Liquid crystal display screen production line operation management method and system

Abstract

The present application relates to the technical field of flexible production and industrial intelligent control, in particular to a liquid crystal display production line operation management method and system. The method comprises: time stamp alignment and denoising processing of multi-source heterogeneous data; inputting a physical state feature vector into a double-channel hybrid inference unit to generate a comprehensive prediction benchmark; calculating the difference between the physical state feature vector and the comprehensive prediction benchmark to generate an original residual vector; topological structure analysis of the original residual vector to generate a residual semantic signal; dynamic tolerance boundary drift judgment to generate corresponding dynamic operation mode instructions; in response to the dynamic operation mode instructions indicating benign line change fluctuation, an evolved cognitive model is constructed; in response to the dynamic operation mode instructions indicating malignant abnormalities, an adaptive closed-loop control from data space to physical space is constructed. The present application provides a reliable basis for subsequent differentiated regulation, ensuring the continuity and stability of the production line operation.

Description

Technical Field

[0001] This invention relates to the field of flexible manufacturing and industrial intelligent control technology, specifically to a method and system for the operation management of liquid crystal display production lines. Background Technology

[0002] In current LCD production lines, especially in flexible production environments with small batches and multiple varieties, the production lines are in a continuous state of non-steady operation for a long time. In order to ensure production quality, the system needs to collect multi-source heterogeneous data such as optical detection images, sensor timing data and equipment status parameters in real time. These data not only come from diverse sources, but also have different collection frequencies and transmission delays. If they are directly fused, it is easy to cause timing deviations and information misalignments, which can lead to false anomalies.

[0003] Existing solutions mostly rely on traditional static models trained on massive amounts of historical data for monitoring. However, under continuous non-steady-state operating conditions, such static models tend to fail because they cannot adapt to frequent process changes and environmental fluctuations. They are unable to distinguish between benign process fluctuations and malignant new defects, resulting in a conflict between the system's generalization ability and sensitivity, a high false alarm rate, and serious disruption to normal production efficiency. At the same time, traditional methods are highly dependent on training on known defect samples and cannot keenly capture novel anomalies that have never been seen before.

[0004] Therefore, how to construct a cognitive model that can self-evolve in a non-steady-state environment, effectively distinguish between benign drift and malignant anomalies, and achieve zero-sample detection of unknown defects has become an urgent technical problem to be solved.

[0005] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a method and system for managing the operation of a liquid crystal display production line. Specifically, the technical solution of this invention includes:

[0007] Please see Figure 1 Step S1: Collect multi-source heterogeneous data from the LCD production line, perform timestamp alignment and noise reduction on the multi-source heterogeneous data, and map the processed data into a physical state feature vector.

[0008] Step S2: Input the physical state feature vector into the dual-channel hybrid inference unit, generate a standard expected value using the static benchmark channel, generate a trend correction value using the dynamic trend channel, and fuse the standard expected value and the trend correction value to generate a comprehensive prediction benchmark.

[0009] Step S3: Calculate the difference between the physical state feature vector and the comprehensive prediction benchmark to generate the original residual vector; perform topological structure analysis on the original residual vector to analyze its clustering features and directionality, thereby generating residual semantic signals;

[0010] Step S4: Based on the residual semantic signal and the preset dynamic tolerance boundary, perform drift determination, determine the current production line status as benign line change fluctuation or malignant anomaly, and generate the corresponding dynamic operation mode instruction.

[0011] Step S5: In response to the dynamic operation mode command indicating benign line-changing fluctuations, activate the online evolution update unit, use the physical state feature vector to fine-tune the parameters of the dynamic trend channel, and construct the evolved cognitive model;

[0012] Step S6: In response to the dynamic operation mode command indicating a malignant anomaly, activate the process reverse control unit, perform physical intervention based on the original residual vector and the residual semantic signal, and construct an adaptive closed-loop control from the data space to the physical space.

[0013] Preferably, step S1 specifically includes:

[0014] The optical inspection images, sensor timing data and equipment status parameters collected from the production line constitute the multi-source heterogeneous data.

