Display process temperature management optimization system

Abstract

The invention relates to the technical field of electric digital data processing, in particular to a display process temperature management optimization system which comprises a data acquisition module, a dynamic modeling module, a cooperative compensation module and an early warning module. The data acquisition module acquires temperature data of a heating area through a multi-type temperature sensor array, and generates a temperature field matrix with a confidence label through multi-source verification; the dynamic modeling module receives the matrix, operates a plurality of machine learning algorithms in parallel to generate a thermal coupling strength distribution diagram, and outputs a region coupling coefficient and a confidence coefficient thermodynamic diagram; the cooperative compensation module calls knowledge base strategies in a layered mode according to the thermodynamic diagram and feeds back early warning signals when abnormity is detected. According to the system, a cross-module closed-loop control mechanism is formed through confidence label driving algorithm weight distribution, confidence thermodynamic diagram control strategy priority, abnormal signal triggering model reconstruction and knowledge base updating, and millisecond-level cooperative suppression of the thermal crosstalk effect of multiple heating areas is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical digital data processing, and in particular to a display process temperature management optimization system. Background Art

[0002] In the manufacturing process of display devices, temperature management is the core technology for maintaining process stability and improving product quality. By precisely controlling the thermal environment at each stage, phase change or interface failure of materials during deposition, curing or annealing can be prevented, thereby avoiding the formation of defects. Specific means include deploying high-precision temperature sensors and closed-loop feedback systems, combined with algorithms to achieve real-time monitoring and dynamic adjustment, achieve uniform distribution of the thermal field, and then optimize molecular arrangement and film consistency, ultimately improving yield and extending component life.

[0003] Existing display process temperature management systems lack the efficiency of synergizing temperature fields across multiple heating zones. In the manufacture of large or complex graphic display panels, the substrate is typically divided into numerous independently temperature-controlled zones to achieve overall thermal uniformity. However, as the number of zones increases, the physical thermal crosstalk effect between heating units (i.e., the mutual temperature influence caused by energy transfer between adjacent zones) significantly increases. This nonlinear coupling makes it difficult for existing control algorithms to predict and compensate for the mutual disturbances between all zones in real time under dynamic process conditions. For example, during the high-temperature phase of the LED evaporation process, adjustments to the set temperature or slight fluctuations in heating power in one zone are rapidly transmitted to adjacent units through heat conduction and radiation. This physical crosstalk is not captured and offset in a timely manner by the control model, causing the actual temperature within the zone to deviate from the set value and generate localized hot or cold spots. This directly leads to uneven molecular packing density, ultimately manifesting as visible mura defects (macroscopic brightness or color unevenness) on the product. This problem is particularly prominent in large-area panel production. The root cause is that existing algorithms lack the adaptability to efficiently manage high-dimensional, strongly coupled, multi-input, multi-output temperature fields. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a display process temperature management optimization system to solve the problem of insufficient real-time compensation of thermal crosstalk effects in multiple heating zones.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows: The present invention provides a display process temperature management optimization system, comprising: a data acquisition module, which acquires temperature sensor data of multiple heating zones, performs multi-source data verification on the temperature sensor data, generates a temperature field matrix with confidence labels, and outputs the matrix to the dynamic modeling module; A dynamic modeling module receives the temperature field matrix, generates a thermal coupling intensity distribution map through parallel processing of multiple machine learning algorithms, and outputs the inter-region coupling coefficient and confidence thermodynamic map to the collaborative compensation module; A collaborative compensation module generates a power compensation strategy based on the thermal coupling intensity distribution map and the confidence thermodynamic map, verifies the strategy convergence by calling historical data from the knowledge base, and outputs a power instruction sequence with phase offset to the early warning module; The early warning module converts the power instruction sequence into an execution signal, monitors the power output characteristics and temperature gradient changes in real time, and triggers an early warning signal when abnormal characteristics are detected and feeds back to the collaborative compensation module.

[0006] Furthermore, in the display process temperature management optimization system of the present invention, the data acquisition module includes: A multi-type temperature sensor array collects raw temperature data and inputs it into a data verification unit; the data verification unit performs: Splitting the original data into a simulation copy, an optimization copy, and a control copy; Input the simulation copy into the digital twin engine for dynamic simulation of heat conduction; The optimized copy is input to the filter processor to perform two-stage processing of wavelet denoising and Kalman filtering; The confidence labels are generated by comparing the differences between the three replica data and are appended to the temperature field matrix.

[0007] Furthermore, in the display process temperature management optimization system of the present invention, the data verification unit is configured to: When it is detected that the temperature difference between the simulation replica, the optimized replica, and the control replica exceeds a preset threshold, the control replica is calibrated using the simulation results output by the digital twin engine; The filtering processor performs a two-stage process of wavelet noise reduction and Kalman filtering to eliminate sensor drift errors and then update the optimized replica.

[0008] Furthermore, in the display process temperature management optimization system of the present invention, the dynamic modeling module is configured to: receiving the temperature field matrix with confidence labels, Run the following algorithms in parallel to process the temperature field matrix: The gradient boosting decision tree algorithm analyzes thermally coupled nonlinear features and outputs feature importance distribution; Long short-term memory networks predict time-varying crosstalk trends over multiple future control cycles; Adaptive particle swarm optimization algorithm dynamically adjusts the frequency of model weight updates; Robust least squares method for online correction of heat transfer function coefficients; The output results of each algorithm are integrated to generate the thermal coupling intensity distribution map.

[0009] Furthermore, in the display process temperature management optimization system of the present invention, the dynamic modeling module is configured to: When the confidence value of any area in the confidence heat map is lower than a preset threshold, a model reconstruction request signal is sent to the data acquisition module; The fault feature coding vector fed back by the early warning module is received, and the long short-term memory network weight is reconstructed by combining the fault feature coding vector and the latest temperature field matrix.

