A paper product thickness on-line prediction system and method

Abstract

The application discloses a paper product thickness online prediction system and method, and relates to the technical field of intelligent control of papermaking industry. The system comprises a data acquisition module, a data storage module, an AI prediction module and an output module, can acquire 42 process parameters of beating and papermaking in real time, and combines historical test data to construct a distributed data warehouse. An LSTM, a random forest and an XGBoost fusion algorithm are used for thickness prediction, and the key parameter weight is enhanced through an attention mechanism. The prediction result is converted into executable instructions through a parameter adjustment rule base, and is bidirectionally interacted with a MES system through an API, so that closed-loop regulation and control are realized. The system supports sliding window iterative training and abnormal automatic optimization mechanism, and ensures long-term stable operation. The scheme realizes the target of high thickness prediction accuracy, low response delay and automatic regulation and control, significantly improves product qualification rate and reduces raw material waste.

Description

Technical Field

[0001] This invention relates to intelligent control and artificial intelligence applications in the papermaking industry, and particularly to an online paper thickness prediction system and method. Background Technology

[0002] In the production of tissue paper, the thickness of the base paper is a key indicator affecting product quality and material utilization. Traditional production typically employs an offline testing model combined with manual adjustment based on experience. This involves periodically sampling and testing to obtain thickness data, which is then adjusted by operators based on their experience. In recent years, some companies have begun to introduce data analysis platforms based on Manufacturing Execution Systems (MES) to monitor certain process parameters in the papermaking process and use simple statistical models or machine learning algorithms for quality prediction. International companies such as Metso in Finland and Valmet in Sweden have developed online monitoring systems integrating sensors and AI models, enabling real-time acquisition and partial prediction of key parameters. However, these systems often focus on a single process, have incomplete data coverage, and rely on costly hardware modifications, making them difficult to adapt to the actual operating conditions of domestic tissue paper production lines. However, existing technologies still have significant shortcomings: First, most systems only collect data from the papermaking stage, lacking key parameters such as fiber wet weight and grinding gap in the pulping process, resulting in incomplete input information; second, models are mostly built based on a single algorithm, making it difficult to adapt to complex nonlinear time-series relationships, and prediction accuracy is generally low; third, prediction results cannot be automatically converted into executable operation instructions, still requiring manual intervention, resulting in severe response lag; finally, the lack of a system update mechanism makes it unable to cope with dynamic changes such as raw material fluctuations and equipment aging, leading to a significant decline in prediction performance after long-term operation. These problems collectively result in untimely thickness control and limited improvement in yield, hindering the further development of intelligent manufacturing. Summary of the Invention

[0003] This invention aims to solve at least one of the technical problems existing in the prior art. To this end, this invention proposes an online paper thickness prediction system and method, which can solve the technical problems of inaccurate thickness control and difficulty in continuously improving the pass rate caused by incomplete data collection, insufficient model accuracy, and delayed control response in the production process of household paper.

[0004] According to a first aspect of the present invention, an online paper thickness prediction system includes: The data acquisition module is configured to collect process parameters of the pulping and papermaking processes in real time. The pulping process includes pulping, refining, and storage, while the papermaking process includes the wire section, press section, drying section, and post-processing section. The data storage module is configured to store process parameters and historical inspection data from the inspection center, and to build a distributed data warehouse. The AI ​​prediction module is built by fusing the Long Short-Term Memory Network (LSTM) with at least one alternative algorithm, such as Random Forest or XGBoost, and is used to predict the thickness of paper products in real time based on process parameters. The output module is configured to output the thickness prediction results and the corresponding process parameter adjustment scheme, and to conduct bidirectional data interaction with the Manufacturing Execution System (MES) through the API interface. According to some embodiments of the present invention, the data acquisition module includes: Multiple online fiber analyzers are configured to monitor fiber length distribution and thickness in real time; Multiple thickness pre-detectors are configured to collect the thickness index of the base paper in real time; The online fiber analyzer and thickness pre-detector have completed interface development and compatibility testing with the existing quality inspection system. According to some embodiments of the present invention, the process parameters include multiple core features such as beating degree, press roller linear pressure, fiber wet weight, grinding disc gap, machine speed, and raw material ratio. These core features are selected from dozens of parameters through correlation analysis and principal component analysis. According to some embodiments of the present invention, the data storage module is further configured as follows: Establish data collection standards, unify data formats, ensure timestamps are accurate to the second, and standardize units; Establish rules for identifying abnormal data, and determine data that exceeds the normal range by several times the standard deviation as abnormal; The data storage period is no less than 3 years, and it supports the access and management of thousands to tens of thousands of real-time data entries per day. According to some embodiments of the present invention, the AI ​​prediction module further includes: The feature engineering unit is configured to select several core features from dozens of parameters through correlation analysis and principal component analysis; Attention mechanism unit, configured to enhance the weighting of key parameters; The model inference engine is configured to process input data and output prediction results in real time, with a prediction latency of less than or equal to 200 milliseconds. According to some embodiments of the present invention, the output module further includes: The parameter adjustment rule library, built based on the experience of process experts, is configured to convert the deviation between the predicted thickness and the target value into specific operation instructions, including the adjustment range of the pulping pressure and the optimization value of the wrinkling rate. The human-computer interaction interface is configured to display parameter curves, predicted thickness, pass rate trends and adjustment suggestions in real time, and supports historical data backtracking and model effect analysis. According to some embodiments of the present invention, the AI ​​prediction module performs real-time iterative training using the sliding window method, with a window size of 3 to 30 days, and regularly conducts model accuracy tests, and performs specific optimizations for scenarios where the deviation exceeds a set threshold. A method for online prediction of paper thickness according to a second aspect of the present invention includes the following steps: Multiple process parameters of the pulping and papermaking processes are collected in real time; Standardize the process parameters by unifying timestamps and units, and identify and remove abnormal data. The standardized data is input into the AI ​​prediction model, which is constructed by fusing LSTM with at least one of the alternative algorithms, such as Random Forest and XGBoost. Output the paper thickness prediction results and corresponding process parameter adjustment schemes, and push them to the Manufacturing Execution System (MES) via API interface; Based on the deviation between the predicted thickness and the target thickness, the corresponding process adjustment instructions are matched from the rule base and pushed to the Manufacturing Execution System (MES) to ensure that the thickness deviation is controlled within the set range. According to some embodiments of the present invention, the training process of the AI ​​prediction model includes: The first stage involves basic training based on historical data and real-time collected data to optimize model parameters. The second stage involves iterative training based on real-time data, updating the training set using a sliding window method with a window size ranging from 3 to 30 days. Regularly conduct multi-scenario verifications, covering different raw material ratios and vehicle speed conditions, to ensure that the model fit is no less than 90% and the prediction error is controlled within the set range. According to some embodiments of the present invention, a method for online prediction of paper thickness further includes: When the model prediction bias exceeds a set threshold, a special optimization mechanism is triggered to retrain or adjust the model parameters. The optimized model is deployed to the production environment to enable online model updates and closed-loop control. The paper thickness online prediction system and method according to embodiments of the present invention have at least the following beneficial effects: By integrating the process parameters of the entire pulping and papermaking process, the key factors affecting thickness are comprehensively captured, solving the problem of inaccurate prediction due to missing data; by adopting a fusion algorithm of LSTM and Random Forest / XGBoost for modeling, the system balances time-series dependence and nonlinear feature expression capabilities, improving prediction accuracy and robustness; by introducing an attention mechanism to strengthen the weight of key parameters, the model is made more consistent with actual production logic; by combining feature engineering and a lightweight inference engine, the system ensures millisecond-level low-latency response, meeting the real-time control requirements of high-speed production lines; by storing expert knowledge through a rule base and automatically converting prediction results into operation instructions, the system establishes a closed loop of "perception-analysis-decision-execution," significantly shortening the control response time, effectively improving the pass rate of raw paper thickness, and reducing resource waste.

