A tempered glass stress on-line detection and process regulation system

Abstract

The application discloses a tempered glass stress online detection and process regulation system and relates to the technical field of glass deep processing and industrial automatic control.The system comprises a multimodal stress detection module arranged at the cold end of a tempering furnace, a line-side edge computing node and a cloud server.The multimodal stress detection module is composed of a polarization camera array and an infrared thermal imager.The edge computing node is configured with a pre-trained stress field inversion model and a process parameter optimization model.The cloud server is configured with a model training unit and a process knowledge base unit.The application realizes real-time inversion of the full-field stress of glass through synchronous acquisition of multimodal data, generates a process adjustment instruction based on stress deviation, realizes closed-loop control and continuous evolution of the model through edge-cloud cooperation, and can be widely applied to continuous physical tempered glass production lines, thereby effectively improving product consistency and production yield.

Description

Technical Field

[0001] This invention relates to the field of glass deep processing and industrial automation control technology, specifically to an online stress detection and process control system for tempered glass. Background Technology

[0002] Tempered glass is a widely used safety glass in applications such as building doors and windows, glass curtain walls, automotive doors and windows, and appliance panels. The physical tempering process involves heating the glass to near its softening point and then rapidly cooling it, forming a compressive stress layer on the glass surface and a tensile stress layer internally, thereby improving the glass's load-bearing capacity and impact resistance. The surface compressive stress value, stress distribution uniformity, and internal tensile stress value of tempered glass are core technical indicators for determining its quality and safety level. Therefore, in the continuous production of tempered glass, conducting online stress testing on each piece of glass and adjusting process parameters such as the heating temperature, quenching air pressure, and cooling time of the tempering furnace in real time based on the test results is a fundamental requirement to ensure product consistency and production yield.

[0003] Some existing technologies collect process data such as glass temperature and cooling air pressure from the quenching section of the tempering furnace, solve the thermal history of the glass using a heat transfer model, and then indirectly calculate the stress distribution using a thermoelastic model. The control system of this approach is deployed locally on the production line and can automatically adjust process parameters based on the calculation results. However, its stress calculation is based on indirect extrapolation from process data, rather than direct physical quantity measurement of the stress field. The control loop uses temperature variables as an intermediary rather than stress variables as the core controlled variable. Furthermore, its cloud connection is only used for data exchange and visualization; the training and updating of the control strategy are all completed locally, lacking the ability for collaborative learning between the edge and cloud sides and continuous model evolution. Other existing technologies install stress sensors on tempered or semi-tempered glass production lines, setting the test area based on glass sheet size and arrangement information, and performing multi-point contact stress detection. This approach achieves online acquisition of stress data, but only outputs the detection results for manual interpretation and adjustment of process parameters, failing to form an automated closed loop between stress detection data and tempering furnace process parameter control.

[0004] As the above analysis shows, the fundamental deficiency of existing technologies lies in the long-standing functional disconnect between online stress detection and closed-loop control of process parameters in tempered glass. Specifically, systems with real-time stress detection capabilities only output detection data, requiring manual interpretation and adjustment of process parameters, lacking automatic closed-loop control capabilities. Systems with automatic process control capabilities rely on indirect stress estimation from process data rather than direct physical measurement of the stress field. This leads to a contradiction between the high real-time requirements of direct stress field inversion and the need for large-scale data training of process optimization models, compromising computational resources and response time. It is impossible to simultaneously achieve high real-time stress measurement closed-loop control and highly intelligent model self-evolution within the same system. This problem severely restricts further breakthroughs in tempered glass production lines in areas such as product consistency, rapid new product commissioning, and long-term process optimization.

[0005] Therefore, it is necessary to provide an industrial control system that can simultaneously meet the dual requirements of real-time direct stress measurement closed loop and continuous evolution of intelligent model at the system architecture level, so as to solve the technical problem that the real-time performance and intelligence of online stress detection and process control of tempered glass cannot be balanced. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an online stress detection and process control system for tempered glass. By deploying a polarization camera array and an infrared thermal imager at the cold end of the tempering furnace to simultaneously collect multimodal data, edge computing nodes invert the full-field stress in real time and generate closed-loop control commands, and cloud servers train and incrementally update the model offline, the system achieves the integration of online direct measurement of tempered glass stress and automatic control of process parameters, taking into account both the real-time response capability of the industrial site and the intelligent continuous evolution of the model.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an online stress detection and process control system for tempered glass, the system comprising a multimodal stress detection module deployed at the cold end of the tempering furnace. The multimodal stress detection module consists of a polarization camera array and an infrared thermal imager, and is used to simultaneously acquire polarization image data and temperature distribution data after the tempered glass is taken out of the furnace. An edge computing node is communicatively connected to the multimodal stress detection module. The edge computing node is deployed on the tempering furnace production line side and is equipped with a pre-trained stress field inversion model and a process parameter optimization model. It is used to invert the stress distribution of the glass in real time based on the polarization image data and temperature distribution data, calculate the deviation between the inverted stress distribution and the preset target stress field, and generate tempering furnace process parameter adjustment instructions based on the deviation calculation results. A cloud server is communicatively connected to the edge computing node. The cloud server is configured with a model training unit and a process knowledge base unit, which are used to perform offline training on the stress field inversion model and the process parameter optimization model using historical stress detection data and historical process data, and to compress the trained model and send it to the edge computing node. The edge computing node is also connected to the tempering furnace programmable logic controller (PLC) to send the generated process parameter adjustment instructions to the PLC. The PLC then drives the heating elements in each temperature zone of the tempering furnace heating section and the air grid actuators in each air pressure zone of the quenching section to perform parameter adjustments, forming a closed-loop control from stress detection to process regulation.