[0010] Furthermore, in the display process temperature management optimization system of the present invention, the collaborative compensation module is configured as follows: Retrieving matching historical optimization strategies from a knowledge base according to the partition confidence value levels of the confidence heat map; The initial power compensation amount is generated by adopting the retrieved history optimization strategy; Verifying the convergence boundary of the initial power compensation amount by Lyapunov exponent analysis; When the convergence boundary meets the stability condition, the compensation amount is decomposed into a power command sequence with phase offset according to the heater thermal response time constant.

[0011] Furthermore, in the display process temperature management optimization system of the present invention, the collaborative compensation module is configured as follows: When the temperature gradient change after the power instruction sequence with phase offset is executed is lower than the set threshold, the current compensation strategy parameters, the thermal coupling intensity distribution map and the compensation effect score are extracted to generate a feature vector; Storing the feature vector in a knowledge base to update the historical optimization strategy; When receiving the strategy freeze instruction from the early warning module, it switches to the reinforcement learning exploration mode to generate a new compensation strategy and overwrite the current strategy.

[0012] Furthermore, in the display process temperature management optimization system of the present invention, the early warning module includes: A power conversion unit receives the power instruction sequence and converts it into a PWM control signal and outputs it to the heater; Harmonic analysis unit collects PWM control signals in real time, performs fast Fourier transform, and outputs harmonic distortion rate; The temperature gradient monitoring unit receives the temperature field matrix and calculates the change of the spatial second-order derivative; The abnormality decision maker compares whether the combined value of the harmonic distortion rate and the spatial second-order derivative change exceeds a threshold.

[0013] Furthermore, in the display process temperature management optimization system of the present invention, the abnormality decision maker is configured to: When the combined value of the harmonic distortion rate and the change in the spatial second-order derivative exceeds a preset threshold, the hierarchical limit control is triggered to reduce the current power output; Extract the current harmonic spectrum characteristics, temperature gradient mutation area and equipment status parameters to generate fault feature coding vector; Feeding back the fault feature coding vector to a dynamic modeling module; A strategy freeze instruction is sent to the collaborative compensation module and the power ramp protection mechanism is activated to gradually reduce the heating power.

[0014] Furthermore, in the display process temperature management optimization system of the present invention, the confidence labels dynamically allocate calculation weights of multiple machine learning algorithms in the dynamic modeling module; The confidence value hierarchy of the partitions of the confidence heat map controls the priority of the collaborative compensation module in calling the knowledge base strategy; The abnormal warning signal triggered by the warning module executes: Drive the knowledge base to update the compensation strategy feature vector, triggering the dynamic modeling module to reconstruct the long short-term memory network weights; The thermal crosstalk effect of multiple heating zones is suppressed by closing the data loop across modules.

[0015] Beneficial effects of the present invention: The display process temperature management optimization system provided by the present invention significantly improves the compensation efficiency of thermal crosstalk effects in multiple heating zones through a multi-module collaborative mechanism: first, the data acquisition module uses multi-source data verification to generate a temperature field matrix with confidence labels, and eliminates sensor drift and thermal conduction hysteresis errors through digital twin simulation and two-stage filtering, providing a high-reliability data foundation for the system; secondly, the dynamic modeling module dynamically allocates algorithm weights according to the confidence labels, and runs the machine learning algorithm in parallel to generate a thermal coupling intensity distribution map, accurately quantifying the nonlinear coupling relationship between regions, and solving the defect of traditional models in insufficient prediction of high-dimensional thermal crosstalk; finally, the collaborative compensation module hierarchically calls the knowledge base strategy based on the confidence thermal map, combines Lyapunov stability verification and phase offset instruction decomposition, and realizes the matching of compensation amount and heat transfer physical delay. At the same time, the abnormal decision-making mechanism of the early warning module triggers hierarchical protection at the millisecond level and drives model reconstruction and strategy update through the joint diagnosis of harmonic distortion rate and temperature gradient changes. This technical solution forms a closed-loop control chain of data acquisition, modeling, compensation and protection, which increases the response speed of thermal crosstalk compensation to sub-second level, effectively suppresses the formation of local hot spots, and improves the manufacturing yield and product consistency of display panels. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0017] Figure 1 A system architecture diagram of a display process temperature management optimization system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the embodiments described 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 work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0019] See also Figure 1 The present invention provides a display process temperature management optimization system, comprising: a data acquisition module, which acquires temperature sensor data of multiple heating zones, performs multi-source data verification on the temperature sensor data, generates a temperature field matrix with confidence labels, and outputs the matrix to the dynamic modeling module; A dynamic modeling module receives the temperature field matrix, generates a thermal coupling intensity distribution map through parallel processing of multiple machine learning algorithms, and outputs the inter-region coupling coefficient and confidence thermodynamic map to the collaborative compensation module; A collaborative compensation module generates a power compensation strategy based on the thermal coupling intensity distribution map and the confidence thermodynamic map, verifies the strategy convergence by calling historical data from the knowledge base, and outputs a power instruction sequence with phase offset to the early warning module; The early warning module converts the power instruction sequence into an execution signal, monitors the power output characteristics and temperature gradient changes in real time, and triggers an early warning signal when abnormal characteristics are detected and feeds back to the collaborative compensation module.

[0020] Redundant thermocouple arrays and dual-band infrared thermal imagers are deployed on the substrate carrier and heater array to simultaneously collect physical temperature and thermal field distribution data in multiple temperature zones. The multi-source data verification specifically involves establishing three-channel data streams: a process simulation replica, a data optimization replica, and an actual control replica. The simulation replica is input into the digital twin platform for dynamic simulation of heat conduction, the optimization replica eliminates sensor drift errors through a wavelet noise reduction and Kalman filter fusion algorithm, and the control replica serves as the actual execution benchmark. By comparing the consistency of the three replica data and the simulation deviation threshold, the optimal data source is dynamically selected and a data reliability label (confidence score) is attached. Finally, a multi-dimensional temperature field matrix containing spatial coordinates, temperature values, and confidence scores is generated and output to the dynamic modeling module. This design solves the problem of insufficient reliability of traditional single sensor data through dual verification of physical models and data-driven.