[0015] The multi-source heterogeneous data is timestamped to eliminate data timing deviations;

[0016] The aligned data is mapped to a high-dimensional space using a pre-trained feature extractor to extract the physical state feature vector.

[0017] Preferably, step S2 specifically includes:

[0018] The static benchmark channel uses a deep learning model trained on historical data to output the standard expected value representing the ideal steady-state process based on the physical state feature vector.

[0019] The dynamic trend channel uses a sliding window regression model to output a trend correction value representing the current process drift trend based on the historical state of the most recent N time steps.

[0020] The standard expected value and the trend correction value are weighted and calculated to generate the comprehensive forecast benchmark.

[0021] Preferably, step S3 specifically includes:

[0022] The numerical difference between the physical state feature vector and the comprehensive prediction benchmark is calculated to obtain the original residual vector;

[0023] Determine whether the original residual vector exhibits clustering features with consistent direction and compact density in the feature space;

[0024] If the clustering features are present, the residual semantic signal is labeled as structural drift semantics;

[0025] If the residual semantic signal exhibits characteristics of being discrete, disordered, or pointing to an unknown region, it is marked as an outlier anomalous semantic.

[0026] Preferably, step S4 specifically includes:

[0027] Obtain the dynamic tolerance boundary set based on expert experience and historical yield statistics;

[0028] When the residual semantic signal is a structural drift semantic and the drift amplitude does not exceed the dynamic tolerance boundary, it is determined to be a benign line-switching fluctuation.

[0029] When the residual semantic signal is an outlier anomalous semantic, or although it is structural drift semantic, the drift amplitude exceeds the dynamic tolerance boundary, it is determined to be a malignant anomaly.

[0030] Preferably, step S5 specifically includes:

[0031] Keep the weight parameters of the static reference channel unchanged;

[0032] The current physical state feature vector is marked as a positive sample;

[0033] The weight parameters of the dynamic trend channel are updated using an incremental learning algorithm, so that the prediction benchmark of the dynamic trend channel moves closer to the current physical state.

[0034] Preferably, step S6 specifically includes:

[0035] The fault handling strategy library is retrieved based on the residual semantic signals.

[0036] If the parameter misalignment type corresponding to the residual semantic signal is abnormal, extract the magnitude and direction sign of the original residual vector;

[0037] Based on the preset sensitivity relationship of process parameters, calculate the amount of physical parameter adjustment required to eliminate the original residual vector;

[0038] The physical parameter adjustment amount is converted into equipment control commands and sent to the production line actuator to perform quantitative reverse compensation.

[0039] A liquid crystal display production line operation management system, comprising:

[0040] The production line physical state perception and feature mapping unit is used to collect multi-source heterogeneous data from the LCD production line, perform timestamp alignment and noise reduction on the multi-source heterogeneous data, and map the processed data into a physical state feature vector.

[0041] The dual-channel hybrid inference unit is used to input the physical state feature vector into the dual-channel hybrid inference unit, generate a standard expected value using the static benchmark channel, generate a trend correction value using the dynamic trend channel, and fuse the standard expected value and the trend correction value to generate a comprehensive prediction benchmark.

[0042] The residual information semantic parsing unit is used to calculate the difference between the physical state feature vector and the comprehensive prediction benchmark to generate the original residual vector; and to perform topological structure analysis on the original residual vector to analyze its clustering features and directionality, thereby generating the residual semantic signal.

[0043] The drift determination and routing distribution unit is used to determine the drift based on the residual semantic signal and the preset dynamic tolerance boundary, determine the current production line status as benign line change fluctuation or malignant anomaly, and generate the corresponding dynamic operation mode instruction.

[0044] An online evolution update unit is used to respond to a dynamic operating mode command indicating benign line-switching fluctuations, activate the online evolution update unit, use the physical state feature vector to fine-tune the parameters of the dynamic trend channel, and construct an evolved cognitive model.