[0008] By synchronously acquiring polarization images and temperature data through a polarization camera array and an infrared thermal imager, edge computing nodes invert the stress distribution across the entire field in real time and generate process parameter adjustment instructions based on the deviation. Combined with offline training and compressed distribution of the model on a cloud server, a fully automatic closed-loop control can be formed from direct stress measurement to process parameter execution, while solving the problem of difficulty in balancing real-time control and model evolution.

[0009] Furthermore, the edge computing node is equipped with a real-time closed-loop control unit. When the stress deviation is detected to exceed a preset threshold, the real-time closed-loop control unit generates a process parameter adjustment command within a preset response time and sends it to the programmable logic controller of the tempering furnace. The real-time closed-loop control unit operates independently of the network connection status between the edge computing node and the cloud server. When the network connection between the edge computing node and the cloud server is interrupted, the real-time closed-loop control unit maintains closed-loop control based on the stress field inversion model and process parameter optimization model stored locally on the edge computing node.

[0010] By configuring a real-time closed-loop control unit independent of network connection status through edge computing nodes, adjustment commands are generated and issued within a preset response time when stress deviation exceeds the limit. Even when the network is interrupted, the control continues based on the local model, which can ensure the real-time continuity and robustness of closed-loop control and avoid control failure due to cloud communication delays or interruptions.

[0011] Furthermore, the cloud server is also equipped with an incremental learning module, which is used to receive incremental stress detection data and incremental process data uploaded by the edge computing node, perform incremental training on the stress field inversion model and process parameter optimization model in the cloud server based on the incremental stress detection data and incremental process data, and send the compressed model after incremental training to the edge computing node through a secure channel. The edge computing node receives and loads the updated model and replaces the original local model.

[0012] By receiving incremental data uploaded from edge nodes through the incremental learning module on the cloud server and incrementally training the model, the compressed data is then distributed to the edge nodes to replace the local model. This enables the stress field inversion model and process parameter optimization model to continuously evolve using new data, maintain their adaptability to new working conditions, and prevent model performance from deteriorating over time.

[0013] Furthermore, the process knowledge base unit is used to construct a cross-production line process knowledge graph, which integrates stress detection data and process parameter data from multiple tempered glass production lines in a structured manner, mines recommended ranges of process parameters and stress anomaly pattern features common to multiple production lines, and distributes the cross-production line process knowledge graph to the edge computing nodes of each production line in the form of a structured knowledge base file. The edge computing nodes adapt the cross-production line process knowledge graph distributed from the cloud according to the characteristics of their local production lines.

[0014] By constructing a cross-production line process knowledge graph through cloud-based process knowledge base units, integrating data from multiple production lines to mine recommended ranges of general process parameters and abnormal pattern characteristics, and distributing them to edge nodes for local adaptation, it is possible to achieve intelligent sharing among multiple production lines. This allows individual production lines to quickly adjust processes using industry experience, shortening the new product debugging cycle.

[0015] Furthermore, the polarization camera array is equipped with an electronically controlled liquid crystal phase delay device to achieve synchronous acquisition of multiple polarization state images within a single frame image acquisition cycle. The polarization camera array is installed above and below the cold end conveyor roller of the tempering furnace outlet to simultaneously acquire polarization images of the upper and lower surfaces of the tempered glass. The infrared thermal imager operates in the 8μm to 14μm band and is synchronously triggered with the polarization camera array using a unified hardware trigger signal. The polarization image data acquired by the polarization camera array and the temperature distribution data acquired by the infrared thermal imager are precisely aligned at the timestamp level.

[0016] By using an electronically controlled liquid crystal phase delay device to achieve synchronous acquisition of multiple polarization states in a single frame, and simultaneously detecting the upper and lower surfaces of the polarization camera array, the infrared thermal imager uses a unified hardware trigger signal to synchronize with the polarization camera, which can obtain polarization images and temperature data with precise timestamp alignment, improve the spatiotemporal consistency of stress inversion and online detection speed, and meet the production line cycle time requirements.

[0017] Furthermore, the edge computing node is also equipped with a data fusion unit. The data fusion unit performs spatial registration on the polarization image data acquired by the polarization camera array and the temperature distribution data acquired by the infrared thermal imager, establishes the mapping relationship between the coordinate systems of each sensor and the physical coordinate system of the tempered glass, and inputs the spatially registered multi-polarization image data and temperature distribution data as joint inputs into the stress field inversion model.

[0018] By spatially registering polarization image data and temperature distribution data through a data fusion unit, a mapping relationship between the sensor coordinate system and the glass physical coordinate system is established. The registered multimodal data is then provided as a joint input to the stress field inversion model, ensuring that different sensor data correspond precisely in spatial dimensions and improving the accuracy and reliability of full-field stress inversion.

