Intelligent thickness control system for calendering high-brightness edge sealing sheet

Abstract

The invention discloses an intelligent thickness control system for calendering a high-brightness edge sealing sheet, and relates to the technical field of intelligent manufacturing and industrial process control. The method sequentially comprises six steps of multi-modal online acquisition, multi-source fusion drift compensation, digital twinning real-time synchronization, model prediction closed-loop control, high-speed execution fault self-diagnosis and reinforcement learning adaptive optimization. Thickness data are synchronously obtained through laser, terahertz, ultrasound and environment quantity, trusted thickness flow is output through a drift model and a confidence matrix, future thickness is predicted through digital twinning rolling to give uncertainty, first-step control quantity is generated through quadratic programming, millisecond-level execution is conducted through a distributed clock, and self-healing is conducted under shadow driving hot standby and cloud diagnosis. And reinforcement learning agent online iteration improves the energy-saving and stable performance, and full-life-cycle closed-loop control with accurate thickness, low energy consumption and long service life of equipment under complex working conditions is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial process control, in particular to an intelligent thickness control system for calendering high-brightness edge sealing sheet. BACKGROUND

[0002] High-brightness polyvinyl chloride edge sealing sheet produced by high-speed calendering is usually continuously formed by a four-roller or five-roller vertical or inclined large-scale calender at a linear speed of more than 100 meters per minute, and the workshop environment is kept at a roller temperature of more than 200 degrees Celsius and high humidity for a long time to ensure the melt flowability and surface mirrorization. The existing production line is usually equipped with a laser displacement thickness gauge or a beta / gamma ray gauge at the exit of the roll gap, and the single-path thickness data is sent to the upper computer through an industrial Ethernet, and then the average thickness is maintained through open-loop roll gap or roll bending adjustment. For layer thickness distribution or local depression detection of the edge sealing strip, some manufacturers have begun to introduce terahertz or ultrasonic single-point probes for spot-checking scanning, but multi-source synchronous online monitoring has not yet been achieved.

[0003] At the control level, traditional PID and piecewise model predictive control have been used for thickness and flatness adjustment of plastic, rubber and electrode sheet calendering, which can maintain millimeter-level error under stable working conditions. At the same time, the steel strip and aluminum strip industry is exploring the use of reinforcement learning for rolling flatness and thickness control to improve the adaptability to large disturbances. In the field of equipment maintenance, foreign calenders have deployed vibration and temperature rise diagnosis packages to predict the failure window of key bearings and servo screws using residual life algorithms. Overall, existing technologies rely on single physical quantity thickness measurement at the sensing end, are based on fixed mechanism models at the control end, and emphasize offline diagnosis at the operation and maintenance end, and there is still a lack of integrated chain from multi-modal sensing to real-time self-learning adjustment and health closed loop.

[0004] The most prominent technical problem at this stage is: Millisecond-level stable control of the full-width thickness of mirror-level edge sealing sheet cannot be achieved under the high-temperature, high-humidity and high-speed calendering scene. When the sheet surface glossiness increases, the laser thickness gauge is prone to signal clipping due to spot saturation, the terahertz pulse is attenuated due to water vapor absorption, and the ultrasonic sound path drifts with temperature-humidity coupling, so single-channel thickness measurement will appear instantaneous off-target when the working condition changes suddenly; at the same time, the actual opening of the roll gap deviates from the set value due to the thermal expansion of the roller and the fluctuation of the traction tension, and the model predictive control with fixed parameters cannot compensate in time; if the thickness deviation accumulates to the upper limit of the tolerance within a few seconds, it will directly cause the edge sealing strip to fail to fit, the edge to crack or a large number of scrap, causing raw material waste and delivery date violation risks for furniture board manufacturers.

[0005] Due to the lack of multi-source thickness fusion, dynamic confidence evaluation and online self-learning control chain in existing systems, it is difficult to maintain continuous, high-precision and low-energy consumption of thickness closed loop under mirror reflection saturation, high humidity sound speed drift, power grid pressure drop or raw material viscosity jump, and an overall solution is urgently needed to realize real-time accurate thickness adjustment of high-brightness edge sealing sheet under extreme working conditions and long-period safe operation of equipment.

[0006] To this end, the present application provides an intelligent thickness control system for calendering high-brightness edge sealing sheet. SUMMARY

[0007] (I) Technical problems solved In view of the deficiencies of the prior art, the present application provides an intelligent thickness control system for calendering high-brightness edge sealing sheet, which synchronously acquires thickness data by laser, terahertz, ultrasound and environmental quantity, outputs reliable thickness flow by drift model and confidence matrix, gives uncertainty by digital twin rolling prediction of future thickness, generates first-step control quantity by quadratic programming, executes by distributed clock at millisecond level, self-heals under shadow drive hot standby and cloud diagnosis, online iteration improves energy saving and stability performance by reinforcement learning agent, realizes accurate thickness, low energy consumption and long service life of equipment in complex working conditions, and realizes whole life cycle closed loop control; The technical problems described in the background art are solved.

[0008] (II) Technical solutions To achieve the above purpose, the present application is implemented by the following technical solutions: The intelligent thickness control system for calendering high-brightness edge sealing sheet comprises, A multi-modal acquisition module synchronously acquires the network by multiple physical principles, combines mirror reflection suppression and dynamic range splicing, performs same frequency and same phase acquisition of laser, terahertz and ultrasound data, and synchronously records environmental disturbance signals; A drift compensation fusion module models the drift mechanism driven by environmental disturbance, extracts drift vectors and noise distribution of each channel, adaptively adjusts the gain matrix by Bayesian-Kalman fusion algorithm, and outputs a fusion thickness vector with confidence; A digital twin synchronization module builds a digital twin model containing roll system elasticity, material rheology and thermal field coupling, synchronizes by Kalman-particle hybridization algorithm and measured data, and realizes millisecond-level prediction of sheet thickness evolution; A closed loop execution module generates control instructions that meet the speed, energy consumption and safety constraints of the production line based on the rolling prediction of digital twin, and sends them to the actuator through high-speed scheduling; A health assessment module establishes a multi-domain health matrix, performs real-time health assessment on acquisition, fusion and control link, adopts exponential entropy weight scoring and Markov residual life prediction, triggers shadow drive hot standby and self-healing strategy; A policy fusion module fuses the reinforcement learning policy with the model predictive control, optimizes the energy consumption and control accuracy under the premise of meeting the hard constraints through online fine-tuning and safety monitoring, and executes adaptive intelligent decision-making.

[0009] Further, the multi-source synchronous acquisition includes adopting a multi-physical principle synchronous acquisition network, capturing thickness signals in real time in the same space profile and time granularity through the upper and lower biaxial laser displacement sensors, terahertz pulse probes and air-coupled ultrasonic probes, and improving signal quality through mirror reflection suppression and control and double-redundancy calibration mechanism.

[0010] Further, the multi-source synchronous acquisition further includes using atomic clocks and optical fiber synchronization technology to ensure millisecond-level phase synchronization of different sensing channels, realizing time reference unification through master-slave clock structure and least squares bias correction, and realizing a priori alignment of thickness data and coordinate system through full-amplitude air running self-checking and mirror reflection target calibration.

[0011] Further, the drift compensation fusion includes using the internal correlation between environmental disturbance and historical thickness sequence, constructing a drift prediction function through linear-nonlinear mixed kernel regression, extracting slow drift vectors and instantaneous noise distribution of each channel in real time, and adaptively adjusting the gain matrix of the Bayesian-Kalman fusion through the confidence matrix.

[0012] Further, the drift compensation fusion further includes triggering an abnormal replacement mechanism through Kalman residual and Mahalanobis distance threshold value judgment when a single channel is abnormal, seamlessly replacing the abnormal channel with the predicted value of the digital twin model, and maintaining the continuity and stability of the fused thickness vector.

[0013] Further, the digital twin synchronization includes constructing a digital twin body containing three-layer mechanism models of roll system elasticity-geometry, material rheology-stress and thermal field-solidification, performing state assimilation with measured data through Kalman-particle hybrid assimilation algorithm, and updating model boundary conditions in real time by ingesting process parameter flow.

[0014] Further, the digital twin synchronization further includes using a four-order explicit Runge-Kutta-Cash-Karp variable step strategy for rolling prediction, outputting a thickness prediction curve and an uncertainty band, and encapsulating and forwarding the prediction data to the model predictive controller through the message bus.

[0015] Further, the closed-loop execution control includes using the rolling prediction sequence of the digital twin, adopting a rolling quadratic programming model predictive control algorithm with confidence constraints, solving the control amount sequence that minimizes the weighted square sum of thickness deviation, and issuing instructions to the actuator through a high-speed scheduling mechanism.