[0045] The process reverse control unit is used to respond to dynamic operating mode commands indicating malignant anomalies, activate the process reverse control unit, perform physical intervention based on the original residual vector and the residual semantic signal, and construct adaptive closed-loop control from data space to physical space.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This invention achieves accurate qualitative analysis of production line status by introducing a dual-channel hybrid inference unit and semantic parsing of residual information. Unlike traditional methods that struggle to distinguish between process switching and actual faults, this invention can accurately identify benign line changeover fluctuations and malignant anomalies through topology analysis and dynamic tolerance boundaries. This differentiation mechanism significantly reduces false alarms and unnecessary downtime caused by process switching, avoids frequent production line interventions, provides a reliable basis for subsequent differentiated control, and ensures the continuity and stability of production line operation.

[0048] 2. This invention designs an online evolutionary update unit for benign process changeover fluctuations, enabling the model to adaptively learn. When the system determines that a normal process changeover or benign drift has occurred on the production line, this unit is activated, using the current physical state feature vector to fine-tune the parameters of the dynamic trend channel. This allows the cognitive model to autonomously adapt to new stable process states without the need for costly offline global model retraining. This achieves online self-evolution of the model, greatly improving the system's rapid response and adaptability to new products and processes.

[0049] 3. This invention constructs an adaptive closed-loop control system from the data space to the physical space to address malignant anomalies, achieving rapid, accurate, and quantitative correction of faults. When a malignant anomaly is detected, the process reverse control unit is activated. This is not merely a simple alarm; it automatically calculates and executes quantitative adjustments to physical parameters based on the value, direction, and semantics of the original residual vector, searching the strategy library and combining it with the sensitivity relationships of process parameters. This reverse compensation mechanism ensures immediate and accurate correction of process misalignments, significantly improving the production line's autonomy and yield control level.

[0050] 4. This invention constructs a highly robust comprehensive prediction benchmark through deep fusion of multi-source heterogeneous data and dual-channel inference. It solves the challenges of timestamp alignment and feature mapping for production line optical inspection images and sensor time-series data. It innovatively integrates a static benchmark channel representing an ideal steady state and a dynamic trend channel reflecting real-time process drift. This design ensures that the prediction benchmark maintains anchorage to standard processes while also considering short-term trend corrections. Its accuracy and reliability far exceed those of a single model, laying a solid foundation for subsequent precise residual analysis. Attached Figure Description

[0051] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0052] Figure 1 This is a flowchart of the method of the present invention.

[0053] Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0055] Example 1

[0056] The LCD production line operation management method includes: Step S1: Collect multi-source heterogeneous data from the LCD production line, perform timestamp alignment and noise reduction on the multi-source heterogeneous data, and map the processed data into a physical state feature vector.

[0057] Step S2: Input the physical state feature vector into the dual-channel hybrid inference unit, use the static benchmark channel to generate the standard expected value, use the dynamic trend channel to generate the trend correction value, and fuse the standard expected value and the trend correction value to generate a comprehensive prediction benchmark.

[0058] Step S3: Calculate the difference between the physical state feature vector and the comprehensive prediction benchmark to generate the original residual vector; perform topological structure analysis on the original residual vector to analyze its clustering features and directionality, thereby generating the residual semantic signal;

[0059] Step S4: Based on the residual semantic signal and the preset dynamic tolerance boundary, perform drift judgment, determine the current production line status as benign line change fluctuation or malignant anomaly, and generate the corresponding dynamic operation mode instruction.

[0060] Step S5: In response to the dynamic operation mode command indicating benign line-changing fluctuations, activate the online evolution update unit, use the physical state feature vector to fine-tune the parameters of the dynamic trend channel, and construct the evolved cognitive model;

[0061] Step S6: In response to the dynamic operation mode command indicating a malignant anomaly, activate the process reverse control unit, perform physical intervention based on the original residual vector and residual semantic signal, and construct an adaptive closed-loop control from the data space to the physical space.