[0005] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Detailed Implementation

[0006] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0007] Example 1: To address the technical challenges of large fluctuations in paper thickness and low yield rates in the production of tissue paper due to reliance on offline detection and delayed control responses, current technologies often employ a reactive, manual adjustment approach. This approach fails to achieve dynamic sensing and real-time prediction of key parameters throughout the pulping and papermaking processes. Particularly in high-speed paper production lines, the thickness of the base paper is influenced by the coupled effects of dozens of process parameters, including fiber state, pressing pressure, and drying temperature. Traditional single-stage data acquisition and simple statistical models struggle to fully capture these changing patterns, resulting in insufficient prediction accuracy, delayed control command generation, and severely impacting product quality stability and raw material utilization.

[0008] This application makes the following: An online paper thickness prediction system includes: The data acquisition module is configured to collect process parameters of the pulping and papermaking processes in real time. The pulping process includes pulping, refining, and storage, while the papermaking process includes the wire section, press section, drying section, and post-processing section. The data storage module is configured to store process parameters and historical inspection data from the inspection center, and to build a distributed data warehouse. The AI ​​prediction module is built by fusing the Long Short-Term Memory Network (LSTM) with at least one alternative algorithm, such as Random Forest or XGBoost, and is used to predict the thickness of paper products in real time based on process parameters. The output module is configured to output the thickness prediction results and the corresponding process parameter adjustment scheme, and to conduct bidirectional data interaction with the Manufacturing Execution System (MES) through the API interface.

[0009] This system integrates end-to-end process parameter acquisition, distributed data management, multi-algorithm fusion modeling, and MES system linkage control to form a closed-loop architecture from data perception to decision feedback. The data acquisition module covers the two core processes of pulping and papermaking, encompassing pulping, refining, storage, and multiple sub-stages including the wire section, press section, drying section, and post-processing. It can simultaneously acquire key operating parameters reflecting fiber processing status and forming process, ensuring the completeness and timeliness of input information. This module can be deployed on existing production line sensor networks, is compatible with both analog and digital signal input, and supports high-frequency sampling (time resolution ≤ 1 second), thus providing a high-fidelity raw data stream for subsequent analysis.

[0010] The data storage module adopts a distributed architecture design, possessing horizontal scalability and suitable for efficient writing and querying of massive time-series data. It not only receives real-time process parameters from the data acquisition module but also integrates historical quality testing data accumulated by the inspection center (such as paper thickness, density, and tensile strength measured in the laboratory), achieving spatiotemporal alignment between production process data and final quality data. By building a unified data warehouse, the system can complete cross-process and cross-time period data retrieval and correlation analysis within minutes, supporting model training and backtesting verification needs. This module supports mixed storage of structured and semi-structured data, is compatible with mainstream database protocols (such as SQL and NoSQL), and has access control, backup and recovery, and security auditing functions.

[0011] The AI ​​prediction module is based on a Long Short-Term Memory (LSTM) network framework, combined with at least one machine learning algorithm, such as Random Forest or eXtreme Gradient Boosting, for fusion modeling. LSTM excels at capturing long-term dependencies in time series, making it suitable for describing dynamic processes such as the evolution of pulp density over time and the transmission of vehicle speed fluctuations. Random Forest and eXtreme Gradient Boosting, on the other hand, perform well in handling nonlinear feature interactions and outlier robustness, and can be used to identify the risk of sudden thickness changes under the combined effects of multiple variables. The fusion strategy can employ weighted averaging, stacking, or gating mechanisms to adaptively adjust the contribution weights of each sub-model according to actual working conditions, thereby improving the overall generalization ability. This module runs on a high-performance computing platform, supports hot model loading and version switching, and ensures the continuity of prediction services.

[0012] The output module is responsible for converting the results generated by the AI ​​prediction module into actionable information output. It not only displays the predicted paper thickness values ​​for the current and short-term future, but also generates suggested process parameter adjustment schemes based on preset logic, such as prompting to increase refining energy, adjust press roll linear pressure, or optimize creping blade angle. All output content achieves bidirectional communication with the Manufacturing Execution System (MES) through a standardized API interface: on the one hand, it receives real-time process parameters from the MES as model input; on the other hand, it pushes prediction results and control suggestions to the MES for operator reference or to directly trigger automatic control commands. The interface follows RESTful or OPC UA specifications, supports JSON / XML format data exchange, and has breakpoint resume and encrypted transmission capabilities to ensure stable and reliable interaction in industrial environments.

[0013] The aforementioned modules coordinate their operation through a unified data bus and scheduling engine, forming a complete technology chain of "perception-storage-reasoning-feedback". The data acquisition module provides the system with source input, the data storage module ensures the persistence and availability of data assets, the AI ​​prediction module undertakes the core intelligent decision-making function, and the output module realizes the transformation of intelligent results into the production system. The four working together enable the system to significantly enhance its quality control capability for paper thickness without changing existing production equipment.

[0014] Through the above technical solutions, this application achieves the following: without adding major process equipment, it comprehensively captures key factors affecting paper thickness by systematically collecting parameters from the entire pulping and papermaking process, avoiding prediction bias caused by local data blind spots; by introducing a fusion algorithm architecture of LSTM and Random Forest / XGBoost, it takes into account both the dynamic evolution characteristics of time series and the ability of multivariate nonlinear mapping, improving the prediction accuracy and stability of the model under complex working conditions; and by breaking down information silos through bidirectional data interaction with the MES system via API interface, it realizes the automated flow of prediction results to production control instructions, significantly shortening the manual intervention cycle. Therefore, while ensuring system compatibility and low-cost deployment, it effectively improves the process control level of paper thickness, reduces raw material waste and rework costs caused by thickness deviations, and meets the comprehensive needs of modern paper manufacturing enterprises for high quality, high efficiency, and low loss.

[0015] Example 2: Based on the above embodiments, this embodiment further provides: The data acquisition module includes: multiple online fiber analyzers configured to monitor fiber length distribution and thickness in real time; multiple thickness pre-detectors configured to collect paper thickness indicators in real time; and interface development and compatibility testing of the online fiber analyzers and thickness pre-detectors with the existing quality inspection system.

[0016] This embodiment achieves continuous, non-contact acquisition of key physical parameters during pulping and papermaking by introducing multiple online fiber analyzers and thickness pre-detectors. The online fiber analyzers acquire core fiber characteristics affecting the uniformity and compactness of the paper structure, while the thickness pre-detectors capture the thickness state of the paper sheet in its early stages of forming. The combination of these two provides highly timely and high-dimensional input data support for the AI ​​prediction model, and seamless integration with existing quality inspection systems is achieved through standardized interfaces, ensuring the integrity and consistency of the entire data chain.

[0017] Online fiber analyzers are deployed in pulp delivery pipelines or open channels between the pulping end and the refining outlet. They can employ sensor devices based on image recognition or laser scattering principles, such as fiber morphology analyzers, to continuously measure microscopic morphological parameters such as fiber length distribution (mm) and fiber thickness (mg / 100m). These parameters directly affect fiber interlacing density and network structure formation ability, and are important precursor variables determining paper bulk, softness, and thickness stability. The online fiber analyzer automatically extracts small pulp samples for analysis at a fixed sampling frequency (e.g., once every 30 seconds), outputting digital signals and uploading them to a data storage module via industrial communication protocols (e.g., OPC UA or Modbus TCP). As an optional implementation, a near-infrared spectroscopy-based non-destructive testing device can be used to replace the traditional sampling analyzer, achieving high-frequency, low-maintenance continuous monitoring.

