Display panel production line production supervision system based on artificial intelligence

Abstract

The invention relates to the technical field of display panel manufacturing, and particularly discloses a display panel production line production supervision system based on artificial intelligence, and the system comprises a data collection module which carries out the deployment according to a production line, so as to achieve the data collection; the data transmission module comprises edge computing equipment and is used for carrying out primary processing and data transmission on the original data acquired by the data acquisition equipment; the data processing center is used for preprocessing the transmitted original data and extracting key features from the preprocessed data; based on the extracted features, performing real-time analysis and evaluation on the production state by using a multi-modal fusion artificial intelligence model; the supervision decision module is used for establishing a grading early warning mechanism according to an analysis result; a scheduling strategy is set according to the early warning level; and the execution control module generates a control instruction according to the scheduling strategy. According to the invention, through multi-dimensional data acquisition and intelligent analysis, tiny defects and potential anomalies which are difficult to find by a traditional method can be detected.

Description

Technical Field

[0001] This invention relates to the field of display panel manufacturing technology, and more specifically to an artificial intelligence-based display panel production line monitoring system. Background Technology

[0002] Modern display panel manufacturing involves numerous precision processes, including glass substrate cutting, coating, photolithography, etching, lamination, and encapsulation. Even minor deviations in any step can lead to decreased product quality, or even a large number of defective products, severely impacting the company's economic benefits. Traditional display panel production line monitoring primarily relies on manual inspection and automated control with fixed parameters. Manual inspection suffers from low efficiency, strong subjectivity, and difficulty in detecting minute defects, and cannot achieve real-time monitoring of the production process. Automated control with fixed parameters lacks flexibility and struggles to adapt to environmental changes and equipment wear during production, resulting in a high rate of missed inspections and an inability to promptly detect potential quality problems and production anomalies. This approach is no longer sufficient to meet the current demands for high-precision, high-efficiency display panel production.

[0003] Market research and industry data statistics show that, in terms of defect detection, traditional monitoring methods result in a defect rate of approximately 8%-12% due to missed inspections on a production line producing 10,000 display panels per day. Furthermore, the efficiency of manual inspection is inversely proportional to production scale; in large-scale production scenarios, the efficiency of manual inspection drops sharply, making it difficult to keep up with production pace. In addition, traditional automated control systems are poorly adaptable to dynamic changes such as ambient temperature fluctuations (±2℃) and equipment wear (accuracy decreases by 5%-10% after 1000 hours of cumulative operation), leading to an average deviation of production parameters from standard ranges of 30-60 minutes, severely impacting product quality stability. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based production monitoring system for display panel production lines to solve the aforementioned technical problems.

[0005] The objective of this invention can be achieved through the following technical solutions: An artificial intelligence-based production monitoring system for display panel production lines mainly consists of the following modules: Data acquisition module: Multispectral industrial camera array: Covering multiple spectral ranges such as visible light (380-780nm), near-infrared (780-1100nm), and ultraviolet (200-400nm), it is equipped with 16-32 industrial cameras to form a ring or matrix distributed layout, ensuring all-round three-dimensional coverage of the display panel.

[0006] Let the spatial resolution of the camera array be ρ s(Unit: μm / pixel, representing the actual physical size of each pixel), with a temporal resolution of ρ. t (Unit: frames / second, representing the number of image frames acquired per second), its smallest detectable defect size d min The relationship with spatial resolution is d min =1 / ρ s The relationship between frame rate f, production speed v (unit: mm / s, representing the speed of the production line), and detection accuracy requirement p (unit: μm, representing the minimum feature size to be detected) is f = (v*p) / ρ s It can reach up to 1000 frames per second.

[0007] Microcracks as small as 0.1 μm can be detected in real time during the glass substrate cutting process; In the coating process, it can accurately identify coating thickness deviations (accuracy up to 0.1nm).

[0008] MEMS sensor network: integrates various MEMS sensors such as temperature (measurement range -40℃-200℃, accuracy ±0.1℃), pressure (measurement range 0-10MPa, accuracy ±0.01MPa), humidity (measurement range 0-100%RH, accuracy ±1%RH), vibration (acceleration measurement range ±50g, accuracy ±0.01g), and strain (measurement range 0-5000με, accuracy ±5με). The number of sensors is configured from 50 to 200 depending on the complexity of the production line.

