Glass deep processing full-process line connection system and control method thereof

Abstract

The invention relates to the technical field of glass deep processing automation control, and discloses a glass deep processing full-process connection system and a control method thereof.The glass deep processing full-process connection method comprises the steps that multiple identification processing is conducted on cut glass, and an identification mapping relation is established; performing quality detection on the glass, and performing fusion processing on sensor data through a weighted average algorithm; according to the state of the edge grinding machine, a greedy algorithm based on a heuristic rule is adopted to formulate a wafer matching scheme; decomposing a sheet arrangement problem into sub-problems of size matching, temperature compatibility and time scheduling for processing, and controlling the temperature of the toughening furnace by adopting a regional proportional integral differential controller; the method comprises the following steps: establishing a unified data model by adopting an OPC UA standard architecture, and deploying edge computing nodes to pre-process data; predicting and analyzing the health state of the equipment by adopting a machine learning algorithm; according to the invention, the continuous control of the whole process from cutting to tempering of the glass is realized, and the production efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of automated control technology for deep glass processing, and more specifically, to a complete in-line system for deep glass processing and its control method. Background Technology

[0002] With the rapid development of the glass deep processing industry and the continuous improvement of market demand, traditional glass production models can no longer meet the requirements of modern production. Currently, in glass deep processing production lines, key equipment such as cutting machines, edging machines, and tempering furnaces typically operate independently, requiring glass products to undergo multiple manual handling and loading / unloading operations between processes. This discrete production model is not only inefficient but also presents numerous technical challenges. Especially in the production of high-end products such as automotive glass, higher demands are placed on product quality, production efficiency, and cost control; traditional production methods have become a bottleneck restricting enterprise development.

[0003] In terms of production efficiency, glass products require frequent manual handling between cutting, edging, and tempering processes, severely limiting production cycle time. Manual handling typically takes 2-3 minutes per operation and poses safety hazards. Regarding product quality control, the frequent manual contact and mechanical handling result in a high rate of surface scratches, significantly impacting product quality. Traditional quality inspection methods rely primarily on visual inspection, which is inaccurate and inefficient, failing to achieve comprehensive quality monitoring. In terms of resource utilization, the inability to dynamically optimize glass cutting schemes based on the technical parameters of the edging machine or tempering furnace leads to significant waste of scrap materials. Tempering furnace sheet arrangement relies heavily on manual experience, lacking scientific optimization algorithms, resulting in insufficient loading rates and high energy consumption.

[0004] Furthermore, the lack of effective information exchange and coordinated control mechanisms between various devices makes real-time monitoring and dynamic adjustment of equipment status impossible. When one piece of equipment malfunctions or experiences performance fluctuations, it is difficult to adjust the operating parameters of other equipment in a timely manner, affecting the stability of the entire production line. Simultaneously, the general lack of intelligent sheet storage functionality prevents the temporary storage and automatic recall of glass products, resulting in insufficient production line flexibility and an inability to respond to order changes and equipment maintenance needs. The control system primarily employs traditional control methods, lacking AI-based decision-making and optimization functions. It cannot automatically optimize and adjust production parameters based on historical data and real-time status, resulting in a high proportion of manual intervention throughout the process, severely hindering the improvement of automation levels. Therefore, a fundamental transformation of the production model through technological innovation is urgently needed. Summary of the Invention

[0005] This invention provides a complete glass deep processing line system and its control method, which solves the technical problems in related technologies such as the lack of effective information interaction and coordination control mechanisms between various devices, and the inability to achieve real-time monitoring and dynamic adjustment of equipment status.

[0006] This invention provides a control method for a complete glass deep processing line system, comprising the following steps:

[0007] The cut glass is subjected to multiple identification processes, including RFID tags, laser engraving codes, and visual recognition, to establish an identification mapping relationship and obtain identification data.

[0008] The system receives identification data and uses a visible light camera, an infrared thermal imager, and a structured light scanner to inspect the glass quality. The sensor data is then fused using a weighted average algorithm to obtain the quality inspection results.

[0009] Receive the quality inspection results, formulate a piece matching scheme based on the status of the edge grinding machine using a greedy algorithm based on heuristic rules, optimize the piece matching order, and obtain the piece matching result;

[0010] The wafer arrangement results are received, and the wafer arrangement problem is decomposed into sub-problems of size matching, temperature compatibility and time scheduling for processing. A regional proportional-integral-derivative controller is used to control the temperature of the tempering furnace to obtain the wafer arrangement data.

[0011] Receive scheduling data, establish a unified data model using the OPC UA standard architecture, deploy edge computing nodes to preprocess the data, and obtain monitoring data;

[0012] Receive monitoring data, build an equipment health management platform based on digital twin technology, and use machine learning algorithms to predict and analyze the health status of equipment to obtain maintenance decisions.

[0013] In a preferred embodiment, the multi-identification process of applying RFID tags, laser engraving codes, and visual recognition to the cut glass includes:

[0014] Basic information about the glass is recorded using RFID tags, including dimensions (length L, width W, thickness T), cutting time, and order number.

[0015] The laser-engraved code uses a DataMatrix format QR code, with an engraving depth controlled between 0.1 and 0.2 mm. It can store 128 characters of information and has temperature resistance.

[0016] Visual recognition generates unique visual feature codes by extracting glass edge contours, surface textures, and geometric features, establishing a mapping relationship between three identification methods to accurately identify objects under different environmental conditions.

[0017] In a preferred embodiment, the quality inspection of the glass using a visible light camera, an infrared thermal imager, and a structured light scanner includes:

[0018] The visible light camera has a resolution of 4096×3000 pixels and a pixel size of 2.4μm, and is used to detect surface defects;

[0019] The infrared thermal imager has a resolution of 640×480 pixels and a temperature resolution of 0.1℃, and is used to detect the internal stress distribution of glass.

[0020] The structured light scanner uses a blue LED light source and has a scanning accuracy of 0.005mm. It is used to measure the flatness and thickness uniformity of glass.

[0021] The fusion processing of sensor data using a weighted average algorithm includes:

[0022] Sensor data fusion employs a weighted average algorithm, with the weighting coefficients dynamically adjusted based on environmental conditions.

[0023] Under normal lighting conditions, the weight of the visible light camera is 0.6, the weight of the infrared thermal imager is 0.25, and the weight of the structured light scanner is 0.15.

[0024] When there is insufficient light, the weight of the infrared thermal imager is increased to 0.5, while the weight of the visible light camera is reduced to 0.35.

[0025] Data fusion requires preprocessing: visible light images are enhanced for contrast through histogram equalization, infrared thermal image data are removed for noise through median filtering, and structured light scanning data are smoothed through Gaussian filtering.

[0026] The three types of data have different dimensions and need to be normalized to the 0 to 1 range before being merged and calculated.

[0027] In a preferred embodiment, the step of formulating a piece matching scheme using a heuristic rule-based greedy algorithm based on the edge grinding machine status includes:

[0028] The intelligent glass sheet matching system formulates the optimal glass sheet matching scheme based on the current status of the edging machine and the capacity arrangement of the subsequent tempering furnace. When the edging machine is set to a width of 1.8m, the system automatically selects glass sheets with a width in the range of 1.6 to 1.8m.

[0029] The matching optimization algorithm adopts a greedy algorithm based on heuristic rules, where the matching degree coefficient ranges from 0.8 to 1.0, the priority weight of urgent orders is 1.5, and the priority weight of ordinary orders is 1.0.

[0030] The optimization of the order of dispensing includes a flexible gripping system, which is equipped with a six-degree-of-freedom robotic arm and intelligent grippers.

[0031] The fixture uses a combination of vacuum chuck and mechanical gripper. The vacuum chuck has a diameter of 150mm and the vacuum level is adjustable from -80kPa to -95kPa.

[0032] The mechanical gripper is made of soft material and covered with a 2mm thick polyurethane buffer layer.

[0033] The clamping position is determined by calculating the center of gravity. For rectangular glass, the center of gravity is the midpoint between the length and width.

[0034] For irregularly shaped glass, the system calculates the mass distribution and determines the center of gravity coordinates based on the glass profile shape;

[0035] The clamping accuracy is monitored in real time by a laser position sensor, and the positioning accuracy reaches ±0.2mm.

[0036] In a preferred embodiment, the step of decomposing the scheduling problem into sub-problems of size matching, temperature compatibility, and time scheduling includes:

[0037] The size matching subproblem is solved using a two-dimensional bin packing algorithm. The algorithm aims to maximize space utilization. For one hundred pieces of glass to be tempered, the system sorts them by area from largest to smallest and uses a bottom-left priority algorithm for layout. The spacing between adjacent glass pieces is not less than 10mm to avoid mutual contact during the tempering process.

