Large-size microcrystalline glass surface stress detection device

Abstract

The invention discloses a large-size microcrystalline glass surface stress detection device, and relates to the technical field of microcrystalline glass stress detection, the large-size microcrystalline glass surface stress detection device comprises a detection host and a digital twinning processing unit, the digital twinning processing unit is used for constructing a regional characteristic digital twinning model of large-size microcrystalline glass, and generating a regional characteristic task state collaborative tensor; the multi-source sensing unit is used for collecting surface data of a microcrystalline glass sample and generating a sensing data vector; the digital twin processing unit updates a regional characteristic task state collaborative tensor; the reinforcement learning double-agent module comprises a path planning agent and a parameter adjusting agent; the data interaction unit transmits the path instruction to the motion control unit and transmits the parameter instruction to the parameter adjusting unit; the motion control unit drives the detection host to move according to a path instruction, and the parameter adjusting unit adjusts detection parameters according to a parameter instruction; according to the invention, the adaptation of the path density and the regional characteristics is realized.

Description

Technical Field

[0001] This invention relates to the field of stress detection technology for microcrystalline glass, specifically to a device for detecting surface stress in large-size microcrystalline glass. Background Technology

[0002] When performing automated scanning inspection of large-size microcrystalline glass, the software path planning and parameter configuration of existing stress detection devices are in an independent control mode. The path planning does not take into account the stress distribution characteristics of different regions such as the edge and center, and only uses a uniform grid path, which leads to missed detection in stress concentration areas or repeated measurement in non-critical areas. Meanwhile, the parameter configuration is in a globally unified mode, without a regional parameter preset library. When the path scans to different characteristic areas (such as high noise areas at the edges or stable stress areas at the center), the path execution needs to be manually interrupted to adjust the parameters. This not only reduces efficiency because the path interruption disrupts the scanning continuity, but also easily leads to insufficient detection accuracy because the timing of parameter switching does not match the path area (such as not switching to high-precision parameters in time in stress concentration areas). Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a large-size microcrystalline glass surface stress detection device.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a large-size microcrystalline glass surface stress detection device, including a detection host, a multi-source sensing unit, a digital twin processing unit, a motion control unit, a parameter adjustment unit, a data interaction unit, and a reinforcement learning dual-agent module; The digital twin processing unit is used to construct a digital twin model of the regional characteristics of large-size microcrystalline glass and generate a regional characteristic task state co-current tensor. The multi-source sensing unit is used to collect surface data of the microcrystalline glass sample and generate a sensing data vector; The digital twin processing unit calibrates the regional characteristic digital twin model based on the perceived data vector and updates the regional characteristic task state co-current tensor. The reinforcement learning dual-agent module includes a path planning agent and a parameter adjustment agent. The path planning agent generates path instructions based on the updated regional characteristics, task state, and collaborative tensor. The parameter adjustment agent generates parameter instructions based on the updated regional characteristics, task state, and collaborative tensor. The data interaction unit transmits the path command to the motion control unit and the parameter command to the parameter adjustment unit; The motion control unit drives the detection host to move according to the path instruction, and the parameter adjustment unit adjusts the detection parameters according to the parameter instruction. The detection host performs stress detection and feeds back the detection data to the digital twin processing unit.

[0005] As a preferred embodiment of the present invention, the construction process of the regional characteristic digital twin model includes: The system invokes a preset historical testing database, which stores historical regional data of different batches of large-size microcrystalline glass, including stress concentration data, noise data, and surface undulation data of the edge and center regions. Geometric virtual models of samples are constructed based on historical data to map the external contours of large-size microcrystalline glass. By fusing the historical data of the region with the geometric virtual model, an initial state of the region's characteristic task state co-current tensor with three dimensions, including physical characteristic vector, task progress vector, and strategy effect vector, is generated.

[0006] As a preferred embodiment of the present invention, the data acquisition process of the multi-source sensing unit includes: Collect images of the sample surface, extract the actual region boundary coordinates, and clarify the division range between the edge and the center; Collect surface shape data of the sample surface along the preset scan line to obtain surface undulation data; It covers the entire area, acquires stress distribution data, and identifies stress concentration coordinates; The multi-source sensing unit fuses multi-source data to generate an integrated data vector that reflects the actual characteristics of the sample. The integrated data vector includes region identifiers, actual edge deviations, measured surface undulations, measured stress concentration data, and measured noise data.

[0007] As a preferred embodiment of the present invention, the virtual physical error calibration process of the digital twin processing unit includes: The deviation values ​​between the measured data and the predicted values ​​of the digital twin model of regional characteristics are calculated. The deviation values ​​include edge coordinate deviation, surface undulation deviation, and stress concentration probability deviation. When the deviation exceeds the preset threshold, the characteristic parameters of the regional characteristic digital twin model are adjusted. When the deviation is within the preset range, update the local characteristic data of the regional characteristic digital twin model.