[0016] Further, the closed-loop execution control further comprises translating the control instructions into servo pulses, roll temperature valve pulse width and traction drive reference frequency within a millisecond network cycle, and ensuring the accuracy and continuity of the execution response through real-time feedback nesting and redundant switching mechanism.

[0017] Further, the health assessment and self-recovery comprises millisecond-level health assessment on the acquisition-fusion-control full link, construction of an electric-mechanical-environmental three-domain health matrix, and real-time discrimination of device and algorithm performance degradation using exponential entropy weight scoring and Markov residual life prediction.

[0018] Further, the health assessment and self-recovery further comprises triggering shadow drive hot standby, cloud operation and maintenance and self-recovery strategy when the health degree is lower than the threshold, and realizing long-period high availability and low downtime risk of the system through software and hardware mixed redundancy preempt.

[0019] Further, the strategy fusion optimization comprises designing a reinforcement learning agent architecture compatible with the edge sheet calendering scene, pre-training a high-performance initial strategy through digital twin offline simulation for deep reinforcement learning, and providing a safe and controllable starting point for online fine-tuning.

[0020] Further, the strategy fusion optimization further comprises iterative reinforcement learning strategy in real production in a controlled manner. Through adjustable weight gating and model predictive control output fusion, and using safety monitoring and versioned verification to realize robust adaptive intelligent decision-making.

[0021] (Three) beneficial effects The application provides an intelligent thickness control system for calendering high-brightness edge sealing sheet, which has the following beneficial effects: The thickness signals of laser, terahertz and ultrasonic three physical principles are cross-mapped with temperature, humidity, vibration and other environmental quantities in the acquisition layer, and clock unification, range splicing and mirror surface suppression are used to make the basic data have the properties of redundancy, traceability and comparability, avoid misjudgment caused by single source distortion, and realize less noise, less drift and less blind area in the perception chain from the source.

[0022] Through drift mechanism modeling, confidence adaptive weighting and Kalman Bayesian fusion, the multi-source data is compressed into a high-confidence thickness stream with confidence within a millisecond scale, and subsequent digital twin only needs a small amount of correction to synchronize the physical production line, greatly reducing the model throughput pressure, forming a positive closed loop of high-quality input driving high-fidelity model.

[0023] The three-layer mechanism digital twin and Kalman particle assimilation real-time feedback each other, the rolling prediction output thickness trend and uncertainty band, and then the dynamic weight and confidence constraint quadratic programming generates the optimal control quantity, so that the planning layer, the execution layer and the health layer make decisions in the same time domain, significantly compressing the adjustment overshoot and steady-state error.

[0024] Through distributed synchronous clocks, dual-buffer incremental interpolation, and shadow drive hot standby, the roll gap, roll temperature, and pull-off speed actions are synchronized at the physical layer, with failover within seconds. Combined with multi-domain health vectors, remaining life prediction, and remote diagnostics, equipment maintenance, quality control, and energy conservation and consumption reduction are unified within the same control loop, resulting in safer control. Information and energy flows create positive feedback at every level, ultimately ensuring that edge-banding sheets maintain thickness accuracy, energy savings, and extended equipment life even under high-temperature, high-humidity, and high-speed operating conditions.

[0025] After the reinforcement learning agent introduces multi-dimensional states, compound rewards and safety projections, it is integrated with model predictive control according to health gate weights, inheriting the robustness of optimal determinism while gaining the evolutionary nature of self-learning; online fine-tuning, experience-first replay and grayscale verification mechanisms ensure that innovative strategies are continuously optimized without compromising production safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The figure is a schematic structural diagram of the intelligent thickness control system for calendering high-gloss edge-sealed sheets according to the present invention. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0028] See also Figure 1 The present invention provides an intelligent thickness control system for calendering high-gloss edge-sealed sheets, comprising: On the continuous production line of high-speed calendered high-gloss edge-banding sheets, the sheet surface has a mirror-like gloss, and the production workshop is often in a high-temperature and high-humidity environment. In this scenario, a single sensor is very likely to experience problems such as reflection saturation, temperature drift, and lack of real-time performance, resulting in drastic fluctuations in thickness data and an inability to provide high-reliability support for downstream control systems.

[0029] If real-time and reliable thickness information cannot be accurately and stably captured at the source, all subsequent model synchronization, predictive control and execution compensation will be accumulated and amplified due to input errors, ultimately manifesting in negative consequences such as uneven sheet thickness, edge sealing failure, material waste and production downtime.

[0030] Step one, realize the fast, synchronous and high-density acquisition of multi-modal thickness and environmental data on the production line, provide high-resolution and disturbance-labeled original information basis for subsequent modeling, fusion and control, and ensure the same phase of different sensing channels within milliseconds through atomic clock and optical fiber synchronization to avoid data misplacement caused by time drift, while using mirror reflection suppression and double-redundancy calibration mechanism to improve signal quality in harsh working conditions.

[0031] Step 101, construction of multi-physical principle synchronous acquisition network In the same space profile and the same time granularity, real-time capture of laser thickness signal, terahertz thickness signal, ultrasonic thickness signal and environmental disturbance signal provides cross-source, same-frequency and same-phase data base for thickness fusion-compensation; To suppress the mirror saturation phenomenon, the laser displacement sensor uses double-axis transmission up and down and installs a polarization filter cover. When the normal reflectivity of the sheet surface increases and the receiver current approaches the saturation threshold , the transmission power is adjusted adaptively according to the following formula:

[0032] Where: the transmission power : the current output power of the laser; is the transmission power issued by the laser in the previous control period, serving as a reference quantity; power adjustment coefficient : a dimensionless proportional coefficient between zero and one, used to limit the rapid decrease of power; receiving current : real-time photodiode output; saturation threshold : the upper limit current obtained according to the sensor range calibration.

[0033] Through the linear proportional pressure reduction mechanism, the transmission power is quickly adjusted downward when the saturation trend is detected, thereby avoiding signal clipping distortion. Therefore, the laser thickness original signal still maintains linear response in the mirror scene, realizing the true thickness of the entire survey area.

[0034] Based on the complementarity of different sensing principles to materials and working conditions, the terahertz pulse probe and air-coupled ultrasonic probe are set up at intervals on the same measurement cross section. In order to ensure the alignment of the two channels on the time axis, double PLL clocks are used to synchronize and correct the trigger delay . There is a difference in propagation speed between ultrasonic ranging and terahertz time-of-flight in physical mechanism. In order to make the output thickness of the two comparable, the equivalent conversion coefficient is introduced in real time:

[0035] In the formula: the conversion coefficient : Obtained by scanning the standard sample in the self-checking stage, used to eliminate the coupling effect of medium elastic modulus and strain rate difference on acoustic velocity; , the converted ultrasonic thickness sequence, consistent with the terahertz thickness in terms of dimension, reference coordinate system, and value range, used for subsequent fusion and compensation; ultrasonic converted thickness : The thickness sequence after coefficient correction, which is of the same dimension and reference coordinate system as the terahertz thickness. Through this conversion, the two signals are unified in terms of dimension, coordinate system, and value range, laying a prior consistency for subsequent Kalman fusion.

[0036] The high temperature and humidity fluctuations in the production site have a huge impact on the performance of each sensor and the physical properties of the sheet. A thermocouple array and a MEMS humidity-vibration composite are arranged around the measuring frame to output the environmental temperature , relative humidity , and roll surface vibration acceleration in real time. The above disturbance signals and thickness original signals are timestamped to form a quadruple :

[0037] In the formula: quadruple vector : A six-dimensional feature vector composed of thickness three-source and environmental three-source data, and bound to a unified time label. Through ring network clock drift compensation technology, the sampling error of each dimension is less than microseconds, so that it can be regarded as the same time slice in multi-source fusion. Establishes a real-time mapping between thickness and environmental disturbance, providing a horizontal correlation basis for subsequent drift compensation, and avoids misjudging the working condition fluctuations as thickness abnormalities.

[0038] The thickness fluctuation amplitude of the high-speed production line may exceed the single sensor range. To avoid clipping distortion, both the laser and terahertz sensors are equipped with dual ADC links: one with high gain and small range for capturing small fluctuations, and one with low gain and large range for covering extreme overage. A range splicing algorithm is used:

[0039] In the formula: spliced thickness : The thickness value used for subsequent calculation after dual-link selection; high-gain thickness : High-sensitivity ADC reading; low-gain thickness : Wide-range ADC reading; baseline thickness : The target thickness calibration value of the current batch; threshold : Range switching criterion, used to prevent frequent jitter.

[0040] Through dynamic range splicing, the system takes into account both high resolution and wide dynamic range, and can still maintain linear distortion when the thickness suddenly changes or fluctuates. The above mechanism ensures that even if the sheet thickness is temporarily out of tolerance, the system can still capture the signal completely without saturation and truncation, and reserves the data integrity for subsequent filtering and compensation.