[0009] Let x be the data sequence collected by each sensor. i (t)(i=1,2,...),x i (t) represents the sequence of measurements taken by the i-th sensor at time t, and the total amount of data collected by the sensor network within the time interval [t1, t2]. (D is the sum of the absolute values ​​of all sensor data over the time interval, representing the total amount of data). Through sensor networks, the operating status of key components of equipment (such as the precision transmission mechanism of lithography equipment) can be monitored in real time, and the probability of equipment failure can be predicted 72 hours in advance.

[0010] Acoustic monitoring system: Composed of an array of 8-16 high-sensitivity microphones, it collects sound signals during equipment operation. The sampling frequency of the sound signals is f. s =48kHz, sampling precision is 24-bit, the relationship between the minimum identifiable sound intensity change ΔI and the sampling precision is ΔI= (I max -I min ) / 2 b (where I) max , I minHere, b represents the maximum and minimum sound intensities, and b is the sampling precision in bits (b=24 here). In the packaging process, sound characteristic analysis can detect the wear condition of the ultrasonic welding head of the packaging equipment with an accuracy of 95%.

[0011] Data transmission module: A multi-layered, multi-redundant communication architecture is adopted to ensure reliable data transmission, building a low-latency, high-bandwidth, and highly secure data transmission channel. Edge computing nodes: Deployed on the production line, using high-performance edge computing devices such as NVIDIA Jetson AGX Orin, to perform preliminary processing and feature extraction on the collected raw data, reducing the amount of data transmission.

[0012] Let the original data size be D0 (unit: GB, original uncompressed data size), and the processed data size be D1 (unit: GB, compressed data size). The data compression ratio η = D1 / D0 (η < 1, representing the degree of data compression, typical value 1 / 10-1 / 20). The processing latency t of the edge computing node... e With data volume D0 and computing power C e (Unit: TOPS, trillion operations per second, characterizing computing device performance) The relationship is t e =D0 / C e The single-node processing latency is no more than 10ms.

[0013] Industrial-grade 5G / WiFi 6 hybrid network: Provides high-speed, low-latency data transmission channels. Network bandwidth B is dynamically configured from 1-10Gbps. The relationship between data transmission rate R and bandwidth, and transmission protocol efficiency ε (values ​​ranging from 0.7 to 0.9, characterizing the effective bandwidth utilization of the protocol) is R=Bε. Network latency t n Satisfy t n ≤t max (t) max The maximum allowable latency is set to 50ms. SD-WAN (Software-Defined Wide Area Network) technology is used to achieve intelligent switching and load balancing between 5G and WiFi 6 networks, with a network packet loss rate of P. loss Satisfy P loss ≤P threshold (P) threshold The packet loss rate threshold is set to 0.1%.

[0014] Quantum-encrypted transmission: Sensitive production data is encrypted using quantum key distribution (QKD) technology, with an encryption key length of l. k It is 256 bits, with encryption security S e The relationship with the key length is S e =2 lk (This represents the size of the key space and can resist quantum computing attacks.)

[0015] Data Processing Center: Built in the cloud or a private enterprise data center, it adopts a distributed computing architecture and includes the following core units: Data preprocessing unit: Employs multi-dimensional data cleaning algorithms to perform preprocessing operations such as cleaning, noise reduction, and normalization on the transmitted raw data. For image data, adaptive median filtering is used to remove salt-and-pepper noise. Let the probability of noise points be p. noise (Percentage of noise points in the image before filtering), probability p of noise points after filtering noise' Satisfy p noise' =p noise *(1-α) (α is the filtering coefficient, ranging from 0.6 to 0.9, representing the noise suppression strength). Histogram equalization is used to enhance image contrast. Grayscale values ​​are converted using the formula g(x,y)=(b'-a') / (ba)(f(x,y)-a)+a', where f(x,y) is the original image grayscale value, g(x,y) is the processed image grayscale value, [a,b] is the original grayscale value range, and [a',b'] is the grayscale value range after equalization.

[0016] For sensor data, a Kalman filter algorithm is used for noise reduction. Let the state estimation error covariance be P. k The prediction error covariance is P. k|k-1 It satisfies the Kalman filter recursive formula: predict: k|k-1 =F k k-1|k-1 +B k u k ( k|k-1 Let F be the predicted state value at time k. k Let B be the state transition matrix. k To control the input matrix, u k (for control input). P k|k-1 =F k P k-1|k-1 +Q k (Q) k (This is the process noise covariance matrix). Update: K k =P k|k-1 (H k P k|k-1 +R k ) -1(K) k For Kalman gain, H k Let R be the observation matrix. k (for the observation noise covariance matrix). k|k = k|k-1 +K k (z k -H k k|k-1 )(z k (The observation value at time k) P k =(IK k H k )P k|k-1 (I is the identity matrix).