[0038] The temperature compatibility sub-problem is grouped according to the tempering temperature requirements of different thicknesses of glass: 4mm glass has a tempering temperature of 630 to 650℃, 6mm glass has a tempering temperature of 650 to 670℃, and 12mm glass has a tempering temperature of 670 to 690℃. The system prioritizes arranging glass of the same thickness in the same batch. When different thicknesses must be mixed, a zoned temperature control strategy is adopted.

[0039] In a preferred embodiment, establishing a unified data model using the OPC UA standard architecture includes:

[0040] The data integration system adopts an industrial internet architecture based on the OPC UA standard to establish a unified data model and communication protocol;

[0041] The deployment of edge computing nodes for data preprocessing includes: each production line is equipped with an edge server responsible for real-time data acquisition, preprocessing and local decision-making. The edge server is configured with an Intel Xeon E5-2680v4 processor, 32GB of memory and 1TB SSD storage.

[0042] Data standardization processing employs a three-step process of data extraction, transformation, and loading, supporting automatic conversion of various data formats. Data preprocessing removes noise through filtering, fills in missing values ​​using linear interpolation, and normalizes data of different dimensions to the 0-1 range using the maximum-minimum standardization method, ensuring data quality and consistency.

[0043] In a preferred embodiment, the step of formulating a matching scheme based on a heuristic rule-based greedy algorithm according to the edge grinding machine status further includes an intelligent matching algorithm based on deep reinforcement learning:

[0044] The intelligent matching algorithm based on deep reinforcement learning learns the optimal matching strategy through interaction with the environment and adopts a deep Q-network structure.

[0045] The state space includes the properties of the glass to be fitted, the status of the edge grinding machine, and the historical fitting effects, while the action space contains preset fitting schemes.

[0046] The reward function calculates the comprehensive reward value by normalizing the batching efficiency, equipment utilization rate, and product quality indicators, and then using a weighted summation.

[0047] In a preferred embodiment, the quality inspection of glass using a visible light camera, an infrared thermal imager, and a structured light scanner further includes a Transformer model based on a self-attention mechanism:

[0048] The Transformer model, which employs a self-attention mechanism, replaces the traditional convolutional neural network, significantly improving processing speed while maintaining detection accuracy.

[0049] The Transformer model can better capture the spatial relationships and contextual information of defects on the glass surface, thus improving the accuracy of defect identification.

[0050] In a preferred embodiment, the process of decomposing the scheduling problem into sub-problems of size matching, temperature compatibility, and time scheduling further includes virtual scheduling verification based on digital twin technology:

[0051] Before the actual scheduling, simulation verification is carried out in the digital twin model to predict the feasibility and effectiveness of the scheduling plan;

[0052] The digital twin model is constructed based on computational fluid dynamics and finite element analysis. The computational fluid dynamics model uses the Reynolds-averaged Navier-Stokes equations to solve the gas flow state inside the furnace.

[0053] The finite element analysis model uses the heat conduction equation to calculate the temperature distribution. The two models are solved by coupling the heat transfer boundary conditions to accurately simulate the temperature distribution, airflow state and glass stress changes inside the tempering furnace.

[0054] In a preferred embodiment, a glass deep processing end-to-end system is provided for executing the aforementioned control method for a glass deep processing end-to-end system, comprising:

[0055] The multi-marking tracking module is used to perform multi-marking processing on the cut glass using RFID tags, laser engraving codes, and visual recognition, establish mark mapping relationships, and obtain mark data;

[0056] The quality inspection module receives the identification data, uses a visible light camera, an infrared thermal imager, and a structured light scanner to inspect the glass quality, and uses a weighted average algorithm to fuse the sensor data to obtain the quality inspection results.

[0057] The intelligent piece matching module receives quality inspection results, formulates a piece matching scheme based on a heuristic rule-based greedy algorithm according to the status of the edge grinding machine, optimizes the piece matching order, and obtains the piece matching result.

[0058] The wafer arrangement optimization module receives the wafer arrangement results, decomposes the wafer arrangement problem into sub-problems of size matching, temperature compatibility and time scheduling for processing, and uses a regional proportional-integral-derivative controller to control the temperature of the tempering furnace to obtain wafer arrangement data;

[0059] The data integration module is used to receive scheduling data, establish a unified data model using the OPC UA standard architecture, and deploy edge computing nodes to preprocess the data to obtain monitoring data.

[0060] The predictive maintenance module receives monitoring data, builds an equipment health management platform based on digital twin technology, and uses machine learning algorithms to predict and analyze the health status of the equipment to obtain maintenance decisions.

[0061] The beneficial effects of this invention are as follows:

[0062] By organically combining multi-identifier tracking technology and intelligent sheet-matching algorithms, seamless connection and coordinated control of the entire glass deep processing process are achieved. The system employs a distributed control architecture and edge computing technology, significantly shortening the production cycle, improving equipment utilization, and substantially enhancing the automation level of the production line. The application of intelligent sheet-matching optimization and flexible clamping control effectively reduces manual intervention and waiting time, resulting in a significant improvement in overall production efficiency.

[0063] The system employs multi-sensor fusion quality inspection technology, integrating a visible light camera, an infrared thermal imager, and a structured light scanner. Deep learning algorithms enable precise defect identification and quality assessment. Multiple marking schemes ensure end-to-end product traceability, while flexible clamping technology avoids damage to the glass surface caused by traditional rigid clamps. The system achieves industry-leading accuracy in quality inspection and product pass rates, significantly improving customer satisfaction and effectively guaranteeing product quality stability.

[0064] By optimizing the loading rate of the tempering furnace through intelligent racking algorithms and adopting a zoned temperature control strategy, high-efficiency energy utilization was achieved. The application of frequency conversion control technology in the transmission system and intelligent sleep mode for the detection equipment further reduced the overall system energy consumption. Full utilization of peak-valley electricity pricing and effective reduction of carbon emissions brought considerable economic and environmental benefits to the enterprise, achieving the goal of green manufacturing.

[0065] The system's predictive maintenance platform, built on digital twin technology, accurately predicts equipment health status and remaining service life, enabling a shift from reactive to proactive maintenance. The integrated application of multiple machine learning algorithms improves the accuracy and timeliness of fault prediction. The expert knowledge base and reinforcement learning algorithms integrated into the intelligent decision-making system provide a scientific basis for optimizing maintenance plans, effectively reducing equipment maintenance costs and unplanned downtime losses. Attached Figure Description

[0066] Figure 1 This is a flowchart of a control method for a glass deep processing full-process connection system according to the present invention;

[0067] Figure 2 This is a module diagram of a complete glass deep processing line system according to the present invention;

[0068] Figure 3 This is a comparison chart of the production efficiency improvement of the present invention;

[0069] Figure 4 This is a comparison chart of the quality control effects of the present invention;

[0070] Figure 5 This is a graph showing the energy consumption optimization effect of the present invention. Detailed Implementation

[0071] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0072] At least one embodiment of the present invention discloses a control method for a complete glass deep processing line system, such as... Figure 1 As shown, it includes the following steps:

[0073] Step 1: Perform multiple identification processing on the cut glass using RFID tags, laser engraving codes, and visual recognition to establish an identification mapping relationship and obtain identification data;

[0074] After the glass is cut, the system immediately implements a multi-marking scheme for each piece of glass; basic information about the glass is recorded through RFID tags, including dimensions (length L, width W, thickness T), cutting time, and order number. When the glass size is 1.8m × 1.2m × 4mm, the data stored in the RFID tag is in the format "GLZ-20241201-001-1800-1200-4", where GLZ represents the glass identification, 20241201 is the date, 001 is the serial number, and the last three digits represent the length, width, and thickness in millimeters, respectively.

[0075] Simultaneously, laser engraving equipment engraves QR codes on the glass edge, with the engraving depth controlled between 0.1 and 0.2 mm to ensure that the glass strength is not affected. The QR code uses the DataMatrix format and can store 128 characters of information. To ensure the reliability of the label, a visual recognition system captures the surface features of the glass using an industrial camera and generates a visual feature code using a deep learning-based feature extraction method.

[0076] The visual feature code generation process includes: converting the glass surface image captured by the industrial camera to grayscale, removing noise using Gaussian filtering, and uniformly adjusting the image size to 512×512 pixels; using the SIFT algorithm to detect key feature points on the glass surface, including edge contours, surface texture, and local geometric features, extracting 128 feature points for each piece of glass; calculating a 128-dimensional SIFT descriptor vector for each feature point, the descriptor containing gradient direction and magnitude information of the region surrounding the feature point; using principal component analysis to compress the 128-dimensional feature vector into 64 dimensions, retaining more than 95% of the feature information; finally, converting the 64-dimensional feature vector into a 32-bit binary hash code using the locality-sensitive hashing algorithm, which serves as the unique visual identifier for the glass.