[0018] The thickness pre-measurement instrument is installed on the wet paper web transport path between the wire section outlet and the press section inlet of the paper machine, or in the dry paper area at the front end of the drying section, to measure the initial thickness of the base paper (unit: μm) in real time. It can employ capacitive, beta-ray transmission, or laser triangulation sensing technologies, featuring non-contact operation, fast response, and strong anti-interference capabilities. This instrument can obtain a thickness baseline value before the paper undergoes significant drying shrinkage and press compaction, providing an early reference signal for dynamic compensation in subsequent processes. As a variant, the thickness pre-measurement instrument can be configured with multiple detection points distributed along the width direction to achieve synchronous monitoring of transverse thickness uniformity, thereby identifying local anomalies such as edge thickness differences and segregation.

[0019] Both the online fiber analyzer and thickness pre-inspector require interface development and compatibility testing with the factory's existing quality inspection system. This includes standardizing data formats (e.g., using JSON or XML encapsulation), timestamp alignment (accurate to the second), unit standardization (e.g., converting "beating degree" from different manufacturers to °SR), and communication protocol adaptation. The debugging process encompasses hardware and software integration, data consistency verification, and fault redundancy testing to ensure that the addition of equipment does not cause system conflicts or data loss. Furthermore, necessary middleware services are configured to bridge data between heterogeneous systems, such as using an edge computing gateway for protocol conversion and caching, improving overall system stability and scalability.

[0020] Through the above technical solution, this application achieves high-frequency acquisition and system fusion of key leading parameters affecting paper thickness without interrupting the production process. By introducing fiber-level microscopic features and initial thickness signals during the forming process, the AI ​​prediction model can detect quality fluctuation trends earlier. Simultaneously, by designing interface compatibility between the new and old detection systems, the problem of information silos is avoided, and data flow efficiency is improved. Therefore, it solves the problems of traditional acquisition methods, such as reliance on offline sampling, delayed feedback, and insufficient parameter coverage, providing a reliable data foundation for achieving high-precision online thickness prediction.

[0021] Example 3: Based on the above embodiments, this embodiment further provides: The process parameters include several core features such as beating degree, press roller linear pressure, fiber wet weight, grinding disc gap, machine speed, and raw material ratio. These core features were selected from dozens of parameters through correlation analysis and principal component analysis.

[0022] This embodiment provides a method for screening key process parameters in an online paper thickness prediction system. Through mathematical modeling and statistical analysis of a large number of originally collected process parameters, the core characteristic variables that have the most significant impact on paper thickness are extracted. Among these, freeness reflects the degree to which fibers are cut and swollen, directly affecting the compactness of the paper structure; press roll linear pressure determines the stress intensity of the wet paper sheet during dewatering, thus affecting its density and thickness consistency; fiber wet weight is an important indicator of fiber bundle quality, related to fiber interlacing ability and forming uniformity; grinding disc gap controls the mechanical treatment intensity of fibers during refining, adjusting the degree of fiber fibrillation; machine speed, as a key variable in the production line's operating rhythm, affects the dwell time and energy input of each process; and raw material ratio determines the physicochemical properties of the base pulp, directly affecting paper performance. These parameters collectively constitute the set of key factors affecting paper thickness.

[0023] To identify and retain the most representative input variables from the dozens of original process parameters, this embodiment employs a two-stage feature selection strategy: First, based on Pearson correlation analysis, the linear correlation strength between each parameter and the actual paper thickness is evaluated, and a correlation coefficient threshold is set (e.g., absolute value greater than 0.4). Only highly correlated candidate parameters are retained, thereby eliminating redundant or noisy variables that have no significant relationship with the target output. Then, based on the retained highly correlated parameters, Principal Component Analysis (PCA) is applied for dimensionality reduction. Through orthogonal transformation, multiple potentially correlated original variables are converted into a few uncorrelated composite principal components. Each principal component is a linear combination of the original variables, and they are sorted according to their proportion of explained variance. The top few principal components with a cumulative contribution rate of over 85% are selected as the final model input. This method not only effectively reduces the model input dimensionality and computational complexity but also preserves the main information structure in the original data, improving the training efficiency and stability of the subsequent AI prediction module.

[0024] Furthermore, this feature selection process can be dynamically adjusted according to different product specifications or production conditions. For example, when producing thin facial tissues, the correlation between fiber wet weight and freeness is more prominent, and the system can automatically increase their weight priority; while under high-speed operating conditions, the coupling effect between machine speed and pressing pressure is enhanced, and PCA can capture new principal component combinations to adapt to changes. As an optional implementation, Spearman rank correlation analysis can be introduced to replace Pearson correlation to handle data scenarios with nonlinear monotonic relationships; or nonlinear dimensionality reduction algorithms such as t-SNE (t-distributed random neighborhood embedding) and UMAP (uniform manifold approximation and projection) can be used as supplementary schemes to PCA to explore potential clustering structures in high-dimensional space. Sparse feature selection can also be combined with L1 regularization (such as Lasso regression) to further improve model interpretability.

[0025] Through the above technical solution, this application achieves effective extraction and optimized expression of massive process parameters. Correlation analysis eliminates interfering parameters that are unrelated to or weakly correlated with paper thickness, reducing noise interference during model learning. Principal component analysis integrates highly correlated variables into a few comprehensive indicators, avoiding multicollinearity and improving model convergence speed. Furthermore, focusing on mathematically validated core features enhances the generalization ability and robustness of the prediction model. Therefore, this technical approach solves the problems of heavy computational burden, model instability, and low prediction accuracy caused by too many input parameters in traditional prediction methods, achieving the technical effect of improving the response speed and accuracy of AI prediction modules and ensuring the feasibility of online real-time control.

[0026] Example 4: Based on the above embodiments, this embodiment further provides: Establish data collection standards, unify data formats, ensure timestamps are accurate to the second, and standardize units; establish rules for identifying abnormal data, and determine data that exceeds the normal range by several times the standard deviation as abnormal; the data storage period is no less than 3 years, and it supports the access and management of thousands to tens of thousands of real-time data entries per day.

[0027] This embodiment improves the consistency of multi-source process parameters in both time and numerical dimensions through a systematic and standardized data management mechanism, ensuring the quality and stability of the data upon which subsequent AI model training and inference rely. Specifically, all data collected from the pulping and papermaking processes must follow a unified data processing workflow before entering the distributed data warehouse, ensuring that information from different equipment and systems is comparable and time-series aligned.

[0028] "Establishing data acquisition standards, unifying data formats, ensuring timestamps are accurate to the second, and standardizing units" refers to setting clear technical standards during the data acquisition phase to eliminate data heterogeneity issues caused by differences in equipment manufacturers or inconsistent system interfaces. For example, fiber length data from online fiber analyzers is standardized to millimeters (mm), press roller linear pressure is standardized to kilonewtons per meter (kN / m), and machine speed is standardized to meters per minute (m / min). Simultaneously, all data timestamps are recorded with second-level precision in the UTC+8 time zone to avoid misalignment of multi-source signals caused by sub-millisecond precision loss, thus ensuring that time-series models such as LSTM can accurately capture parameter change trends. As an optional implementation, timestamps can also be recorded at the nanosecond level and downsampled to the second level after being stored, balancing high-precision acquisition with system performance; or, in specific high-frequency monitoring scenarios, some key parameters (such as grinding disc gap fluctuations) can retain millisecond-level timestamps and be modeled and processed separately.

[0029] "Establishing rules for identifying abnormal data and classifying data exceeding the normal range by several times the standard deviation as abnormal" refers to building an automated data quality control mechanism based on statistical methods. Specifically, the system periodically calculates the historical mean and standard deviation (σ) of each process parameter. When the real-time collected value deviates from the mean by more than a preset multiple (such as 3 or 5 times the standard deviation), it is marked as suspected abnormal data. For example, if historical data from a production line shows a standard deviation of 2g for fiber wet weight and a mean of 18g, then when a fiber wet weight reading exceeding 26g is detected at a certain moment, the system will trigger an alarm and, based on the configured strategy, choose automatic rejection, manual review, or temporary storage for further investigation. This rule supports dynamic adjustment of the threshold multiple and is suitable for robust judgment under different operating conditions. As a variant, box plot method (IQR) or isolation forest algorithm can also be introduced as supplementary means to improve the ability to identify outliers in non-normally distributed data, especially suitable for scenarios where frequent raw material changes or equipment aging leads to data distribution shifts.