[0041] Step 101 establishes a multi-source thickness raw data network across physical principles, ranges, and working conditions through four technical features: mirror surface inhibition of laser biaxial arrangement, terahertz-ultrasonic redundancy cooperation, synchronous side recording of environmental disturbance, and dynamic range expansion, providing data input for the next step of signal drift compensation and weighted fusion that is consistent in time and space, consistent in dimension, and error-describable.

[0042] The construction of the multi-physical principle acquisition network makes the thickness perception independent of a single optical echo: the laser path avoids high-energy beams through adaptive polarization when the mirror reflection fluctuates violently; the terahertz channel provides independent ranging capability for different layer interfaces, and the ultrasonic channel compensates for the blind area of electromagnetic wave signal attenuation in high-temperature and high-humidity extreme scenarios. The three-way complementation makes the inevitable defects of a single channel be orderly covered by the remaining channels. Dynamic range splicing ensures that the system still maintains complete data link during the thickness step caused by sudden increase of process speed or switching of raw material batches, without interruption due to saturation and truncation. Synchronous side recording of environmental disturbance not only provides a basis for subsequent algorithms in drift compensation, but also provides a continuous reference for the cooperative optimization of heat, humidity, and vibration in the production process, realizing the first connection between the perception link and the process link.

[0043] Step 102, time reference unification and prior calibration link On the basis of ensuring that the cross-source thickness data is the same in frequency and phase, further with the help of hardware clock-fiber time reference, full-width self-checking, and mirror reflection target calibration, the thickness data and the coordinate system are aligned in advance, providing error-measurable thickness raw signals for subsequent Kalman filtering-confidence adaptive fusion output.

[0044] A master-slave clock structure is adopted, with the master clock being an atomic time base , and the slave clock being an embedded crystal oscillator in the sensor . The pps signal is distributed through a ring fiber network, and each slave clock is corrected in a least squares manner :

[0045] In the formula: corrected slave clock : the correction time axis of the th sensor : the original measurement time (delay) sequence of the th frame (sampling point) : the reference (master) channel in the Frame time (delay) Deviation amount : The minimum offset to be solved, the optimal constant offset relative to the main channel; Discrete sampling points : Sampling sequence number within the calibration time window; Variable in the solving process (offset to be optimized); Sample number : Total amount of data points for deviation estimation.

[0046] The deviation is solved by least squares, ensuring that the timeline of each sensor is synchronized with the master clock with a deviation less than microseconds, thereby ensuring that each signal is in the same time domain during Kalman filtering-dynamic fusion.

[0047] Before starting each day, let the calender run empty and place the standard thickness target sheet on the measuring frame Activate the laser, terahertz, and ultrasonic three-source automatic scanning and record the measurement values, and perform one-dimensional polynomial fitting on the scanning track:

[0048] Wherein: is the polynomial model output of the channel; is the channel identifier; Scan position : Original measurement input, spatial coordinates along the width direction of the sheet; Polynomial coefficients : The order coefficient of the channel, which maps to the polynomial coefficient of the real thickness; Polynomial order : The highest order determined according to the device resolution, controlling the balance between approximation accuracy and overfitting risk.

[0049] If for any input, , there is , then it is determined that the zero point is consistent; otherwise, update the zero point compensation table according to the deviation curve .

[0050] In the formula: is the target thickness (target value); the zero point consistency threshold : The upper limit of the static error allowed by the device.

[0051] By full-width empty running self-checking, eliminate static mechanical assembly error and sensor zero point offset, so that each sensor output is aligned in spatial coordinates, providing a mechanical-optical balanced prior basis for subsequent fine filtering and fusion.

[0052] At both ends of the production line, mirror reflection targets are arranged, when the sheet head and tail pass through the measuring frame, the laser sensor will scan the surface of the target with super high reflectivity, thereby triggering the rapid linearization calibration:

[0053] In the formula: linearization coefficient : real-time adjustment of laser gain factor; Real-time laser thickness value; initial thickness : the first frame thickness of the sheet entering the measurement interval; mirror thickness : the theoretical extreme thickness corresponding to the mirror target reflection.

[0054] By using the target reflection as a dynamic slope calibration point, the gain-linearity of the laser signal can be corrected without stopping, ensuring that the laser data is in the linear working area throughout the production cycle. Mirror target calibration belongs to the online linearity maintenance method of production and calibration, which ensures that the laser thickness can remain calibrated consistently without manual calibration for a long period of time, significantly improving the system's online continuous operation capability.

[0055] In order to realize parameter consistency in subsequent steps, a unified coordinate identifier is used Encapsulate the thickness multi-source data and environmental data in the same collection period:

[0056] In the formula: frame package : complete data frame object; Identifier : unique number generated based on the main clock and space index mixing.

[0057] Through the binding of the unique identifier, the same identifier is used when referring to thickness data in all subsequent steps, eliminating the ambiguity of multiple names for the same object or multiple objects with the same name, providing strong consistency for the vertical running of the process.

[0058] Step 102 takes fiber synchronization-inlaid clock calibration, empty running self-test-zero point fitting, mirror target-dynamic linearization, and cross-source coordinate unique identification as the core, and completes the four-dimensional alignment of multi-source thickness signals in time, space, gain, and naming system. Through this series of processing, the thickness original signal is transformed into error measurable, frame-level unique basic data, providing a completely isomorphic input matrix for the Kalman fusion-drift compensation of step two.

[0059] The design of time reference unification and prior calibration link realizes the dual solidification of measuring time domain and space domain. The fiber-inlaid clock scheme avoids the common timing drift in the high electromagnetic noise environment of the production workshop, so that the multi-channel data frame has a natural consistent time label; the empty running self-check completes the static space curve fitting and thermal expansion compensation in advance, so that the thickness measurement has excluded most mechanical assembly errors before the real production; the mirror target provides a dynamic linearization path that is continuously connected to production, so that the laser gain is always in a controllable interval from start to stop. The unique identification of cross-source coordinates establishes a unambiguous primary key for all thickness-environment-time multidimensional data, so that the subsequent data fusion, twin modeling and predictive control do not misplace when calling data objects.

[0060] As the first step of the intelligent thickness control method, multi-modal online acquisition realizes the reliable landing of thickness data in high-brightness light, high-temperature and high-humidity, and high-speed calendering scenes through steps 101 and 102. The former focuses on the sensing layer and simultaneously captures thickness from laser, terahertz and ultrasonic multi-channels and side records environmental disturbances, ensuring the spatio-temporal co-frame of multi-source information; the latter focuses on clock alignment, zero calibration and gain linearization, so that each signal is completely consistent in sampling dimension and coordinate dimension.

[0061] High-brightness light edge sealing sheet production site continuously outputs hundreds of frames The multi-source thickness raw data stream contains not only the laser thickness signal attenuated by the dynamic reflection of the mirror surface, but also the terahertz pulse penetration echo, air-coupled ultrasonic time delay conversion value, and six-dimensional environmental disturbance vector. Although step one has ensured the co-frame consistency of all signals in the time axis and the space axis, the long-term superposition of sensor sensitive element aging, thermal drift, wet expansion nonlinearity and vibration coupled noise will continuously erode the thickness measurement reliability. If structured and traceable multi-source fusion-drift compensation is not implemented immediately after acquisition, the thickness data will be amplified due to asynchronous noise and drift terms, resulting in input distortion of the digital twin model, misjudgment of the model predictive controller, and even reverse regulation.

[0062] Step two, through drift mechanism modeling and Bayesian-Kalman fusion algorithm, the multi-source thickness measurement stream is compressed into a single high-credibility state vector with confidence. This link estimates the drift trend and measurement noise in real time, adaptively adjusts the gain matrix, and still outputs stable and quantifiable uncertainty limits when a single channel fails or noise bursts, providing reliable prediction input for digital twins and back-end control.

[0063] Step 201, thickness drift mechanism modeling and confidence adaptive measurement The inherent correlation between environmental disturbance and historical thickness sequence is utilized to extract the slow drift vector and instantaneous noise distribution of each channel, and the confidence matrix is calculated in real time to provide quantitative weights for subsequent fusion.

[0064] The laser thickness signal is particularly sensitive to thermal drift, while the terahertz signal is more susceptible to humidity absorption attenuation; the ultrasonic channel produces sound path deviation due to the temperature-viscosity coupling of the material. First, the rolling window inner convergence 、 、 is used to aggregate the three-source thickness signal, and a linear-nonlinear mixed kernel regression is used to construct the drift prediction function , each thickness signal is expressed as the sum of instantaneous true thickness and drift term , so by minimizing:

[0065] In the formula: the drift vector is fitted and updated in real time.

[0066] Real-time thickness of the first machine stand; environmental temperature , relative humidity , roll surface vibration acceleration ; Drift vector : describes the slow error of the first sensor under environmental disturbance; Regression function : a nonlinear kernel model that maps three-dimensional environmental disturbance to thickness drift space; Window size : matches the time of one lateral width of the sheet passing through the sampling stand.