[0017] Feature extraction unit: Utilizes multiple deep learning algorithms to extract key features from preprocessed data. A 3D Convolutional Neural Network (3DCNN) is used to process temporal image data. The input data dimension is T*H*W*C (T is the number of time frames, H is the image height, W is the image width, and C is the number of channels). After l convolutional layers, the output feature map dimension is T'*H'*W'*C'. The convolutional kernel size is k*k*k (the spatiotemporal 3D convolutional kernel size), and the stride is s (the convolutional movement step size). , , (H', W', and T' represent the spatial and temporal dimensions of the output feature map), and C' is determined by the convolutional layer settings. In defect detection during photolithography, 3DCNN can effectively extract the three-dimensional morphological features of photoresist patterns, improving the detection accuracy to 98%.

[0018] Using Graph Neural Networks (GNNs) to mine synergistic effects between devices, the graph structure has N nodes. node (Number of device nodes), number of edges is N edge (Number of edges connecting devices), the node feature vector is hi (feature vector of the i-th device), and after m layers of graph convolution, the node feature update formula is: (where N(i) is the set of neighboring nodes of node i, c) ij W is the normalization coefficient. (m) , Let be the weight matrix of the m-th layer, and σ be the activation function. Through GNN analysis, production anomalies caused by parameter mismatch between equipment can be predicted in advance, with a prediction accuracy of 92%.

[0019] The Transformer architecture is used to process long sequence sensor data. The input sequence length is L (the length of the sensor data time series). After passing through a multi-head attention mechanism, the output feature sequence length remains unchanged. The attention weights are calculated as follows: Attention(Q,K,V)=softmax(QK T / (where Q, K, and V are the query, key, and value matrices, and d) k (The dimension of the key is used to scale the dot product). In device failure prediction, Transformer can accurately capture long-term dependencies in sensor data, extending the failure prediction lead time to 7 days.

[0020] Intelligent Analysis Unit: Based on extracted features, it utilizes a multimodal fusion artificial intelligence model to perform real-time analysis and evaluation of production status. When the anomaly detection model employs an autoencoder, the reconstruction error is set to e. r (The error between the input data and the reconstructed data), when e r >E th (E) th A threshold (representing the threshold for anomaly detection) is used to classify a defect as an anomaly; the defect classification model is based on a deep learning algorithm, and the classification accuracy is set to A. c The confusion matrix is ​​M(M ii M is the number of correctly classified categories. ij (Number of misclassifications) ; The quality prediction model employs time series prediction algorithms such as LSTM and GRU, assuming the predicted value is... The actual value is y i Mean square error (n is the number of samples).

[0021] Through multimodal fusion, the system comprehensively evaluates the accuracy P. total =ω1P acc1 +ω2P acc2 +ω3P acc3 (Where ω1+ω2+ω3=1 is the weighting coefficient, with typical values ​​of ω1=0.3, ω2=0.4, ω3=0.3, P) acc1 、P acc2 、P acc3 (The accuracy of each modality assessment) can reach a comprehensive assessment accuracy of 99.5%.

[0022] Regulatory decision-making module: Based on the output of the intelligent analysis unit, corresponding regulatory decisions are generated using algorithms such as decision trees and reinforcement learning, constructing an intelligent decision support system: a tiered early warning mechanism. Let the severity index of the anomaly be S. a(Dimensionless, the larger the value, the more severe the anomaly), when S a When S1 ≥ S1, it is a Level 1 warning (emergency), triggering an immediate production line shutdown and activation of the emergency plan; when S2 ≤ S a <S1 indicates a Level 2 warning (important), automatically adjusting production parameters and notifying technical personnel; when S a When S2 < S2, it is a Level 3 warning (general), recording abnormal information and continuously monitoring (S1 and S2 are warning thresholds, S1 > S2). Dynamic scheduling strategy: When using reinforcement learning algorithm, let the state space be S (the set of all possible production states), the action space be A (the set of all possible control actions), the reward function be R(s,a) (the immediate reward for performing action a in state s), and the optimal policy π. * By Bellman equations: Solve for γ, where γ is the discount factor, ranging from 0.8 to 0.99, and P(s'|s,a) is the state transition probability.

[0023] In the event of equipment failure, the system can redistribute production tasks within 30 seconds, ensuring continuous operation of the production line.

