity, temperature and conductivity representations, and resampled at the same time step, so that the reference state trajectory and the real-time online monitoring data are consistent in the time dimension and variable dimension, thereby forming a directly comparable online reference state. The numerical deviation, gradient deviation, and stage deviation between the real-time online state and the reference state are calculated. At each time sampling point, the real-time online monitoring data and the mapped reference state data are compared point by point, and three types of deviation indices are calculated respectively: the first is the numerical deviation, which is the difference between the online viscosity, temperature, and conductivity and the reference value; the second is the gradient deviation, which is the difference between the rate of change within the current time window and the rate of change of the reference trajectory; and the third is the stage deviation, which is the offset difference between the reaction stage to which the current data belongs and the corresponding stage of the reference trajectory. By uniformly normalizing the three types of deviations, a multi-dimensional deviation dataset corresponding to the stage is formed for subsequent structured analysis. The deviation direction is calibrated based on the positive and negative relationship of the deviation, and the deviation intensity is calibrated based on the magnitude of the absolute value of the deviation. Direction and intensity analysis are performed on multidimensional deviation data. The deviation direction is calibrated based on the positive and negative difference between the online data and the reference data. When the online value is greater than the reference value, it is marked as a positive deviation direction, and when the online value is less than the reference value, it is marked as a negative deviation direction. The deviation intensity is quantified and graded based on the magnitude of the absolute value of the deviation, and standardized by combining the distribution range of similar deviations under historical stable working conditions, so as to obtain the deviation intensity value under a unified scale, thereby realizing the comparability of the degree of deviation between different variables. The deviation data from each stage are combined to form a stage residual matrix, a deviation direction matrix, and a deviation intensity matrix, and an aggregated state residual constraint structure is constructed. The multidimensional deviation data is structured according to the reaction stage, and stage residual matrices, deviation direction matrices, and deviation intensity matrices are constructed respectively. The stage residual matrix is ​​used to describe the overall deviation distribution under different reaction stages, the deviation direction matrix is ​​used to record the consistency relationship of the deviation direction of each variable, and the deviation intensity matrix is ​​used to quantify the magnitude of the deviation at each stage. The above three types of matrices are jointly stored and linked to form an aggregated state residual constraint structure, which is used for continuous tracking and constraint analysis of the state offset trajectory of the aggregation process.

[0028] The process of determining the state offset trajectory of the aggregation process in S4 is as follows: At the offline sampling time, the phase shift between the online observation signal and the offline detection result is calculated. During the polymerization reaction, according to the preset offline sampling time point, the online monitoring data and the offline detection result at the corresponding time are compared point by point. The phase shift of each reaction stage is calculated. The time series of the online signal is aligned with the time interval of the reaction stage where the offline detection value is located. The missing points of the online data are filled in by interpolation. The time difference between the online signal and the offline value reaching the same viscosity level or the stage marker point is calculated by using the time shift function. The phase shift is obtained to reflect the time misalignment between online observation and actual reaction progress. By combining the characteristics of viscosity, temperature, and conductivity changes at corresponding stages, a correlation between stage phase shift and process parameters is established. For each stage phase shift, a mathematical correlation model is established by combining process parameters such as the online viscosity growth rate, temperature change rate, and conductivity change rate at that stage. Linear regression or polynomial fitting can be used, with the phase shift as the dependent variable and the viscosity, temperature, and conductivity change rates as independent variables. The model is trained and validated using historical stable batch data to obtain the mapping relationship between stage phase shift and each process parameter. This can be used to analyze the process causes of the shift and provide a basis for calculating the state shift trend. The polymerization state shift trend is calculated based on the phase shift change amplitude between adjacent reaction stages. The phase shift difference between consecutive stages is taken as the shift increment, and the shift rate is calculated in combination with the stage time interval. The viscosity, temperature and conductivity change characteristics are used as weights to form the polymerization state shift trend value for each time step, which is used to guide model adjustment or process optimization. Based on historical stable operating condition reference trajectories, the Kalman filter method is used to correct the state offset. To eliminate the influence of online monitoring noise and occasional anomalies on the state offset calculation, the Kalman filter method is used to correct the stage phase offset. The offset trend is used as the state prediction, the online observation deviation is used as the measurement value, the state estimate is updated using Kalman gain, and the corrected offset at each time step is calculated iteratively to obtain a smooth, continuous and noise-suppressed stage phase offset sequence. Combined with the historical stable operating condition trajectory as prior information, the estimation accuracy and reliability are improved. By combining the corrected stage phase offset changes, the state offset trajectory of the polymerization process is obtained through fitting. The corrected stage phase offset changes are then organized in chronological order, and a continuous state offset trajectory of the polymerization process is generated using a curve fitting method. During the fitting process, appropriate fitting parameters are selected based on the reaction characteristics of each stage to ensure that the trajectory is smooth while retaining the key offset node information of the actual reaction. The generated state offset trajectory can be used for further dynamic reference benchmark updates, online status monitoring, and early warning of polymerization reaction anomalies.