[0021] The dynamic modeling module receives the temperature field matrix with confidence labels and processes it using a four-algorithm parallel architecture: a gradient boosting decision tree algorithm analyzes the nonlinear thermal coupling characteristics of the heating interval and outputs a feature importance distribution; a long-short-term memory network predicts the time-varying crosstalk trend over the next five control cycles; an adaptive particle swarm algorithm dynamically adjusts the model weight update frequency based on the confidence labels; and a robust least squares method is used to online correct the heat transfer function coefficients to compensate for sensor nonlinear errors. An ensemble learning framework performs a weighted fusion of the four algorithm outputs to generate a thermal coupling intensity distribution map containing inter-region coupling coefficients and confidence thermodynamic maps, which is then output to the collaborative compensation module. The confidence labels directly drive the algorithm weight assignment, with high-confidence regions prioritizing the predicted model output and low-confidence regions strengthening the error correction algorithm weight.

[0022] The collaborative compensation module determines the thermal crosstalk intensity based on the inter-region coupling coefficients in the thermal coupling intensity distribution map. It then uses the confidence values ​​of the confidence thermodynamic map to hierarchically invoke a knowledge-based historical optimization strategy: a combined model predictive control and sliding mode control strategy is used in high-confidence zones, while a reinforcement learning exploration strategy is employed in low-confidence zones. Lyapunov exponent analysis is used to verify the convergence bounds of the generated power compensation strategy. When the convergence bounds meet stability requirements, the compensation is decomposed into a power command sequence with a phase offset based on the heater's thermal response time constant and output to the early warning module. The phase offset of the power command sequence is dynamically calculated based on thermal inertia parameters to align the compensation action with the heat transfer delay.

[0023] The early warning module converts the power command sequence into a PWM control signal to drive the heater. It also analyzes the harmonic distortion of the power output spectrum through a fast Fourier transform and calculates the change in the spatial second-order derivative using the temperature field matrix. When the combined value of the harmonic distortion and temperature gradient change exceeds a preset threshold, a thermal shock risk is identified and a three-level response is triggered: 1) Initiating graded limiting control to reduce power output; 2) Extracting harmonic spectrum features, temperature gradient mutation areas, and device status parameters to generate a fault feature encoding vector that is fed back to the dynamic modeling module; and 3) Sending a policy freeze command to the collaborative compensation module and initiating the power ramp protection mechanism. This early warning signal triggers the collaborative compensation module to switch to reinforcement learning mode, while simultaneously driving the knowledge base to update the historical policy feature vector.

[0024] The confidence labels generated by the data acquisition module drive the algorithm weight allocation strategy of the dynamic modeling module. The confidence heat map output by the dynamic modeling module guides the collaborative compensation module in selecting the control strategy hierarchy for each partition. The abnormal warning signal triggered by the early warning module simultaneously triggers the knowledge base strategy update and model reconstruction. When the confidence value of a region in the confidence heat map falls below a threshold, the dynamic modeling module sends a model reconstruction request to the data acquisition module, triggering sensor calibration and data replica reconstruction. Simultaneously, the fault feature vectors fed back by the early warning module accelerate the reconstruction of the long-short-term memory network weights. Through data label transmission, abnormality feedback, and strategy switching, each module forms a closed-loop control chain, achieving millisecond-level collaborative suppression of thermal crosstalk effects.

[0025] Specifically, in the display process temperature management optimization system of the present invention, the data acquisition module includes: A multi-type temperature sensor array collects raw temperature data and inputs it into a data verification unit; the data verification unit performs: Splitting the original data into a simulation copy, an optimization copy, and a control copy; Input the simulation copy into the digital twin engine for dynamic simulation of heat conduction; The optimized copy is input to the filter processor to perform two-stage processing of wavelet denoising and Kalman filtering; The confidence labels are generated by comparing the differences between the three replica data and are appended to the temperature field matrix.

[0026] The multi-type temperature sensor array includes PT100-grade thermocouples embedded in the gaps between the heating elements and a gantry-mounted dual-band infrared thermal imager. The thermocouples are spaced no more than 5mm apart, and the infrared bands cover the 3-5μm and 8-14μm spectral ranges. These sensors simultaneously collect data on the substrate surface temperature distribution and the radiated heat field from the heaters. Raw temperature data is transmitted to the data verification unit via an electromagnetically shielded CAN FD bus, with a sampling frequency of at least 100Hz to prevent data distortion caused by high-frequency electromagnetic interference.

[0027] The data verification unit splits the original temperature data into three independent copies: the simulation copy is input into the digital twin engine for dynamic simulation of heat conduction. The engine constructs a three-dimensional thermodynamic model based on the finite element method. The input parameters include the specific heat capacity of the substrate material, the thermal diffusivity of the carrier and the radiation angle coefficient of the heater; the optimization copy is input to the filter processor for two-stage processing. The first stage uses the db4 wavelet basis function to decompose the high-frequency noise of the temperature signal. The second stage uses the Kalman filter to fuse the device vibration compensation parameters to eliminate the sensor drift error; the control copy is temporarily stored in the buffer register as the actual control reference data stream.

[0028] The digital twin engine outputs dynamic heat conduction simulation results, and the filter processor outputs optimized data after noise reduction. These two are compared with the control replica in real time. When the absolute temperature difference between any two replicas exceeds a threshold of 0.5°C, an arbitration mechanism is triggered: the digital twin engine's simulation results are preferentially used to calibrate the control replica. Simultaneously, the filter processor updates the filter coefficients of the optimized replica based on the latest device state parameters, forming a dynamic error compensation closed loop.

[0029] The three replicas are compared to generate confidence score labels. Specifically, the root mean square error (RMS) between the simulation results and the optimized data is calculated as the baseline confidence value. When the arbitration mechanism is triggered, the confidence value of the region is downgraded. Ultimately, a multi-dimensional temperature field matrix containing spatial coordinates, temperature values, and confidence scores is generated. The confidence labels are directly linked to the algorithm weight allocation strategy of subsequent modules, enabling the mapping of data reliability to control priority.