[0024] Root Cause Analysis Engine: Employs a causal inference algorithm, assuming the causal relationship strength is C. r (Dimensionless, between 0 and 1, the larger the value, the stronger the causal relationship), when C r ≥C th (C) th A causal relationship is defined as a threshold of 0.7, which can quickly pinpoint the root cause of production anomalies, with an average root cause analysis time of no more than 5 minutes. 5. The execution control module converts the control commands generated by the monitoring decision module into specific operation signals, which are then sent to the execution equipment on the production line through control interfaces such as industrial buses (e.g., Profinet, EtherCAT) and PLCs, to achieve real-time control of the production process.

[0025] Execution control module: The control commands generated by the regulatory decision-making module are converted into specific operation signals and sent to the execution equipment on the production line through control interfaces such as industrial buses (e.g., Profinet, EtherCAT) and PLCs, thereby achieving real-time control of the production process. Equipment collaborative control: Let the control parameter of equipment i be p. i The objective function for the collaborative optimization of the parameters of each device is: (w) i f is the weighting coefficient. i (p i ( ) is the performance function of device i.

[0026] In the bonding process, by coordinating and controlling bonding pressure, temperature, and speed parameters, the bonding yield can be increased from 92% to 98%.

[0027] Adaptive control algorithm: Let the adjustment step size of the control parameter be Δp, and its relationship with the error e be Δp=αe, where α is the adjustment coefficient (with a value of 0.1-0.5), which can realize the dynamic optimization of production parameters.

[0028] Safety interlock mechanism: Let the danger detection signal be D. s When D s When =1, an emergency stop command is triggered, and the response time from the detection of danger to the complete shutdown of the equipment does not exceed 500ms.

[0029] The beneficial effects of this invention are as follows: Through multi-dimensional data collection and intelligent analysis, comprehensive and real-time monitoring of the production process is achieved, enabling the detection of minute defects and potential anomalies that are difficult to detect using traditional methods; regulatory decisions generated based on artificial intelligence algorithms can quickly respond to production anomalies, provide accurate handling suggestions, and reduce manual intervention; the decision response time is shortened compared to traditional methods, with the system's average decision response time being less than 10 seconds, significantly improving production efficiency; production quality is improved: through dynamic parameter optimization and equipment collaborative control, the product yield rate is increased, effectively reducing enterprise production costs; combining sensor data and deep learning algorithms, the probability of equipment failure can be predicted in advance, and production anomalies can be predicted in advance, realizing a shift from passive maintenance to proactive prevention and reducing downtime losses. Attached Figure Description

[0030] The invention will now be further described with reference to the accompanying drawings.

[0031] Figure 1 This is a schematic diagram of the production monitoring system for a display panel production line based on artificial intelligence, according to the present invention. Detailed Implementation

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0033] Please see Figure 1 As shown, this invention is a production monitoring system for display panel production lines based on artificial intelligence, which mainly consists of the following modules: Data acquisition module: Multispectral industrial camera array: Covering multiple spectral ranges such as visible light (380-780nm), near-infrared (780-1100nm), and ultraviolet (200-400nm), it is equipped with 16-32 industrial cameras to form a ring or matrix distributed layout, ensuring all-round three-dimensional coverage of the display panel.

[0034] Let the spatial resolution of the camera array be ρ s (Unit: μm / pixel, representing the actual physical size of each pixel), with a temporal resolution of ρ. t (Unit: frames / second, representing the number of image frames acquired per second), its smallest detectable defect size d min The relationship with spatial resolution is d min =1 / ρ s The relationship between frame rate f, production speed v (unit: mm / s, representing the speed of the production line), and detection accuracy requirement p (unit: μm, representing the minimum feature size to be detected) is f = (v*p) / ρ s It can reach up to 1000 frames per second.

[0035] Microcracks as small as 0.1 μm can be detected in real time during the glass substrate cutting process; In the coating process, it can accurately identify coating thickness deviations (accuracy up to 0.1nm).

[0036] MEMS sensor network: integrates various MEMS sensors such as temperature (measurement range -40℃-200℃, accuracy ±0.1℃), pressure (measurement range 0-10MPa, accuracy ±0.01MPa), humidity (measurement range 0-100%RH, accuracy ±1%RH), vibration (acceleration measurement range ±50g, accuracy ±0.01g), and strain (measurement range 0-5000με, accuracy ±5με). The number of sensors is configured from 50 to 200 depending on the complexity of the production line.

[0037] Let x be the data sequence collected by each sensor. i (t)(i=1,2,...),x i (t) represents the sequence of measurements taken by the i-th sensor at time t, and the total amount of data collected by the sensor network within the time interval [t1, t2]. (D is the sum of the absolute values ​​of all sensor data over the time interval, representing the total amount of data). Through sensor networks, the operating status of key components of equipment (such as the precision transmission mechanism of lithography equipment) can be monitored in real time, and the probability of equipment failure can be predicted 72 hours in advance.