[0008] As a preferred embodiment of the present invention, the dynamic density path generation method of the path planning agent includes: When a region is identified as an edge region, a high-density path coverage scheme is adopted, and the spacing between adjacent detection points is the first spacing value. When the central region is identified, a low-density path scheme is adopted, and the spacing between adjacent detection points is the second spacing value, which is greater than the first spacing value. Once the path planning is complete, a path instruction containing the regional path sequence, movement speed, and path node coordinates is generated.

[0009] As a preferred embodiment of the present invention, the parameter matching process of the parameter adjustment agent includes: For regions where noise data exceeds the threshold, the light source power is set to the first power value and the integration time is set to the first duration value. For regions where stress concentration data exceeds a threshold, the polarizer angle is set to a first angle value; for the central region, the light source power is set to a second power value, the integration time to a second duration value, and the polarizer angle to a second angle value, wherein the first power value is greater than the second power value, and the first duration value is greater than the second duration value. After parameter matching is completed, a parameter command containing parameter values, switching timing, and parameter suitability evaluation is generated; The parameter fit evaluation is achieved by calculating the matching value between the parameters and the regional characteristics. When the matching value is lower than a preset threshold, a feedback flag is written into the regional characteristic task state collaborative tensor.

[0010] As a preferred embodiment of the present invention, the interaction verification mechanism of the dual-agent module includes: The parameter adjustment agent calculates the parameter fit value. When the fit value reaches the preset standard, it outputs path instructions and parameter instructions. When the fitness value is lower than the standard value, the parameter adjustment agent writes a feedback flag into the regional characteristic task state collaborative tensor. After the path planning agent detects this feedback flag, it increases the path density value of the corresponding area.

[0011] As a preferred embodiment of the present invention, the closed-loop control process includes: The detection host transmits the detection data of a single area to the digital twin processing unit; The digital twin processing unit analyzes the accuracy of the area, compares the measured stress value with the reference value of the standard sample, and counts the proportion of detection points that meet the accuracy requirements. When the accuracy compliance rate reaches the preset standard, the current strategy effect data will be written into the regional characteristic task state collaborative tensor. When the accuracy compliance rate is lower than the standard, the area is marked as a high uncertainty area, and the multi-source sensing unit is instructed to increase the sampling frequency of the area. Re-perform virtual physical error calibration and update the physical property vector of the region; Based on the calibrated physical characteristic vector, the dual-agent module reconstructs the path and parameter instructions for the region and re-executes the detection.

[0012] The beneficial effects of this invention are: 1. In this invention, through the collaborative work of a regional characteristic digital twin model and a multi-source sensing unit, the digital twin processing unit calibrates the regional characteristic digital twin model based on real-time sensing data vectors and updates the regional characteristic task state collaborative tensor. The path planning agent generates dynamic density path instructions based on the updated regional characteristic task state collaborative tensor, so that the edge region adopts a high-density path coverage scheme and the central region adopts a low-density path scheme, thereby achieving the adaptation and matching of path density and regional characteristics.

[0013] 2. In this invention, through the cooperative mechanism of the dual-agent module of reinforcement learning, the parameter adjustment agent generates detection parameter instructions based on the regional characteristics and task state cooperative tensor. According to different regional characteristics, the light source power, integration time, and polarizer angle parameter values ​​are automatically set. The parameter switching is completed before the detection host reaches the target area, realizing the regional characteristic adaptation of detection parameters and the continuous execution of the detection process.

[0014] 3. In this invention, through a closed-loop control process and a dual-agent interactive verification mechanism, when the detection accuracy does not meet the requirements, the system automatically marks high uncertainty areas, increases the sampling frequency, and recalibrates the digital twin model of the area characteristics. The dual agents reconstruct the path and parameter instructions of the area and re-execute the detection. At the same time, the detection strategy is optimized in real time through the update formula of the regional characteristic task state collaborative tensor and the evaluation process of the strategy effect vector, thus achieving a dynamic balance between detection accuracy and efficiency. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0016] Figure 1 This is a schematic diagram of the overall structure of the large-size microcrystalline glass surface stress detection device of the present invention. Detailed Implementation

[0017] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0018] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0019] Example 1 like Figure 1 As shown, a large-size microcrystalline glass surface stress detection device includes a detection host, a multi-source sensing unit, a digital twin processing unit, a motion control unit, a parameter adjustment unit, a data interaction unit, and a reinforcement learning dual-agent module. In this embodiment, the detection host is the core execution component of the entire system. It consists of an LD520nm, 30mW light source, a Glass3B prism, and a polarization analysis module. It is used to emit a stable detection beam, collect scattered light signals generated by the sample, and calculate the stress value and stress layer depth of the glass surface by analyzing the polarization characteristics. The detection host is fixed on a three-axis moving platform and can move precisely along the planned path under the drive of the motion control unit to perform point-by-point scanning detection on the surface of large-size microcrystalline glass.