[0067] Through drift estimation driven by environmental variables, the error trend of each channel with changing working conditions is obtained, providing a quantitative basis for subsequent weight reduction.

[0068] The instantaneous residual obtained after removing the drift term contains high-frequency noise and random jumps, and to characterize its statistical distribution, the residual covariance matrix is estimated on the window . For this matrix, a Shannon entropy threshold is introduced to judge the information content abundance, so as to dynamically adjust the upper limit of noise variance to adapt to the change of sampling density caused by the change of production rhythm.

[0069] The real-time update of the residual noise covariance matrix makes the Kalman filter have a noise prior that is synchronized with the current working condition, so that the filter gain is not weakened by outdated statistics.

[0070] Jointly project the drift vector and noise covariance into the confidence space to generate a diagonal confidence matrix

[0071] wherein: confidence matrix : synthetically depicts the confidence level of each channel at the current time, and the confidence is automatically reduced when the drift is larger or the noise is stronger. , i.e. , , , the dynamic compensation amount / deviation amount of the first channel; is an environmental disturbance / measurement noise covariance matrix, representing the statistical characteristics of three-dimensional disturbances (or three-sensor errors); is its precision matrix, used for whitening or dimensionless.

[0072] confidence matrix provides adaptive weights for subsequent Bayesian-Kalman fusion, and integrates slow drift and instantaneous noise into a unified evaluation framework at one time. Through environmental coupling drift factor extraction, instantaneous noise covariance estimation and confidence matrix normalization, the error source identification-quantification-weight mapping closed loop is completed, providing an updateable and iterative prior information flow for multi-source thickness fusion in step 202.

[0073] By closing the three links of environmental coupling drift, instantaneous noise statistics and confidence quantification in the same rolling window, the system forms a layered capture ability for slow error and fast noise. The drift model introduces the cross check of time derivative and prediction residual, so that the system not only corrects the error, but also gives an early warning in the error acceleration stage; noise covariance adaptive update gives the filter flexibility to follow the sudden changes in working conditions; the confidence matrix maps the drift and noise to a unified evaluation scale, so that the weight distribution is theoretically interpretable.

[0074] Step 202, Bayesian-Kalman dynamic fusion and abnormal replacement With the help of Bayesian-Kalman gain online update driven by confidence matrix, three thickness signals are fused into a high-confidence thickness vector, and model prediction value is seamlessly enabled to replace in single-channel abnormality, ensuring the stability and continuity of thickness data flow.

[0075] At the starting time of the rolling window, the state of the digital twin at the last time is called and the confidence matrix , the thickness prior distribution is generated by the Bayes formula ​​. This distribution is input to the prediction step of the Kalman filter, giving the filter a shape-accurate prior before it observes new frame data. The prior distribution is modulated by a confidence matrix, with higher-confidence channels having more weight in the prior, allowing the prediction to also discriminate between good and bad sources.

[0076] After obtaining the observation vector (measured (observed) thickness vector, corresponding to the real-time fused output before laser, terahertz, and ultrasonic channels), the Kalman gain calculation uses an improved formula:

[0077] Kalman gain : a matrix that balances the prediction covariance and the observation noise covariance , a fusion weight matrix, with a proportional coefficient matrix within the value range interval; prediction covariance : generated from the prior distribution, indicating the strength of the prediction error, a positive definite matrix, with a positive real number domain; observation noise covariance : directly uses the real-time update of step 201 , indicating the observation noise strength, a positive definite matrix; By introducing a confidence matrix into the gain formula, the system dynamically suppresses channels with large drift or strong noise, ensuring that the fusion result is more dependent on high-confidence sources.

[0078] Update the thickness estimate using the Kalman gain:

[0079] In the formula: fused thickness : the final output thickness vector; , prior predicted thickness (derived from the previous time state through a physical / data model to the current time); Kalman residual : the difference between the observation and the fusion; If the Kalman residual experiences a three times squared Mahalanobis distance jump in any channel, mark the channel as abnormal and remove the corresponding row and column from the confidence matrix and the observation noise covariance , and fill in the gap with the predicted , until the abnormal signal returns to within the drift-noise threshold, and the abnormal replacement strategy ensures that any single-source failure does not interrupt the thickness data stream, and the fused thickness remains smooth and continuous. The Mahalanobis distance threshold is three times the square value, which is the abnormality decision criterion.

[0080] Output thickness vector frame packaging and database landing: Final fused thickness and updated balanced prediction covariance encapsulated into corresponding data packets and written into a dedicated time-series database along with residual statistics. This database provides an API interface for the digital twin step to call, enabling cross-step parameter passing.

[0081] The thickness fusion result and quality metrics are persisted synchronously, providing direct key-value queries for the twin synchronization algorithm in step three, and improving the overall throughput of the system.

[0082] The Bayesian-Kalman dynamic fusion completes the mapping of multi-physical-source thickness signals to a high-confidence thickness vector with the help of a confidence matrix; the abnormal replacement mechanism ensures that any single-source failure can be completed by the remaining channels and prediction, and the fused thickness meets the digital twin model's requirements in terms of time series and quality, realizing real-time closed-loop between the perception layer and the model layer.

[0083] Step two successfully realizes the real-time conversion from multi-source thickness raw signals to a high-confidence thickness vector through a two-level structure of drift mechanism modeling + confidence quantization and Bayesian-Kalman dynamic fusion + abnormal replacement.

[0084] The confidence matrix and residual covariance output by step 201 provide accurate priors for the Kalman gain iteration in step 202, enabling the fusion process to adaptively allocate weights in fluctuating conditions; the thickness vector encapsulated by step 202 not only meets the dual requirements of real-time and continuity, but also synchronously writes the covariance and residual into the database, providing complete state correction basis for the digital twin.

[0085] The cleverness of Bayesian-Kalman dynamic fusion lies in injecting real-time confidence into the Kalman gain, while the abnormal replacement strategy ensures that any single-source mutation will not break the data chain. In this framework, the prediction covariance, observation noise covariance, and confidence matrix form a prediction-observation-quality three-element closed loop, and the system adjusts the trust level in real time according to the quality of the signal, avoiding good signals from being drowned or bad signals from being amplified. Multi-frame confirmation of abnormal processing not only highlights the tolerance of instantaneous spikes in high-speed scenarios, but also takes into account the stability and safety of fault persistence. The positive definiteness check of covariance before disk writing eliminates the risk of mathematical anomalies to the twin model, controlling data integrity to the secondary verification level.

[0086] In the production process of calendered high-gloss edge-lit sheeting, the digital twin model not only needs to be similar to the real production line, but also must be synchronized with the production line on the time scale within milliseconds, otherwise the future thickness trajectory obtained by the model predictive controller will be out of phase with the physical process, directly leading to inaccurate timing of roll gap adjustment. The core task of step three is to make the virtual model continuously fit the physical world under dynamic environment and equipment aging conditions: on the one hand, it uses the fused thickness vector and covariance The model state is assimilated; on the other hand, the process parameter stream is taken in real time, including the roll temperature sequence, the tension sequence, the melt viscosity curve, etc., which is used as the driving boundary condition to update the model kernel, forming a virtual-real interactive closed loop.

[0087] Step three, build a digital twin that is real-time and same frequency with the production line, through the three-layer mechanism model of roll system elasticity-geometry, material rheology-stress, thermal field-curing, and Kalman-particle hybrid assimilation with measured data, to realize millisecond-level prediction of plate thickness evolution and equipment state. The model boundary condition is refreshed with new data, so that the virtual production line still maintains physical consistency under extreme working conditions, providing dynamic constraints and priori for MPC optimization.

[0088] Step 301, physical-data assimilation framework construction A multi-layer digital twin model is constructed, which contains the roll-material coupling mechanism, the thermal-flow-stress multi-field mutual feedback, and the real-time data interface, and the state vector initialization and parameter implantation are realized through the Kalman-particle hybrid assimilation algorithm.

[0089] The digital twin model adopts a three-layer nested structure: the outermost layer is the roll system elasticity-geometry sub-model, which describes the force-deformation-gap evolution of the upper and lower rolls; the middle layer is the material rheology-stress sub-model, which describes the viscoelastic-viscosity coupling flow of the melt in the roll gap; the innermost layer is the thermal field-curing sub-model, which simulates the process of polymer from thermal state to solid state. The three layers share the core variables including roll gap displacement, melt shear rate, polymer specific heat, etc., so that the thermal-flow-force information is transmitted synchronously in time domain. This structure ensures the scalability while maintaining the details, and if new cooling air curtain or roll friction is added in the future, it can be mounted as an attached physical block without destroying the overall coupling framework.