[0029] S5: Based on the state offset trajectory and the residual convergence characteristics in the aggregated state residual constraint structure, calculate the degree of proximity to the reaction endpoint, and combine the stage consistency change characteristics between online observation signals and offline detection results to generate an online and offline consistency index. When the online and offline consistency index is lower than the preset threshold, generate an endpoint risk warning signal.

[0030] The process for calculating the proximity of the reaction endpoint in S5 is as follows: The convergence rate of state offset in each stage is extracted based on the state offset trajectory. Based on the state offset trajectory, the state offset in each reaction stage is decomposed into a time series. The rate of change in each stage is calculated by the offset difference between adjacent sampling time points. The rate series is smoothed by a sliding window in the stage dimension to eliminate the influence of instantaneous fluctuations. Thus, the convergence rate of state offset in each stage is obtained, which is used to characterize the dynamic change trend of the state of the aggregation process approximating the reference trajectory. By combining the residual reduction magnitude and the duration of residual stability in the aggregation state residual constraint structure, the degree of state convergence is calculated. By combining the stage residual data in the aggregation state residual constraint structure, the change process of residual over time in each stage is continuously tracked and analyzed. The magnitude change during the residual decrease process and the duration of residual maintenance in the low fluctuation range are statistically analyzed. The residual reduction magnitude and the duration of stability are normalized and weighted to obtain the degree of state convergence index that reflects the degree of stable approximation of the aggregation state in the current stage. The endpoint reference proximity is calculated based on the normalized distance between the current stage residual and the historical endpoint reference residual. The historical endpoint reference residual originates from the output of the residual constraint structure at the corresponding moment when historically stable qualified batches reached the reaction endpoint judgment conditions. This residual is calculated by aligning online monitoring data and offline viscosity detection results at the endpoint stage, and then using the average or weighted average of multiple batches of endpoint samples to form a standard endpoint residual benchmark. During the calculation process, the current stage residual and the historical endpoint reference residual are compared dimension-by-dimensionally. The absolute difference is calculated for each dimension of the state variable, and normalization is performed based on the maximum fluctuation range of the historical endpoint residual in each dimension to obtain the standardized distance value for each dimension. The normalization results of all dimensions are then weighted and summarized to form the endpoint reference proximity between the current aggregation state and the historical endpoint stable state, which characterizes the consistency between the current reaction state and the endpoint state. The convergence degree of the state and the endpoint reference proximity are weighted and fused to quantify the current endpoint proximity of the polymerization reaction. The convergence degree of the state and the endpoint reference proximity are linearly or piecewise weighted and fused. According to the importance of the endpoint judgment in different reaction stages, corresponding weight coefficients are configured to calculate a unified scalar form endpoint proximity index, which is used to characterize the degree of approach of the polymerization reaction to the endpoint state in real time, and serves as the basis for reaction endpoint judgment and process control adjustment.