[0019] Furthermore, the stress field inversion model is a deep learning model. The training data of the deep learning model is constructed by mixing digital twin simulation data, laboratory offline precision measurement data, and production line measured and labeled data. The stress field inversion model uses multi-polarization state light intensity images and temperature distribution images as joint inputs and outputs a three-dimensional full-field stress distribution data matrix of the tempered glass surface. The three-dimensional full-field stress distribution data matrix includes the compressive stress value of the tempered glass surface, the estimated value of the internal tensile stress, and the location markers of the abnormal stress spatial distribution areas.

[0020] By using a deep learning model with multi-polarization light intensity images and temperature distribution images as joint inputs, and through mixed training with digital twin simulation data, laboratory precision measurement data, and production line measured and labeled data, the model outputs a three-dimensional stress field data matrix containing surface compressive stress values, estimated internal tensile stress values, and abnormal area markers, enabling non-contact, full-field, and quantitative real-time stress inversion.

[0021] Furthermore, the target stress field is preset based on the specifications and application information of the tempered glass. The edge computing node automatically obtains the glass specifications and application information through the production line manufacturing execution system interface, and selects the corresponding target stress field from the built-in target stress field model library. The target stress field model library covers the stress standards for architectural tempered glass, automotive tempered glass, and household appliance tempered glass.

[0022] By automatically acquiring glass specification and application information from the manufacturing execution system interface through edge computing nodes, the corresponding target stress field is selected from the built-in model library. The model library covers stress standards for tempered glass used in construction, automobiles, and home appliances, and can automatically match stress control benchmarks for different varieties, achieving flexible and precise control when multiple varieties are produced on the same line.

[0023] Furthermore, the edge computing node is also equipped with a control decision engine, which performs hierarchical processing on the tempering furnace process parameters, classifies the quenching air pressure parameters into the fast response level, and prioritizes adjusting the quenching air pressure parameters after detecting stress deviation. The temperature parameters of the heating furnace zones are classified into the slow adjustment stage, and a model predictive control strategy is used to adjust the temperature parameters of the heating furnace zones step by step to avoid over-adjustment of the temperature parameters of the heating furnace zones.

[0024] By controlling the decision engine, the quenching air pressure parameter is prioritized for adjustment in the fast response level, while the heating furnace zone temperature parameter is prioritized for adjustment in the slow adjustment level and adjusted step by step using a model predictive control strategy. This approach can balance the immediate correction of cooling parameters with the stable transition of heating parameters, avoid stress fluctuations caused by over-adjustment of heating temperature, and improve the stability of process adjustment.

[0025] Furthermore, the edge computing node is also equipped with a safety interlock module, which operates independently of the main control loop of the edge computing node. When the edge computing node experiences a stress field inversion model inference timeout or a communication failure, the safety interlock module automatically switches the tempering furnace to preset safety process parameters and sends an interlock protection signal to the tempering furnace programmable logic controller.

[0026] By using a safety interlocking module that operates independently of the main control loop, the system automatically switches to preset safety process parameters and sends an interlocking protection signal when the stress field inversion model inference times out or a communication failure occurs. This prevents equipment malfunction and batch scrap caused by control system abnormalities, ensuring the stability of continuous production and the product qualification rate.

[0027] Compared with existing technologies, this online stress detection and process control system for tempered glass has the following advantages: I. This invention, by deploying a multimodal stress detection module consisting of a polarization camera array and an infrared thermal imager at the cold end of the tempering furnace, combined with a system architecture that integrates production line-side edge computing nodes and a cloud server, enables direct physical measurement and real-time inversion of the stress field after tempered glass exits the furnace. Simultaneously, it generates process parameter adjustment instructions based on the deviation between the stress distribution and the target value and sends them to the tempering furnace actuator. This completely breaks down the functional barriers between online stress detection and closed-loop control of process parameters, solving the problems of fragmented stress detection and control, insufficient accuracy of indirect stress estimation, and the inability to simultaneously achieve real-time closed-loop control and intelligent model evolution in existing technologies. It realizes full-process automated closed-loop control with stress as the core controlled variable, effectively improving product consistency and production yield in tempered glass production.

[0028] Second, this invention configures independently operating real-time closed-loop control units, hierarchical control decision engines, and safety interlocking modules on edge computing nodes. Combined with incremental learning modules and process knowledge base units on cloud servers, it can maintain stable operation of local closed-loop control in network interruption scenarios, achieve hierarchical and smooth adjustment of process parameters and safety interlocking protection under abnormal system conditions, and simultaneously complete continuous iterative optimization of models and cross-production line process knowledge sharing. This significantly improves the robustness and production safety of system operation, shortens the debugging cycle of new product processes, and ensures the long-term stability and continuous optimization capability of tempered glass continuous production.