[0077] To achieve unified management of the three identification methods, the system establishes a distributed mapping table to store identification relationships. The mapping table adopts a distributed hash table design, using the glass unique identifier (128-bit UUID format) as the primary key to ensure global uniqueness. Each record includes RFID identification fields (tag ID, data content, read timestamp, signal strength), laser engraving code fields (QR code content, engraving location coordinates, engraving depth, recognition confidence level), visual feature code fields (feature hash code, feature point coordinate array, feature descriptor, matching threshold), status information fields (creation time, last update time, access count, synchronization status flag), and redundancy backup fields (backup node address, backup timestamp, data verification code).

[0078] Considering the complexity of the production environment, the system is designed with a robust arbitration mechanism for tag synchronization failures. Every 100ms, the system checks the synchronization status of the three types of tags. When data inconsistency is detected, an arbitration process is triggered: the three tags are scored based on signal quality, recognition confidence, and historical success rate, with RFID tags having a confidence weight of 0.4, laser-engraved codes 0.35, and visual feature codes 0.25. When two or more of the three tags have consistent data, a majority vote is used to determine the correct data. When the three tags have completely inconsistent data, the tag with the highest confidence score is selected as the benchmark, and a manual intervention process is triggered. Based on the selected benchmark tag, the data for the other two tags is regenerated, the mapping table is updated, and the arbitration log is recorded to ensure data consistency is restored. During the arbitration process, the system monitors the environmental conditions for tag reading in real time, including factors such as temperature, humidity, and electromagnetic interference intensity, dynamically adjusting the confidence weight of each tag method to improve the accuracy of the arbitration decision.

[0079] Data transmission is achieved via industrial Ethernet at a rate of 1Gbps, with data packets using the UDP protocol to ensure real-time performance. When an RFID tag fails in a high-temperature environment, the system automatically switches to laser-engraved code recognition. A redundancy switching algorithm automatically selects the optimal identification method based on ambient temperature, humidity, and electromagnetic interference intensity.

[0080] Furthermore, blockchain technology can be used to establish an immutable chain of identification records. Each reading and update of identification information generates a new block in the blockchain, ensuring data integrity and traceability. The blockchain network adopts a consortium blockchain format, maintained by key nodes within the enterprise. The consensus algorithm uses a practical Byzantine fault-tolerant algorithm to ensure that the system can still operate normally even if some nodes fail.

[0081] Step 2: Receive the identification data, use a visible light camera, an infrared thermal imager, and a structured light scanner to perform quality inspection on the glass, and use a weighted average algorithm to fuse the sensor data to obtain the quality inspection result;

[0082] The quality inspection system employs multi-sensor fusion technology, integrating a visible light camera, an infrared thermal imager, and a structured light scanner. The visible light camera has a resolution of 4096×3000 pixels and a pixel size of 2.4μm, capable of detecting surface defects as small as 0.01mm. The infrared thermal imager has a resolution of 640×480 pixels and a temperature resolution of 0.1℃, used to detect the internal stress distribution of the glass. The structured light scanner uses a blue LED light source and achieves a scanning accuracy of 0.005mm, used to measure the flatness and thickness uniformity of the glass.

[0083] Sensor data fusion employs a weighted average algorithm, with weight coefficients dynamically adjusted based on environmental conditions. Under normal lighting conditions, the visible light camera has a weight of 0.6, the infrared thermal imager has a weight of 0.25, and the structured light scanner has a weight of 0.15. When lighting is insufficient, the weight of the infrared thermal imager increases to 0.5, while the weight of the visible light camera decreases to 0.35.

[0084] Data fusion requires preprocessing: visible light images are enhanced for contrast through histogram equalization, infrared thermal image data are removed for noise through median filtering, and structured light scanning data are smoothed through Gaussian filtering. Since the three types of data have different dimensions, they need to be normalized to the 0-1 range before fusion calculation.

[0085] The defect identification algorithm employs a deep learning model based on a convolutional neural network (CNN). The network structure includes 5 convolutional layers, 3 pooling layers, and 2 fully connected layers. The CNN model is trained using 150,000 labeled samples: 120,000 for training, 20,000 for validation, and 10,000 for testing. The labeling rules categorize defects into three levels: Class A (severe defects, scratch depth greater than 0.1 mm, bubble diameter greater than 2 mm), Class B (moderate defects, scratch depth 0.05–0.1 mm, bubble diameter 1–2 mm), and Class C (minor defects, scratch depth less than 0.05 mm, bubble diameter less than 1 mm). The network hyperparameters are set as follows: learning rate 0.001, batch size 32, training epochs 200, Adam optimizer, cross-entropy loss function, and regularization coefficient 0.0001.

[0086] The quality score was calculated using a weighted average method, and the defect classification threshold after fusion was determined based on statistical analysis: Analysis of the quality score distribution of 5000 qualified glass samples revealed a normal distribution with a mean of 92.3 points and a standard deviation of 4.2 points. Based on the statistical analysis results, 85 points was set as the acceptable threshold, and 70 points as the unacceptable threshold. The specific steps are as follows:

[0087] Quality scores from three sensors—visible light camera, infrared thermal imager, and structured light scanner—are obtained, with each score ranging from 0 to 100. Weighting coefficients for visible light detection, infrared thermal imager detection, and structured light scanning are determined based on environmental conditions, with their sum equal to 1.0. Under normal lighting conditions, the weight for the visible light camera is 0.6, for the infrared thermal imager 0.25, and for the structured light scanner 0.15. A comprehensive quality score is calculated by multiplying each sensor score by its corresponding weighting coefficient and then summing the results, with a value ranging from 0 to 100. Quality is judged based on the score: a comprehensive quality score greater than 85 indicates a qualified product; a score between 70 and 85 initiates a manual re-inspection process; and a score below 70 indicates a non-qualified product. A defect localization system is also established, using binocular stereo vision technology to convert detected defects from the image coordinate system to the world coordinate system. This coordinate transformation is performed using camera calibration parameters, where the world coordinates corresponding to the image coordinates are calculated using the camera intrinsic parameter matrix and depth information.

[0088] To achieve precise defect localization and subsequent processing, the system established a calibration method from world coordinates to robotic arm coordinates: a 9-point calibration method was adopted, with 9 calibration points evenly distributed within the glass inspection area. The robotic arm's end effector touched the calibration points to obtain the robotic arm coordinates, while simultaneously recording the corresponding world coordinates, thus establishing a coordinate transformation relationship. Calibration accuracy was optimized using the least squares method, with reprojection error controlled within 0.5mm. Coordinate transformation was achieved through rotation matrices and translation vectors, converting the defect's world coordinates into the corresponding robotic arm coordinates.

[0089] Furthermore, a Transformer model with a self-attention mechanism can be used to replace the traditional convolutional neural network, significantly improving processing speed while maintaining detection accuracy. The Transformer model's training sample size was expanded to 200,000 images, and data augmentation techniques, including rotation, flipping, brightness adjustment, and contrast variation, were employed to increase sample diversity. The network hyperparameters were set as follows: embedding dimension 512, attention heads 8, encoder layers 6, learning rate 0.0001, batch size 16, and training epochs 150. The Transformer model is better able to capture the spatial relationships and contextual information of glass surface defects, especially demonstrating stronger recognition capabilities for defects with complex shapes.

[0090] Step 3: Receive the quality inspection results, formulate a piece matching scheme using a greedy algorithm based on heuristic rules according to the status of the edge grinding machine, optimize the piece matching order, and obtain the piece matching result;

[0091] The intelligent glass blending system formulates the optimal glass blending plan based on the current status of the edging machine and the capacity arrangement of the subsequent tempering furnace. It obtains the current width parameters of the edging machine; for example, when the edging machine's set width is 1.8m, the system automatically selects glass sheets with a width between 1.6 and 1.8m. The selection algorithm considers factors such as glass priority, quality grade, and processing difficulty.

[0092] The chip allocation optimization algorithm adopts a greedy algorithm based on heuristic rules, and the optimization decision process is as follows:

[0093] The matching degree between the glass and the edging machine is evaluated. Based on the degree of conformity between the glass dimensions and the current settings of the edging machine, a matching degree coefficient is calculated. The formula for the matching degree coefficient is: Matching Degree Coefficient = 0.6 × Size Matching Degree + 0.3 × Thickness Matching Degree + 0.1 × Shape Matching Degree. The size matching degree is calculated by the overlap ratio between the glass length and width and the edging machine's processing range. It is 1.0 when the glass dimensions are completely within the edging machine's processing range, and decreases proportionally when they exceed the range. The thickness matching degree is calculated based on the deviation between the glass thickness and the edging machine's optimal processing thickness. It is 1.0 if the deviation is within ±0.5mm, and decreases exponentially when the deviation exceeds this range. The shape matching degree is determined for irregularly shaped glass based on the matching degree between the edge complexity and the edging machine's processing capacity: 1.0 for rectangular glass, 0.9 for rounded edges, and 0.8 for complex irregular shapes. The value ranges from 0.8 to 1.0, with 1.0 for a perfect match; determine the glass processing priority, setting the priority weight for urgent orders to 1.5 and for regular orders to 1.0 to ensure urgent orders are processed first; check equipment compatibility, assessing the compatibility between the glass and the equipment, with a compatibility factor of 1.0 for perfect compatibility and 0.8 for adjustments; estimate processing time and cost, predicting processing time based on glass specifications and process requirements, and setting a time penalty coefficient of 0.01 to avoid selecting solutions with excessively long processing times; comprehensively evaluate and select the optimal solution, using the product of matching degree, priority, and compatibility as the benefit item, subtracting the time cost as the penalty item, and selecting the matching solution with the highest comprehensive score.