[0062] To address the failure of traditional static models caused by the continuous non-steady-state operation of production lines in flexible, small-batch, multi-variety production environments, this embodiment constructs an adaptive management mechanism based on residual-driven mechanisms. Sensing devices deployed at key nodes of the production line capture real-time basic data reflecting the production status. The physical state feature vector is a high-dimensional data set that has undergone dimensionality reduction and abstraction. The aim is to unify scattered optical, electrical, and mechanical signals with different dimensions into a mathematical representation that can be processed by a computer. It originates from the output of preprocessing and feature extraction of multi-source heterogeneous data. Abandoning the single-model prediction method, it adopts a dual-channel architecture to capture long-term steady-state laws and short-term drift trends respectively, generating a comprehensive prediction benchmark that includes both historical experience and current trends. ;

[0063] The system executes core sensing logic, without directly comparing the absolute error between the measured value and the standard value, and calculates the physical state feature vector. Compared with the comprehensive forecast benchmark The difference between them, i.e., the original residual vector This vector not only represents numerical deviation but also implies the physical root cause of the deviation. Topological analysis of this residual vector distinguishes whether the deviation originates from systematic drift or sudden anomalies, thereby generating a residual semantic signal. ;

[0064] Decisions are made using preset boundary conditions to accurately characterize the production line status; when it is determined to be a benign fluctuation, an online evolution mechanism is triggered to update the model’s understanding of the trend using current data, achieving positive coupling between perception and evolution; when it is determined to be a malignant anomaly, physical intervention is triggered to directly adjust the production line equipment parameters, achieving reverse coupling between perception and control.

[0065] A closed-loop system is constructed that includes perception, reasoning, analysis, judgment, evolution, and counter-control. It abandons the traditional binary opposition between model training and use and regards the residual between predicted values ​​and physical measured values ​​as the core information source. This mechanism enables the system to effectively distinguish between benign process fluctuations and malignant new defects without relying on a large number of defect samples for training, reduces the false alarm rate in HMLV mode, and achieves keen capture of unknown novel anomalies.

[0066] Example 2

[0067] Step S1 specifically includes:

[0068] Collect optical inspection images, sensor timing data and equipment status parameters from the production line to form multi-source heterogeneous data;

[0069] Timestamp alignment is performed on multi-source heterogeneous data to eliminate data timing deviations;

[0070] The aligned data is mapped to a high-dimensional space using a pre-trained feature extractor to extract physical state feature vectors.

[0071] To address the issue of false anomalies caused by asynchronous data sources, the production line physical state perception and feature mapping unit collects data from all dimensions, including optical inspection images reflecting product appearance quality, sensor time-series data reflecting the environment and process, and equipment status parameters reflecting equipment operational health. Given the varying collection frequencies and transmission delays of these data, direct fusion can lead to information misalignment. Therefore, strict timestamp alignment is performed before feature extraction, using the reference clock of a master device. Using this as a benchmark, eliminate timing biases between data. ;

[0072] By invoking a pre-trained feature extractor, such as a convolutional neural network (CNN) for optical images or an autoencoder for temporal data, the aligned raw physical data is mapped to a high-dimensional feature space, outputting a physical state feature vector. ;in Each dimension corresponds to the key process indicators of the production line, such as the estimated film thickness and surface defect density, thus preserving physical interpretability;

[0073] By aligning timestamps and mapping high dimensions, the spatiotemporal barriers between multi-source heterogeneous data are effectively eliminated, transforming the complex fluctuations of the physical world into a unified mathematical expression that computers can understand, providing pure and semantically rich high-quality input for subsequent high-precision inference and residual analysis.

[0074] Example 3

[0075] Step S2 specifically includes:

[0076] The static benchmark channel uses a deep learning model trained on historical data to output a standard expected value representing the ideal steady-state process based on the physical state feature vector.

[0077] The dynamic trend channel uses a sliding window regression model to output a trend correction value representing the current process drift trend based on the historical state of the most recent N time steps.

[0078] A comprehensive forecast benchmark is generated by weighting the standard expected value and the trend correction value.