[0038] Acoustic monitoring system: Composed of an array of 8-16 high-sensitivity microphones, it collects sound signals during equipment operation. The sampling frequency of the sound signals is f. s=48kHz, sampling precision is 24-bit, the relationship between the minimum identifiable sound intensity change ΔI and the sampling precision is ΔI= (I max -I min ) / 2 b (where I) max , I min Here, b represents the maximum and minimum sound intensities, and b is the sampling precision in bits (b=24 here). In the packaging process, sound characteristic analysis can detect the wear condition of the ultrasonic welding head of the packaging equipment with an accuracy of 95%.

[0039] Data transmission module: A multi-layered, multi-redundant communication architecture is adopted to ensure reliable data transmission, building a low-latency, high-bandwidth, and highly secure data transmission channel. Edge computing nodes: Deployed on the production line, using high-performance edge computing devices such as NVIDIA Jetson AGX Orin, to perform preliminary processing and feature extraction on the collected raw data, reducing the amount of data transmission.

[0040] Let the original data size be D0 (unit: GB, original uncompressed data size), and the processed data size be D1 (unit: GB, compressed data size). The data compression ratio η = D1 / D0 (η < 1, representing the degree of data compression, typical value 1 / 10-1 / 20). The processing latency t of the edge computing node... e With data volume D0 and computing power C e (Unit: TOPS, trillion operations per second, characterizing computing device performance) The relationship is t e =D0 / C e The single-node processing latency is no more than 10ms.

[0041] Industrial-grade 5G / WiFi 6 hybrid network: Provides high-speed, low-latency data transmission channels. Network bandwidth B is dynamically configured from 1-10Gbps. The relationship between data transmission rate R and bandwidth, and transmission protocol efficiency ε (values ​​ranging from 0.7 to 0.9, characterizing the effective bandwidth utilization of the protocol) is R=Bε. Network latency t n Satisfy t n ≤t max (t) max The maximum allowable latency is set to 50ms. SD-WAN (Software-Defined Wide Area Network) technology is used to achieve intelligent switching and load balancing between 5G and WiFi 6 networks, with a network packet loss rate of P. loss Satisfy P loss ≤P threshold (P) threshold The packet loss rate threshold is set to 0.1%.

[0042] Quantum-encrypted transmission: Sensitive production data is encrypted using quantum key distribution (QKD) technology, with an encryption key length of l. k It is 256 bits, with encryption security S e The relationship with the key length is S e =2 lk (This represents the size of the key space and can resist quantum computing attacks.)

[0043] Data Processing Center: Built in the cloud or a private enterprise data center, it adopts a distributed computing architecture and includes the following core units: Data preprocessing unit: Employs multi-dimensional data cleaning algorithms to perform preprocessing operations such as cleaning, noise reduction, and normalization on the transmitted raw data. For image data, adaptive median filtering is used to remove salt-and-pepper noise. Let the probability of noise points be p. noise (Percentage of noise points in the image before filtering), probability p of noise points after filtering noise' Satisfy p noise' =p noise *(1-α) (α is the filtering coefficient, ranging from 0.6 to 0.9, representing the noise suppression strength). Histogram equalization is used to enhance image contrast. Grayscale values ​​are converted using the formula g(x,y)=(b'-a') / (ba)(f(x,y)-a)+a', where f(x,y) is the original image grayscale value, g(x,y) is the processed image grayscale value, [a,b] is the original grayscale value range, and [a',b'] is the grayscale value range after equalization.

[0044] For sensor data, a Kalman filter algorithm is used for noise reduction. Let the state estimation error covariance be P. k The prediction error covariance is P. k|k-1 It satisfies the Kalman filter recursive formula: predict: k|k-1 =F k k-1|k-1 +B k u k ( k|k-1 Let F be the predicted state value at time k. k Let B be the state transition matrix. k To control the input matrix, u k (for control input). P k|k-1 =F k P k-1|k-1 +Q k (Q) k(This is the process noise covariance matrix). Update: K k =P k|k-1 (H k P k|k-1 +R k ) -1 (K) k For Kalman gain, H k Let R be the observation matrix. k (for the observation noise covariance matrix). k|k = k|k-1 +K k (z k -H k k|k-1 )(z k (The observation value at time k) P k =(IK k H k )P k|k-1 (I is the identity matrix).