[0090] The three-layer physical variables are compressed into the state vector , and then mapped to the measurement domain through the observation projection matrix to obtain the thickness prediction:

[0091] The observation projection matrix : The matrix that linearly maps the multi-field variables to the thickness space, the dimension is determined by the number of mechanism layer variables and the dimension of thickness observation. In the formula: the state vector : contains roll gap displacement, polymer temperature field average, melt density, etc., the value changes continuously with time; the predicted thickness : the thickness estimate output by the digital twin model, the dimension is consistent with the fusion thickness vector .

[0092] In order to take into account the uncertainty of linear high-dimensional state and nonlinear parameters, the system adopts the Kalman-particle hybrid assimilation algorithm: Use unscented Kalman filter update, particle filter weight resampling for material parameters that are difficult to linearize (such as yield modulus and temperature sensitivity), and state assimilation update formula:

[0093] Where: predicted state : The prior state obtained after the model advances in time, which is obtained from the state equation / physical model Deduced to Typical predictions include: thickness field state, material / equipment parameter drift, environmental compensation coefficient, etc. Calibration status : Assimilated state, the final state obtained after fusing the current observations, used for output and next step prediction; Unscented Kalman Gain : Update weights calculated based on the state-observation covariance; Particle collection: Indicates the Particle material parameter vector, updated by residual weight resampling.

[0094] is the observed thickness vector of the three channels (laser, terahertz, and ultrasound) after weight / robust processing, or the measured value output by the Kalman observation model in the previous section; is the predicted measurement (prior measurement) obtained from the state prediction; This allows the linear and nonlinear variables to be divided and conquered, achieving rapid convergence of high-dimensional states without sacrificing nonlinear parameter adaptability.

[0095] Roller temperature sequence , traction tension sequence , melt viscosity curve Push to the twin model in real time to update the boundary vector . Refresh the boundary vector immediately after each assimilation , and the correction status With boundary vector Together, they serve as input for the next prediction step, forming a bidirectional closed loop between external drive and internal state. Process parameter embedding maps real-world operational actions to the model, ensuring that predictions align with actual operating conditions and preventing spurious responses due to lagging boundary conditions.

[0096] Through the three-layer mechanism nesting, state projection matrix, Kalman-particle hybrid assimilation and real-time refresh of boundary conditions, the digital twin model obtains the dynamic heartbeat of the same frequency resonance with the physical production line, laying a solid data-model integrated foundation for the next step of prediction-uncertainty quantification-output broadcast.

[0097] Through the deep coupling of three-layer mechanism sub-models and data assimilation, the digital twin model breaks away from the traditional offline modeling and online correction mode, and evolves online in real time. The Kalman-particle hybrid strategy divides linear high-dimensional states and nonlinear material parameters, allocates computing resources according to weights, and thus realizes accurate tracking of complex coupled processes within the hard real-time computing budget. At the same time, the embedding of process parameter flow enables the model to respond immediately to roll temperature, tension and raw material batch disturbances, and no longer appears the situation of gradually moving away from the physical production line.

[0098] Step 302, rolling prediction-uncertainty quantification-data broadcast Use the assimilated state to perform multi-step prediction in the rolling time domain, output the thickness prediction curve and uncertainty band, and encapsulate and forward it to the model predictive controller.

[0099] The model time advancement adopts the fourth-order explicit Runge-Kutta-Cash-Karp variable step strategy, dynamically adjusts the step size according to the covariance estimation at the previous time, shrinks the step size when the state disturbance accelerates, to prevent numerical explosion; relaxes the step size in the stable stage, improves the calculation throughput, and keeps pace with the production rhythm.

[0100] For each step of the prediction step Output the thickness prediction And extrapolate the covariance Covariance extrapolation uses Lyapunov equation to ensure positive definiteness; if the extrapolated value shows a divergence trend, automatically trigger step 301 to re-assimilate, forming a self-converging closed loop.

[0101] Map the thickness prediction And the extrapolated covariance To the upper and lower confidence bounds in the credibility channel space If the confidence bounds touch the edge of the thickness tolerance band, an early warning flag is attached before data broadcast, so that the model predictive controller can tighten the optimization constraints in advance according to the risk.

[0102] Finally, the predicted thickness sequence, uncertainty band and warning flag are packaged into an uncertainty matrix and pushed to the real-time message bus. The interface description follows the same data schema as step two, ensuring that the controller can be invoked without secondary parsing at the subscription end, completing the parameter cross-mapping. Rolling prediction-uncertainty quantification-data broadcasting converts the assimilated digital twin model results into a thickness-confidence-warning triplet that can directly drive the model predictive controller, enabling seamless transmission of the virtual model to the decision logic.

[0103] Through the physical-data assimilation framework of step 301, the state of the digital twin model is successfully aligned with the real-time conditions, and then in step 302, the model cognition is converted into decision-making information through rolling prediction and uncertainty quantification. The former is responsible for alignment, and the latter is responsible for foresight, both of which provide the model predictive closed-loop controller of step four with a high-fidelity, high-real-time, risk-labeled thickness prediction sequence. The injection of digital twin model self-learning, self-converging, and self-warning capabilities marks the transition of intelligent thickness control systems from passive correction to active prevention.

[0104] The rolling prediction module not only provides future thickness values, but also synchronously provides covariance bands and warning signs, visualizing and quantifying uncertain futures for the controller to choose a robust strategy; variable step integration and covariance positive definiteness guardianship ensure numerical stability of the model over a long period of time, avoiding distortion when encountering high gradient areas with traditional fixed step methods. The dual-theme message broadcasting framework enables the control system to obtain current thickness and future trends simultaneously through the same interface, reducing system integration complexity.

[0105] High-gloss edge sealing sheets have extremely narrow tolerance for thickness tolerance, and any millisecond-level control lag can push the sheet to the edge of the tolerance. Step three has output a rolling thickness prediction sequence with uncertainty, warning signs, and frame-level consistent process parameters, providing the decision layer with a future-oriented process preview. The mission of step four is to integrate this prediction information with device physical limits, energy consumption targets, and product quality weights into an integrated optimization framework to calculate roll gap displacement, roll temperature adjustment, and traction speed three types of execution instructions in real time; then, with the help of high-speed scheduling mechanism, the instructions are accurately broadcast to each execution mechanism within milliseconds, realizing thickness feedforward-feedback coupled adjustment. The entire closed loop needs to remain robust in an environment where the production line runs at high speed, uncertainty changes dynamically, and hardware constraints coexist, ensuring that control output does not produce dramatic overshoot due to model errors or delays, and leaving room for energy consumption and mechanical life in multi-objective trade-off.

[0106] Step four, after obtaining predictable thickness and uncertainty information, uses a rolling quadratic programming model predictive control (MPC) algorithm with confidence constraints to generate optimal control instructions that meet line speed, energy consumption, and safety constraints in real time, and through instruction quantization-pulse mapping and IEEE-1588 nanosecond-level network synchronization, the instructions are delivered to the driving execution mechanism within milliseconds.

[0107] Step 401, rolling horizon optimal control quantity solving Using the digital twin rolling prediction sequence and uncertainty information, the control quantity sequence that minimizes the weighted square sum of thickness deviation is solved under the given physical and process constraints, and the first-step control vector of the current period is output.

[0108] The future provided by digital twin Step thickness prediction sequence And target thickness Form a deviation vector. The system defines a rolling cost function:

[0109] In the formula: objective function : Total cost in the rolling window; The thickness prediction value of the future step ; Weight coefficient : Thickness accuracy weight, value range ; Control input column vector : Contains roll gap displacement increment , roll temperature adjustment , traction speed increment ; Input weight matrix : Diagonal positive definite matrix, parameters are calibrated by energy consumption and mechanical wear indicators; Weight coefficient : Energy consumption weight, value range , Control input column vector ; Thickness deviation square term ensures that the control strategy focuses on size accuracy; input weighting term suppresses drastic adjustment, controls energy consumption and equipment impact. Prediction covariance band Provide the future fluctuation range of thickness, in order to avoid the optimal solution falling in the high-risk area, introduce confidence constraint, let the optimization solution search near the center of the prediction band, reduce the risk of exceeding the boundary:

[0110] In the formula: thickness safety margin : Loose tolerance margin defined according to process standards; The minimum predicted thickness allowed for step ; The maximum predicted thickness allowed for step ; Confidence parameter : Value , the smaller the more conservative.