[0031] The process of generating online and offline consistency metrics in S5 is as follows: Based on the phase offset between the online observation signal and the offline detection result, the phase consistency attenuation degree is calculated. During the polymerization process, the online observation signal and the corresponding offline viscosity detection result are time-aligned for each stage. The peak points and characteristic change inflection points of the online signal and the offline result are compared to calculate the phase offset for each stage. Based on the rate and magnitude of the offset change over time, the phase consistency attenuation curve is obtained. To improve reliability, a sliding time window can be used for smoothing to reduce the impact of measurement noise on the offset calculation, thereby quantifying the consistency attenuation degree of online data relative to offline data. The cumulative phase shift and abnormal duration of the continuous reaction stage are statistically analyzed. The phase shift sequence of the continuous reaction stage is cumulatively statistically analyzed, and the length of time that the phase shift exceeds the preset threshold in each stage is recorded. The continuous intervals where the shift exceeds the threshold are marked as abnormal segments, and the duration of the abnormal segments is statistically analyzed. By analyzing the cumulative shift and abnormal duration of different stages, the degree of inconsistency between the online monitoring signal and the offline detection results and its continuous characteristics can be reflected, providing a quantitative basis for weighted calculation. By combining the stage deviation values ​​in the aggregation state residual constraint structure, and weighting them according to the phase consistency attenuation degree, anomaly duration, and stage deviation values, an online and offline consistency index is generated. The phase shift sequence of the continuous reaction stage is cumulatively statistically analyzed, recording the length of time the phase shift exceeds a preset threshold in each stage. Continuous intervals with shifts exceeding the threshold are marked as anomaly segments, and the duration of these anomaly segments is statistically analyzed. By analyzing the cumulative shift and anomaly duration at different stages, the degree of inconsistency and its persistence characteristics between online monitoring signals and offline detection results can be reflected. The phase consistency attenuation degree, cumulative phase shift, anomaly duration, and stage deviation values ​​are normalized to eliminate dimensional differences. Each index is weighted and fused according to preset weighting coefficients to obtain the online and offline consistency score for each stage. Finally, all stage scores are comprehensively summarized to form the final online and offline consistency index, used to quantify the overall consistency level between the online monitoring results and offline detection results of the current aggregation process.

[0032] The process of generating the endpoint risk warning signal in S5 is as follows: Based on the online and offline consistency indices and the stage deviation changes in the residual constraint structure of the aggregation state, the degree of instability at the endpoint of the aggregation reaction is calculated. The online and offline consistency indices are used as the global consistency benchmark input. At the same time, the stage deviation change sequence of each stage in the residual constraint structure is extracted, including the rate of change of numerical deviation, the magnitude of the gradient expansion, and the drift of the deviation direction. By performing time difference calculation on the continuous stage deviation changes, the deviation expansion rate is obtained. Combined with the decrease magnitude of the consistency index, a weighted fusion is performed to form a quantitative value representing the degree of instability at the endpoint of the current aggregation process that deviates from the stable endpoint state. When the online and offline consistency index is lower than the preset threshold and the stage deviation continues to expand, it is determined that the current aggregation process has a risk of endpoint instability. During the determination process, when the online and offline consistency index is lower than the preset threshold, this condition is marked as a basic abnormal trigger condition. At the same time, trend analysis is performed on the stage deviation change sequence. The mean and standard deviation of the deviation are calculated through a sliding time window to determine whether the stage deviation shows a continuous expansion trend. When the mean deviation continues to rise and the standard deviation increases synchronously within multiple consecutive time windows, it is determined that the stage deviation is in an expansion state. Combined with the condition that the consistency index is lower than the threshold, it is confirmed that the current aggregation process has a risk of endpoint instability, thus entering the risk assessment stage. Risk levels are categorized based on the degree of instability and the duration of abnormalities at the endpoint. A risk warning signal is generated according to the corresponding risk level, triggering control commands such as delayed discharge, increased offline re-inspection frequency, or reduced reaction temperature. During risk level categorization, a joint assessment is performed based on the degree of instability and the duration of abnormalities. The duration of abnormalities is obtained by measuring the continuous time the statistical deviation exceeds the stability interval threshold. The degree of instability is mapped to different levels and combined with the duration of abnormalities in a two-dimensional manner, resulting in three risk levels: low, medium, and high. Higher instability and longer duration of abnormalities correspond to higher risk levels, thus achieving hierarchical identification of unstable states at the polymerization endpoint. A corresponding risk warning signal is generated based on the determined risk level: low risk corresponds to a warning signal indicating guidance, medium risk to a warning signal indicating process adjustment, and high risk to a warning signal indicating emergency control. Control commands are triggered according to the risk level: increasing the offline re-inspection frequency to increase detection density for low risk; delaying discharge to extend reaction observation time for medium risk; and reducing the reaction temperature to suppress the polymerization rate and prevent system instability for high risk. This achieves tiered response control for abnormal risks at the polymerization endpoint.