[0045] Feature extraction unit: Utilizes multiple deep learning algorithms to extract key features from preprocessed data. A 3D Convolutional Neural Network (3DCNN) is used to process temporal image data. The input data dimension is T*H*W*C (T is the number of time frames, H is the image height, W is the image width, and C is the number of channels). After l convolutional layers, the output feature map dimension is T'*H'*W'*C'. The convolutional kernel size is k*k*k (the spatiotemporal 3D convolutional kernel size), and the stride is s (the convolutional movement step size). , , (H', W', and T' represent the spatial and temporal dimensions of the output feature map), and C' is determined by the convolutional layer settings. In defect detection during photolithography, 3DCNN can effectively extract the three-dimensional morphological features of photoresist patterns, improving the detection accuracy to 98%.

[0046] Using Graph Neural Networks (GNNs) to mine synergistic effects between devices, the graph structure has N nodes. node (Number of device nodes), number of edges is N edge (Number of edges connecting devices), the node feature vector is hi (feature vector of the i-th device), and after m layers of graph convolution, the node feature update formula is: (where N(i) is the set of neighboring nodes of node i, c) ijW is the normalization coefficient. (m) , Let be the weight matrix of the m-th layer, and σ be the activation function. Through GNN analysis, production anomalies caused by parameter mismatch between equipment can be predicted in advance, with a prediction accuracy of 92%.

[0047] The Transformer architecture is used to process long sequence sensor data. The input sequence length is L (the length of the sensor data time series). After passing through a multi-head attention mechanism, the output feature sequence length remains unchanged. The attention weights are calculated as follows: Attention(Q,K,V)=softmax(QK T / (where Q, K, and V are the query, key, and value matrices, and d) k (The dimension of the key is used to scale the dot product). In device failure prediction, Transformer can accurately capture long-term dependencies in sensor data, extending the failure prediction lead time to 7 days.

[0048] Intelligent Analysis Unit: Based on extracted features, it utilizes a multimodal fusion artificial intelligence model to perform real-time analysis and evaluation of production status. When the anomaly detection model employs an autoencoder, the reconstruction error is set to e. r (The error between the input data and the reconstructed data), when e r >E th (E) th A threshold (representing the threshold for anomaly detection) is used to classify a defect as an anomaly; the defect classification model is based on a deep learning algorithm, and the classification accuracy is set to A. c The confusion matrix is ​​M(M ii M is the number of correctly classified categories. ij (Number of misclassifications) ; The quality prediction model employs time series prediction algorithms such as LSTM and GRU, assuming the predicted value is... The actual value is y i Mean square error (n is the number of samples).

[0049] Through multimodal fusion, the system comprehensively evaluates the accuracy P. total =ω1P acc1 +ω2P acc2 +ω3P acc3 (Where ω1+ω2+ω3=1 is the weighting coefficient, with typical values ​​of ω1=0.3, ω2=0.4, ω3=0.3, P) acc1 、P acc2 、P acc3 (The accuracy of each modality assessment) can reach a comprehensive assessment accuracy of 99.5%.

[0050] Regulatory decision-making module: Based on the output of the intelligent analysis unit, corresponding regulatory decisions are generated using algorithms such as decision trees and reinforcement learning, constructing an intelligent decision support system: a tiered early warning mechanism. Let the severity index of the anomaly be S. a (Dimensionless, the larger the value, the more severe the anomaly), when S a When S1 ≥ S1, it is a Level 1 warning (emergency), triggering an immediate production line shutdown and activation of the emergency plan; when S2 ≤ S a <S1 indicates a Level 2 warning (important), automatically adjusting production parameters and notifying technical personnel; when S a When S2 < S2, it is a Level 3 warning (general), recording abnormal information and continuously monitoring (S1 and S2 are warning thresholds, S1 > S2). Dynamic scheduling strategy: When using reinforcement learning algorithm, let the state space be S (the set of all possible production states), the action space be A (the set of all possible control actions), the reward function be R(s,a) (the immediate reward for performing action a in state s), and the optimal policy π. * By Bellman equations: Solve for γ, where γ is the discount factor, ranging from 0.8 to 0.99, and P(s'|s,a) is the state transition probability.

[0051] In the event of equipment failure, the system can redistribute production tasks within 30 seconds, ensuring continuous operation of the production line.