[0111] Roll gap displacement, roll temperature, and pulling speed are all subject to mechanical limits and thermal inertia, ensuring that optimization does not output instructions that the equipment cannot execute or cause material instability:

[0112] Where: lower limit of roller gap Capped with roller gap Converted from roller contact safety distance; is the current roller gap displacement, For the Step roller gap increment; Roller temperature upper and lower limits 、 Determined by the polymer thermal degradation threshold and the roller material temperature resistance limit; is the current roller temperature; For the Step roller temperature increment; Speed ​​upper and lower limits 、 By matching the traction drive capability with the material cooling cycle, is the current traction speed, For the Step speed increment, is the prediction step index; Using a sequential quadratic planner with prediction covariance linearization, The optimal control vector sequence is obtained . Only capture the first step The system sends the remaining steps to the scheduling layer, caching them as warm standby trajectories. If the solution for the next cycle is not complete, the cached trajectory can be temporarily used to ensure uninterrupted control. The solution for the optimal control variable in the rolling horizon utilizes multi-objective confidence-constrained quadratic programming, providing the lower-level scheduling with a risk-assessed first-step control vector. The dual cache design of first-step and warm standby ensures that solution delays do not lead to control windows.

[0113] With the introduction of a dynamic energy-accuracy allocator, the objective function automatically shifts its focus between fluctuating and stable operating conditions, ensuring neither sacrificing thickness quality nor forcing the equipment to operate at unconstrained high energy consumption. Furthermore, by linking the main eigenvalues ​​with a safety margin, the optimizer can proactively reduce the amount of operation when the predicted mean is safe but volatility is high, avoiding the blind spots of traditional confidence zones that overlook fluctuation trends. Dynamic physical constraints allow the safety margin to slowly adjust based on equipment health and ambient temperature, eliminating the need for manual hard-coding. The FPGA coprocessor pipeline budget significantly reduces solution time, providing algorithmic support for millisecond-level scheduling.

[0114] Step 402: High-speed instruction scheduling and real-time feedback nesting The first step control vector is translated into servo pulses, roll temperature valve pulse width and traction drive reference frequency within the millisecond network cycle, and the execution response is monitored in real time, and the feedback is embedded in the next cycle optimization.

[0115] Control vector element roll gap increment , roll temperature increment , speed increment Convert to executable pulses through device calibration curve: roll gap displacement quantization table maps roll gap increment To the number of stepper motor pulses with an acceleration ramp; roll temperature valve uses PWM modulation roll temperature increment Convert to duty cycle; traction speed increment Directly converted into VFD frequency increment. All quantization curves are linearly interpolated to ensure high resolution output.

[0116] In order to ensure that the three-way execution action synchronously reaches the effect layer, the zero-point alignment pulse is distributed based on the IEEE1588 distributed time protocol on the real-time industrial Ethernet. After network static calibration, the clock deviation is controlled within microseconds, so that the roll gap adjustment, roll temperature regulation and traction speed frequency conversion are started at the same physical time, avoiding control coupling and causing thickness secondary fluctuation.

[0117] During the movement of the actuator, the grating ruler reads the roll gap displacement in real time, the infrared probe reads the roll temperature in real time, and the new magneto-electric encoder reads the traction speed. These feedbacks are returned to the scheduling layer in the form of observation residuals Within half a control cycle. Use interpolation algorithm to calculate the remaining time period that has not been executed in the current cycle, continue to send increment instructions, form increment-correction microcycle, and ensure that the execution value seen by the controller and the expected deviation are compressed within one adjustment cycle.

[0118] If the read value deviates from the target by more than the threshold value within two cycles, the system quickly locates the fault according to the failure mode diagnosis table: roll gap channel deviation triggers hydraulic redundancy reduction, roll temperature channel deviation triggers parallel electric heating block, and traction speed channel deviation triggers standby servo driver. Redundant devices have the same instruction interface, and the switching process is completed within the network cycle level, avoiding thickness control interruption.

[0119] High-speed instruction scheduling uses three safety measures of synchronous clock, incremental interpolation and redundant switching to convert the optimal control vector into stable, traceable and uninterrupted physical action, and writes the response value back to the next optimization cycle, forming an optimization-execution-feedback closed loop.

[0120] The instruction-effect mapping sampling uses cubic spline interpolation to eliminate linear splicing jitter, and improves the smoothness of high-precision roll gap and roll temperature regulation. The double-layer synchronous gate lock execution end real starting time solves the old difficult problem of network synchronization but endpoint asynchronization, eliminating the weak link for high-speed production. The double-loop incremental interpolation strategy divides the passband into two levels of servo and main control, allowing local equipment to be transient and stable, and giving the optimizer enough time to analyze trends, avoiding mutual interference between servo response and main control sluggishness in reality. Bayesian networks help diagnose complex compound faults, and let the redundancy switching select the most targeted backup path.

[0121] Step four uses the rolling horizon optimal control quantity to solve the high-speed instruction scheduling and real-time feedback nested two-level mechanism, which maps the digital twin prediction to physical action, realizes the last hop from data to execution. Through confidence constrained quadratic programming, the system takes the digital twin uncertainty into the control decision, avoiding overlooking risks; through synchronous clock and incremental interpolation, the system converts the calculation results into stable output on the millisecond scale, filling the model-execution gap; through failure diagnosis and redundancy switching, the system brings potential execution layer faults into closed-loop control, extending the intelligent tentacle to the edge of hardware.

[0122] High-speed calendering production puts forward strict requirements on the thickness control execution chain: on the one hand, the first control vector has been optimized in step four, but if any link in the servo, hydraulic, temperature control or frequency conversion chain is delayed, drifted or stuck, the thickness error will be immediately amplified; on the other hand, long-term operation in high temperature and humidity will accelerate the aging of the actuator, and if there is no real-time health assessment and self-healing closed loop, the gradually accumulated micro-faults will eventually evolve into a sudden shutdown.

[0123] Step five, millisecond-level health assessment of the acquisition-fusion-control whole chain, construction of the electrical-mechanical-environmental three-domain health matrix, use of exponential entropy weight scoring and Markov residual life prediction to real-time judge the performance degradation of devices and algorithms; if the health degree is lower than the threshold, the shadow drive hot standby, cloud operation and maintenance and self-healing strategies are triggered, realizing the hybrid redundancy occupation of software and hardware, and ensuring the system to maintain high availability and low downtime risk in long-term operation.

[0124] Step 501, real-time health assessment and early warning identification In each control cycle, the health index vector of the key nodes of the execution chain is extracted, the drift trend and dynamic reliability are quantified, and the decision basis for fault prediction and redundancy switching is provided.

[0125] The execution chain is divided into four nodes: roll gap servo, electro-hydraulic pressing, roll temperature loop and traction frequency conversion, and three-domain signals are collected for each node: electrical domain signals (current waveform, power factor), mechanical domain signals (displacement residual, velocity ripple), ring domain signal, (oil temperature, bearing vibration), through normalized mapping function :

[0126] form node health vector , spliced into a full-chain health matrix , health matrix provides real-time health quantification basis for full-chain, unified scale:

[0127] where: normalized mapping function , the first node three-domain signal is dimensionally compressed and dimensionally unified linear-nonlinear hybrid mapping; node health vector , The value is 1 to 4, including power distortion factor, motor temperature rise rate, residual root mean square, etc.

[0128] On the sliding window , the full-chain health matrix is first-order differentiated to calculate the drift matrix , and then the exponential entropy weight method is used to calculate the node health score :

[0129] where: node health score : output interval , the value is closer to 1, indicating a higher degree of health; exponential decay factor : according to the long-term reliability of the node, take the value of positive real number; weight matrix : equal to the entropy weight vector After diagonalization, left multiply the inverse of the node Mahalanobis covariance matrix to ensure double weighting of high information indicators and covariance compression dimensions; first-order drift vector : first-order difference of node health vector; second-order drift vector : first-order difference of , used to capture drift acceleration; second-order weighting coefficient : take , adaptively increase with production speed to strengthen early sensitivity to high-speed misalignment; two norms : Euclidean length after matrix multiplication, used to quantify the overall drift energy.

[0130] By giving higher weight to the drift index with high information content, the score is sensitive to true failure precursors but not overly responsive to noise.

[0131] Node health score sequence Discretized into several health level states, Markov transition matrix is constructed Rolling prediction of the first arrival probability of the failure absorbing state is made using the Chapman-Kolmogorov equation, and the remaining life of the node is calculated When is lower than the preset health level threshold, the system generates a preventive maintenance work order in the cloud.

[0132] Health score Synchronize the input weight matrix of step four Let the high health node bear more control amplitude, and the low health node automatically reduce the load to prolong the total chain life cycle.

[0133] Through multi-domain health vector, exponential entropy weight drift score, Markov residual life and health-control collaborative weighting, the system diagnoses the execution chain life trend in milliseconds, and embeds the health information closed loop into the optimizer.

[0134] The real-time health assessment system is based on multi-domain signal fusion, making the previous black box executor become a transparent white box. The second-order drift and entropy weight amplification strategy makes the system extremely sensitive to early micro-defects; the Markov RUL mechanism gives the remaining available time on the probability level, so that the decision is no longer dependent on the head. Health-control collaborative weighting decouples the operation and maintenance information from production optimization, realizing the dynamic symbiosis of equipment and process.