[0030] The requirement of a data storage period of at least three years signifies that the system design meets the needs of long-term data archiving, providing ample historical samples to support the iterative optimization of AI models. Over three years of continuous operational data covers various production scenarios, including seasonal raw material changes, equipment maintenance cycles, and product specification switching, which helps improve the model's generalization ability. Data is stored in a distributed database cluster, partitioned by time, supporting rapid retrieval and batch access. As an alternative, cold data can be migrated to a low-cost object storage system (such as HDFS or cloud storage) according to the enterprise's actual needs, maintaining access transparency through metadata indexing.

[0031] The ability to "support the access and management of thousands to tens of thousands of real-time data entries per day" demonstrates the system's data throughput capability under high concurrency. Taking Vinda Paper's No. 3 production line as an example, it generates approximately 800 valid process records per hour, with a daily data volume of 19,000 records. The system uses message queues (such as Kafka) for buffering and combines them with streaming computing frameworks (such as Flink) to achieve real-time cleaning and writing, ensuring stable operation even during peak periods. The database architecture adopts a sharding strategy, sliced ​​by production line number and time dimension to improve query efficiency. In addition, the system reserves horizontal expansion interfaces, allowing for flexible expansion when more production lines are connected or the data collection frequency is increased in the future.

[0032] Through the aforementioned technical solutions, this application achieves standardized data governance across the entire data chain, from source to storage. By establishing unified data collection standards and enforcing timestamp and unit standardization, model misjudgments caused by unit confusion or time asynchrony are avoided. The establishment of an anomaly identification mechanism based on standard deviation effectively filters erroneous data caused by sensor drift, communication interruptions, etc., improving the reliability of input data. The setting of a data retention period of no less than three years provides sufficient historical evidence for model backtracking training, deviation analysis, and root cause analysis. The ability to continuously access tens of thousands of data points daily meets the stringent requirements of industrial-grade continuous operation for system stability and scalability. Overall, this data management system lays a solid foundation for the high-precision, low-latency operation of the AI ​​prediction module.

[0033] Example 5: Based on the above embodiments, this embodiment further provides: The AI ​​prediction module also includes: The feature engineering unit is configured to select several core features from dozens of parameters through correlation analysis and principal component analysis; Attention mechanism unit, configured to enhance the weighting of key parameters; The model inference engine is configured to process input data and output prediction results in real time, with a prediction latency of less than or equal to 200 milliseconds. This embodiment relates to a functional enhancement scheme for the AI ​​prediction module in an online paper thickness prediction system. It aims to address issues in existing technologies such as low prediction accuracy and lagging control caused by redundant input parameters, ineffective identification of key process variables, and insufficient model response speed. In actual production, the pulping and papermaking processes involve as many as 42 process parameters. Directly inputting all of these into the model not only increases the computational burden but may also introduce noise interference, leading to a decrease in the model's generalization ability. Furthermore, under high-speed paper machine operation conditions (speeds can reach 1000m / min), if the prediction delay exceeds 200ms, it will be difficult to support real-time closed-loop control, resulting in accumulated thickness deviations and affecting product quality stability.

[0034] The feature engineering unit is used to preprocess and reduce the dimensionality of dozens of raw process parameters. Specifically, this unit first uses Pearson Correlation to assess the linear correlation between each parameter and paper thickness, eliminating weakly correlated parameters with correlation coefficients below a set threshold (e.g., |ρ|<0.3). Then, Principal Component Analysis (PCA) is applied to orthogonally transform the retained highly correlated parameters, extracting principal components that can explain ≥85% of the cumulative variance contribution rate as core features input to the subsequent model. This process achieves a shift from "full parameter-driven" to "key feature-driven," significantly reducing the model input dimensionality and improving training efficiency and prediction stability. For example, in a certain production scenario, the original 42 parameters can be simplified to 27 core features after feature engineering, including beating degree, press roll linear pressure, fiber wet weight, and grinding disc gap, retaining the main process information while avoiding redundant interference. As an alternative implementation, recursive feature elimination (RFE) or tree-based feature importance ranking can be used to replace PCA to adapt to data distributions with stronger nonlinear relationships.

[0035] The attention mechanism unit is integrated into the internal structure of the AI ​​fusion model. Its function is to dynamically adjust the attention weights of different input features at different time points. This unit constructs a learnable attention weight matrix, enabling the model to automatically focus on key parameters that have a greater impact on thickness changes when processing sequential data. For example, in the grinding stage, the system tends to assign higher attention scores to "grinding gap" and "motor current"; while in the pressing stage, it increases the weight ratio of "pressing roller linear pressure" and "machine speed". This mechanism gives the model stronger time-varying adaptability, allowing it to autonomously adjust parameter sensitivity according to the current working conditions, thus more closely reflecting the actual production process logic. In terms of implementation, the attention mechanism can adopt additive attention or scaled dot-product attention structures and be embedded in the hidden state propagation process of the LSTM network. It can also apply weight allocation to the output results of each sub-model at the multi-algorithm fusion level, achieving cross-model level attention control.

[0036] The model inference engine receives the standardized data stream after feature engineering, calls the trained AI prediction model to perform forward inference, and returns the thickness prediction result within a limited time. To meet industrial-grade real-time requirements, the engine adopts a lightweight deployment strategy, including but not limited to model pruning, quantization compression, and operator fusion optimization techniques, reducing computational resource consumption while ensuring prediction accuracy. At the hardware level, it supports GPU acceleration or deployment on edge computing devices (such as industrial AI boxes), ensuring end-to-end prediction latency does not exceed 200 milliseconds. Furthermore, the inference engine has high concurrency processing capabilities, capable of simultaneously responding to data requests from multiple production lines, supporting a continuous inflow and instant output of hundreds of data records per second. As an optional implementation, a caching mechanism can be used to preload typical pattern response tables under high-frequency operating conditions, achieving near real-time inference in scenarios with extremely low latency requirements.

[0037] The three components work together: the feature engineering unit filters out invalid information before input, improving input quality; the attention mechanism unit selectively amplifies key parameters within the model; and the model inference engine ensures the overall system's response speed and operational efficiency. Together, they form a highly efficient, accurate, and low-latency AI prediction execution chain.

[0038] Through the above technical solution, this application achieves rapid and accurate prediction of paper thickness under complex process environments. By scientifically screening the original parameters, the computational load on the model is reduced and the effectiveness of feature representation is improved. The introduction of an attention mechanism enables the model to dynamically focus on key control variables according to different process stages, enhancing the consistency between prediction logic and actual production. The optimized inference engine architecture ensures that prediction results can be fed back to the control system within milliseconds, meeting the real-time control requirements of high-speed paper machine continuous production. Therefore, this embodiment effectively solves the technical challenges of large interference from multi-source heterogeneous parameters, difficulty in identifying key factors, and slow prediction response, providing reliable support for early intervention and closed-loop control of thickness deviations.

[0039] Example 6: Based on the above embodiments, this embodiment further provides: The output module also includes: a parameter adjustment rule library, built based on the experience of process experts, configured to convert the deviation between the predicted thickness and the target value into specific operation instructions, including the adjustment range of the grinding pressure and the optimization value of the wrinkling rate; and a human-machine interface, configured to display parameter curves, predicted thickness, pass rate trends and adjustment suggestions in real time, and support historical data backtracking and model effect analysis.