[0029] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0031] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a schematic diagram of the installation and synchronous acquisition structure of the multimodal stress detection module of the present invention; Figure 3 This is a schematic diagram of the functional units and closed-loop control process of the edge computing node of the present invention. Detailed Implementation

[0032] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0033] Example This embodiment discloses a specific implementation of an online stress detection and process control system for tempered glass, applied to a continuous physical tempered glass production line. It fully realizes the synchronous acquisition of multimodal stress data after tempered glass exits the furnace, real-time stress inversion across the entire field, intelligent optimization of process parameters, and closed-loop automatic control throughout the entire process. This embodiment achieves high real-time local closed-loop control through edge computing nodes, and realizes offline model training, incremental iteration, and cross-production line process knowledge sharing through a cloud server. Simultaneously, it configures an independent safety interlock mechanism to ensure stable production, solving the problems of fragmented stress detection and process control functions and the difficulty in balancing real-time performance and intelligence in related technologies. This provides a stable and implementable end-to-end industrial control solution for continuous tempered glass production.

[0034] In this implementation, the overall system architecture is as follows: like Figure 1As shown, the online stress detection and process control system for tempered glass in this embodiment comprises four core components: a multimodal stress detection module, edge computing nodes, a cloud server, and a tempering furnace programmable logic controller (PLC). The multimodal stress detection module is deployed at the cold end of the tempering furnace and establishes a communication connection with the edge computing nodes via an industrial Ethernet network. The edge computing nodes are deployed on the tempering furnace production line side and establish bidirectional communication links with both the cloud server and the tempering furnace PLC. The cloud server is deployed on an industrial cloud platform and can simultaneously connect to edge computing nodes from multiple production lines. The tempering furnace PLC establishes hardware connections with the actuators in the heating and quenching sections of the tempering furnace, responsible for executing process parameter adjustment commands.

[0035] Specifically, the continuous physical tempered glass production line used in this embodiment includes an upper section, a heating section, a quenching section, a cold-end conveyor rollers, and a lower section connected sequentially. The production line is equipped with a tempering furnace programmable logic controller (PLC) to achieve basic control of equipment actions and process parameters, and a manufacturing execution system (MES) to achieve unified management of production plans, product specifications, and production data. A multimodal stress detection module is installed at the connection point between the quenching section outlet and the cold-end conveyor rollers. This location is where the tempered glass first enters a stable conveying zone free from strong wind interference after completing the quenching process, maximizing the stability and accuracy of the collected data and avoiding the impact of production line vibration and strong wind convection on the accuracy of the data acquisition.

[0036] The implementation details of the multimodal stress detection module are as follows: like Figure 2 As shown, the multimodal stress detection module consists of a polarization camera array and an infrared thermal imager. Its core function is to simultaneously acquire polarization image data and temperature distribution data after the tempered glass is taken out of the furnace.

[0037] Specifically, the polarization camera array is equipped with an electronically controlled liquid crystal phase delay unit (ECL). This ECL is hardware-level linked with the image sensor of the polarization camera, enabling phase modulation and synchronous image acquisition of multiple polarization states within a single frame image acquisition cycle. This avoids motion blur issues caused by time-division polarization acquisition in related technologies, perfectly adapting to the continuous conveying cycle of the production line and ensuring that the tempered glass surface in motion can achieve motion-free polarization image acquisition. For example, the polarization camera array can simultaneously acquire images of four polarization states within a single frame image acquisition cycle, covering polarization directions of 0°, 45°, 90°, and 135°, meeting the basic data requirements for stress inversion. The polarization camera array consists of an upper array and a lower array. The upper array is installed above the cold-end conveyor rollers, with its lens perpendicularly facing the upper surface of the tempered glass. The lower array is installed below the cold-end conveyor rollers, with its lens perpendicularly facing the lower surface of the tempered glass. The acquisition field of view of both arrays completely covers the entire surface of the tempered glass, eliminating blind spots and enabling simultaneous acquisition of polarization images of both the upper and lower surfaces of the tempered glass, providing complete data support for double-sided stress detection. For example, the acquisition frame rate of the polarization camera array can be set from 30fps to 60fps, which is linked and matched with the production line conveyor speed to ensure that each piece of tempered glass completes at least 10 full-panel acquisitions when passing through the inspection area.

[0038] The infrared thermal imager operates in the 8μm to 14μm wavelength range, which allows for precise acquisition of the infrared radiation signal from the tempered glass itself, unaffected by ambient visible light, production line lighting, or background radiation. This ensures stable output of temperature distribution data for the tempered glass surface. The infrared thermal imager and polarization camera array are synchronized using a unified hardware trigger signal, generated by a synchronization control board built into the edge computing node. The trigger frequency is linked in real-time with the production line conveyor speed, ensuring complete data acquisition across the entire tempered glass surface as each piece passes through the inspection area. For example, when the production line conveyor speed is 15m / min, the trigger frequency is set to 30Hz, guaranteeing seamless and uninterrupted data acquisition. Under this unified hardware triggering mechanism, the polarization image data acquired by the polarization camera array and the temperature distribution data acquired by the infrared thermal imager have identical acquisition timestamps, achieving precise alignment in the time dimension. This eliminates stress inversion errors caused by acquisition time differences, providing spatiotemporally consistent foundational data for subsequent accurate stress field inversion.