[0052] Root Cause Analysis Engine: Employs a causal inference algorithm, assuming the causal relationship strength is C. r (Dimensionless, between 0 and 1, the larger the value, the stronger the causal relationship), when C r ≥C th (C) th A causal relationship is defined as a threshold of 0.7, which can quickly pinpoint the root cause of production anomalies, with an average root cause analysis time of no more than 5 minutes. 5. The execution control module converts the control commands generated by the monitoring decision module into specific operation signals, which are then sent to the execution equipment on the production line through control interfaces such as industrial buses (e.g., Profinet, EtherCAT) and PLCs, to achieve real-time control of the production process.

[0053] Execution control module: The control commands generated by the regulatory decision-making module are converted into specific operation signals and sent to the execution equipment on the production line through control interfaces such as industrial buses (e.g., Profinet, EtherCAT) and PLCs, thereby achieving real-time control of the production process. Equipment collaborative control: Let the control parameter of equipment i be p. i The objective function for the collaborative optimization of the parameters of each device is: (w) i f is the weighting coefficient. i (p i ( ) is the performance function of device i.

[0054] In the bonding process, by coordinating and controlling bonding pressure, temperature, and speed parameters, the bonding yield can be increased from 92% to 98%.

[0055] Adaptive control algorithm: Let the adjustment step size of the control parameter be Δp, and its relationship with the error e be Δp=αe, where α is the adjustment coefficient (with a value of 0.1-0.5), which can realize the dynamic optimization of production parameters.

[0056] Safety interlock mechanism: Let the danger detection signal be D. s When D s When =1, an emergency stop command is triggered, and the response time from the detection of danger to the complete shutdown of the equipment does not exceed 500ms.

[0057] This system was deployed in the intelligent monitoring of an 8.5-generation TFT-LCD production line, with the following specific applications: Data acquisition module: 24 multispectral industrial cameras (spatial resolution ρ) are configured in the glass substrate cutting process. s =0.1μm / pixel), real-time detection of 0.1μm level microcracks, frame rate f=500 frames / second (production speed v=200mm / s, detection accuracy p=2.5μm).

[0058] Eighty MEMS vibration sensors (accuracy ±0.01g) were deployed on the photolithography equipment, and the total data was analyzed using the formula... By monitoring vibration data of precision transmission mechanisms, the probability of guide rail wear failure can be predicted 72 hours in advance.

[0059] A 12-microphone array is deployed in the packaging process, using ΔI = (I max -I min ) / 2 24 The algorithm can identify wear on ultrasonic welding heads with an accuracy rate of 95%.

[0060] Data transmission module: The edge computing nodes use NVIDIA Jetson AGX Orin. The original data size D0 = 200GB, and the compressed data size D1 = 10GB, with a compression ratio η = 1 / 20 and a processing latency t. e =200 / 200=1ms (Computational capability C) e =200TOPS).

[0061] Industrial-grade 5G network bandwidth B=5Gbps, protocol efficiency ε=0.8, transmission rate R=4Gbps, latency t n=20ms < 50ms, which meets the real-time requirements.

[0062] Data Processing Center: Image preprocessing employs adaptive median filtering (α=0.8), reducing the probability of noise points from p{noise}=5% to p{noise}'=1%; sensor data undergoes Kalman filtering for noise reduction, resulting in a 60% reduction in state estimation error.

[0063] A 3DCNN was used to detect photoresist defects. The input data dimensions were T = 10 frames * H = 2000 pixels * W = 2000 pixels * C = 3 channels, the convolution kernel k = 3, the stride s = 1, and the output feature map dimensions were T' = 8 frames * H' = 1998 pixels * W' = 1998 pixels * C' = 64. The detection accuracy reached 98.2%.

[0064] Multimodal fusion evaluation accuracy P total =ω1P acc1 +ω2P acc2 +ω3P acc3 =0.3×98%+0.4×95%+0.3×99%=97.1%, which is improved to 99.5% after comprehensive optimization.

[0065] Regulatory decision-making and implementation: When the MEMS sensor detects the device vibration value x i When (t) > 3σ (σ is the standard deviation), a level 2 warning is triggered. The system dynamically adjusts the exposure power parameters through reinforcement learning and completes the redistribution of production tasks within 30 seconds to avoid downtime losses.

[0066] By optimizing pressure (p1), temperature (p2), and speed (p3) through collaborative control of the objective function F(p1,p2,p3)=min(0.5f1+0.3f2+0.2f3) in the bonding process, the bonding yield rate was increased from 92% to 98%.