[0079] The dual-channel hybrid inference unit resolves the contradiction between steady-state assumptions and unsteady-state environments; the static baseline channel serves as the system's ballast, utilizing a deep learning model trained offline on massive historical big data, based on the input feature vector. The output represents the standard expected value that should be presented under ideal steady-state process conditions. The dynamic trend channel acts as a regulator for the system, utilizing a lightweight sliding window regression model that focuses only on the most recent... A historical state with a time step;

[0080] parameter This is a quantitative indicator used to define the reference range for short-term trends. The value is based on empirical values ​​set according to the production line changeover frequency and process stabilization cycle; the trend correction value output by this channel... Reflecting the drift trend of current process parameters relative to historical averages; using a weighted logic formula. By integrating the two, a comprehensive forecasting benchmark can be generated. ,in These are weighting coefficients; in some implementations, These are fixed values ​​obtained through offline optimization based on historical data, for example, through grid search on historical validation datasets. exist The value of the interval is selected to make the comprehensive forecast benchmark. With physical state feature vector The one with the smallest mean square error Values, or variables that are dynamically adjusted based on the current production line status;

[0081] The hybrid architecture of static benchmark and dynamic trend not only retains the high recognition accuracy of deep models for known patterns, but also gives the system the flexibility to capture short-term process drift. The prediction benchmark is no longer a rigid fixed value, but a reference system that can dynamically float with the production line status, which greatly reduces misjudgments caused by normal process fluctuations.

[0082] Example 4

[0083] Step S3 specifically includes:

[0084] The numerical difference between the physical state feature vector and the comprehensive prediction benchmark is calculated to obtain the original residual vector;

[0085] Determine whether the original residual vector exhibits a clustering feature with consistent direction and compact density in the feature space;

[0086] If clustering characteristics are observed, the residual semantic signal is labeled as structural drift semantics;

[0087] If the residual semantic signal exhibits characteristics of being discrete, disordered, or pointing to unknown regions, it is labeled as an outlier semantic signal.

[0088] The residual information semantic parsing unit deciphers the physical essence beyond the data appearance; it calculates the difference between the measured features and the prediction benchmark to generate the original residual vector. Perform topological analysis on the vector, for example, using density-based clustering algorithms such as DBSCAN or manifold learning methods such as t-SNE to analyze its distribution characteristics and examine the geometric distribution of the residuals in the feature space.

[0089] The residual vectors exhibit clustering characteristics with consistent direction and compact density, implying that the biases are systematic and regular, which can be interpreted as structural drift semantics. This corresponds to benign process changes; the residual vector is discrete, disordered, or points to unknown blank areas in the feature space, meaning that the deviation violates known physical laws, which can be interpreted as outlier anomaly semantics. This corresponds to sudden failures or defect patterns never seen before.

[0090] The topological structure of the residuals is analyzed to achieve zero-sample detection of unknown and novel anomalies. Whenever a defect that has never been seen before appears, its residuals will inevitably disrupt the original structural rules, thereby issuing an alarm. There is no need to collect samples of this type of defect for training in advance, thus completely solving the problem of OOD detection.

[0091] Example 5

[0092] Step S4 specifically includes:

[0093] Obtain a dynamic tolerance boundary based on expert experience and historical yield statistics;

[0094] When the residual semantic signal is structural drift semantic and the drift amplitude does not exceed the dynamic tolerance boundary, it is judged as benign line-switching fluctuation.

[0095] When the residual semantic signal is an outlier anomalous semantic, or although it is a structural drift semantic, the drift amplitude exceeds the dynamic tolerance boundary, it is judged as a malignant anomaly.

[0096] The drift determination and routing distribution unit establishes precise decision-making logic; a dynamic tolerance boundary is introduced. The threshold index used to distinguish between normal fluctuations and potential risks was determined after numerical synthesis of prior knowledge from process experts and statistical analysis of historical yield data from the production line; the residual was identified as structural drift semantics. And its amplitude The current state is determined to be a benign line changeover fluctuation, and the deviation is within an acceptable range for process adjustments; where, the symbol The Euclidean norm of the vector is used to quantify the overall magnitude or strength of the original residual vector in the feature space; the residual is identified as an outlier semantic. Or, although there is a regular drift, its amplitude... It was decisively determined to be a malignant abnormality;

[0097] The judgment logic effectively resolves the conflict between generalization ability and sensitivity; it is tolerant of benign fluctuations to avoid false alarms, and highly sensitive to malignant anomalies to prevent missed detections; the hierarchical processing mechanism ensures that the production line management system maintains a high yield rate while not interfering with production efficiency due to frequent false alarms.