[0030] This multi-copy verification mechanism addresses data distortion caused by sensor drift and thermal conduction hysteresis in high-temperature environments through a triple-path approach: physical model simulation, signal processing optimization, and real-time arbitration calibration. Dynamic generation of confidence labels enables downstream modules to allocate computing resources based on data reliability, providing a data foundation for real-time control of complex thermal fields.

[0031] Specifically, in the display process temperature management optimization system of the present invention, the data verification unit is configured to: When it is detected that the temperature difference between the simulation replica, the optimized replica, and the control replica exceeds a preset threshold, the control replica is calibrated using the simulation results output by the digital twin engine; The filtering processor performs a two-stage process of wavelet noise reduction and Kalman filtering to eliminate sensor drift errors and then update the optimized replica.

[0032] The data verification unit monitors the temperature difference between the simulation replica, optimization replica, and control replica in real time, with a preset temperature difference threshold of 0.5°C in absolute value. When the temperature difference between any two replicas exceeds this threshold, it is determined to be a data reliability anomaly. At this time, the dynamic simulation results of heat conduction output by the digital twin engine are preferentially used to overwrite the current control replica data. This operation is directly written into the control replica register through a high-priority data channel, overwriting the original collected data. The physical model of the digital twin engine is constructed based on the thermal expansion coefficient of the substrate material and the mechanical deformation parameters of the equipment, and its output results serve as the source of thermodynamic benchmark data.

[0033] Simultaneously, the filter processor continuously performs two-stage signal processing on the optimized replica. The first stage uses the Daubechies 4 wavelet basis function to decompose the temperature signal, filtering out high-frequency electromagnetic interference and random noise. The second stage integrates the heater vibration sensor data through a Kalman filter to compensate for thermocouple drift errors caused by mechanical stress. After completing this two-stage processing, the optimized data immediately updates the original optimized replica, forming a dynamic error correction closed loop. The updated optimized replica participates in the next round of temperature difference comparison, forming a continuous optimization mechanism.

[0034] This calibration and update mechanism improves data reliability through two pathways: digital twin calibration addresses physical model deviations caused by thermal conduction hysteresis, while filter updates eliminate sensor drift errors. Control replica calibration directly influences subsequent power control command generation, while optimization replica updates provide a real-time data foundation for confidence scoring. These two mechanisms work together to ensure that the confidence labels in the temperature field matrix accurately reflect the actual thermal state of multiple heating zones, providing physically consistent input data for the dynamic modeling module.

[0035] This technical solution breaks through the traditional single-calibration model. By combining physical models with signal processing, it reduces temperature data acquisition errors under the high-temperature conditions of large-scale panel manufacturing. The historical records generated by calibration and update operations are simultaneously entered into a knowledge base, providing characteristic sample data for subsequent model reconstruction.

[0036] Specifically, in the display process temperature management optimization system of the present invention, the dynamic modeling module is configured to: receiving the temperature field matrix with confidence labels, Run the following algorithms in parallel to process the temperature field matrix: The gradient boosting decision tree algorithm analyzes thermally coupled nonlinear features and outputs feature importance distribution; Long short-term memory networks predict time-varying crosstalk trends over multiple future control cycles; Adaptive particle swarm optimization algorithm dynamically adjusts the frequency of model weight updates; Robust least squares method for online correction of heat transfer function coefficients; The output results of each algorithm are integrated to generate the thermal coupling intensity distribution map.

[0037] After receiving the temperature field matrix with confidence labels, the dynamic modeling module allocates four computing cores to execute the algorithm in parallel. The gradient boosting decision tree algorithm uses the spatial coordinates and temperature values ​​of the temperature field matrix as input features. Through recursive splitting, it constructs a decision tree with a depth of no more than 15. This algorithm analyzes the nonlinear coupling relationship between heat transfer between heating zones and outputs the importance distribution of the thermal crosstalk intensity features in each zone. This distribution quantifies the influence of energy transfer between adjacent heating units, providing a basis for compensation priority.

[0038] The Long Short-Term Memory (LSTM) network uses a 128-hidden-layer architecture to process historical temperature data labeled with confidence levels in a time series manner, predicting time-varying crosstalk trends in each region over the next five control cycles. The predictions include information on the amplitude and phase of temperature fluctuations. The dynamic modeling module adjusts the prediction model's learning rate based on the confidence levels, allocating more computing resources to high-confidence regions to improve prediction accuracy.

[0039] An adaptive particle swarm algorithm monitors the convergence of feature importance distribution and time-varying crosstalk trends in real time, dynamically adjusting the particle swarm inertia weight within a range of 0.4 to 0.9 to optimize the frequency of model parameter updates. When confidence labels indicate a decrease in data reliability, the particle swarm search step size is increased to accelerate model response. A robust least squares algorithm simultaneously corrects the heat transfer function coefficients online, using a forgetting factor of λ = 0.98 to mitigate the impact of sensor nonlinearity on the model.

[0040] The ensemble learning framework uses a random forest algorithm to perform a weighted fusion of the outputs of the four algorithms. Feature importance distribution serves as the basis for regional thermal coupling strength, time-varying crosstalk trends provide a dynamic correction factor, model update frequency controls the fusion calculation period, and the corrected heat transfer function coefficients ensure physical consistency. Ultimately, a thermal coupling strength distribution map is generated, including inter-regional coupling coefficients and a confidence heatmap. The confidence heatmap inherits the input data confidence labels and overlays the model processing confidence assessment. This is then output to the collaborative compensation module to guide strategy selection.

[0041] This parallel processing architecture combines physical models with data-driven approaches through algorithmic division of labor. A resource allocation mechanism driven by confidence labels enables the system to maintain prediction accuracy despite fluctuations in data reliability. The multi-dimensional parameters of the thermal coupling intensity distribution map provide a quantitative basis for downstream compensation strategies, addressing the difficulty of traditional single models in handling high-dimensional thermal crosstalk.