[0040] This embodiment, by introducing a parameter adjustment rule base and a human-machine interface into the output module, realizes the transformation from AI prediction results to executable control strategies, and improves the visualization level and decision support capabilities of system operation. The parameter adjustment rule base is a knowledge mapping mechanism built upon long-term accumulated process expert experience. Its core function is to automatically match the deviation between the thickness prediction value output by the AI ​​model and the production target value into a series of specific, executable process parameter adjustment instructions. For example, when the predicted thickness exceeds the target value by ±0.01mm, the system will determine, based on preset experience rules, which key parameter should be adjusted first—such as reducing the pressure of the press rollers to reduce compression, or adjusting the grinding wheel gap in the refiner to change the fiber bonding state. Similarly, in the production of thin tissue paper, if high fiber wet weight is detected, leading to increased basis weight and affecting thickness, the rule base can trigger a combined adjustment scheme of "reducing pulp concentration + increasing beating degree." These operation instructions are not simple linear relationships, but rather nonlinear response strategies formed after comprehensively considering the coupling effects of multiple processes, ensuring that the adjustment actions effectively correct deviations without causing other quality fluctuations.

[0041] The human-machine interface (HMI) serves as an information bridge between operators and the AI ​​system, integrating functions such as dynamic data visualization, trend alerts, suggestion pushes, and historical traceability. The interface uses a timeline as its main axis, synchronously displaying the change curves of key process parameters (such as machine speed, beating degree, and drying temperature), real-time predicted thickness trends, current batch pass rate statistics, and system-recommended adjustment suggestions. All data displays can be filtered and replayed by shift, date, product specification, and other dimensions, facilitating managers to review and analyze the production process within a specific time period. Furthermore, the interface includes a built-in model performance monitoring module, which displays recent prediction error distribution, model confidence changes, and rule base call frequency, assisting technicians in evaluating the stability and adaptability of the AI ​​system. To improve usability, the interface layout adopts a segmented design: the left side is the parameter monitoring area, the middle is the trend chart area, and the right side is the suggested operation area. Key alarm information is highlighted with color coding (e.g., red indicates out-of-tolerance, yellow indicates approaching the threshold), ensuring operators can identify anomalies and respond immediately.

[0042] The two components work together to form a closed-loop logical chain of "perception—analysis—decision—presentation". When the AI ​​prediction module outputs a new thickness prediction value, the system first calculates the deviation from the set target value, then queries the mapping table stored in the parameter adjustment rule base to retrieve the most matching set of operation strategies. After priority sorting and conflict detection, this strategy generates standardized operation instructions and pushes them to the MES system through the API interface. At the same time, the human-machine interface prompts the operator to confirm the execution through pop-ups or highlights. The entire process can complete the transformation from data to action without human intervention, significantly shortening the response cycle of the traditional "manual report review—experience judgment—manual adjustment". Especially in the operating environment of high-speed paper machines (speed ≥ 800m / min), this automated decision-making mechanism can achieve millisecond-level response, avoiding the accumulation of quality fluctuations caused by human delay.

[0043] Through the above technical solution, this application realizes the transformation of abstract AI prediction results into specific and controllable production process actions, solving the problem of the disconnect between prediction and execution in existing systems. The parameter adjustment rule base encapsulates the knowledge and experience of senior process engineers, enabling stable production quality to be maintained even in the absence of highly skilled operators; the human-machine interface enhances the transparency and interpretability of the system, allowing operators to not only "know what" but also "know why," thereby increasing their trust in and willingness to use the AI ​​system. Simultaneously, the function of supporting historical data backtracking provides valuable feedback data for subsequent model optimization, forming a virtuous cycle mechanism of "application-feedback-iteration." This embodiment is particularly suitable for multi-variety, small-batch tissue paper production lines, maintaining high control accuracy and adaptability even under conditions of frequent brand changes and large fluctuations in raw materials.

[0044] Example 7: Based on the above embodiments, this embodiment further provides: The AI ​​prediction module uses a sliding window method for real-time iterative training, with a window size ranging from 3 to 30 days. Model accuracy is tested regularly, and specific optimizations are performed for scenarios where the deviation exceeds a set threshold.

[0045] This embodiment provides a technical means for dynamically updating an AI prediction model, aiming to improve the model's adaptability to changes in the production environment during long-term operation. Specifically, the AI ​​prediction module uses a sliding window mechanism to continuously update the training dataset, combined with periodic accuracy evaluation and targeted optimization strategies, to ensure the model consistently maintains good predictive performance.

[0046] The sliding window method organizes historical and real-time acquired process parameter data chronologically, retaining only the most recent data as the current training set, and continuously discarding outdated data and incorporating new data over time. The window's time span is set to "3 to 30 days," for example, 7 days, 14 days, or 30 days, with the specific value dynamically adjusted based on factors such as production line stability, raw material change frequency, and equipment aging trends. In this way, the training data continuously reflects the current actual production conditions, avoiding model lag due to the use of outdated data. As an optional implementation, window updates can use fixed-step scrolling (e.g., sliding forward one day daily) or event-triggered updates (e.g., immediately resetting the window when a raw material batch change is detected).

[0047] Regular model accuracy testing refers to evaluating the performance of the current AI prediction model on the latest validation set at preset intervals (e.g., every 24 hours, every 72 hours, or weekly). Key metrics include goodness of fit (R²), mean absolute error (MAE), and maximum deviation. Test data comes from recent independent samples that do not overlap with the training set, covering different combinations of machine speed, beating degree, and fiber ratio to ensure the representativeness of the evaluation results. When test results show that the overall prediction error has not increased significantly, but the deviation exceeds a set threshold under specific operating conditions (e.g., thickness prediction error greater than ±0.015mm), the system determines that there is a local degradation phenomenon and initiates a special optimization process.

[0048] For scenarios where deviations exceed a set threshold, specific optimizations are performed. These include: identifying key operating condition features that cause errors to exceed limits (e.g., high pulp ratio + low linear pressing); extracting complete process data sequences for the corresponding time period from the data warehouse to construct a dedicated sub-training set; subsequently, enhancing this subset (e.g., adding noise perturbations, interpolating to fill missing points); and fine-tuning relevant parameter layers in the original model using transfer learning or incremental training, especially the weight coefficients associated with key parameters in the attention mechanism unit. After optimization, the new model must undergo regression testing to confirm that the error in this scenario has fallen back to within the allowable range before deployment. Furthermore, as a variant, an independent auxiliary prediction model can be established for frequently occurring abnormal scenarios, automatically switching to it when the main model outputs anomalies, forming a redundant fault-tolerance mechanism.

[0049] The synergistic effect of each technical component is manifested in the following ways: the sliding window mechanism ensures the freshness and representativeness of the training data, allowing the model to evolve naturally with the production pace; regular accuracy testing provides a quantitative feedback channel, promptly exposing potential performance degradation; and specialized optimization enables problem-oriented, refined repairs, avoiding the resource consumption and interruption risks associated with global retraining. The combination of these three elements forms a closed-loop learning framework, enhancing the adaptability and robustness of the AI ​​prediction module.

[0050] Through the above technical solutions, this application achieves the continuous evolution capability of the AI ​​prediction model. By employing the sliding window method, it ensures that training data always focuses on recent production conditions, effectively avoiding model mismatch problems caused by raw material fluctuations or equipment wear and tear. Regular accuracy testing enables early detection of signs of model performance degradation, preventing the accumulation and amplification of prediction bias. The introduction of a dedicated optimization mechanism for scenarios exceeding standards allows for precise repair of local defects without affecting overall stability. Therefore, this embodiment significantly extends the effective service life of the model and enhances its practical value in complex and ever-changing industrial environments.

[0051] Example 8: Based on the above embodiments, this embodiment further provides: A method for online prediction of paper thickness includes the following steps: real-time acquisition of multiple process parameters from the pulping and papermaking processes; data standardization of the process parameters, unifying timestamps and units, and identifying and removing abnormal data; inputting the standardized data into an AI prediction model, which is constructed based on the fusion of LSTM with at least one alternative algorithm among random forest and XGBoost; outputting the paper thickness prediction result and the corresponding process parameter adjustment scheme, and pushing it to the Manufacturing Execution System (MES) via an API interface; based on the deviation between the predicted thickness and the target thickness, matching the corresponding process adjustment instructions from the rule base, and pushing the adjustment instructions to the MES to ensure that the thickness deviation is controlled within the set range.