[0098] Example 6

[0099] Step S5 specifically includes:

[0100] Keep the weight parameters of the static baseline channel unchanged;

[0101] Mark the current physical state feature vector as a positive sample;

[0102] The weight parameters of the dynamic trend channel are updated using an incremental learning algorithm, so that the prediction benchmark of the dynamic trend channel is closer to the current physical state.

[0103] The online evolutionary update unit enables the model to quickly adapt to new process environments; when the current state is determined to be a benign fluctuation, retraining the entire model is not preferred, and a fine-tuning strategy is adopted; the weight parameters of the static baseline channel are forcibly locked. This ensures that the model remains unchanged, preventing catastrophic forgetting and avoiding the loss of historical general rules when adapting to new environments; the current physical state feature vector is used to... As new positive samples, the incremental learning algorithm is used to update only the weight parameters of the dynamic trend channel. The prediction baseline of the dynamic channel rapidly converges to the current physical state.

[0104] Selective online evolution rapidly absorbs process changes brought about by line changes and transforms them into new knowledge for the model; the residuals generated in the next inference are reduced, eliminating persistent false alarms caused by normal process fluctuations at the source, and realizing the self-iteration and upgrading of the system's cognition.

[0105] Example 7

[0106] Step S6 specifically includes:

[0107] Fault handling strategy library is retrieved based on residual semantic signals;

[0108] If the parameter misalignment type corresponding to the residual semantic signal is abnormal, extract the magnitude and direction sign of the original residual vector;

[0109] Based on the preset sensitivity relationship of process parameters, calculate the amount of physical parameter adjustment required to eliminate the original residual vector;

[0110] The physical parameter adjustment is converted into equipment control commands and sent to the production line actuators to perform quantitative reverse compensation.

[0111] The process reverse control unit constructs a closed loop from digital analysis to physical action; when a malignant anomaly is determined, it is based on the residual semantic signal. The system retrieves the preset fault handling strategy library; if the anomaly is identified as a parameter misalignment type, quantitative reverse compensation is executed; if the anomaly is identified as a hardware failure or a new type of defect that cannot be repaired online, the system activates the alarm unit and generates a shutdown or manual intervention instruction.

[0112] Extracting the original residual vector The magnitude and direction sign of the value; sensitivity relationship using preset process parameters. The model can be regressed using historical data. The specific steps include: Step a: During the production line debugging phase, the physical parameters are... Apply a series of small, known perturbations Step b: Synchronously acquire the residual vector caused by the disturbance. Step c: Collect N sets Data pairs; Step d: Fit the data using the least squares method and solve... Sensitivity matrix in ;

[0113] Should That is A linear implementation can be pre-constructed through mechanistic simulation or experimental design methods, i.e., a mapping model describing how changes in physical parameters affect eigenvectors, which is then solved through numerical optimization or backpropagation algorithms. For example, in For the above linear matrix In this case, the specific solution steps are as follows: by calculating pseudo-reversal The physical parameter adjustment amount is obtained. This adjustment amount can make Maximum offset Physical parameter adjustment amount This adjustment is then converted into specific equipment control commands and directly sent to the production line actuators.

[0114] This mechanism not only identifies problems but also solves them automatically. Quantitative reverse compensation forcibly pulls the deviated process state back to a steady state before defects cause batch scrap, improving the yield and automation control level of the production line and realizing the physical implementation of industrial artificial intelligence.

[0115] Example 8

[0116] Please see Figure 2 The production line physical state perception and feature mapping unit is used to collect multi-source heterogeneous data from the LCD production line, perform timestamp alignment and noise reduction on the multi-source heterogeneous data, and map the processed data into physical state feature vectors.

[0117] The dual-channel hybrid inference unit is used to input the physical state feature vector into the dual-channel hybrid inference unit, generate standard expected values ​​using the static benchmark channel, generate trend correction values ​​using the dynamic trend channel, and fuse the standard expected values ​​and trend correction values ​​to generate a comprehensive prediction benchmark.