[0094] When implementing the piece-matching optimization algorithm, all possible piece-matching schemes are enumerated, and then the comprehensive score of each scheme is calculated according to the optimization decision-making process described above. The scheme with the highest score is selected as the optimal solution. To ensure the accuracy of the calculation, each parameter needs to undergo data preprocessing, unifying indicators with different dimensions to the same numerical range for comparison.

[0095] The flexible gripping system is equipped with a six-degree-of-freedom robotic arm and intelligent grippers. The grippers combine a vacuum suction cup and a mechanical claw. The vacuum suction cup has a diameter of 150mm and the vacuum level is adjustable from -80kPa to -95kPa. The mechanical claw is made of soft material and covered with a 2mm thick polyurethane buffer layer to prevent damage to the glass surface.

[0096] The clamping force is determined using the following steps:

[0097] Identify the glass material type: set the material coefficient of ordinary glass to 1.2 and that of tempered glass to 0.9, as different materials require different clamping forces; measure the glass thickness parameter, converting the thickness value from millimeters to meters, and then calculate the thickness to the power of 1.5; the greater the thickness, the greater the clamping force required; calculate the effect of glass weight, using glass with a density of 2500 kg / m³. 3 And gravitational acceleration 9.8 m / s² 2 Based on the contact area between the clamp and the glass, the basic clamping force requirement is calculated; a safety factor is determined, and to ensure safe and reliable clamping, a safety factor of 1.5 is set, meaning the actual clamping force is 1.5 times the theoretically calculated value; the final clamping force is calculated by multiplying the material coefficient, thickness effect, weight effect, and safety factor to obtain the final clamping force value, in Newtons.

[0098] In actual operation, the system will automatically calculate the clamping force required for each piece of glass according to the above operation steps, and transmit the calculation results to the robotic arm control system to ensure that the clamping process is both safe and reliable and will not damage the glass.

[0099] The gripping position is determined through center of gravity calculation. For rectangular glass, the center of gravity is the midpoint of its length and width; for irregularly shaped glass, the system calculates the mass distribution based on the glass's outline to determine the center of gravity coordinates. Gripping accuracy is monitored in real time by a laser position sensor, achieving a positioning accuracy of ±0.2mm.

[0100] The robotic arm motion control employs a geometry-based inverse kinematics algorithm. For the kinematic equations of a six-DOF robotic arm, analytical methods are used to calculate joint angles. The inverse algorithm steps are as follows: First, the coordinates of the wrist center point are calculated based on the target position and orientation of the end effector. The first three joint angles are then calculated using geometric relationships, employing the cosine theorem and arctangent function. Finally, the last three joint angles are calculated based on the orientation requirements, obtained through rotation matrix decomposition. When multiple solutions exist, the solution with the smallest deviation from the current joint angle is selected to ensure smooth motion.

[0101] The collision detection strategy employs a hierarchical bounding box method, representing each link of the robotic arm and obstacles in the working environment using axis-aligned bounding boxes. The detection algorithm consists of two stages: coarse detection and fine detection. The coarse detection stage uses AABB bounding boxes to quickly eliminate obviously non-intersecting object pairs; the fine detection stage performs detailed geometric calculations on potentially colliding object pairs and uses the separating axis theorem to determine actual collisions. The system establishes a 3D environment model, including static obstacles such as the edging machine, conveyor belt, and glass stacks, as well as dynamic obstacles from other moving equipment. The collision detection frequency is 100Hz. When a potential collision is detected, the system automatically triggers path replanning, using a fast random tree algorithm to generate a new collision-free path.

[0102] Furthermore, an intelligent glass matching algorithm based on deep reinforcement learning can be adopted, which learns the optimal matching strategy through interaction with the environment. The reinforcement learning algorithm adopts a deep Q-network structure, and the state space includes the attributes of the glass to be matched, the device state, and the historical matching effects, while the action space is the selectable matching scheme.

[0103] The state-space encoding uses a vectorized representation method, discretizing continuous states into 128-dimensional feature vectors. Glass attribute encoding includes size information (length, width, and thickness each occupying 4 bits of binary code), quality grade (3 bits representing excellent, good, medium, and poor grades), and priority (2 bits representing urgent and normal). Equipment status encoding includes the current load rate of the edging machine (8 bits representing 0%–100% load), the remaining capacity of the tempering furnace (8 bits), and the occupancy status of the conveyor belt (16 bits representing 16 conveying positions). Historical piece matching effect encoding includes recent success rate (8 bits), average processing time (8 bits), and equipment utilization rate (8 bits).

[0104] Action space discretization transforms the continuous piece-matching decision problem into a finite set of discrete actions. The system predefines 64 standard piece-matching schemes, each corresponding to an action number. The action encoding method is as follows: the first two bits represent the piece-matching strategy type (00 - efficiency priority, 01 - quality priority, 10 - balanced strategy, 11 - emergency handling), the middle three bits represent the target device selection (000 - edge grinding machine 1, 001 - edge grinding machine 2, and so on), and the last bit represents whether the caching mechanism is enabled (0 - direct processing, 1 - temporary caching). Action selection adopts an ε-greedy strategy, with an initial exploration rate ε = 0.3, which decays to 0.05 as the training progresses.

[0105] The weight parameters of the deep Q-network were determined based on extensive experimental verification and theoretical analysis. Analysis of historical data from 1000 batches of wafer matching tasks revealed that matching efficiency has the most significant impact on overall production cycle time, accounting for 40% of the overall performance indicators; equipment utilization directly relates to production costs and resource allocation, also accounting for 40% of the weight; while product quality is important, its impact on the wafer matching stage is relatively small, accounting for 20% of the weight. During the experimental verification process, different weight combinations (0.5 / 0.3 / 0.2, 0.4 / 0.4 / 0.2, 0.3 / 0.5 / 0.2, etc.) were tested. Comparative analysis showed that the weight configuration of 0.4 / 0.4 / 0.2 performed optimally in terms of comprehensive performance indicators, balancing production efficiency with equipment utilization and product quality.

[0106] The reward signal is calculated using the following decision-making logic:

[0107] The process involves calculating the batch processing efficiency by comparing the current batch processing time with the historical best processing time; shorter processing time indicates higher efficiency. Equipment utilization is assessed by comparing the current space occupancy rate with the equipment's maximum capacity; higher occupancy rate indicates better utilization. Product quality is measured by comparing the actual defect detection rate with the target detection rate; a detection rate closer to the target value indicates better quality. Weighting coefficients are set: batch processing efficiency is weighted at 0.4, equipment utilization at 0.4, and product quality at 0.2, ensuring a balance between efficiency and utilization. A comprehensive reward signal is calculated by multiplying each of the three indicators by its corresponding weight and summing the results. This summation is then mapped to a range of -1 to 1 using a sigmoid function, serving as the reward signal for reinforcement learning.

[0108] Through this reward mechanism, reinforcement learning algorithms can improve film matching efficiency and equipment utilization while ensuring product quality and achieving intelligent optimization of film matching strategies.

[0109] Step 4: Receive the sheet arrangement results, decompose the sheet arrangement problem into sub-problems of size matching, temperature compatibility and time scheduling, and use a regional proportional-integral-derivative controller to control the temperature of the tempering furnace to obtain the sheet arrangement data;

[0110] The intelligent sheet-laying system for tempering furnaces employs a layered optimization strategy, breaking down the complex sheet-laying problem into three sub-problems: size matching, temperature compatibility, and time scheduling. Taking a 2.4m×6m tempering furnace as an example, the effective heating area of ​​the furnace chamber is 2.2m×5.8m, and the support platform uses ceramic roller conveyors with a roller spacing of 150mm.

[0111] The size matching subproblem is solved using a two-dimensional bin packing algorithm, with the goal of maximizing space utilization. For the 100 pieces of glass to be tempered, they are sorted by area from largest to smallest; a bottom-left-fill algorithm is used for layout. The algorithm considers the minimum spacing requirement between glass panes, ensuring a minimum distance of 10mm between adjacent panes to avoid contact during the tempering process.