[0039] The implementation details of edge computing nodes are as follows: like Figure 3As shown, the edge computing node is deployed in the industrial control cabinet on the tempering furnace production line side. It employs industrial-grade edge computing hardware, featuring wide-temperature operation, resistance to electromagnetic interference, and high stability, making it adaptable to the complex environment of industrial production sites. The edge computing node is equipped with a pre-trained stress field inversion model and process parameter optimization model. It also integrates a data fusion unit, a real-time closed-loop control unit, a control decision engine, a safety interlock module, and corresponding industrial communication interfaces. Its core function is to invert the full-field stress distribution of the glass in real time based on polarization image data and temperature distribution data, calculate the deviation between the inverted stress distribution and the preset target stress field, and generate tempering furnace process parameter adjustment commands based on the deviation calculation results.

[0040] Specifically, the core function of the data fusion unit is to complete the spatial registration of multimodal data, providing accurate joint input data for the stress field inversion model. After receiving polarization image data and temperature distribution data uploaded by the multimodal stress detection module, the data fusion unit first completes the spatial coordinate matching of the two sets of data. It is understandable that there are differences in the installation position and lens parameters of the polarization camera array and the infrared thermal imager, and there is a mapping deviation between the pixel coordinate system of the two sets of devices and the physical coordinate system of the tempered glass. Spatial registration is needed to achieve a one-to-one correspondence between polarization information and temperature information at the same physical location. Before production line deployment, the data fusion unit completes the pre-calibration of the sensors. The calibration process consists of three steps. The first step is single-sensor calibration, where the upper and lower cameras of the polarization camera array and the infrared thermal imager are calibrated separately to obtain the intrinsic parameter matrix and distortion coefficient of each sensor, eliminating acquisition errors caused by lens distortion. The second step is joint calibration. A standard calibration board with a checkerboard pattern is horizontally fixed within the detection area of ​​the cold-end conveyor roller, ensuring that the surface of the calibration board completely overlaps with the conveying plane of the tempered glass. Images of the calibration board are acquired using a polarization camera array and an infrared thermal imager, and the checkerboard corner features are extracted. Using the physical coordinate system of the calibration board as a reference, the rotation matrix and translation vector between each sensor coordinate system and the glass physical coordinate system are calculated to construct a spatial mapping relationship. The third step is calibration verification. Using a standard-sized tempered glass sample, data is acquired through a multi-modal stress detection module to verify the accuracy of spatial registration, ensuring that the deviation between the polarization image and the temperature distribution image at the same physical position does not exceed a single pixel, thus completing the calibration process. In actual production operation, the data fusion unit, based on the pre-constructed mapping relationship, performs spatial registration on each set of synchronously acquired polarization image data and temperature distribution data. The spatially registered multi-polarization image data and temperature distribution data are integrated into a joint input dataset and transmitted to the stress field inversion model.

[0041] The stress field inversion model employs a deep learning model, specifically a convolutional neural network model with an encoder-decoder structure. Its core function is to invert the three-dimensional full-field stress distribution of tempered glass in real time based on multimodal input data. The model's training data is constructed from a mixture of digital twin simulation data, offline precision measurement data from the laboratory, and actual measured and labeled data from the production line, ensuring the model simultaneously possesses full-condition coverage and field adaptability. Specifically, the digital twin simulation data is generated through a tempering process digital twin system. Based on the glass thermal conduction equation, thermoelasticity equation, and stress optics law, it simulates the complete tempering process under different heating temperatures, quenching pressures, and glass specifications, generating corresponding polarization-state light intensity images, temperature distribution images, and true values ​​of the full-field stress distribution. This covers sample scenarios across all specifications and process ranges, providing ample basic training samples for the model. Offline precision measurement data from the laboratory is collected using a laboratory-grade glass stress meter. Point-by-point scanning measurements are performed on tempered glass samples prepared under different process conditions to obtain surface compressive stress, internal tensile stress, and stress distribution data. Simultaneously, polarization images and temperature distribution data of the corresponding samples are acquired, forming high-precision labeled samples and improving the model's inversion accuracy. The measured and labeled data of the production line is obtained through on-site sampling and testing. After online data collection of the tempered glass produced on the production line, offline measurement and verification are carried out using professional stress testing equipment to supplement the sample data under the actual working conditions of the production line and improve the adaptability of the model to the on-site environment.

[0042] After the model completes full training, it undergoes model compression and quantization processing via a cloud server to reduce computational load and inference latency, ensuring millisecond-level real-time inference on edge computing nodes. The model's input consists of spatially registered multi-polarization light intensity and temperature distribution images, while the output is a three-dimensional full-field stress distribution data matrix of the tempered glass surface. This matrix includes the surface compressive stress value, estimated internal tensile stress value, and location markers for areas of abnormal stress distribution at each coordinate point on the tempered glass surface, enabling quantitative, non-contact, and non-destructive online detection of full-field stress in tempered glass.

[0043] Edge computing nodes establish a bidirectional connection with the production line's manufacturing execution system via industrial communication interfaces, acquiring real-time specifications and application information for the current batch of tempered glass and automatically matching the corresponding target stress field. Specifically, specifications include the glass's thickness, length, and width, while application information categorizes applications such as construction, automotive, and home appliances. The edge computing nodes have a built-in target stress field model library, covering stress standards for architectural, automotive, and home appliance tempered glass, as well as stress control ranges for different glass thicknesses. Based on the acquired glass specifications and application information, the edge computing nodes automatically select the corresponding target stress field from the model library as the benchmark for stress deviation calculation. For tempered glass in different application scenarios, the target stress field is configured with corresponding surface compressive stress control ranges, internal tensile stress control ranges, and stress distribution uniformity control requirements, enabling flexible and precise control during the co-production of different types of tempered glass.