[0033] Example 2: Please refer to Figure 2 A deep learning-based data processing system for papermaking additives production includes: Data acquisition and aggregation module: Real-time acquisition of online viscosity, temperature, and conductivity data during the polymerization reaction of papermaking additives, and simultaneous acquisition of offline viscosity detection results to form multi-source heterogeneous process status information; Constraint Modeling Module: Analyzes the evolution patterns of online and offline data under historical stable operating conditions, and constructs the initial state constraint space of the aggregation process by combining the coupling changes of circulating flow rate with viscosity and conductivity. State prediction module: Establishes an online and offline deviation correlation model, and predicts the aggregated state under the initial state constraint space and physical information deep learning iterative estimation model to generate a dynamic reference aggregated baseline state; The residual mapping module maps the dynamic reference aggregation baseline state to the online measurement space, calculates the real-time parameter deviation and trend deviation, constructs the aggregation state residual constraint structure, and determines the state offset trajectory of the aggregation process. Risk warning module: Based on the state offset trajectory and residual convergence characteristics, calculate the degree of proximity to the reaction endpoint, generate online and offline consistency indicators by combining stage consistency changes, and output endpoint risk warning signals when the indicators are below the threshold.

[0034] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0035] 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 method for processing papermaking additive production data based on deep learning, characterized in that, Includes the following steps: Online viscosity, temperature and conductivity data were collected during the polymerization reaction of papermaking additives, and offline viscosity detection results of corresponding batches at different reaction times were obtained to form multi-source heterogeneous process state information. Under historically stable polymerization conditions, the temporal correlation between online viscosity, temperature, conductivity and offline viscosity evolution is extracted, and the coupling change characteristics of online viscosity and conductivity at different flow rates during the cyclic conveying process are combined to construct the initial state constraint space of the polymerization process. Based on the stage deviation between online observation signals and offline viscosity detection results, an online and offline deviation correlation model is established. Under the initial state constraint space constraint, a physical information deep learning iterative estimation model with embedded polymerization reaction kinetic equations and mass conservation constraints is used to predict the polymerization state and generate a dynamic reference polymerization baseline state. Based on the preset observation mapping relationship, the dynamic reference aggregation benchmark state is mapped to the online measurement space. Based on the parameter deviation and trend deviation between the mapped reference state and the real-time multi-source heterogeneous process state information, the aggregation state residual constraint structure is constructed, and the state offset trajectory of the aggregation process is determined. Based on the state offset trajectory and the residual convergence characteristics in the aggregated state residual constraint structure, the degree of proximity to the reaction endpoint is calculated. Combined with the stage consistency change characteristics between online observation signals and offline detection results, an online and offline consistency index is generated. When the online and offline consistency index is lower than a preset threshold, an endpoint risk warning signal is generated.

2. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of forming state information for multi-source heterogeneous processes is as follows: Viscosity, temperature and conductivity acquisition units are installed in the polymerization reactor, circulation pipeline and online detection nodes to collect real-time data of the entire polymerization process. To address the issue of sensor clogging due to high viscosity of papermaking additives, abnormal sampling segments are identified and eliminated based on the characteristics of abnormal conductivity fluctuations and stagnant viscosity changes within a continuous sampling period. Offline sampling and detection are performed according to the time nodes corresponding to the preset reaction stages to obtain offline viscosity data for each stage. By aligning online continuous monitoring data with offline stage detection data in terms of time and stage, multi-source heterogeneous process status information for corresponding batches is formed.

3. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of constructing the initial state constraint space for the aggregation process is as follows: Retrieve historical stable and qualified polymerization production data, and extract online viscosity growth rate, temperature change rate and conductivity change trend based on the reaction stage; Based on the differences in online viscosity changes and conductivity responses corresponding to different flow rates during the cyclic transport process, the shear-thinning behavior characteristics of the system are extracted; By combining the offline viscosity evolution results of each reaction stage, a stage correspondence between online monitoring characteristics and offline detection results is established; The allowable fluctuation range of characteristic parameters at each stage, the stage change gradient, and the stage transition boundary are combined to form a stage constraint matrix, and the initial state constraint space of the aggregation process is constructed based on the stage constraint matrix. The initial state constraint space of the aggregation process is used to limit the state search range of the physical information deep learning iterative estimation model.

4. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of establishing a correlation model between online and offline deviations is as follows: Based on the stage phase shift, gradient difference, and numerical deviation between the online observation signal and the offline viscosity detection result, which are collected in real time by the online detection unit arranged in the papermaking additive polymerization reaction system, the online observation deviation coefficient is calculated. Calculate the offline detection deviation coefficient based on the degree of deviation of the offline viscosity test results from the historical stable operating condition reference value; By combining the changes in the online observation deviation coefficient and the offline detection deviation coefficient at different reaction stages, an online and offline deviation correlation model is established.

5. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of generating a dynamic reference aggregation baseline state is as follows: Based on the online and offline deviation correlation model, the stage deviation range between online monitoring data and offline detection data is determined; By combining the stage constraint matrix in the initial state constraint space of the aggregation process, the state range of the current aggregation state is limited; A physical information deep learning iterative estimation model with embedded polymerization reaction kinetics equations, mass conservation constraints, and energy balance constraints is used to iteratively estimate the polymerization chain growth state, system flow state, and stage transition state. The input tensor of the physical information deep learning iterative estimation model includes online viscosity, temperature, conductivity and offline viscosity stage data, and the output tensor includes aggregated state parameters and stage transition parameters. By combining the estimation results with the historical stable operating condition reference trajectory, the reference state trajectory is dynamically corrected to generate a dynamic reference aggregated baseline state.

6. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of constructing the aggregated state residual constraint structure is as follows: Extract the reference state trajectory corresponding to the dynamic reference aggregation baseline state; Based on the preset observation mapping relationship, the reference state trajectory is mapped to the online measurement space; Calculate the numerical deviation, gradient deviation, and stage deviation between the real-time online state and the reference state; The direction of the deviation is determined based on the sign of the deviation, and the intensity of the deviation is determined based on the magnitude of the absolute value of the deviation. The deviation data of each stage are combined to form the stage residual matrix, deviation direction matrix and deviation intensity matrix, and an aggregated state residual constraint structure is constructed.

7. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of determining the state offset trajectory of the aggregation process is as follows: Calculate the phase offset between the online observation signal and the offline detection result at the offline sampling time; By combining the characteristics of viscosity, temperature and conductivity changes at the corresponding stages, the correlation between stage phase shift and process parameters is established. The trend of polymerization state shift is calculated based on the phase shift change between adjacent reaction stages; Based on the historical stable operating condition reference trajectory, the Kalman filter method is used to correct the state offset; By combining the corrected stage phase offset changes, the state offset trajectory of the aggregation process is obtained through fitting.

8. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of calculating the degree of proximity to the reaction endpoint is as follows: Extract the convergence rate of state migration at each stage based on the state migration trajectory; The degree of state convergence is calculated by combining the residual reduction magnitude and residual stability duration in the aggregated state residual constraint structure. Calculate the endpoint reference proximity based on the normalized distance between the current stage residual and the historical endpoint reference residual; The convergence degree of the state and the endpoint reference proximity are weighted and fused to quantify the endpoint proximity of the current polymerization reaction.

9. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of generating online and offline consistency metrics is as follows: The phase consistency attenuation degree is calculated based on the phase offset between the online observation signal and the offline detection result. The cumulative phase shift and duration of anomalies during the continuous reaction phase were statistically analyzed. By combining the stage deviation values ​​in the aggregated state residual constraint structure, and weighting them according to the phase consistency attenuation degree, anomaly duration, and stage deviation values, an online and offline consistency index is generated.

10. The method for processing papermaking auxiliary production data based on deep learning according to claim 1, characterized in that, The process of generating the endpoint risk warning signal is as follows: Based on the consistency index between online and offline processes and the stage deviation changes in the residual constraint structure of the aggregation state, the degree of instability at the endpoint of the aggregation reaction is calculated. When the consistency index between online and offline processes falls below a preset threshold and the stage deviation continues to expand, it is determined that the current aggregation process has a risk of endpoint instability. Risk levels are classified based on the degree of instability at the endpoint and the duration of the abnormality. An endpoint risk warning signal is generated according to the corresponding risk level, and control instructions such as delaying material discharge, increasing the frequency of offline re-inspection, or reducing the reaction temperature are triggered.