[0112] The temperature compatibility sub-problem is grouped according to the tempering temperature requirements of glass of different thicknesses. 4mm glass requires a tempering temperature of 630-650℃, 6mm glass 650-670℃, and 12mm glass 670-690℃. The system prioritizes grouping glass of the same thickness within the same batch. When it is necessary to mix different thicknesses, a zoned temperature control strategy is employed.

[0113] The physical isolation and airflow isolation structure design for the zoned temperature control strategy is as follows:

[0114] Physical partition structure: The furnace interior employs an adjustable ceramic partition system. The partitions are made of alumina ceramic, 20mm thick, and resistant to temperatures up to 1200℃. The partitions are vertically raised and lowered via a pneumatic drive, with a lifting stroke of 200mm and a response time of 3 seconds. The bottom of the partition is 50mm from the roller conveyor surface, and the top is 150mm from the furnace top, creating effective temperature zones. The partition surface is coated with an infrared reflective coating, achieving a reflectivity of 95%, effectively blocking radiative heat transfer.

[0115] Airflow isolation structure: Layered airflow isolation technology is adopted, with an airflow distribution chamber set in the middle of the furnace, utilizing multi-stage centrifugal fans (power 15kW, air volume 8000m³ / h). 3 / h) generates directional airflow. The airflow distribution chamber is equipped with an adjustable guide vane, the angle of which can be adjusted within the range of 0-45° to achieve precise control of the airflow direction.

[0116] The airflow isolation system comprises a three-layer structure: The upper isolation airflow has a temperature of 650℃ and a flow rate of 2.5m / s, flowing along the length of the furnace to prevent heat from the high-temperature zone from diffusing to the low-temperature zone; the middle working airflow has a temperature adjusted according to the glass requirements, a flow rate of 1.8m / s, and flows vertically downwards, directly acting on the glass surface; the lower return airflow has a temperature of 580℃ and a flow rate of 2.0m / s, flowing in the opposite direction below the roller conveyor to form a closed-loop circulation. The airflow isolation effect is verified through temperature gradient control: the temperature difference between adjacent sections is controlled within ±8℃, and the temperature uniformity within each section reaches ±3℃, effectively achieving differentiated temperature control for glass of different thicknesses.

[0117] The decision-making process for optimizing film scheduling is as follows:

[0118] Assess space utilization by calculating the ratio of the effective area occupied by glass to the total area of ​​the tempering furnace in the current layout scheme, aiming for a space utilization rate of over 0.9 to maximize equipment capacity. Analyze energy efficiency by reducing the number of temperature adjustments and shortening the temperature adjustment time, prioritizing the arrangement of glass with the same thickness and similar tempering temperature requirements in the same batch. Control temperature uniformity by minimizing the temperature gradient within the furnace to ensure tempering quality, and rationally distributing the positions of glass of different thicknesses to avoid excessive local temperature differences. Set weighting coefficients: space utilization rate with a weighting coefficient of 0.5, energy efficiency with a weighting coefficient of 0.3, and temperature uniformity with a weighting coefficient of 0.2, with the sum of the three equaling 1.0. Comprehensively evaluate and select the optimal scheme by multiplying the three evaluation indicators by their corresponding weighting coefficients and summing the results, selecting the layout scheme with the highest comprehensive score as the final implementation scheme.

[0119] During the decision-making process, the system will traverse all possible screen layout schemes, score each scheme according to the above-mentioned screen layout optimization decision-making process, and finally select the optimal screen layout scheme that can simultaneously meet the requirements of high space utilization, low energy consumption and good temperature uniformity.

[0120] The temperature control system employs a zoned proportional-integral-derivative (PID) controller. The furnace is divided into nine temperature control zones, each equipped with an independent heating element and temperature sensor. The controller parameters are adjusted in real-time using a fuzzy self-tuning algorithm, with the proportional coefficient varying from 1.2 to 2.8, the integral coefficient from 0.3 to 0.8, and the derivative coefficient from 0.1 to 0.5. Temperature control accuracy reaches ±3℃, and temperature uniformity is controlled within ±5℃.

[0121] Furthermore, virtual scheduling verification based on digital twin technology can be employed. Before actual scheduling, simulation verification can be performed in the digital twin model to predict the feasibility and effectiveness of the scheduling scheme. The construction and application of a digital twin CFD / FEM simulation system follows these logical steps:

[0122] Simulation Model Architecture Design: A coupled simulation model based on Computational Fluid Dynamics (CFD) and Finite Element Analysis (FEM) was established. The CFD model uses the Reynolds-averaged Navier-Stokes equations to solve for the gas flow state inside the furnace, including the velocity field, pressure field, and turbulence characteristics. The FEM model uses the heat conduction equations to calculate the temperature distribution and stress field, including the temperature gradient and thermal stress distribution inside the glass. The two models are coupled through heat transfer boundary conditions to ensure the accuracy of the simulation results.

[0123] Mesh Generation and Quality Control: The CFD model uses a structured hexahedral mesh with a total of 2.8 million meshes. The mesh size for the main furnace area is 20mm × 20mm × 15mm. A boundary layer mesh is used near the glass surface, with an initial mesh height of 0.5mm, a growth rate of 1.2, and a total of 8 layers. The mesh around the heating elements is locally refined to 5mm × 5mm × 5mm. The airflow inlet / outlet uses a transition mesh that gradually changes from 5mm to 20mm. The FEM model uses a tetrahedral mesh with a total of 1.5 million nodes. The glass body mesh size is 3mm to ensure at least 3 mesh layers in the thickness direction. The roller contact area is locally refined to 1mm, and the furnace wall mesh size is 10mm. Mesh quality checks ensure a minimum orthogonal mass > 0.3 and a maximum skewness < 0.85.

[0124] Boundary conditions and material property settings: CFD boundary conditions include: the airflow inlet is set as a mass flow inlet (flow rate 8000 kg / h, temperature 650℃), the airflow outlet is set as a pressure outlet (gauge pressure 0 Pa, reflux temperature 600℃), the furnace wall is set as an isothermal boundary (temperature 680℃, emissivity 0.85), the glass surface is set as a coupled heat transfer boundary for data exchange with the FEM model, and the roller conveyor surface is set as an isothermal boundary (temperature 620℃). FEM boundary conditions include: convective heat transfer on the upper surface of the glass (heat transfer coefficient 25 W / (m²)). 2 • K), ambient temperature (from CFD calculation), contact heat transfer on the lower surface of the glass (contact thermal resistance 0.001 m). 2 The heat transfer method involves radiative heat exchange along the glass side (emissivity 0.9, ambient temperature 680℃), with initial temperatures set at 20℃ for the glass and 650℃ for the furnace. Material properties are set as follows: glass density 2500 kg / m³. 3 Specific heat 830 J / (kg·K), thermal conductivity 1.2 W / (m·K).

[0125] Solver configuration and computational strategy: The CFD solver adopts the SIMPLE algorithm and a second-order upwind scheme. The k-ε standard model is selected as the turbulence model (turbulence intensity 5%), and the P1 radiation model (absorption coefficient 0.1m) is used. -1 The convergence criterion is set as residual < 10. -4 Furthermore, the energy balance error is <1%, the time step is 0.1 seconds, and the total simulation time is 600 seconds. The FEM solver uses transient thermal analysis and the Newmark-β method. The time integration uses an implicit scheme with a time step of 0.5 seconds. The convergence criteria are temperature change <0.1℃ and heat flow balance error <0.5%.

[0126] Model Coupling and Data Exchange: A loosely coupled approach is used to exchange data between the CFD and FEM models, with data transfer occurring every 10 time steps. Exchanged variables include surface temperature, heat flux density, and convective heat transfer coefficient, ensuring the continuity of physical quantities between the two models at the heat transfer boundary. The coupling convergence criterion is that the surface temperature change is less than 1°C between two consecutive iterations; when the convergence condition is met, the next time step is initiated.

[0127] Simulation Results Verification and Application: The accuracy of the simulation model was verified through experimental data. Temperature field verification involved comparing measured data at nine measurement points, with an average error of ±2.8℃ and a maximum error of ±4.5℃. Airflow field verification used PIV measurements, with a velocity field correlation coefficient of 0.92 and turbulence intensity error <15%. Stress field verification used strain gauge measurements, with stress prediction accuracy of ±8MPa and stress distribution trend consistency >90%. A single simulation took 45 minutes (64-core parallel computing), meeting real-time decision-making requirements. This digital twin system can predict temperature distribution uniformity, glass stress levels, and energy consumption before the execution of the glass arrangement scheme, achieving a prediction accuracy of over 95%, effectively avoiding product quality problems and energy waste caused by unreasonable glass arrangement schemes.