[0042] Specifically, in the display process temperature management optimization system of the present invention, the dynamic modeling module is configured to: When the confidence value of any area in the confidence heat map is lower than a preset threshold, a model reconstruction request signal is sent to the data acquisition module; The fault feature coding vector fed back by the early warning module is received, and the long short-term memory network weight is reconstructed by combining the fault feature coding vector and the latest temperature field matrix.

[0043] The dynamic modeling module monitors the regional confidence values ​​of the confidence heat map in real time, with a preset threshold of 0.85. If the confidence value of any region falls below this threshold, the thermal coupling model for that region is deemed unreliable. A model reconstruction request signal is generated, carrying the region's coordinates, confidence value, and timestamp, and transmitted to the data acquisition module via a high-speed data bus. This signal triggers the data acquisition module to initiate the sensor calibration process and data replica reconstruction mechanism, thereby improving the quality of the original data acquisition at the hardware level.

[0044] The fault feature encoding vector fed back by the early warning module contains characteristic data from three dimensions: harmonic spectrum features mapping power controller anomalies, coordinates of temperature gradient abrupt changes identifying locations where heat transfer is blocked, and device status parameter records including heater current fluctuations and radiator efficiency. The dynamic modeling module receives this vector and spatially aligns it with the latest temperature field matrix to extract the thermal coupling feature dataset corresponding to the fault period.

[0045] The process of reconstructing the long-short-term memory network weights utilizes a transfer learning mechanism: the parameters of the fully connected layers in the network infrastructure are retained, while the weight matrices from the output layer to the hidden layer are reset. Fault signature datasets are prioritized as training samples and the initial learning rate is increased to three times the standard value to accelerate feature extraction. The reconstructed network improves its ability to recognize sudden thermal crosstalk patterns, and the predicted output is synchronously fed back to the adaptive particle swarm algorithm to adjust the model update frequency.

[0046] This reconstruction mechanism shortens the prediction recovery time in low-confidence areas by synergizing hardware-level data calibration with software-level model updates. Operation logs generated during the reconstruction process are simultaneously stored in the knowledge base's fault case collection, providing strategic reference for subsequent similar anomalies. The combined processing of fault feature vectors and temperature field matrices enables the system to adaptively map physical anomalies to model parameters.

[0047] Specifically, in the display process temperature management optimization system of the present invention, the collaborative compensation module is configured as follows: Retrieving matching historical optimization strategies from a knowledge base according to the partition confidence value levels of the confidence heat map; The initial power compensation amount is generated by adopting the retrieved history optimization strategy; Verifying the convergence boundary of the initial power compensation amount by Lyapunov exponent analysis; When the convergence boundary meets the stability condition, the compensation amount is decomposed into a power command sequence with phase offset according to the heater thermal response time constant.

[0048] The collaborative compensation module prioritizes control based on the confidence values ​​of the regions in the confidence heatmap. In the high-confidence zone (confidence value ≥ 0.85), a combined model predictive control and sliding mode control strategy from the knowledge base is invoked. In the medium-confidence zone (0.7 ≤ confidence value < 0.85), a fuzzy PID control strategy is activated. In the low-confidence zone (confidence value < 0.7), historical reinforcement learning exploration cases are retrieved. The retrieval process uses a feature hashing algorithm to match the regional coupling coefficients of the current thermal coupling intensity distribution map and returns the most similar historical optimization strategy set.

[0049] Initial power compensation values ​​are generated using a retrieved historical optimization strategy. The specific steps are: The compensation gain coefficient matrix in the analytical strategy is multiplied with the regional coupling coefficients of the current thermal coupling intensity distribution map to output a baseline compensation value for each heating unit. The baseline compensation value is then superimposed with a proportional differential correction term for the temperature gradient rate of change to form the initial power compensation value set. This process retains the core parameter framework of the historical strategy while introducing real-time dynamic correction factors.

[0050] The convergence boundary of the initial power compensation is verified through Lyapunov exponent analysis: a differential equation for the compensation is constructed and the divergence rate of its trajectory in phase space is calculated. Stability conditions are considered met when the maximum Lyapunov exponent is less than zero and the convergence radius does not exceed the heater safety threshold. Verification results are categorized into three levels: full convergence (exponent < -0.5), boundary convergence (-0.5 ≤ exponent ≤ 0), and divergence (exponent > 0). Only the first two levels trigger subsequent instruction generation.

[0051] The compensation is decomposed into a power command sequence with a phase offset based on the heater's thermal response time constant. The time constant is pre-calibrated through step response testing (range: 0.2-1.5 seconds). The phase offset is calculated using the formula Δφ = τ·ω (ω is the angular frequency of thermal fluctuation), synchronizing the compensation command with the heat transfer process. The command sequence is broken down into millisecond-level power adjustment segments according to the control cycle to avoid thermal shock caused by sudden power surges.

[0052] This technology chain addresses the response lag and overshoot issues of traditional compensation methods through confidence-graded strategy selection, real-time dynamic correction, mathematical stability verification, and fourth-order timing matching. A phase-shifting mechanism within the power command sequence improves thermal crosstalk suppression by 40%, while Lyapunov analysis eliminates the risk of control instability. After strategy execution, the knowledge base automatically records compensation parameters and performance scores, forming a continuous optimization closed loop.

[0053] Specifically, in the display process temperature management optimization system of the present invention, the collaborative compensation module is configured as follows: When the temperature gradient change after the power instruction sequence with phase offset is executed is lower than the set threshold, the current compensation strategy parameters, the thermal coupling intensity distribution map and the compensation effect score are extracted to generate a feature vector; Storing the feature vector in a knowledge base to update the historical optimization strategy; When receiving the strategy freeze instruction from the early warning module, it switches to the reinforcement learning exploration mode to generate a new compensation strategy and overwrite the current strategy.