[0052] Step 1: Real-time acquisition of multiple process parameters for the pulping and papermaking processes. in, "Real-time data acquisition" refers to the continuous acquisition of dynamically changing process parameters during the production process at a granular level down to the second, ensuring the temporal continuity and spatial coverage of the data. This acquisition process covers the pulping, refining, and storage stages in the pulping process, as well as the wire section, press section, drying section, and post-processing stages in the papermaking process, forming a complete data chain. The acquired process parameters include, but are not limited to, key variables such as freeness, press roll linear pressure, fiber wet weight, grinding wheel gap, machine speed, and raw material ratio. These parameters directly or indirectly affect fiber distribution, moisture evaporation, and structural density during the paper forming process, thus determining the final paper thickness. The data acquisition equipment can be a distributed sensor network, integrated into the existing DCS (Distributed Control System), supporting industrial communication protocols such as Modbus / TCP and OPC UA, achieving seamless integration with the central data platform. For example, high-precision torque sensors can be deployed in the grinding section to monitor the grinding disc load, hydraulic pressure transmitters can be installed in the pressing section to measure linear pressure fluctuations, and infrared moisture meters can be deployed in the mesh section to track the dewatering rate. Furthermore, online fiber analyzers can be used to obtain fiber length distribution and thickness information as supplementary input features to enhance the model's characterization capabilities. As an optional implementation method, the sampling frequency can be configured differently according to the variation characteristics of different parameters: for fast-response parameters (such as vehicle speed and pressure), the sampling period is set to 1 second; for slowly changing parameters (such as raw material ratio and pulp concentration), the average value of 5-10 seconds can be uploaded, which ensures data representativeness and reduces communication load. Step 2: Standardize the process parameters by unifying timestamps and units, and identify and remove outlier data. in, "Data standardization" refers to the process of converting raw data from different sources with different dimensions and orders of magnitude into a unified format and scale. The aim is to eliminate model training bias caused by inconsistent units or differences in numerical ranges. Specific operations include: aligning the timestamps of all parameters to the UTC+8 time zone, accurate to the second, and using normalization methods (such as Min-Max Scaling or Z-score standardization) to map them to the [0,1] or standard normal distribution interval. For example, vehicle speed (unit: m / min) is scaled to the 0–1 interval after Min-Max transformation, while slurry density (°SR) is processed using Z-score to achieve a mean of 0 and a variance of 1. Identifying and removing outlier data is a crucial step in ensuring the quality of model input. This application establishes an anomaly detection mechanism based on a combination of statistics and domain knowledge. On one hand, a general threshold rule is set, such as marking any parameter exceeding its historical mean ± 3 times the standard deviation as a suspected anomaly. On the other hand, process logic constraints are introduced; for example, when the vehicle speed exceeds the equipment's rated upper limit but the pressure does not rise accordingly, it is judged as signal drift or sensor failure. Identified outliers can be processed using forward and backward sliding window interpolation, KNN imputation, or direct discarding strategies. Cubic spline interpolation can be optionally used to maintain curve smoothness. As an alternative, lightweight anomaly detection models (such as IsolationForest or Autoencoder) can be introduced to run online to achieve intelligent identification of atypical working conditions, which is especially suitable for data cleaning during the trial production of new raw materials or equipment commissioning. Step 3: Input the standardized data into the AI ​​prediction model. The AI ​​prediction model is constructed by fusing LSTM with at least one of the following alternative algorithms: Random Forest and XGBoost. in, The "AI Prediction Model" is a multi-algorithm fusion architecture for modeling time-series data. Its core is dominated by a Long Short-Term Memory (LSTM) network, supplemented by ensemble learning algorithms such as Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) to form a hybrid prediction framework. LSTM excels at capturing the long-term dependencies of process parameters over time, effectively modeling the dynamic lag effects throughout the entire chain from pulping to paper drying. Meanwhile, RF and XGBoost, with their strong nonlinear fitting capabilities and anti-overfitting properties, compensate for LSTM's shortcomings in small samples or under noise interference, improving overall robustness. Model fusion can be achieved using weighted averaging, stacking, or gating mechanisms. For example, during the inference phase, each sub-model outputs a predicted value, and weights are dynamically allocated based on the RMSE performance on the historical validation set; alternatively, a meta-learner can be designed to synthesize the outputs of each model with current operating conditions (such as machine speed range and pulp type) to make a final decision. The model input dimension consists of a set of core parameters selected through feature engineering, typically 25–30 items, covering key factors such as beating degree, press line pressure, fiber wet weight, mill gap, machine speed, and basis weight. As an optional implementation, an attention mechanism module can also be embedded in the model structure to automatically learn the importance weights of different parameters at different times. For example, during the stage of frequent adjustment of wrinkle rate, the system automatically enhances the attention scores of "scraper angle" and "drying cylinder temperature," thereby improving local prediction sensitivity. In addition, the model supports incremental learning mode, which can periodically load new data to fine-tune parameters without affecting online services, maintaining long-term effectiveness. Step 4: Output the paper thickness prediction results and the corresponding process parameter adjustment plan, and push them to the Manufacturing Execution System (MES) via API interface. in, "Output paper thickness prediction results" refers to the estimated thickness of the raw paper at the next moment or in the current batch, generated after each inference iteration of the model. The unit is micrometers (μm), and it is usually presented in the form of rolling predictions, such as updating the thickness trend curve for the next 30 seconds every 5 seconds. This prediction result not only includes a single value, but may also include a confidence interval (such as 95% CI) for operators to assess the level of uncertainty. The "corresponding process parameter adjustment scheme" is a set of operation suggestions automatically generated based on prediction deviations. Essentially, it's an intermediate layer of logic that transforms AI output into executable control actions. For example, if the predicted thickness is 0.015mm higher than the target value, the system suggests "reducing grinding energy consumption by 5%", "increasing press roller linear pressure by 3%", and "fine-tuning the feed concentration by 0.1g / L". These schemes are not simply derived from experience, but are collaboratively optimized by a rules engine based on the current operating context (such as raw material batch, temperature, and humidity). "Pushing data to the Manufacturing Execution System (MES) via API interface" enables data connectivity between the IT and OT layers. The API adopts a RESTful architecture, supports HTTPS encrypted transmission, and the message body is in JSON format, containing fields such as `timestamp`, `predicted_thickness`, `adjustment_suggestions`, and `confidence_level`. After receiving the data, the MES can automatically trigger alarms, generate work orders, or push the data to the shop floor HMI terminal, achieving closed-loop management without human intervention. As an alternative implementation, the push mechanism can also use the MQTT protocol to achieve low-latency message broadcasting, which is particularly suitable for centralized monitoring scenarios of multiple production lines; or it can be configured to trigger a notification only when the deviation exceeds a preset threshold (such as ±0.012mm) to avoid information overload. Step 5: Based on the deviation between the predicted thickness and the target thickness, match the corresponding process adjustment instructions from the rule base and push the adjustment instructions to the Manufacturing Execution System (MES) to ensure that the thickness deviation is controlled within the set range. in, The deviation between the predicted thickness and the target thickness is the trigger condition for the adjustment decision, calculated as the absolute value or relative error of the difference between the two. The target thickness is defined by the production order and is usually between 0.08 and 0.16 mm, with an allowable fluctuation range generally set at ±0.01 mm. When the deviation exceeds this limit, the adjustment process is initiated. The "Rule Base" is a knowledge graph-based database built upon the experience of papermaking process experts, storing hundreds of "condition-action" mapping rules. Each rule includes prerequisites (such as "machine speed > 800 m / min and wood pulp content < 70%), recommended actions (such as "increase the upstream pressing pressure by 2 bar"), priority level, and applicable scenario tags. Rules can be maintained through a graphical editor and support version management and canary releases. The "matching of corresponding process adjustment instructions" strategy combines fuzzy matching with priority ranking. The system first filters all candidate rules that meet the current operating conditions, and then selects the optimal option based on the real-time feedback score of the control effect. For example, in high-speed operation, it prioritizes fast-response adjustment methods (such as changing the vacuum chamber negative pressure) rather than time-consuming actions (such as replacing grinding discs). The final step in the control loop is to "push the adjustment command to the MES". The pushed content is a structured control command, including fields such as device ID, parameter name, target value, and execution time limit. After receiving it, the MES can link the PLC to perform automatic adjustment or prompt the operator for manual intervention. The entire process can be completed within 200 milliseconds, which is much faster than traditional manual response (which usually takes 5-10 minutes). As an alternative implementation method, the rule base can also be linked with the reinforcement learning module, enabling the system to continuously explore better strategy combinations during long-term operation, gradually replacing static experience rules, and realizing the evolution from "assisting decision-making" to "autonomous optimization".