[0128] Step 5: Receive the scheduling data, establish a unified data model using the OPC UA standard architecture, deploy edge computing nodes to preprocess the data, and obtain monitoring data;

[0129] The data integration system adopts an industrial internet architecture based on the OPCUA standard, establishing a unified data model and communication protocol. The system deploys edge computing nodes, with each production line equipped with one edge server configured with an Intel Xeon E5-2680v4 processor, 32GB of memory, and 1TB SSD storage. The edge nodes are responsible for real-time data acquisition, preprocessing, and local decision-making, reducing network latency and bandwidth consumption.

[0130] Data standardization processing employs a three-step workflow of data extraction, transformation, and loading, supporting automatic conversion of various data formats. Raw data includes equipment operating parameters, sensor measurements, quality inspection results, and production statistics. Data preprocessing removes noise through filtering, fills in missing values ​​using linear interpolation, and normalizes data of different dimensions to the 0-1 range using a maximum-minimum standardization method, ensuring data quality and consistency.

[0131] Data storage utilizes the time-series database InfluxDB, supporting high-frequency data writing and fast querying. The data retention strategy is as follows: real-time data is retained for 7 days, hourly aggregated data for 3 months, and daily aggregated data for 2 years. The data compression ratio reaches 10:1, effectively saving storage space.

[0132] The real-time monitoring system is built on web technologies, using HTML5, CSS3, and JavaScript to develop the front-end interface, and Node.js and Express frameworks for the back-end. The monitoring interface includes functional modules such as production line overview, equipment status monitoring, quality trend analysis, and alarm management. The interface updates every 1 second to ensure the real-time nature of the monitoring information.

[0133] The early warning system employs a combination of rule engines and machine learning. The rule engine handles explicit threshold warnings, such as exceeding temperature limits or equipment malfunctions. The machine learning model is used for complex pattern recognition, such as equipment performance degradation or abnormal quality trends. Warning levels are divided into four categories: Informational, Warning, Critical, and Emergency, corresponding to different processing procedures and response time requirements.

[0134] Furthermore, federated learning technology can be used to enable collaborative training of models across multiple production lines, thereby improving the accuracy of early warning models while protecting data privacy.

[0135] The weight aggregation process in federated learning is as follows:

[0136] Each production line edge node trains a neural network model using local data. After training, the model weight parameters are uploaded to the central coordination server. Data weights are calculated by dividing the data volume of each node by the total data volume of all nodes, with nodes having larger data volumes receiving higher weights. Performance weights are calculated by dividing the model validation accuracy of each node by the highest accuracy among all nodes, with nodes having higher accuracy receiving higher performance weights. The final aggregate weights are determined by multiplying the data weights by 0.6 and the performance weights by 0.4, then summing them to obtain the final aggregate weight coefficient, ensuring that nodes with large data volumes and good performance have a larger weight in the global model. Weight aggregation is performed by multiplying the local model weight parameters of each node by the corresponding aggregate weight coefficient and then summing them to obtain the global model weight parameters. The global model is then distributed by sending the aggregated global model weight parameters to each edge node for the next round of local training, enabling collaborative learning across multiple production lines.

[0137] Through this federated learning mechanism, each production line can share learning experiences while protecting data privacy, thereby improving the accuracy of the overall system's early warning model.

[0138] Step 6: Receive monitoring data, build an equipment health management platform based on digital twin technology, and use machine learning algorithms to predict and analyze the health status of the equipment to obtain maintenance decisions;

[0139] The predictive maintenance system is built on a digital twin technology platform to manage equipment health. The system employs various machine learning algorithms, including support vector machines, random forests, and long short-term memory networks, to analyze multi-dimensional data such as vibration, temperature, current, and pressure of the equipment, predicting its health status and remaining service life.

[0140] The equipment health assessment model employs a multilayer perceptron neural network with a network structure of input layer (50 nodes), hidden layer 1 (100 nodes), hidden layer 2 (50 nodes), and output layer (1 node). Input features include equipment operating time, load rate, temperature, vibration amplitude, and spectral characteristics. The output is an equipment health index, ranging from 0 to 100. Values ​​above 90 indicate a healthy state, 70-90 a watchful state, 50-70 a warning state, and below 50 a fault state.

[0141] The health index labels are obtained using a hybrid approach combining manual annotation and automatic labeling. The specific implementation steps are as follows:

[0142] A benchmark dataset with expert annotations was established, in which five experts with over 10 years of equipment maintenance experience annotated the historical operational data of 1000 pieces of equipment. Based on actual equipment fault records, maintenance records, and performance, the experts assigned a health index to the equipment at each time point. The annotation rules included: 90-100 points for equipment operating normally with no abnormal indicators; 70-90 points for equipment operating normally but with minor abnormalities; 50-70 points for equipment exhibiting significant abnormalities but still operational; and 0-50 points for equipment experiencing a fault or shutdown. The consistency of the expert annotations was evaluated using a Kappa coefficient, requiring a value of 0.85 or higher.

[0143] An automatic labeling algorithm was constructed, establishing automatic labeling rules based on historical equipment failure data and maintenance records. The algorithm employs a sliding time window method, using the failure occurrence time as a baseline and tracing back 30 days of operational data. Data from 1-3 days prior to the failure is labeled as 0-30 points (severe failure risk), from 4-7 days as 30-50 points (high failure risk), from 8-15 days as 50-70 points (medium failure risk), from 16-30 days as 70-90 points (minor anomaly), and during normal operation as 90-100 points (healthy status). The automatic labeling algorithm also considers factors such as equipment type, working environment, and load intensity, employing differentiated labeling strategies for different equipment types.

[0144] A hybrid annotation quality control system is implemented, using expert-annotated data as the "gold standard" to train and validate the accuracy of the automatic annotation algorithm. Cross-validation is used to assess the consistency between automatic and expert annotations. Samples with an automatic annotation confidence level below 0.8 are automatically submitted to experts for manual review. An annotation quality feedback mechanism is established to continuously optimize the automatic annotation rules and expert annotation standards based on the comparison between model predictions and actual equipment conditions.

[0145] A dynamic annotation update mechanism is established to regularly update the annotation dataset as new equipment failure cases and maintenance data accumulate. New equipment operation data is collected monthly, initially annotated using an automatic annotation algorithm, and then validated by experts through sampling. For new equipment or special operating conditions, expert annotation is prioritized to establish baseline data before gradually transitioning to automatic annotation. The annotation dataset is kept dynamically balanced to ensure a reasonable distribution of sample sizes across different health status levels, avoiding data imbalance that could negatively impact model training performance.

[0146] The fault prediction algorithm employs an LSTM-based time-series prediction model. The model structure consists of three LSTM layers, each with 128 neurons, and is finally connected to a fully connected layer to output the prediction result. Training data includes two years of historical equipment operation data, sampled at a frequency of one minute.

[0147] The maintenance plan optimization employs a mixed-integer programming approach. The decision objective is to minimize the total maintenance cost, which comprises three parts: equipment maintenance cost, downtime loss cost, and failure risk cost. Constraints include maintenance resource limitations, production plan constraints, and equipment dependencies.

[0148] The intelligent decision-making system integrates expert systems and machine learning models to establish a production optimization knowledge base. This knowledge base contains over 3,000 production experience rules, covering aspects such as process parameter settings, quality control, equipment maintenance, and anomaly handling. The system employs a forward inference engine to automatically generate operational suggestions based on the current production status and historical experience.

[0149] Furthermore, an adaptive maintenance strategy based on reinforcement learning can be adopted, in which the system continuously learns and optimizes maintenance decisions through interaction with the environment. The reinforcement learning environment includes elements such as equipment status, production plans, and maintenance resources.

[0150] The reward mechanism for maintaining the strategy adopts the following decision-making logic:

[0151] The evaluation process involves assessing the impact of maintenance costs by comparing current maintenance costs with the historical maximum maintenance costs and calculating the cost ratio; lower costs result in higher rewards. It also analyzes productivity losses by comparing current productivity losses with the maximum possible losses and calculating the loss ratio; smaller losses result in higher rewards. Furthermore, it measures equipment reliability improvement by dividing the equipment health index by 100 to obtain the reliability index; higher equipment reliability results in higher rewards. Weighting coefficients are set: maintenance cost is weighted at -0.3, productivity loss at -0.4, and equipment reliability at 0.7. Negative weights are used for maintenance costs and productivity losses to minimize the objective, while positive weights are used for equipment reliability to maximize the objective. Finally, a comprehensive reward value is calculated by multiplying maintenance costs, productivity losses, and equipment reliability by their respective weights and summing the results. This summation yields the final maintenance strategy reward value, which guides the decision optimization of the reinforcement learning algorithm.