[0067] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A production monitoring system for display panel production lines based on artificial intelligence, characterized in that, include: Data acquisition module: Deployed according to each key process node of the production line to achieve data acquisition, thereby constructing a multi-source heterogeneous data sensing network; Data transmission module: includes edge computing devices deployed on the production line, which perform preliminary processing on the raw data collected by the data acquisition devices, and transmit the processed raw data to the next level through a preset communication protocol; Data Processing Center: Preprocesses the transmitted raw data and extracts key features from the preprocessed data; based on the extracted features, it uses a multimodal fusion artificial intelligence model to perform real-time analysis and evaluation of the production status. Regulatory decision-making module: Based on the analysis results, establish a tiered early warning mechanism; and set scheduling strategies according to the early warning level; Execution control module: Generates control commands based on the scheduling strategy and sends them to the execution equipment on the production line.

2. The artificial intelligence-based display panel production line monitoring system according to claim 1, characterized in that, The data acquisition module includes a multispectral industrial camera array, a MEMS sensor network, and an acoustic monitoring system. The multispectral industrial camera array covers the visible, near-infrared, and ultraviolet spectral ranges. Multiple industrial cameras are configured in a distributed layout, and the minimum detectable defect size satisfies the formula: d min =1 / ρ s ; Where, d min For the minimum defect size, ρ s Spatial resolution; Frame rate satisfies the formula: ； Where f is the frame rate, v is the production speed, and p is the detection accuracy requirement.

3. The artificial intelligence-based display panel production line monitoring system according to claim 1, characterized in that, The data transmission module includes an edge computing node, an industrial-grade 5G / WiFi6 hybrid network, and a quantum-encrypted transmission link; The edge computing node compresses the original data with a compression ratio η = D1 / D0, and the processing latency satisfies t. e =D0 / C e Where D0 is the original data volume, D1 is the processed data volume, and C e For computing power; The data transmission rate of an industrial-grade 5G / WiFi 6 hybrid network satisfies R=Bε, where B is the network bandwidth and ε is the transmission protocol efficiency, with a value of 0.7-0.

9.

4. The artificial intelligence-based display panel production line monitoring system according to claim 1, characterized in that, The data processing center includes a data preprocessing unit, a feature extraction unit, and an intelligent analysis unit; The feature extraction unit employs a 3D convolutional neural network, graph neural network, and Transformer architecture, wherein the output feature map dimension of the 3D convolutional neural network is determined by... , Calculate H, W, and T, which are the height, width, and time dimension of the input image, respectively; k is the convolution kernel size; and s is the stride. The intelligent analysis unit uses a multimodal fusion model P total =ω1P acc1 +ω2P acc2 +ω3P acc3 The production status is assessed, where ω1+ω2+ω3=1 is the weighting coefficient, and P acc1 P acc2 P acc3 The accuracy of each modality is evaluated.

5. The artificial intelligence-based display panel production line monitoring system according to claim 1, characterized in that, The regulatory decision-making module includes a tiered early warning mechanism, a dynamic scheduling strategy, and a root cause analysis engine. The graded early warning mechanism is based on the severity index S of the anomaly. a Triggering different responses, when S a A level 1 warning is triggered when S1 ≥ S2, and when S2 ≤ S3. a A level 2 warning is triggered when S < S1, and when S a <S2 triggers a level 3 warning; The dynamic scheduling strategy employs a reinforcement learning algorithm, based on the Bellman equation. Solving for the optimal strategy , where γ is the discount factor, ranging from 0.8 to 0.99, P(s'|s,a) is the state transition probability, and R(s,a) is the reward function.

6. The artificial intelligence-based display panel production line monitoring system according to claim 1, characterized in that, The execution control module sends control commands through an industrial bus or PLC interface to achieve coordinated equipment control and adaptive parameter adjustment; The objective function for equipment collaborative control is: ； Where p i For equipment control parameters, w i f is the weighting coefficient. i (p i ) represents the device performance function; The adaptive control algorithm adjusts the parameters by Δp=αe, where Δp is the adjustment step size, α is the adjustment coefficient with a value of 0.1-0.5, and e is the error value.

7. The artificial intelligence-based display panel production line monitoring system according to claim 2, characterized in that, The MEMS sensor network integrates temperature, pressure, humidity, vibration, and strain sensors, and uses the formula... Calculate the total amount of data, where x i (t) represents the data sequence collected by the i-th sensor, [t1,t2] is the time interval, and n represents the number of sensors.

8. The artificial intelligence-based display panel production line monitoring system according to claim 2, characterized in that, The acoustic monitoring system consists of an array of high-sensitivity microphones, and the minimum sound intensity change ΔI = (I max -I min ) / 2 b , where I max , I min分别 denoted as the maximum and minimum sound intensities, and b represents the sampling precision in bits.

Images