[0135] Step 502, redundancy switching and remote operation and maintenance collaboration When the node health score drops sharply or the actual response continues to be inaccurate, the same type of redundant device is triggered in milliseconds and the operation and maintenance data is uploaded to the cloud and remote diagnosis is completed in the background.

[0136] The redundant executor is in shadow driving mode in normal times: the driver receives the same pulse, PWM or frequency instructions as the main channel, but the output is optically isolated and does not act on the mechanical end, only for health monitoring. In this way, the shadow executor is always in synchronization with the main channel, and only the soft relay switching output path is needed to complete seamless takeover when switching.

[0137] The switching threshold uses the health score and the response residual Dual criteria: when the health score drops below the soft threshold and the response residual exceeds the residual band for three consecutive periods, the shadow drive preempt is triggered, the soft threshold, that is, the preset buffer ensures early machine replacement to avoid hard failure, and the residual band avoids false positives.

[0138] wherein, is the real-time observation value of the first index / channel (such as thickness, temperature, power, etc.); is the instruction / target / reference value corresponding to the first index; the meaning of the residual band is a safety interval that allows the residual to float around zero, and exceeding it is considered abnormal or requires strong compensation; After preemption, the local controller compresses and uploads the latest ten-thousand-frame health vector slices, drift matrix fragments, and event logs of the fault node to the cloud operation platform; the platform calls the deep residual network and compares it with the case library to generate a diagnosis report, which is pushed to the mobile terminal, and the operation personnel decompose tasks according to the report in the next shift.

[0139] If the shadow drive preemption is still detected to be rapidly deteriorating the health of the secondary node in the same chain, enter the safe shutdown state: reduce the traction speed to the technical lower limit, increase the roll temperature conservative margin, and continuously output the thickness large tolerance product until manual intervention is confirmed, to ensure that the main equipment is not damaged due to cascading failure.

[0140] The shadow drive hot standby, health-response double threshold triggering, edge-cloud collaborative operation and maintenance, and safe shutdown bottom-up together build the execution layer fault precursor-hot standby preemption-remote diagnosis-risk downshift self-healing closed loop.

[0141] The shadow drive hot standby design breaks the traditional cold standby charging or manual machine replacement lag mode, realizes seamless second-level plugging; the double threshold triggering allows the switching decision to be checked in two dimensions of fault signal and quality signal, further reducing the risk of misplacement; edge-cloud collaboration upgrades empirical diagnosis to data-driven expert system, shortening fault positioning time; the safe shutdown bottom-up strategy replaces shutdown by downgrading production, winning key buffer for order delivery.

[0142] Step five realizes millisecond monitoring-millisecond preemption-hour repair-safe downshift full-time execution chain guardian with real-time health assessment as the outpost and redundancy switching and remote maintenance as the backup; the health matrix and the remaining life continuously write back the weight matrix , completing the horizontal needle of data between step four and step five. At this point, the intelligent thickness control system completes the closed loop closure from multi-modal acquisition to health self-healing, ensuring that the high-brightness edge-sealed sheet thickness control can still be stable, economical, and long-period running under complex working conditions and equipment aging conditions.

[0143] Model predictive control provides a deterministic optimal solution for the production line, but when it faces sudden disturbances, dramatic changes in raw material batches, or equipment aging beyond the mechanism model description, it often shows parameter adjustment lag and precision decline.

[0144] ​​Step 6: Fusion reinforcement learning (RL) strategies with the MPC controller. Through adjustable weight gating, energy consumption is reduced while ensuring compliance with hard constraints, while improving thickness control accuracy. The RL agent utilizes historical experience replay and online reward recalibration for safe fine-tuning. When confidence intervals or healthiness breach warning lines, it automatically rolls back to the MPC baseline strategy, ensuring a real-time and explainable balance between performance gains from exploration and industrial safety.

[0145] Step 601: Reinforcement Learning Agent Architecture and Offline Pre-training A state-action-reward space is designed to fully match the edge-banding sheet calendering scenario, enabling deep reinforcement learning pre-training based on digital twin offline simulation to provide a high-performance initial strategy for launch.

[0146] The agent's observation vector In digital twin state As the main body, splicing the main eigenvalue of the health matrix , energy consumption real-time power and thickness prediction covariance principal eigenvalues , and obtain the observation vector :

[0147] Among them: the main eigenvalue vector :Extract the dimensional signals with the most concentrated energy of the health matrix through singular value decomposition; Instant power : The sum of roller heating, servo drive and cooling fan power.

[0148] is the posterior state vector (corrected state): thickness field state, material / equipment parameter drift terms, etc.; Observation / Feature Vector Sliding standard deviation or uncertainty estimate; Motion Vector With control vector increment Same dimension: Roll gap displacement increment, Roll temperature adjustment, and traction speed increment. To ensure that the agent output is within the hard constraint range, the action is output through a differentiable projection layer after the network:

[0149] For strategic networks at all times Given the original motion vector; The control input vector that is the final output of the reinforcement learning (RL) module; Projection operator : Project the original action to the physical-safety feasible domain ; the feasible region is synchronized with the fourth step constraint to keep consistent.

[0150] The reward function is mixedly designed to build the reinforcement learning instant reward :

[0151] wherein: thickness deviation weight , energy consumption weight , health penalty weight , production rhythm reward are positive real numbers; health vector one norm : health degradation comprehensive penalty; is the fused real-time thickness estimation; is the target thickness, is the energy consumption index; is the standard deviation vector of the normalized feature vector ; and is the production rhythm reward term. When used, the composite reward guides the agent to seek the Pareto optimal solution among thickness, energy consumption, equipment health and production capacity, rather than simply pursuing dimensional accuracy.

[0152] A fast inference environment is constructed using a digital twin model: the step three three-layer mechanism model and the step five health-failure model are encapsulated as a Gym-like interface, and the simulation of a physical process takes no more than seconds. In this environment, a Soft-Actor-Critic (continuous action) algorithm is used for pre-training; the policy network and the Q network share a two-layer 256-node ReLU backbone, and output Gaussian policy parameters. During the training process, dynamic random injection of raw material viscosity fluctuations, increased roll gap friction and power grid voltage drop disturbances are used to make the strategy have generalization ability.

[0153] Offline simulation enables the agent to master the response strategy for common disturbances before going online, reducing the risk of online detection. Through multi-dimensional state design, projected action safety mapping, composite reward and digital twin fault simulation environment, offline pre-training obtains an initial intelligent agent, laying a high-performance, safe and controllable starting point for online fine-tuning.

[0154] By adding shift and batch number codes to the state vector, the agent has advanced perception of both human and material variables; shell projection and acceleration constraints make the continuous output of deep strategy and mechanical safety boundaries have a differentiable track, enabling large-scale deep networks to work with sub-millimeter roll gap requirements for the first time; the steady-state window of the composite reward allows the agent to actively switch to energy-saving mode, reflecting the dynamic balance of multi-objective optimization; multi-distribution Monte Carlo disturbance training makes the strategy resilient to black swan extreme conditions.

[0155] Step 602, online safety fine-tuning and decision fusion In real production, the reinforcement learning strategy is iteratively refined in a controlled manner and fused with the model predictive controller output, forming a dual decision-making loop. The model predictive control is the first step With the agent action After weighted fusion:

[0156] In the formula: The final issued hybrid control input vector (roll gap increment, roll temperature adjustment amount, traction speed increment, etc.); The action output by the reinforcement learning strategy; The reference control action, which can be the MPC optimization result, traditional PID / rule control, or the safety action of the previous period; Weight coefficient Increases with the amount of online sampling , mainly relying on MPC at the starting stage, gradually leaning towards agent decision-making with experience accumulation, The maximum weight in steady state The rising rate constant.

[0157] Define safety monitoring indicators If the safety monitoring indicators Exceed twice the tolerance or energy consumption Exceeds the baseline , the rollback window is triggered: the weight coefficient Forced to zero , and the weight slowly rises again during this period to prevent the agent from continuously outputting undesirable actions under unknown extreme disturbances.

[0158] Real running data is stored in the form of a priority experience pool, with priority Scoring according to TD error and thickness deviation events; every Step background resampling updates network parameters, learning rate automatically reduces by half with the mean value of the health matrix, reducing excessive exploration during device aging, focusing learning on difficult samples and abnormal samples, and improving sample efficiency.

[0159] Every Online training freezes the strategy generation version , and alternates with the current running version for 10 minutes AB comparison; if the thickness mean square error and energy consumption are better than the old version, upgrade; otherwise, rollback and write failed samples to a low-priority pool to avoid shocks. When using, versioning and small window AB testing ensure that every online upgrade is quantifiable.

[0160] Add a direction consistency factor :

[0161] When and the included angle exceeds ninety degrees, the weight is immediately lowered to (typical value 0.2), at the same time, a directional conflict event is recorded once and the conflict sample is placed in high priority into the experience pool, preventing the mutual cancellation of the two sets of control quantities from causing thickness jitter.