[0053] Through the above-described steps, this application achieves an end-to-end automated process from raw data acquisition to final control command output. By integrating all process parameters and performing standardized preprocessing, the problem of highly heterogeneous and inconsistent input data quality is solved, thereby improving the generalization ability of the AI ​​model. By employing LSTM combined with multiple machine learning algorithms for modeling, the limitations of single models in complex nonlinear systems are overcome, achieving high-precision, low-latency thickness prediction. Furthermore, by establishing a rule-based deviation response mechanism and deeply integrating it with MES via API, a complete closed loop of "perception-analysis-decision-execution" is established, significantly shortening the control response time and effectively suppressing thickness fluctuations. Ultimately, this method not only provides a replicable and scalable technical path for paper quality prediction but also offers a typical example for the implementation of intelligent manufacturing in the traditional light industry.

[0054] Example 9: Based on the above embodiments, this embodiment further provides: The training process of the AI ​​prediction model includes: the first stage is basic training based on historical data and real-time collected data to optimize model parameters; the second stage is iterative training based on real-time data, updating the training set through the sliding window method, with a window size of 3 to 30 days; and regular multi-scenario verification is carried out, covering different raw material ratios and vehicle speed conditions to ensure that the model fit is not less than 90% and the prediction error is controlled within the set range.

[0055] Step 1: The first stage involves basic training based on historical data and real-time collected data to optimize model parameters; The "first stage" refers to the initial modeling phase during model deployment. Its purpose is to build an AI model with basic predictive capabilities using 1 million historical data points accumulated over 5 years and real-time process parameters from the integrated system. This stage primarily relies on these 1 million historical data points and core process parameters such as beating degree, press roller linear pressure, fiber wet weight, grinding disc gap, machine speed, and raw material ratio continuously integrated from the MES system. This data, after standardization (unifying timestamps to the second level and unit normalization) and outlier removal (values ​​exceeding ±3 standard deviations are considered outliers), serves as the training input set. Basic training employs supervised learning, using actual detected thickness as labels. The weights of the LSTM neural network are adjusted through backpropagation, and ensemble learning methods such as random forest or XGBoost are combined to compensate for and optimize nonlinear features, thereby improving the overall model robustness. During this process, hyperparameters such as learning rate, hidden layer dimension, and batch size are fine-tuned through grid search and cross-validation, ultimately obtaining a basic model with high initial accuracy. Optionally, a transfer learning strategy can be introduced into the basic training. This involves pre-training the model on similar production lines or under historical operating conditions, and then fine-tuning it for the current No. 3 paper production line to accelerate convergence and reduce dependence on local data. Another variation is to use a semi-supervised learning mechanism, which utilizes a large amount of unlabeled real-time data for autoencoder pre-training when labeled data is insufficient, thereby enhancing feature representation capabilities.

[0056] Step 2: The second stage involves iterative training based on real-time data, updating the training set using a sliding window method with a window size ranging from 3 to 30 days. The "second phase" refers to the dynamic update phase after the model goes live, designed to address issues such as process drift, equipment aging, and raw material fluctuations during production, preventing performance degradation due to environmental changes. "Iterative training" means the model is not fixed once, but periodically retrained or its parameters fine-tuned using the latest collected data, achieving online evolution. The adopted "sliding window method" is an incremental learning strategy. Specifically, each training session retains only data from the most recent N days as the training set. N is set to 7, 14, or 30 days depending on actual production stability, with 7 days being a typical value. When new data arrives, the oldest batch is automatically removed, ensuring the training set always reflects the current operating conditions. This mechanism effectively controls computational resource consumption while improving the model's responsiveness to short-term trends. In implementation, the sliding window can be automatically managed through a time-partitioned table in a distributed data warehouse, triggering a model retraining process daily in conjunction with a scheduling task. Training modes can be selected as full retraining or incremental fine-tuning (e.g., updating only the weights of the last few layers). The latter is more suitable for high-frequency update scenarios, reducing the impact of latency. Furthermore, to prevent misjudgments caused by concept drift, a data distribution consistency check (e.g., KS test) can be added before training, initiating the full training process only when the differences between the old and new data are significant.

[0057] Step 3: Conduct regular multi-scenario verifications, covering different raw material ratios and vehicle speed conditions, to ensure that the model fit is not less than 90% and the prediction error is controlled within the set range. Among these, "multi-scenario validation" is a crucial step in ensuring the model's generalization ability. Its core lies in simulating various complex combinations of working conditions that may occur in real production, evaluating the model's performance under non-ideal conditions. "Different raw material ratios" cover a range of wood pulp content from 60% to 80%, corresponding to different fiber structures and paper characteristics; "vehicle speed conditions" cover an operating range of 600m / min to 1000m / min, representing different dewatering efficiencies and drying intensities. Validation tests are conducted monthly, with at least 10 typical working condition samples selected each time. These samples are injected into the model for prediction, and the model's fit (R²) and absolute error (MAE) are compared with actual quality inspection results. If the fit is below 90% or the error exceeds ±0.01mm in a particular scenario, a dedicated optimization mechanism is triggered, supplementing training data or adjusting feature weights for that specific working condition. To further improve validation coverage, a virtual simulation environment can be constructed to generate boundary condition data based on the mechanistic model for extreme testing. For example, the forming process of thin paper under a combination of high machine speed and low beating degree can be simulated to verify whether the model can still accurately capture the thickness change trend. All validation results are recorded in the model lifecycle management system as an important basis for subsequent version iterations.

[0058] Through the above-described steps, this application achieves a leap in AI prediction model capabilities from static modeling to dynamic evolution. This is because the model initially accumulates knowledge by integrating historical and real-time data, laying a solid foundation for prediction; because a sliding window mechanism is used during operation to continuously absorb new data, the model can adapt to long-term process fluctuations; and because performance bottlenecks are actively identified and deviations are corrected promptly through regular multi-scenario verification, the technical problem of prediction inaccuracies caused by data drift in traditional one-time training models is solved. This achieves the technical effect of long-term stable model operation and consistently meeting prediction accuracy standards. This application can be applied to thickness prediction systems in tissue paper production, as well as to other industrial AI applications in continuous manufacturing processes that require long-term high-precision prediction.

[0059] Example 10: Based on the above embodiments, this embodiment further provides: When the model prediction deviation exceeds a set threshold, a special optimization mechanism is triggered to retrain or adjust the model parameters; the optimized model is then deployed to the production environment to achieve online model updates and closed-loop control.