[0054] After the power command sequence executes, the temperature gradient monitoring unit calculates the change in the spatial second-order derivative in real time. A successful compensation event is determined when the absolute value of this change remains below a set threshold (e.g., 0.5°C / mm²) for three consecutive control cycles. Three sets of core parameters are then extracted to generate a feature vector: compensation strategy parameters, including the gain coefficient matrix and the weights of the differential correction term; a thermal coupling intensity distribution map, capturing the regional coupling coefficients and confidence thermograms during the execution period; and a compensation effectiveness score, calculated based on the temperature gradient decrease rate and the duration of steady-state maintenance.

[0055] The feature vectors are uniquely indexed using a 128-bit hash code and stored in the knowledge base's historical optimization strategy tree. This storage process performs a two-level optimization: the first level uses a K-means clustering algorithm to group similar feature vectors, and the second level uses decision tree pruning techniques to merge redundant strategy nodes. The historical optimization strategy tree is stored hierarchically by material type, with separate strategy partitions for LEDs and LED substrates, and thickness parameters used as the basis for subnode classification.

[0056] The early warning module's policy freeze command carries a fault level code (levels 1-3). When the collaborative compensation module receives this command, it immediately terminates the current policy execution and switches to reinforcement learning exploration mode. Reinforcement learning utilizes a Q-learning framework, using the thermal coupling intensity distribution map as the state input, the power compensation amount as the action space, and the compensation effect score as the reward function. The exploration process uses an ε-greedy strategy (ε=0.3), and the generated new compensation strategy directly overwrites the current policy parameter table.

[0057] This mechanism drives knowledge base evolution through quantitative evaluation of compensation effectiveness. Structured storage of feature vectors improves policy retrieval efficiency by 50%. A reinforcement learning exploration model provides innovative policy paths within safety boundaries during fault conditions. Policy overlay operations preserve the optimal parameter architecture while updating local weights. This closed loop of knowledge base optimization and new policy generation enables the system to adapt to the thermal management requirements of new display materials.

[0058] Specifically, the display process temperature management optimization system of the present invention, the early warning module includes: A power conversion unit receives the power instruction sequence and converts it into a PWM control signal and outputs it to the heater; Harmonic analysis unit collects PWM control signals in real time, performs fast Fourier transform, and outputs harmonic distortion rate; The temperature gradient monitoring unit receives the temperature field matrix and calculates the change of the spatial second-order derivative; The abnormality decision maker compares whether the combined value of the harmonic distortion rate and the spatial second-order derivative change exceeds a threshold.

[0059] The power conversion unit receives a phase-shifted power command sequence and converts the voltage amplitude and timing parameters into a pulse-width modulated signal via the IGBT module. The PWM control signal has a duty cycle resolution of 0.1% and a drive frequency range of 1-10kHz. This signal is then output to the power controller for each zone heater. This unit monitors the output current feedback in real time and automatically adjusts the carrier frequency to maintain signal stability when a sudden change in load impedance is detected.

[0060] The harmonic analysis unit uses a high-speed ADC to acquire the PWM control signal waveform, with a sampling frequency of at least 100kHz. It performs a 1024-point fast Fourier transform to calculate the 0-1kHz frequency spectrum distribution, extracting the total harmonic distortion (THD) and the amplitudes of characteristic subharmonic components. This harmonic distortion calculation includes fundamental amplitude correction to eliminate systematic errors caused by measurement line impedance.

[0061] The temperature gradient monitoring unit receives the latest temperature field matrix and calculates the spatial second-order derivative using a five-point difference method. In the xy coordinate system, it traverses the substrate surface with a 5mm step size and outputs the temperature change curvature value at each coordinate point. The spatial second-order derivative is Gaussian filtered to eliminate noise interference, and the global thermal field gradient distribution map is updated every 200ms.

[0062] The anomaly decision maker receives the harmonic distortion rate and the change in the spatial second-order derivative and calculates a combined value: the harmonic distortion rate is normalized to a risk factor α between 0 and 1, and the change in the spatial second-order derivative is mapped to a risk factor β between 0 and 1. The combined value is weighted according to the formula γ = 0.6α + 0.4β. When the γ value exceeds the preset threshold of 0.75, a thermal shock risk is determined and a third-level response instruction is generated. This decision-making mechanism simultaneously references safety margin data in the equipment process parameter library to avoid false triggering.

[0063] Each unit forms a closed detection loop: the real-time output signal of the power conversion unit is the target of harmonic analysis, temperature gradient monitoring results are aligned with the power status in time and space, and the anomaly decision maker integrates power and thermal field characteristics to achieve joint diagnosis. The combined value judgment model addresses the insufficient sensitivity of traditional single-metric detection, identifying local overheating risks 200ms in advance.

[0064] Specifically, in the display process temperature management optimization system of the present invention, the abnormality decision maker is configured to: When the combined value of the harmonic distortion rate and the change in the spatial second-order derivative exceeds a preset threshold, the hierarchical limit control is triggered to reduce the current power output; Extract the current harmonic spectrum characteristics, temperature gradient mutation area and equipment status parameters to generate fault feature coding vector; Feeding back the fault feature coding vector to a dynamic modeling module; A strategy freeze instruction is sent to the collaborative compensation module and the power ramp protection mechanism is activated to gradually reduce the heating power.

[0065] When the weighted combination of the harmonic distortion rate and the change in the spatial second-order derivative exceeds a threshold of 0.75, the anomaly decider triggers a three-level response mechanism. The first level initiates graded limit control, reducing the current power output by 10%, 20%, and 30%. The limit amplitude is dynamically calculated based on the percentage of combined values ​​exceeding the limit. This power reduction is achieved by adjusting the PWM signal duty cycle in real time, with a response delay of less than 50ms, immediately preventing the risk of thermal shock from escalating.