[0162] Online fine-tuning realizes a safe adaptive closed loop of stable takeover-progressive learning-index guardianship-continuous iteration through dynamic fusion of weights, safety monitoring, priority experience playback and version verification of strategies.

[0163] The weight health gate valve realizes parameter-level safety to strategy-level safety, which is a difficult problem in the academic field; the extreme deviation mode enables the monitoring logic to go beyond the threshold cut-off and can pinch the soft landing according to the deviation direction and speed; the multi-scale experience pool focuses on abnormal samples, so that limited online computing power contributes more to the value interval; the gray upgrade path increases along the order tail section-shift-whole day, avoiding the direct impact of new strategies on high-value orders.

[0164] Step 601 provides an available initial strategy, step 602 ensures that online learning is safe and controllable, and is fused with the MPC output through adjustable weights, which not only maintains the original stability but also gradually injects creative scheduling capabilities. The intelligent thickness control system completes the ultimate closed loop from multi-modal perception to adaptive intelligent decision-making, laying the foundation for the evolution of strategies for continuous optimization of pressure-laminated high-gloss light edge sheet in wider working conditions, longer service life and lower energy consumption.

[0165] The high-gloss edge strip of the prior art is coated with a UV coating on the surface to make the surface of the decorative edge strip smooth, and the gloss reaches 85GU or more, but it cannot be applied to the pull handle-free edge process. The shape of the pull handle-free edge strip process has different shapes, and the edge strip needs to be bent into different shapes. The high-gloss edge strip produced by surface coating of a UV coating is prone to cracking and whitening on the surface when bent, which affects the appearance.

[0166] Further, the present application provides a soft-formable high-gloss edge decorative strip, which starts from the selection of raw materials, selects SG-5 or SG-8 type PVC resin powder, and controls the vinyl chloride monomer to be less than 1.0 mg / kg. The lubricant is selected from high-melting-point and high-molecular-weight PE wax, and the filler is selected from nano calcium powder. An environmentally friendly high-performance calcium-zinc stabilizer is used.

[0167] The process uses mirror roller smoothing forming technology, and the surface does not need to be printed, so that the gloss of the decorative strip reaches 85GU (gloss unit) or more.

[0168] The paint-free high-gloss edge sealing decoration strip formula of the present application is as follows: PVC resin (SG-8) 100 parts; nano calcium carbonate 5-10 parts; processing aid: 1-1.5 parts; toughening agent: 7-10 parts; high molecular weight lubricant: 0.5-1.5 parts; antioxidant: 0.5-1.5 parts; high-performance calcium-zinc stabilizer: 3.5-5.5 parts; brightener: several; plasticizer: 2-4 parts; color powder: several; The main purpose of the present application is to fill the gap of the demand for high-gloss products in the soft-molded edge decoration strip, and to solve the abnormal problems such as cracking of the edge sealing material or bubbling of the surface paint film layer in the soft-molding process of the printed high-gloss product.

[0169] Through the high-gloss edge sealing decoration strip of the present application, the gloss can reach more than 85GU; Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0170] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0171] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be realized by other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0172] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e. they can be located in one place or distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the present embodiment.

[0173] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An intelligent thickness control system for calendering high-gloss edge-sealed sheets, comprising: The multimodal acquisition module uses a multi-physics principle synchronous acquisition network, combined with mirror reflection suppression and dynamic range stitching, to perform the same-frequency and same-phase acquisition of laser, terahertz, and ultrasonic data, and simultaneously record environmental disturbance signals; The drift compensation fusion module uses the drift mechanism driven by environmental disturbances to model, extracts the drift vector and noise distribution of each channel, adaptively adjusts the gain matrix through the Bayesian-Kalman fusion algorithm, and outputs a fused thickness vector with confidence; The digital twin synchronization module builds a digital twin model that includes roller system elasticity, material rheology, and thermal field coupling. It uses the Kalman-particle hybrid assimilation algorithm to assimilate measured data to achieve millisecond-level prediction of sheet thickness evolution. The closed-loop execution module, based on rolling predictions of digital twins, uses a model predictive control algorithm with confidence constraints to generate control instructions that meet production line speed, energy consumption, and safety constraints, and then sends them to the actuator through high-speed scheduling; The health assessment module establishes a multi-domain health matrix and performs real-time health assessments on the acquisition, fusion, and control links. It uses exponential entropy weight scoring and Markov residual life prediction to trigger shadow-driven hot standby and self-healing strategies. The strategy fusion module integrates reinforcement learning strategies with model predictive control. Through online fine-tuning and safety monitoring, it optimizes energy consumption and control accuracy while meeting hard constraints and performs adaptive intelligent decision-making.

2. The intelligent thickness control system according to claim 1, characterized in that: The multi-source synchronous acquisition includes a synchronous acquisition network based on multi-physics principles, which uses upper and lower dual-axis laser displacement sensors, terahertz pulse probes and air-coupled ultrasonic probes to capture thickness signals in real time at the same spatial profile and time granularity, and improves signal quality through mirror reflection suppression control and dual redundant calibration mechanisms.

3. The intelligent thickness control system according to claim 2, characterized in that: The multi-source synchronous acquisition also includes using atomic clocks and fiber optic synchronization technology to ensure millisecond-level phase synchronization of different sensing channels, achieving time base unification through a master-slave clock structure and least squares deviation correction, and using full-scale empty run self-test and mirror reflection target calibration to achieve a priori alignment of thickness data and coordinate system.

4. The intelligent thickness control system according to claim 3, characterized in that: The drift compensation fusion includes utilizing the intrinsic correlation between environmental disturbances and historical thickness series, constructing a drift prediction function through linear-nonlinear hybrid kernel regression, extracting the slow-varying drift vector and instantaneous noise distribution of each channel in real time, and adaptively adjusting the gain matrix of Bayesian-Kalman fusion through the confidence matrix.

5. The intelligent thickness control system according to claim 4, characterized in that: The drift compensation fusion also includes triggering an abnormality replacement mechanism through Kalman residual and Mahalanobis distance threshold judgment when a single channel is abnormal, seamlessly replacing the abnormal channel with the predicted value of the digital twin model, and maintaining the continuity and stability of the fused thickness vector.

6. The intelligent thickness control system according to claim 5, characterized in that: The digital twin synchronization includes constructing a digital twin that includes three-layer mechanism models of roller system elasticity-geometry, material rheology-stress, and thermal field-curing, performing state assimilation with measured data through the Kalman-particle hybrid assimilation algorithm, and ingesting process parameter flow in real time to update the model boundary conditions.

7. The intelligent thickness control system according to claim 6, characterized in that: The digital twin synchronization also includes using a fourth-order explicit Runge-Kutta-Cash-Karp variable step size strategy for rolling prediction, outputting a thickness prediction curve and uncertainty band, and encapsulating and forwarding the prediction data to the model predictive controller through a message bus.

8. The intelligent thickness control system according to claim 7, characterized in that: The closed-loop execution control includes utilizing the rolling prediction sequence of the digital twin and adopting a rolling quadratic planning model predictive control algorithm with confidence constraints to solve the control quantity sequence that minimizes the weighted sum of squares of thickness deviations, and issuing instructions to the actuator through a high-speed scheduling mechanism.

9. The intelligent thickness control system according to claim 8, characterized in that: The closed-loop execution control also includes translating control instructions into servo pulses, roller temperature valve pulse widths and traction drive reference frequencies within a millisecond network cycle, and ensuring the accuracy and continuity of execution responses through real-time feedback nesting and redundant switching mechanisms.

10. The intelligent thickness control system according to claim 9, characterized in that: The health assessment and self-healing include millisecond-level health assessment of the entire acquisition-fusion-control link, construction of a three-domain health matrix of motor-machine-environment, and use of exponential entropy weight scoring and Markov remaining life prediction to identify the performance degradation of equipment and algorithms in real time.

11. The intelligent thickness control system according to claim 10, characterized in that: The health assessment and self-healing also include triggering shadow drive hot standby, cloud operation and maintenance, and self-healing strategies when the health level is lower than a threshold, and achieving long-term high availability and low downtime risk of the system through hybrid software and hardware redundancy preemption.

12. The intelligent thickness control system according to claim 11, characterized in that: The strategy fusion optimization includes designing a reinforcement learning agent architecture that fits the edge banding sheet calendering scenario, performing deep reinforcement learning pre-training through digital twin offline simulation, generating high-performance initial strategies, and providing a safe and controllable starting point for online fine-tuning.

13. The intelligent thickness control system according to claim 12, characterized in that: The strategy fusion optimization also includes iterating the reinforcement learning strategy in a controlled manner in real production; Robust adaptive intelligent decision-making is achieved by integrating adjustable weight gating with model predictive control output and utilizing safety monitoring and version verification.

Images