[0152] Technical component integration and data flow: The entire predictive maintenance and intelligent decision-making system achieves the organic integration of various technical components through the following data flow:

[0153] Data Acquisition Layer → Feature Engineering Layer: Raw data collected by sensors (temperature, vibration, current, etc.) is cleaned, standardized, and time-series feature constructed to form a 50-dimensional standardized feature vector, providing high-quality input for subsequent models; Feature Engineering Layer → Health Assessment Layer: The standardized 50-dimensional feature vector is input into a multilayer perceptron neural network, combined with a hybrid labeling system of expert annotation and automatic labeling, to output a 0-100 equipment health index, providing a state basis for maintenance decisions; Health Assessment Layer → Fault Prediction Layer: The health index, as one of the key input features of the LSTM model, combined with historical time-series data, predicts the probability of failure in the next 1-2 weeks, providing a time window for maintenance plan formulation; Fault Prediction Layer → Maintenance Optimization Layer The failure probability Pi,t predicted by LSTM is directly used as the risk cost calculation parameter of the mixed-integer programming model. Together with the maintenance cost Ci,t and downtime loss Di,t, it forms the objective function to achieve the cost-optimal maintenance plan. Maintenance optimization layer → reinforcement learning layer: The optimization result of mixed-integer programming is used as the initial policy of the reinforcement learning environment. Through continuous learning by the DQN algorithm, the optimal policy in the 75-dimensional action space is found in the 128-dimensional state space to achieve adaptive maintenance decision-making. Reinforcement learning layer → intelligent decision-making layer: The maintenance actions output by reinforcement learning are combined with the 3000 rule base of the expert system. The forward inference engine generates the final maintenance instructions and operation suggestions to ensure the interpretability and practicality of the decision.

[0154] System Feedback and Continuous Optimization Mechanism: The system establishes a closed-loop feedback mechanism, with actual maintenance effect data fed back to various technical components: Post-maintenance equipment performance data updates health index labels and optimizes labeling quality; actual fault occurrence data corrects the prediction accuracy of the LSTM model and adjusts model parameters; maintenance cost and effect data updates the cost parameters of the mixed integer programming, improving optimization accuracy; long-term operational performance evaluates the effectiveness of the reinforcement learning strategy and adjusts reward function weights; operator feedback improves the expert system rule base, enhancing decision support capabilities.

[0155] Through this integrated technical architecture, the system achieves full-process automation from data acquisition to intelligent decision-making, ensuring logical coherence and data consistency among various technical components, and providing a reliable predictive maintenance solution for glass deep processing production lines.

[0156] Through this reward mechanism, the system can maximize equipment reliability while reducing maintenance costs and production losses, and achieve continuous optimization of maintenance strategies.

[0157] In one embodiment of the present invention, a complete glass deep processing line system is provided, such as... Figure 2 As shown, it includes:

[0158] The multi-marking tracking module applies multiple marking processes to the cut glass, including RFID tags, laser engraving codes, and visual recognition, to establish a marking mapping relationship and obtain marking data.

[0159] The quality inspection module receives the identification data and uses a visible light camera, an infrared thermal imager, and a structured light scanner to inspect the glass quality. It then uses a weighted average algorithm to fuse the sensor data and obtain the quality inspection results.

[0160] The intelligent piece matching module receives the quality inspection results, formulates a piece matching scheme based on a heuristic rule-based greedy algorithm according to the status of the edge grinding machine, optimizes the piece matching order, and obtains the piece matching result.

[0161] The wafer arrangement optimization module receives the wafer allocation results, decomposes the wafer arrangement problem into sub-problems of size matching, temperature compatibility and time scheduling for processing, and uses a regional proportional-integral-derivative controller to control the temperature of the tempering furnace to obtain wafer arrangement data;

[0162] The data integration module receives scheduling data, establishes a unified data model using the OPC UA standard architecture, and deploys edge computing nodes to preprocess the data to obtain monitoring data.

[0163] The predictive maintenance module receives monitoring data, builds an equipment health management platform based on digital twin technology, and uses machine learning algorithms to predict and analyze the health status of the equipment to obtain maintenance decisions.

[0164] In an embodiment of the present invention, an application example of a control method for a full-process connection system for glass deep processing is provided:

[0165] Field tests were conducted for 180 days in a large automotive glass manufacturing enterprise. During the test period, a complete full-process connection system for glass deep processing was deployed, covering three production lines with a total test area of 12,000 square meters. The test team consisted of 15 engineers and 30 operators, and comprehensively verified various functions and performance indicators of the system.

[0166] An example of multi-identifier data collection is shown in Table 1:

[0167] Table 1: Example of multi-identifier data collection;

[0168]

[0169] Table 1 shows the actual operation of the system's multi-identifier data collection. Each piece of glass is equipped with three identification methods: RFID tags, laser engraving codes, and visual features, ensuring accurate identification under different environmental conditions. The RFID tags use a standardized coding format and contain information such as date, serial number, size, and thickness; the laser engraving codes use the DataMatrix format and have high-density information storage capabilities; the visual feature IDs are generated through image processing algorithms and have uniqueness and stability. The redundant design of the three identification methods ensures an identification success rate of over 99.8%.

[0170] An example of multi-sensor fusion quality inspection data is shown in Table 2:

[0171] Table 2: Example of multi-sensor fusion quality inspection data;

[0172]

[0173] Table 2 shows the actual inspection results of the multi-sensor fusion quality inspection system. The system integrates three sensors: a visible light camera, an infrared thermal imager, and a structured light scanner, and calculates a comprehensive quality score through a weighted fusion algorithm. The weight coefficients (0.6, 0.25, 0.15) are determined based on the sensor accuracy and environmental adaptability. The visible light camera has the highest weight, followed by the infrared thermal imager, and the structured light scanner has the smallest weight. A comprehensive quality score greater than 85 points is judged as qualified, 70 to 85 points enter the re-inspection process, and less than 70 points is judged as unqualified. The inspection accuracy rate reaches 98.5%, effectively ensuring product quality.

[0174] Such as Figure 3The diagram illustrates a comparison of various efficiency indicators between the system of this invention and the traditional production mode. The charts clearly show that the system of this invention achieves significant improvements in production cycle time, equipment utilization, and automation level. The production cycle time is shortened from the traditional 8 hours to 5.5 hours, the overall equipment efficiency increases from 72% to 91%, and the automation level increases from 48% to 82%. These improvements are mainly attributed to the application of end-to-end online control and intelligent optimization algorithms.

[0175] like Figure 4 The figure shows the trend of glass product quality indicators over a six-month period. Under traditional production methods, the glass scratch rate fluctuated significantly, with a monthly average between 2.8% and 3.5%. After the implementation of this invention's system, the scratch rate significantly decreased and remained stable, with a monthly average controlled between 0.05% and 0.12%. The quality pass rate increased from 97.2% to 99.9%, and customer satisfaction improved from 85 points to 96 points. This quality improvement was primarily achieved through multi-sensor fusion detection and flexible gripping technology.

[0176] like Figure 5 The diagram illustrates the optimization effect of the system of this invention in energy consumption management. The charts evaluate energy consumption from six dimensions: tempering furnace energy consumption, transmission system energy consumption, testing equipment energy consumption, overall energy efficiency, peak-valley electricity price utilization rate, and carbon emission reduction. The system optimizes the tempering furnace loading rate through intelligent sheet arrangement, reducing the unit energy consumption of the tempering furnace by 18%. The transmission system adopts frequency conversion control technology, reducing energy consumption by 25%. The testing equipment, through intelligent sleep mode, reduces energy consumption by 30%. Overall energy efficiency is improved by 22%, and by fully utilizing the peak-valley electricity price difference, electricity costs are reduced by 15%.

[0177] Through the above application examples and technical effect verification, the glass deep processing full-process connection system and its control method proposed in this invention can effectively solve the technical problems existing in the traditional production mode, and achieve significant results in improving production efficiency, improving product quality, and reducing production costs, providing an effective technical solution for the intelligent transformation and upgrading of glass manufacturing enterprises.

[0178] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A control method for a complete glass deep processing line system, characterized in that, Includes the following steps: The cut glass is subjected to multiple identification processes, including RFID tags, laser engraving codes, and visual recognition, to establish an identification mapping relationship and obtain identification data. The system receives identification data and uses a visible light camera, an infrared thermal imager, and a structured light scanner to inspect the glass quality. The sensor data is then fused using a weighted average algorithm to obtain the quality inspection results. Receive the quality inspection results, formulate a piece matching scheme based on the status of the edge grinding machine using a greedy algorithm based on heuristic rules, optimize the piece matching order, and obtain the piece matching result; The wafer arrangement results are received, and the wafer arrangement problem is decomposed into sub-problems of size matching, temperature compatibility and time scheduling for processing. A regional proportional-integral-derivative controller is used to control the temperature of the tempering furnace to obtain the wafer arrangement data. Receive scheduling data, establish a unified data model using the OPC UA standard architecture, deploy edge computing nodes to preprocess the data, and obtain monitoring data; Receive monitoring data, build an equipment health management platform based on digital twin technology, and use machine learning algorithms to predict and analyze the health status of equipment to obtain maintenance decisions.