[0118] The residual information semantic parsing unit is used to calculate the difference between the physical state feature vector and the comprehensive prediction benchmark to generate the original residual vector; the original residual vector is subjected to topological structure analysis to analyze its clustering features and directionality, thereby generating the residual semantic signal.

[0119] The drift determination and routing distribution unit is used to determine the drift based on the residual semantic signal and the preset dynamic tolerance boundary, determine the current production line status as benign line change fluctuation or malign abnormality, and generate the corresponding dynamic operation mode instruction.

[0120] The online evolution update unit is activated in response to dynamic operating mode commands indicating benign line-changing fluctuations. It uses physical state feature vectors to fine-tune the parameters of the dynamic trend channel and constructs an evolved cognitive model.

[0121] The process reverse control unit is used to respond to dynamic operating mode commands indicating malignant anomalies. The process reverse control unit is activated and physical intervention is performed based on the original residual vector and residual semantic signals to construct an adaptive closed-loop control from the data space to the physical space.

[0122] This invention also provides a liquid crystal display production line operation management system, in which each functional unit works closely together to implement the aforementioned method at the hardware level; specifically, the production line physical state perception and feature mapping unit integrates a high-precision industrial camera, sensor array and edge computing gateway to ensure full capture and standardization of physical data; the dual-channel hybrid inference unit is deployed in a high-performance computing server, internally running a deep neural network model and a lightweight time series regression model in parallel, achieving millisecond-level benchmark prediction through computing power support;

[0123] The residual information semantic parsing unit performs complex vector space calculations, transforming tedious numerical differences into semantic labels with physical meaning; the drift determination and routing distribution unit acts as a command and dispatch center, intelligently switching the system's operational flow based on a pre-set expert rule base and real-time calculation results; the online evolution update unit has online training capabilities, dynamically updating the model's local parameters without interrupting production, maintaining the freshness of the system's cognition; the process reverse control unit is connected to the production line's physical equipment via an industrial bus, transforming digital decisions into physical actions, completing the last mile of closed-loop control;

[0124] The units are deeply coupled to jointly construct an intelligent agent with complete capabilities of perception, cognition, decision-making, evolution and control; it solves the problem of continuous non-steady state faced by LCD production lines under HMLV mode, and significantly reduces false alarm rate and improves production yield by distinguishing between benign drift and malignant anomaly, while taking into account the sensitivity and robustness of detection.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for managing the operation of a liquid crystal display production line, characterized in that, include: Step S1: Collect multi-source heterogeneous data from the LCD production line, perform timestamp alignment and noise reduction on the multi-source heterogeneous data, and map the processed data into a physical state feature vector; Step S2: Input the physical state feature vector into the dual-channel hybrid inference unit, generate a standard expected value using the static benchmark channel, generate a trend correction value using the dynamic trend channel, and fuse the standard expected value and the trend correction value to generate a comprehensive prediction benchmark. Step S3: Calculate the difference between the physical state feature vector and the comprehensive prediction benchmark to generate the original residual vector; perform topological structure analysis on the original residual vector to analyze its clustering features and directionality, thereby generating residual semantic signals; Step S4: Based on the residual semantic signal and the preset dynamic tolerance boundary, perform drift determination, determine the current production line status as benign line change fluctuation or malignant anomaly, and generate the corresponding dynamic operation mode instruction. Step S5: In response to the dynamic operation mode command indicating benign line-changing fluctuations, activate the online evolution update unit, use the physical state feature vector to fine-tune the parameters of the dynamic trend channel, and construct the evolved cognitive model; Step S6: In response to the dynamic operation mode command indicating a malignant anomaly, activate the process reverse control unit, perform physical intervention based on the original residual vector and the residual semantic signal, and construct an adaptive closed-loop control from the data space to the physical space. Step S2 specifically includes: The static benchmark channel uses a deep learning model trained on historical data to output the standard expected value representing the ideal steady-state process based on the physical state feature vector. The dynamic trend channel uses a sliding window regression model to output a trend correction value representing the current process drift trend based on the historical state of the most recent N time steps. The standard expected value and the trend correction value are weighted and calculated to generate the comprehensive forecast benchmark.