[0066] The three-dimensional data source for extracting fault features includes: the harmonic analysis unit provides the amplitude distribution of 16th-order harmonics in the 0-1kHz spectrum; the temperature gradient monitoring unit identifies the coordinates and slope of the sudden change region of the spatial second-order derivative; and the equipment status parameters include the fluctuation of the heater operating current, the cooling fan speed, and the coolant flow rate. These three types of data are normalized to values ​​in the 0-1 range and encoded into a 128-dimensional feature vector in a fixed field order.

[0067] The fault feature encoding vector is fed back to the dynamic modeling module via a high-speed data bus, with a timestamp and regional coordinate labels appended to the vector header. This vector is directly linked to the dynamic modeling module's reconstruction mechanism, providing labeled training samples for the long-short-term memory network. This feedback operation simultaneously triggers the model weight initialization instruction, eliminating the impact of historical accumulated errors on the prediction.

[0068] The policy freeze command sent to the collaborative compensation module involves two levels of control: immediately terminating the current compensation policy and freezing the knowledge base policy retrieval interface. Simultaneously, a power ramp protection mechanism is initiated, gradually reducing the heating power to a safe threshold at a rate of 5% per second. During the ramp protection period, the temperature gradient is continuously monitored, triggering the recovery protocol when the value falls back into the normal range.

[0069] This response chain uses power limiting control to quickly curb risk spread, feature vector feedback to enable fault tracing and model correction, strategy freezing to prevent erroneous compensation, and a power ramp mechanism to prevent material stress damage caused by sudden temperature changes. These four steps form a complete closed loop of "suppression, analysis, correction, and recovery." The structured design of fault feature encoding vectors improves the accuracy of system self-diagnosis.

[0070] Specifically, in the display process temperature management optimization system of the present invention, the confidence labels dynamically assign calculation weights of multiple machine learning algorithms in the dynamic modeling module; The confidence value hierarchy of the partitions of the confidence heat map controls the priority of the collaborative compensation module in calling the knowledge base strategy; The abnormal warning signal triggered by the warning module executes: Drive the knowledge base to update the compensation strategy feature vector, triggering the dynamic modeling module to reconstruct the long short-term memory network weights; The thermal crosstalk effect of multiple heating zones is suppressed by closing the data loop across modules.

[0071] Confidence labels serve as quantitative indicators of data reliability, driving the dynamic modeling module's algorithm resource allocation strategy: when the regional confidence value is ≥0.85, 70% of computing resources are allocated to the Long Short-Term Memory Network prediction algorithm; when the confidence value is between 0.7 and 0.85, the gradient boosting decision tree algorithm receives 50% of the computing weight; when the confidence value is less than 0.7, the error correction weight of the robust least squares method is increased to 60% of the total resources. This allocation mechanism, implemented through a real-time resource scheduler, prioritizes prediction accuracy in high-reliability areas with limited computing power.

[0072] The confidence heatmap's partitioned confidence levels control the collaborative compensation module's knowledge base call priority: high-confidence regions (confidence values ​​≥ 0.85) prioritize the model predictive control policy library, with response latency less than 10ms; medium-confidence regions (0.7 ≤ confidence value < 0.85) search the fuzzy PID policy library; and low-confidence regions (confidence values ​​< 0.7) enable the reinforcement learning policy exploration queue. Priority parameters are written to the policy selector register to control the scheduling order of search threads.

[0073] The abnormal warning signal triggered by the early warning module includes a three-level linkage response: the first level drives the knowledge base update engine to extract the current compensation strategy feature vector, which includes the power compensation amount, thermal coupling coefficient matrix and temperature gradient change rate. The feature vector is stored in the historical case library after generating an index identifier through 128-bit hash coding; the second level triggers the long-term and short-term memory network weight reconstruction of the dynamic modeling module. The reconstruction process adopts the transfer learning mechanism, retains the fully connected layer parameters, and resets the output layer weight matrix; the third level starts the cross-module data synchronization protocol, and packages the thermal field data, control parameters and equipment status during the abnormal period into a diagnostic data set.

[0074] The above mechanism forms a closed-loop control chain: confidence labels optimize model computational efficiency, confidence value hierarchies enhance policy matching accuracy, and anomaly signals drive system self-evolution. Replica calibration in the data acquisition module, weight allocation in dynamic modeling, strategy selection for collaborative compensation, and anomaly handling in the early warning module achieve millisecond-level data synchronization through shared storage. This design addresses the compensation lag caused by module fragmentation in traditional temperature control systems and improves the efficiency of suppressing thermal crosstalk effects across multiple heating zones.

[0075] The display process temperature management optimization system provided by this invention addresses the issue of insufficient real-time compensation for thermal crosstalk effects in multiple heating zones through a modular architecture and collaborative mechanisms. The system first collects temperature sensor data from multiple heating zones through a data acquisition module, performs multi-source data verification, and generates a temperature field matrix with confidence labels. The confidence labels quantify data reliability, for example, by calibrating data deviations through replica comparison and digital twin simulation, providing high-precision input for downstream processing. This step eliminates noise caused by sensor drift and thermal conduction hysteresis, establishing a reliable data foundation.

[0076] After receiving the temperature field matrix, the dynamic modeling module runs multiple machine learning algorithms in parallel, including a gradient boosting decision tree to analyze the nonlinear characteristics of thermal coupling and a long-short-term memory network to predict time-varying crosstalk trends. The algorithm outputs are fused to generate a thermal coupling intensity distribution map, along with inter-region coupling coefficients and a confidence thermogram. The collaborative compensation module uses the knowledge base's historical strategy to generate initial power compensation based on the partitioned confidence levels of the thermal coupling intensity distribution map and confidence thermogram. The module then verifies the convergence boundaries through Lyapunov exponent analysis. Once stability conditions are met, the compensation is decomposed into a phase-shifted power command sequence based on the heater's thermal response time constant to match the physical delay of heat transfer.