[0044] The real-time closed-loop control unit is integrated within the edge computing node. Its core function is to achieve real-time judgment of stress deviation and rapid response to process adjustment commands, ensuring the continuity and stability of closed-loop control. The real-time closed-loop control unit receives full-field stress distribution data output from the stress field inversion model, calculates the deviation between the actual stress distribution and the preset target stress field, and generates stress deviation data. This stress deviation data includes the overall average stress deviation of the plate surface, the stress distribution uniformity deviation, and the deviation values ​​of stress anomaly areas. The real-time closed-loop control unit pre-sets a stress deviation threshold. When a stress deviation exceeding the preset threshold is detected, a process parameter adjustment command is generated within a preset response time and sent to the tempering furnace programmable logic controller. For example, the preset stress deviation threshold can be set to: average surface compressive stress deviation not exceeding 10% of the target value; stress distribution uniformity deviation not exceeding 15% of the target value; and single-point deviation in stress anomaly areas not exceeding 20% ​​of the target value. For example, the preset response time of the real-time closed-loop control unit can be set to no more than 200ms, ensuring that process parameter adjustments are completed after the inspection of a single piece of glass and before the production of the next piece. The real-time closed-loop control unit operates independently of the network connection between the edge computing node and the cloud server. All core calculations and control logic are completed locally on the edge computing node, without relying on the computing power of the cloud server. When the network connection between the edge computing node and the cloud server is interrupted, the real-time closed-loop control unit continues to complete the entire process of stress detection, deviation calculation, and command generation based on the stress field inversion model and process parameter optimization model stored locally on the edge computing node. This maintains the normal operation of closed-loop control and ensures that the control function will not fail due to the interruption of cloud communication, thus guaranteeing the continuity and stability of production line production.

[0045] The control decision engine is built into the edge computing node and carries a process parameter optimization model. Its core function is to generate scientific and reasonable process parameter adjustment instructions based on stress deviation data, balancing the speed and stability of adjustment. The control decision engine performs hierarchical processing of the tempering furnace process parameters, dividing them into fast-response and slow-adjustment levels according to their response speed to stress. Specifically, the quenching air pressure parameters of each air pressure zone in the quenching section are classified as fast-response level. Adjusting these parameters can change the cooling rate of the glass in a short time, rapidly affecting the stress distribution. Therefore, upon detecting a stress deviation, the control decision engine prioritizes adjusting the quenching air pressure parameters to quickly correct the stress deviation. For example, the maximum adjustment range of the quenching air pressure parameter in a single instance is set to no more than 15% of the current set value to avoid sudden stress changes caused by abrupt changes in air pressure. The heating temperature parameters of each temperature zone in the heating section are classified as slow-adjustment level. Adjusting these parameters requires a certain amount of time to achieve a stable transition of the furnace temperature, resulting in a significant lag in their impact on stress distribution. Therefore, the control decision engine adopts a model predictive control strategy to adjust the temperature parameters of each zone in the heating furnace in stages. For example, the step size for adjusting the temperature parameters of the heating furnace zones is set to no more than 5°C, and the interval between two adjacent adjustments is set to no less than 3 minutes. When using a model predictive control strategy, the prediction step size is set to 3 to 5 control cycles to ensure a smooth transition of furnace temperature, avoid furnace temperature fluctuations caused by over-adjustment of the heating furnace zone temperature parameters, prevent the introduction of new stress deviations, and ensure the stability of process adjustment. After the process parameter adjustment command is generated, the edge computing node sends the command to the tempering furnace programmable logic controller (PLC) through the industrial Ethernet interface. The PLC then drives the heating elements of each temperature zone in the heating section and the air grid actuators of each air pressure zone in the quenching section to perform the corresponding parameter adjustment actions, forming a complete closed-loop control from stress detection to process regulation.

[0046] The safety interlock module is built into the edge computing node and operates independently of the node's main control loop. Its core function is to provide safety protection in abnormal system conditions, preventing equipment malfunction and the generation of batch scrap. The safety interlock module employs an independent hardware watchdog and operational status monitoring logic to monitor the entire operational status of the edge computing node in real time. Specifically, it monitors the inference status of the stress field inversion model, the communication status between the edge computing node and the multimodal stress detection module, and the communication status between the edge computing node and the tempering furnace's programmable logic controller (PLC). When fault conditions such as stress field inversion model inference timeout, core communication link interruption, or main control loop malfunction are detected, the safety interlock module immediately triggers interlock protection, automatically switching the tempering furnace to preset safe process parameters and simultaneously sending an interlock protection signal to the PLC. For example, the threshold for stress field inversion model inference timeout can be set to exceed 500ms; reaching this threshold triggers the safety interlock protection. The preset safety process parameters are those that have been verified over a long period of time on the production line and can guarantee the basic quality of tempered glass and the safety of equipment. They can provide stable basic operation support for the production line in the event of control system malfunctions, avoiding production accidents and batch quality problems.