2. The liquid crystal display production line operation management method according to claim 1, characterized in that, Step S1 specifically includes: The optical inspection images, sensor timing data and equipment status parameters collected from the production line constitute the multi-source heterogeneous data. The multi-source heterogeneous data is timestamped to eliminate data timing deviations; The aligned data is mapped to a high-dimensional space using a pre-trained feature extractor to extract the physical state feature vector.

3. The liquid crystal display production line operation management method according to claim 1, characterized in that, Step S3 specifically includes: The numerical difference between the physical state feature vector and the comprehensive prediction benchmark is calculated to obtain the original residual vector; Determine whether the original residual vector exhibits clustering features with consistent direction and compact density in the feature space; If the clustering features are present, the residual semantic signal is labeled as structural drift semantics; If the residual semantic signal exhibits characteristics of being discrete, disordered, or pointing to an unknown region, it is marked as an outlier anomalous semantic.

4. The liquid crystal display production line operation management method according to claim 1, characterized in that, Step S4 specifically includes: Obtain the dynamic tolerance boundary set based on expert experience and historical yield statistics; When the residual semantic signal is a structural drift semantic and the drift amplitude does not exceed the dynamic tolerance boundary, it is determined to be a benign line-switching fluctuation. When the residual semantic signal is an outlier anomalous semantic, or although it is structural drift semantic, the drift amplitude exceeds the dynamic tolerance boundary, it is determined to be a malignant anomaly.

5. The liquid crystal display production line operation management method according to claim 1, characterized in that, Step S5 specifically includes: Keep the weight parameters of the static reference channel unchanged; The current physical state feature vector is marked as a positive sample; The weight parameters of the dynamic trend channel are updated using an incremental learning algorithm, so that the prediction benchmark of the dynamic trend channel moves closer to the current physical state.

6. The liquid crystal display production line operation management method according to claim 1, characterized in that, Step S6 specifically includes: The fault handling strategy library is retrieved based on the residual semantic signals. If the parameter misalignment type corresponding to the residual semantic signal is abnormal, extract the magnitude and direction sign of the original residual vector; Based on the preset sensitivity relationship of process parameters, calculate the amount of physical parameter adjustment required to eliminate the original residual vector; The physical parameter adjustment amount is converted into equipment control commands and sent to the production line actuator to perform quantitative reverse compensation.

7. A liquid crystal display production line operation management system, characterized in that, The method for implementing the liquid crystal display production line operation management method as described in any one of claims 1 to 6 includes: The production line physical state perception and feature mapping unit is used to collect multi-source heterogeneous data from the LCD production line, perform timestamp alignment and noise reduction on the multi-source heterogeneous data, and map the processed data into a physical state feature vector. The dual-channel hybrid inference unit is used to input the physical state feature vector into the dual-channel hybrid inference unit, generate a standard expected value using the static benchmark channel, generate a trend correction value using the dynamic trend channel, and fuse the standard expected value and the trend correction value to generate a comprehensive prediction benchmark. The residual information semantic parsing unit is used to calculate the difference between the physical state feature vector and the comprehensive prediction benchmark to generate the original residual vector; and to perform topological structure analysis on the original residual vector to analyze its clustering features and directionality, thereby generating the residual semantic signal. The drift determination and routing distribution unit is used to determine the drift based on the residual semantic signal and the preset dynamic tolerance boundary, determine the current production line status as benign line change fluctuation or malignant anomaly, and generate the corresponding dynamic operation mode instruction. An online evolution update unit is used to respond to a dynamic operating mode command indicating benign line-changing fluctuations, activate the online evolution update unit, use the physical state feature vector to fine-tune the parameters of the dynamic trend channel, and construct an evolved cognitive model. The process reverse control unit is used to respond to the dynamic operation mode command indicating a malignant anomaly, activate the process reverse control unit, perform physical intervention based on the original residual vector and the residual semantic signal, and construct an adaptive closed-loop control from the data space to the physical space.

Images