[0060] The phrase "triggering a special optimization mechanism when the model prediction deviation exceeds a set threshold" refers to the system continuously monitoring the error between the predicted paper thickness output by the AI ​​prediction module and the actual inspection data, and setting a configurable deviation threshold (e.g., ±0.015mm or a relative error of 5%). When the prediction deviation exceeds this threshold for multiple consecutive sampling periods, the system automatically determines that the model performance has deteriorated, possibly caused by external disturbances such as sudden changes in raw material ratios, abnormal equipment status, or sensor drift. In this case, instead of relying on the conventional periodic iterative training process, the special optimization mechanism is immediately activated as an emergency response measure. This mechanism includes, but is not limited to, initiating high-frequency data acquisition, enhancing the weight of historical similar working condition samples, dynamically reconstructing the feature engineering process, introducing incremental learning strategies for local parameter fine-tuning, or using Bayesian optimization methods to quickly optimize key hyperparameters (such as the number of LSTM layers, learning rate, number of random forest trees, etc.). The deviation threshold can be flexibly configured according to different product specifications (e.g., tissue paper, facial tissue, toilet paper rolls) and process stages to ensure a balance between sensitivity and robustness.

[0061] "Retraining or adjusting model parameters" refers to selecting the appropriate optimization path based on the specific causes of the current anomaly. If the deviation originates from short-term noise interference, only lightweight parameter calibration (such as bias correction and weight decay) is performed on the model. If it is a structural shift (such as changes in fiber properties due to a change in wood pulp supplier), a new round of full training based on sliding window expansion is initiated, integrating high-quality real-time data from the last 3 to 30 days with matching historical datasets. The training process can be completed collaboratively on edge computing nodes or in the cloud, utilizing a distributed computing framework to accelerate model convergence. After training, the new model must pass an accuracy verification test (such as a fit ≥90% and a maximum error ≤±0.012mm) before entering the deployment queue to avoid the risk of reverse adjustment caused by deploying a substandard model.

[0062] "Deploying the optimized model to the production environment" refers to achieving hot updates or canary releases through an automated model deployment pipeline. The system supports a dual-model parallel operation mechanism. After the new model outputs stable results and its consistency with the old model reaches a preset standard, traffic is gradually switched until the original model is completely replaced. The entire deployment process does not interrupt the production line operation or affect the normal interaction of the MES system, ensuring production continuity and system stability. After deployment, the system automatically records the version number, training time, performance indicators, and operation logs for easy traceability and auditing.

[0063] Among them, "realizing online model updates and closed-loop control" embodies a complete feedback chain from monitoring to intervention. This closed loop not only includes a control loop of "perception-analysis-decision-execution," but also embeds self-evolutionary capabilities: each specific optimization generates new knowledge accumulation, which is used to expand the rule base and improve subsequent training strategies. For example, for a thickness fluctuation event caused by wear of grinding discs, the system can mark this pattern as a typical failure case after completing this optimization, and can respond more quickly when similar signals occur in the future. In addition, the human-machine collaboration mechanism in the closed loop allows process engineers to review and intervene in the automatic optimization results, ensuring an effective combination of technical autonomy and human experience.

[0064] Through the above-described steps, this application enables the system to proactively identify anomalies, autonomously initiate model retraining, and safely update the system in the face of sudden process disturbances or model degradation, thereby significantly improving the robustness and timeliness of the AI ​​prediction system. By setting quantifiable deviation threshold trigger conditions, it achieves a shift from passive maintenance to proactive early warning. By supporting hot model updates and seamless switching, it avoids the downtime and debugging required for traditional system upgrades, reducing operation and maintenance costs. Finally, it constructs an intelligent prediction closed-loop system with self-diagnosis and self-repair capabilities, enhancing the long-term availability and adaptability of the system in complex and ever-changing production environments.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A paper thickness online prediction system, characterized in that, include: The data acquisition module is configured to collect process parameters of the pulping and papermaking processes in real time. The pulping process includes pulping, refining, and storage, and the papermaking process includes the wire section, press section, drying section, and post-processing section. The data storage module is configured to store the process parameters and historical inspection data from the inspection center, and to build a distributed data warehouse; The AI ​​prediction module is constructed by fusing the Long Short-Term Memory Network (LSTM) with at least one alternative algorithm, such as Random Forest or XGBoost, and is used to predict the paper thickness in real time based on the process parameters. The output module is configured to output the thickness prediction results and the corresponding process parameter adjustment scheme, and to conduct bidirectional data interaction with the Manufacturing Execution System (MES) through the API interface.

2. The online paper thickness prediction system according to claim 1, characterized in that, The data acquisition module includes: Multiple online fiber analyzers are configured to monitor fiber length distribution and thickness in real time; Multiple thickness pre-detectors are configured to collect the thickness index of the base paper in real time; The online fiber analyzer and thickness pre-detector have completed interface development and compatibility testing with the existing quality inspection system.

3. The online paper thickness prediction system according to claim 1, characterized in that, The process parameters include several core features such as beating degree, press roller linear pressure, fiber wet weight, grinding disc gap, machine speed, and raw material ratio. These core features are selected from dozens of parameters through correlation analysis and principal component analysis.

4. The online paper thickness prediction system according to claim 1, characterized in that, The data storage module is also configured to: Establish data collection standards, unify data formats, ensure timestamps are accurate to the second, and standardize units; Establish rules for identifying abnormal data, and determine data that exceeds the normal range by several times the standard deviation as abnormal; The data storage period is several years, and it supports the access and management of thousands to tens of thousands of real-time data entries per day.

5. The online paper thickness prediction system according to claim 1, characterized in that, The AI ​​prediction module also includes: The feature engineering unit is configured to select several core features from dozens of parameters through correlation analysis and principal component analysis; Attention mechanism unit, configured to enhance the weighting of key parameters; The model inference engine is configured to process input data and output prediction results in real time, with a prediction latency of less than or equal to 200 milliseconds.

6. The online paper thickness prediction system according to claim 1, characterized in that, The output module also includes: The parameter adjustment rule library, built based on the experience of process experts, is configured to convert the deviation between the predicted thickness and the target value into specific operation instructions, including the adjustment range of the pulping pressure and the optimization value of the wrinkling rate. The human-computer interaction interface is configured to display parameter curves, predicted thickness, pass rate trends and adjustment suggestions in real time, and supports historical data backtracking and model effect analysis.

7. The online paper thickness prediction system according to claim 1, characterized in that, The AI ​​prediction module is trained in real time using a sliding window method, with a window size ranging from 3 to 30 days. The model accuracy is tested regularly, and specific optimizations are performed for scenarios where the deviation exceeds a set threshold.

8. A method for online prediction of paper thickness, characterized in that, Includes the following steps: Multiple process parameters of the pulping and papermaking processes are collected in real time; The process parameters are standardized by unifying timestamps and units, and abnormal data is identified and removed. The standardized data is input into the AI ​​prediction model, which is constructed based on the fusion of LSTM with at least one alternative algorithm, such as Random Forest or XGBoost. Output the paper thickness prediction results and corresponding process parameter adjustment schemes, and push them to the Manufacturing Execution System (MES) via API interface; Based on the deviation between the predicted thickness and the target thickness, the corresponding process adjustment instructions are matched from the rule base and pushed to the Manufacturing Execution System (MES) to ensure that the thickness deviation is controlled within the set range.

9. The method for online prediction of paper thickness according to claim 8, characterized in that, The training process of the AI ​​prediction model includes: The first stage involves basic training based on historical data and real-time collected data to optimize model parameters. The second stage involves iterative training based on real-time data, updating the training set using a sliding window method with a window size ranging from 3 to 30 days. Regularly conduct multi-scenario verifications, covering different raw material ratios and vehicle speed conditions, to ensure that the model fit is no less than 90% and the prediction error is controlled within the set range.

10. The method for online prediction of paper thickness according to claim 8, characterized in that, The method further includes: When the model prediction bias exceeds a set threshold, a special optimization mechanism is triggered to retrain or adjust the model parameters. The optimized model is deployed to the production environment to enable online model updates and closed-loop control.