[0047] The implementation details of the cloud server are as follows: The cloud server is deployed on an industrial cloud platform, possessing capabilities for large-scale data storage, high-performance model training, and multi-node collaborative management. It establishes bidirectional connections with edge computing nodes across multiple production lines via encrypted and secure communication links. The cloud server is equipped with a model training unit, a process knowledge base unit, and an incremental learning module. Its core function is to utilize historical data to complete offline training and continuous iteration of the model, while simultaneously building a cross-production line process knowledge system to achieve collective intelligent sharing across multiple production lines.

[0048] Specifically, the model training unit receives historical stress detection data and historical process data uploaded by edge computing nodes, cleans, standardizes, and labels the data to construct a standardized model training dataset, and performs offline full-scale training on the stress field inversion model and the process parameter optimization model. The offline training process is completed using large-scale cloud computing resources, enabling full parameter optimization and accuracy improvement for complex model structures. After training, the model is compressed and quantized, and then distributed to the corresponding edge computing nodes through an encrypted secure channel, completing the initial deployment and full update of the edge-side model.

[0049] The incremental learning module receives incremental stress detection data and incremental process data uploaded by edge computing nodes in real time. This incremental data consists of newly collected operating condition data from edge computing nodes during daily production that has not been used in full model training. Based on this incremental data, the module performs incremental training on the stress field inversion model and process parameter optimization model stored on the cloud server. The incremental training process only updates some parameters of the model, retaining the general features already learned by the model. This allows the model to quickly adapt to new operating conditions, new glass specifications, and new process scenarios without affecting its basic performance. For example, the trigger condition for incremental learning can be set to allow edge computing nodes to upload more than 1000 sets of valid production samples, or to continuously produce more than 50 batches of new glass products. For example, the model compression uses INT8 quantization levels, significantly reducing model inference latency while ensuring that the loss of model inference accuracy does not exceed 1%, thus adapting to the computing power level of the edge computing nodes. After incremental training, the model is compressed and quantized, and then distributed to the corresponding edge computing nodes through an encrypted secure channel. The edge computing nodes receive and load the updated model, replacing the original local model, thus enabling continuous iteration and performance optimization of the model and preventing the model performance from declining due to changes in production line conditions and equipment aging.

[0050] The process knowledge base unit is used to construct a cross-production line process knowledge graph. It structurally integrates stress detection data and process parameter data from multiple tempered glass production lines connected to the cloud server, building a standardized process data warehouse. Based on structured data, the process knowledge base unit uses data mining algorithms to mine universally applicable recommended ranges of process parameters and stress anomaly pattern features across production lines. Specifically, the recommended process parameter ranges are combinations of process parameters that can stably achieve qualified stress distribution under different glass specifications and application scenarios, providing a reliable benchmark reference for production line process debugging. Stress anomaly pattern features identify the root causes of process deviations corresponding to different stress anomalies, enabling rapid location and root cause analysis of stress anomalies. The construction process of the cross-production line process knowledge graph consists of four steps. The first step is data standardization, which involves classifying and cleaning stress detection data and process parameter data uploaded from different production lines according to glass specifications, application scenarios, and equipment models, unifying the data format and units. The second step is feature extraction, which uses data mining algorithms to extract the correlation features between stress distribution and process parameters under different operating conditions, identifying stress anomaly patterns and corresponding process root causes. The third step is knowledge integration, which involves constructing structured process knowledge nodes and related edges from the extracted features and relationships, forming a cross-production line process knowledge graph. The fourth step is knowledge distribution, which transforms the completed knowledge graph into a structured knowledge base file that can be directly accessed by edge computing nodes and distributes it to the edge computing nodes of each production line. After receiving the structured knowledge base file, the edge computing nodes of each production line adapt and adjust the cross-production line process knowledge graph distributed from the cloud according to the equipment characteristics, process habits, and production scenarios of their local production line. This combines general industry experience with local production line characteristics, achieving collective intelligent sharing across multiple production lines, significantly shortening the process debugging cycle for new tempered glass products, and improving the process optimization efficiency of the production lines.

[0051] This embodiment fully discloses a feasible implementation of an online stress detection and process control system for tempered glass, covering the entire process details from system hardware deployment, software module implementation, data processing flow, closed-loop control logic, and cloud-edge collaboration mechanism. This embodiment achieves direct, full-field, non-contact online stress detection of tempered glass through synchronous acquisition by a polarization camera array and an infrared thermal imager, solving the problem of insufficient accuracy in indirect stress estimation in related technologies. High real-time local closed-loop control is achieved through edge computing nodes, while offline model training, incremental iteration, and cross-production line process knowledge sharing are realized through a cloud server, balancing the real-time and intelligent requirements of industrial control. A hierarchical control strategy and independent safety interlocking mechanism ensure the stability of process adjustments and the safety of production operation.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A tempered glass stress online detection and process control system, characterized in that, The system consists of a multimodal stress detection module deployed at the cold end of the tempering furnace: The multimodal stress detection module consists of a polarization camera array and an infrared thermal imager, and is used to simultaneously acquire polarization image data and temperature distribution data after the tempered glass is taken out of the furnace. An edge computing node is communicatively connected to the multimodal stress detection module. The edge computing node is deployed on the tempering furnace production line side and is equipped with a pre-trained stress field inversion model and a process parameter optimization model. It is used to invert the stress distribution of the glass in real time based on the polarization image data and temperature distribution data, calculate the deviation between the inverted stress distribution and the preset target stress field, and generate tempering furnace process parameter adjustment instructions based on the deviation calculation results. A cloud server is communicatively connected to the edge computing node. The cloud server is configured with a model training unit and a process knowledge base unit, which are used to perform offline training on the stress field inversion model and the process parameter optimization model using historical stress detection data and historical process data, and to compress the trained model and send it to the edge computing node. The edge computing node is also connected to the tempering furnace programmable logic controller (PLC) to send the generated process parameter adjustment instructions to the PLC. The PLC then drives the heating elements in each temperature zone of the tempering furnace heating section and the air grid actuators in each air pressure zone of the quenching section to perform parameter adjustments, forming a closed-loop control from stress detection to process regulation.

2. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The edge computing node is equipped with a real-time closed-loop control unit. When the stress deviation is detected to exceed a preset threshold, the real-time closed-loop control unit generates a process parameter adjustment command within a preset response time and sends it to the programmable logic controller of the tempering furnace. The real-time closed-loop control unit operates independently of the network connection status between the edge computing node and the cloud server. When the network connection between the edge computing node and the cloud server is interrupted, the real-time closed-loop control unit maintains closed-loop control based on the stress field inversion model and process parameter optimization model stored locally on the edge computing node.

3. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The cloud server is also equipped with an incremental learning module, which is used to receive incremental stress detection data and incremental process data uploaded by the edge computing node, perform incremental training on the stress field inversion model and process parameter optimization model in the cloud server based on the incremental stress detection data and incremental process data, and send the compressed model after incremental training to the edge computing node through a secure channel. The edge computing node receives and loads the updated model and replaces the original local model.

4. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The process knowledge base unit is used to construct a cross-production line process knowledge graph, which integrates stress detection data and process parameter data from multiple tempered glass production lines in a structured manner, mines recommended ranges of process parameters and stress anomaly pattern characteristics that are common across production lines, and distributes the cross-production line process knowledge graph to the edge computing nodes of each production line in the form of a structured knowledge base file. The edge computing nodes adapt the cross-production line process knowledge graph distributed from the cloud according to the characteristics of their local production lines.

5. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The polarization camera array is equipped with an electronically controlled liquid crystal phase delay device, which is used to realize the synchronous acquisition of multiple polarization images within a single frame image acquisition cycle. The polarization camera array is installed above and below the cold end conveyor roller of the tempering furnace outlet, and simultaneously acquires polarization images of the upper and lower surfaces of the tempered glass. The infrared thermal imager operates in the 8μm to 14μm band and is synchronously triggered with the polarization camera array using a unified hardware trigger signal. The polarization image data acquired by the polarization camera array and the temperature distribution data acquired by the infrared thermal imager are precisely aligned at the timestamp level.

6. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The edge computing node is also equipped with a data fusion unit. The data fusion unit performs spatial registration on the polarization image data acquired by the polarization camera array and the temperature distribution data acquired by the infrared thermal imager, establishes the mapping relationship between the coordinate systems of each sensor and the physical coordinate system of the tempered glass, and inputs the spatially registered multi-polarization image data and temperature distribution data as joint inputs into the stress field inversion model.

7. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The stress field inversion model is a deep learning model. The training data of the deep learning model is constructed by mixing digital twin simulation data, laboratory offline precision measurement data and production line measured and labeled data. The stress field inversion model uses multi-polarization state light intensity images and temperature distribution images as joint inputs and outputs a three-dimensional full-field stress distribution data matrix of the tempered glass surface. The three-dimensional full-field stress distribution data matrix includes the compressive stress value of the tempered glass surface, the estimated value of the internal tensile stress, and the location markers of the abnormal stress spatial distribution areas.

8. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The target stress field is preset based on the specifications and application information of tempered glass. The edge computing node automatically obtains the specifications and application information of glass through the production line manufacturing execution system interface, and selects the corresponding target stress field from the built-in target stress field model library. The target stress field model library covers the stress standards of tempered glass for building, tempered glass for automobiles, and tempered glass for home appliances.

9. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The edge computing node is also equipped with a control decision engine. The control decision engine performs hierarchical processing on the tempering furnace process parameters, classifies the quenching air pressure parameters into the fast response level, and prioritizes adjusting the quenching air pressure parameters after detecting stress deviation. The temperature parameters of the heating furnace zones are classified into the slow adjustment stage, and a model predictive control strategy is used to adjust the temperature parameters of the heating furnace zones step by step to avoid over-adjustment of the temperature parameters of the heating furnace zones.

10. The online stress detection and process control system for tempered glass according to claim 1, characterized in that, The edge computing node is also equipped with a safety interlock module, which operates independently of the main control loop of the edge computing node. When the edge computing node experiences a stress field inversion model inference timeout or a communication failure, the safety interlock module automatically switches the tempering furnace to preset safety process parameters and sends an interlock protection signal to the tempering furnace programmable logic controller.

Images