2. The control method for a glass deep processing full-process connection system according to claim 1, characterized in that, The multi-marking process of applying RFID tags, laser engraving codes, and visual recognition to the cut glass includes: Basic information about the glass is recorded using RFID tags, including dimensions (length L, width W, thickness T), cutting time, and order number. The laser-engraved code uses a DataMatrix format QR code, with an engraving depth controlled between 0.1 and 0.2 mm. It can store 128 characters of information and has temperature resistance. Visual recognition generates unique visual feature codes by extracting glass edge contours, surface textures, and geometric features, establishing a mapping relationship between three identification methods to accurately identify objects under different environmental conditions.

3. The control method for a glass deep processing end-to-end system according to claim 1, characterized in that, The quality inspection of glass using a visible light camera, an infrared thermal imager, and a structured light scanner includes: The visible light camera has a resolution of 4096×3000 pixels and a pixel size of 2.4μm, and is used to detect surface defects; The infrared thermal imager has a resolution of 640×480 pixels and a temperature resolution of 0.1℃, and is used to detect the internal stress distribution of glass. The structured light scanner uses a blue LED light source and has a scanning accuracy of 0.005mm. It is used to measure the flatness and thickness uniformity of glass. The fusion processing of sensor data using a weighted average algorithm includes: Sensor data fusion employs a weighted average algorithm, with the weighting coefficients dynamically adjusted based on environmental conditions. Under normal lighting conditions, the weight of the visible light camera is 0.6, the weight of the infrared thermal imager is 0.25, and the weight of the structured light scanner is 0.

15. When there is insufficient light, the weight of the infrared thermal imager is increased to 0.5, while the weight of the visible light camera is reduced to 0.

35. Data fusion requires preprocessing: visible light images are enhanced for contrast through histogram equalization, infrared thermal image data are removed for noise through median filtering, and structured light scanning data are smoothed through Gaussian filtering. The three types of data have different dimensions and need to be normalized to the 0 to 1 range before being merged and calculated.

4. The control method for a glass deep processing end-to-end connection system according to claim 1, characterized in that, The step of formulating a piece matching scheme using a heuristic rule-based greedy algorithm based on the edge grinding machine status includes: The intelligent glass sheet matching system formulates the optimal glass sheet matching scheme based on the current status of the edging machine and the capacity arrangement of the subsequent tempering furnace. When the edging machine is set to a width of 1.8m, the system automatically selects glass sheets with a width in the range of 1.6 to 1.8m. The matching optimization algorithm adopts a greedy algorithm based on heuristic rules, where the matching degree coefficient ranges from 0.8 to 1.0, the priority weight of urgent orders is 1.5, and the priority weight of ordinary orders is 1.

0. The optimization of the order of dispensing includes a flexible gripping system, which is equipped with a six-degree-of-freedom robotic arm and intelligent grippers. The fixture uses a combination of vacuum chuck and mechanical gripper. The vacuum chuck has a diameter of 150mm and the vacuum level is adjustable from -80kPa to -95kPa. The mechanical gripper is made of soft material and covered with a 2mm thick polyurethane buffer layer. The clamping position is determined by calculating the center of gravity. For rectangular glass, the center of gravity is the midpoint between the length and width. For irregularly shaped glass, the system calculates the mass distribution and determines the center of gravity coordinates based on the glass profile shape; The clamping accuracy is monitored in real time by a laser position sensor, and the positioning accuracy reaches ±0.2mm.

5. The control method for a glass deep processing end-to-end system according to claim 1, characterized in that, The process of decomposing the film arrangement problem into sub-problems of size matching, temperature compatibility, and time scheduling includes: The size matching subproblem is solved using a two-dimensional bin packing algorithm. The algorithm aims to maximize space utilization. For one hundred pieces of glass to be tempered, the system sorts them by area from largest to smallest and uses a bottom-left priority algorithm for layout. The spacing between adjacent glass pieces is not less than 10mm to avoid mutual contact during the tempering process. The temperature compatibility sub-problem is grouped according to the tempering temperature requirements of different thicknesses of glass: 4mm glass has a tempering temperature of 630 to 650℃, 6mm glass has a tempering temperature of 650 to 670℃, and 12mm glass has a tempering temperature of 670 to 690℃. The system prioritizes arranging glass of the same thickness in the same batch. When different thicknesses must be mixed, a zoned temperature control strategy is adopted.

6. The control method for a glass deep processing end-to-end system according to claim 1, characterized in that, The establishment of a unified data model using the OPC UA standard architecture includes: The data integration system adopts an industrial internet architecture based on the OPC UA standard to establish a unified data model and communication protocol; The deployment of edge computing nodes for data preprocessing includes: each production line is equipped with an edge server responsible for real-time data acquisition, preprocessing and local decision-making. The edge server is configured with an Intel Xeon E5-2680v4 processor, 32GB of memory and 1TB SSD storage. Data standardization processing employs a three-step process of data extraction, transformation, and loading, supporting automatic conversion of various data formats. Data preprocessing removes noise through filtering, fills in missing values ​​using linear interpolation, and normalizes data of different dimensions to the 0-1 range using the maximum-minimum standardization method, ensuring data quality and consistency.

7. The control method for a glass deep processing end-to-end system according to claim 1, characterized in that, The step of formulating a matching scheme based on a heuristic rule-based greedy algorithm according to the edge grinding machine status also includes an intelligent matching algorithm based on deep reinforcement learning: The intelligent matching algorithm based on deep reinforcement learning learns the optimal matching strategy through interaction with the environment and adopts a deep Q-network structure. The state space includes the properties of the glass to be fitted, the status of the edge grinding machine, and the historical fitting effects, while the action space contains preset fitting schemes. The reward function calculates the comprehensive reward value by normalizing the batching efficiency, equipment utilization rate, and product quality indicators, and then using a weighted summation.

8. The control method for a glass deep processing end-to-end system according to claim 1, characterized in that, The method of using visible light cameras, infrared thermal imagers, and structured light scanners for glass quality inspection also includes a Transformer model based on a self-attention mechanism. The Transformer model, which employs a self-attention mechanism, replaces the traditional convolutional neural network, significantly improving processing speed while maintaining detection accuracy. The Transformer model can better capture the spatial relationships and contextual information of defects on the glass surface, thus improving the accuracy of defect identification.

9. The control method for a glass deep processing full-process connection system according to claim 1, characterized in that, The process of decomposing the scheduling problem into sub-problems of size matching, temperature compatibility, and time scheduling also includes virtual scheduling verification based on digital twin technology: Before the actual scheduling, simulation verification is carried out in the digital twin model to predict the feasibility and effectiveness of the scheduling plan; The digital twin model is constructed based on computational fluid dynamics and finite element analysis. The computational fluid dynamics model uses the Reynolds-averaged Navier-Stokes equations to solve the gas flow state inside the furnace. The finite element analysis model uses the heat conduction equation to calculate the temperature distribution. The two models are solved by coupling the heat transfer boundary conditions to accurately simulate the temperature distribution, airflow state and glass stress changes inside the tempering furnace.

10. A complete glass deep processing line system, characterized in that, A control method for executing a glass deep processing end-to-end system as described in any one of claims 1-9 includes: The multi-marking tracking module is used to perform multi-marking processing on the cut glass using RFID tags, laser engraving codes, and visual recognition, establish mark mapping relationships, and obtain mark data; The quality inspection module receives the identification data, uses a visible light camera, an infrared thermal imager, and a structured light scanner to inspect the glass quality, and uses a weighted average algorithm to fuse the sensor data to obtain the quality inspection results. The intelligent piece matching module receives quality inspection results, formulates a piece matching scheme based on a heuristic rule-based greedy algorithm according to the status of the edge grinding machine, optimizes the piece matching order, and obtains the piece matching result. The wafer arrangement optimization module receives the wafer arrangement results, decomposes the wafer arrangement problem into sub-problems of size matching, temperature compatibility and time scheduling, and uses a regional proportional-integral-derivative controller to control the temperature of the tempering furnace to obtain wafer arrangement data. The data integration module is used to receive scheduling data, establish a unified data model using the OPC UA standard architecture, and deploy edge computing nodes to preprocess the data to obtain monitoring data. The predictive maintenance module receives monitoring data, builds an equipment health management platform based on digital twin technology, and uses machine learning algorithms to predict and analyze the health status of the equipment to obtain maintenance decisions.

Images