[0077] The early warning module converts power command sequences into execution signals, monitoring harmonic distortion and changes in spatial second-order derivatives in real time. When the anomaly decider detects a combined value exceeding a threshold, it triggers hierarchical limiting control and power ramp protection, while simultaneously providing feedback to the collaborative compensation module on a policy freeze command. Confidence labels dynamically assign modeling algorithm weights, and confidence heatmaps control policy invocation priorities. Anomaly signals drive knowledge base updates and model reconstruction. This closed-loop mechanism achieves millisecond-level thermal crosstalk compensation, suppresses local hotspot formation, and improves panel manufacturing yield.

Claims

1. A display process temperature management optimization system, characterized in that: include: a data acquisition module, which acquires temperature sensor data of multiple heating zones, performs multi-source data verification on the temperature sensor data, generates a temperature field matrix with confidence labels, and outputs the matrix to the dynamic modeling module; A dynamic modeling module receives the temperature field matrix, generates a thermal coupling intensity distribution map through parallel processing of multiple machine learning algorithms, and outputs the inter-region coupling coefficient and confidence thermodynamic map to the collaborative compensation module; A collaborative compensation module generates a power compensation strategy based on the thermal coupling intensity distribution map and the confidence thermodynamic map, verifies the strategy convergence by calling historical data from the knowledge base, and outputs a power instruction sequence with phase offset to the early warning module; The early warning module converts the power instruction sequence into an execution signal, monitors the power output characteristics and temperature gradient changes in real time, and triggers an early warning signal when abnormal characteristics are detected and feeds back to the collaborative compensation module.

2. The display process temperature management optimization system according to claim 1, characterized in that: The data acquisition module includes: A multi-type temperature sensor array collects raw temperature data and inputs it into a data verification unit; the data verification unit performs: Splitting the original data into a simulation copy, an optimization copy, and a control copy; Input the simulation copy into the digital twin engine for dynamic simulation of heat conduction; The optimized copy is input to the filter processor to perform two-stage processing of wavelet denoising and Kalman filtering; The confidence labels are generated by comparing the differences between the three replica data and are appended to the temperature field matrix.

3. The display process temperature management optimization system according to claim 2, characterized in that: The data verification unit is configured to: When it is detected that the temperature difference between the simulation replica, the optimized replica, and the control replica exceeds a preset threshold, the control replica is calibrated using the simulation results output by the digital twin engine; The filtering processor performs a two-stage process of wavelet noise reduction and Kalman filtering to eliminate sensor drift errors and then update the optimized replica.

4. The display process temperature management optimization system according to claim 3, characterized in that: The dynamic modeling module is configured to: receiving the temperature field matrix with confidence labels, Run the following algorithms in parallel to process the temperature field matrix: The gradient boosting decision tree algorithm analyzes thermally coupled nonlinear features and outputs feature importance distribution; Long short-term memory networks predict time-varying crosstalk trends over multiple future control cycles; Adaptive particle swarm optimization algorithm dynamically adjusts the frequency of model weight updates; Robust least squares method for online correction of heat transfer function coefficients; The output results of each algorithm are integrated to generate the thermal coupling intensity distribution map.

5. The display process temperature management optimization system according to claim 4, characterized in that: The dynamic modeling module is configured to: When the confidence value of any area in the confidence heat map is lower than a preset threshold, a model reconstruction request signal is sent to the data acquisition module; The fault feature coding vector fed back by the early warning module is received, and the long short-term memory network weight is reconstructed by combining the fault feature coding vector and the latest temperature field matrix.

6. The display process temperature management optimization system according to claim 5, characterized in that: The collaborative compensation module is configured to: Retrieving matching historical optimization strategies from a knowledge base according to the partition confidence value levels of the confidence heat map; The initial power compensation amount is generated by adopting the retrieved history optimization strategy; Verifying the convergence boundary of the initial power compensation amount by Lyapunov exponent analysis; When the convergence boundary meets the stability condition, the compensation amount is decomposed into a power command sequence with phase offset according to the heater thermal response time constant.

7. The display process temperature management optimization system according to claim 6, characterized in that: The collaborative compensation module is configured to: When the temperature gradient change after the power instruction sequence with phase offset is executed is lower than the set threshold, the current compensation strategy parameters, the thermal coupling intensity distribution map and the compensation effect score are extracted to generate a feature vector; Storing the feature vector in a knowledge base to update the historical optimization strategy; When receiving the strategy freeze instruction from the early warning module, it switches to the reinforcement learning exploration mode to generate a new compensation strategy and overwrite the current strategy.

8. The display process temperature management optimization system according to claim 7, characterized in that: The early warning module includes: A power conversion unit receives the power instruction sequence and converts it into a PWM control signal and outputs it to the heater; Harmonic analysis unit collects PWM control signals in real time, performs fast Fourier transform, and outputs harmonic distortion rate; The temperature gradient monitoring unit receives the temperature field matrix and calculates the change of the spatial second-order derivative; The abnormality decision maker compares whether the combined value of the harmonic distortion rate and the spatial second-order derivative change exceeds a threshold.

9. The display process temperature management optimization system according to claim 8, characterized in that: The anomaly decider is configured to: When the combined value of the harmonic distortion rate and the change in the spatial second-order derivative exceeds a preset threshold, the hierarchical limit control is triggered to reduce the current power output; Extract the current harmonic spectrum characteristics, temperature gradient mutation area and equipment status parameters to generate fault feature coding vector; Feeding back the fault feature coding vector to a dynamic modeling module; A strategy freeze instruction is sent to the collaborative compensation module and the power ramp protection mechanism is activated to gradually reduce the heating power.

10. The display process temperature management optimization system according to claim 9, characterized in that: The confidence labels dynamically assign calculation weights to multiple machine learning algorithms in the dynamic modeling module; The confidence value hierarchy of the partitions of the confidence heat map controls the priority of the collaborative compensation module in calling the knowledge base strategy; The abnormal warning signal triggered by the warning module executes: Drive the knowledge base to update the compensation strategy feature vector, triggering the dynamic modeling module to reconstruct the long short-term memory network weights; The thermal crosstalk effect of multiple heating zones is suppressed by closing the data loop across modules.

Images