[0020] The digital twin processing unit is used to construct a digital twin model of the regional characteristics of large-size microcrystalline glass and generate a regional characteristic task state co-current tensor.

[0021] The multi-source sensing unit is used to collect surface data of the microcrystalline glass sample and generate a sensing data vector.

[0022] The multi-source sensing unit fuses the above-mentioned data to generate a real-time sensing data vector that reflects the actual characteristics of the sample, providing accurate input for subsequent regional characteristic identification and path planning.

[0023] The digital twin processing unit calibrates the regional characteristic digital twin model based on the perceived data vector and updates the regional characteristic task state co-current tensor. As the core data processing unit of the system, the digital twin processing unit adopts a high-performance industrial computer configuration and has a built-in regional characteristic digital twin model construction module. At the same time, the digital twin processing unit receives the detection data fed back by the detection host, analyzes the detection accuracy, and triggers the closed-loop optimization control process.

[0024] The reinforcement learning dual-agent module includes a path planning agent and a parameter adjustment agent. The path planning agent generates path instructions based on the updated regional characteristics, task state, and collaborative tensor. The parameter adjustment agent generates parameter instructions based on the updated regional characteristics, task state, and collaborative tensor.

[0025] The data interaction unit transmits the path command to the motion control unit and the parameter command to the parameter adjustment unit; The data interaction unit is built using a high-speed industrial Ethernet communication module, which is responsible for data transmission and synchronization between various modules of the system, and feeds back the detection data collected by the detection host to the digital twin processing unit.

[0026] The motion control unit drives the detection host to move according to the path instruction, and the parameter adjustment unit adjusts the detection parameters according to the parameter instruction. The motion control unit includes a high-precision XYZ three-axis platform and a servo motor system. It receives path instructions transmitted by the data interaction unit and drives the detection host to move precisely along the planned path. The motion control unit can adjust the movement speed according to the characteristics of the area. In high-density path areas, the movement speed is reduced to ensure detection accuracy, and in low-density path areas, the movement speed is increased to improve detection efficiency. The parameter adjustment unit includes a laser power regulator, an integration time controller, and a polarizer rotating stage. It receives parameter instructions transmitted by the data interaction unit and completes parameter switching before the detection host reaches the target area. For areas where noise data exceeds the threshold, the parameter adjustment unit sets a high light source power and a long integration time; for stress concentration areas, it adjusts the polarizer angle to the optimal position; for centrally stable areas, it adopts a medium parameter configuration to balance detection efficiency and accuracy.

[0027] The detection host performs stress detection and feeds back the detection data to the digital twin processing unit, forming a closed-loop control process.

[0028] The working process of this device is as follows: First, the digital twin processing unit constructs a digital twin model of the regional characteristics of a large-size microcrystalline glass and generates an initial regional characteristic task state co-existence tensor. Then, the multi-source sensing unit collects surface data of the microcrystalline glass sample and generates a real-time sensing data vector; The digital twin processing unit calibrates the regional characteristic digital twin model based on real-time perceived data vectors and updates the regional characteristic task state collaborative tensor. In the reinforcement learning dual-agent module, the path planning agent and the parameter tuning agent generate dynamic density path instructions and detection parameter instructions based on the updated regional characteristics and task state co-current tensor, respectively. The data interaction unit transmits path commands to the motion control unit and parameter commands to the parameter adjustment unit; The motion control unit drives the detection host to move according to the path instructions, and the parameter adjustment unit adjusts the detection parameters according to the parameter instructions before the detection host reaches the target area; The main detection unit performs stress detection and feeds back the detection data to the digital twin processing unit, forming a closed-loop control process.

[0029] Through this closed-loop control process, the system can adapt to the characteristic differences of different regions, and achieve high-precision and high-efficiency stress detection on the surface of large-size microcrystalline glass. While ensuring the detection accuracy of high stress concentration areas at the edges, it significantly improves the overall detection efficiency.

[0030] It should be noted that this invention relates to a surface stress detection device for large-size microcrystalline glass, which is particularly suitable for the manufacturing of products with extremely high flatness requirements. The flatness of the product needs to be within 100nm. The stress of microcrystalline glass material has a significant impact on the surface shape. Therefore, it is necessary to accurately detect the surface stress of microcrystalline glass for process improvement. This device is mainly for automated high-precision stress detection of large-size microcrystalline glass with a diameter of 600mm. It solves the core problems in the prior art, such as missed detection of stress concentration areas, repeated measurement of non-critical areas, and process interruption caused by parameter switching, due to independent control of path planning and parameter configuration.

[0031] Furthermore, the construction process of the regional characteristic digital twin model includes: The system invokes a preset historical testing database, which stores historical regional data of different batches of large-size microcrystalline glass, including stress concentration data, noise data, and surface undulation data of the edge and center regions. The historical testing database is a collection of data that has been accumulated and verified over a long period of time. It stores regional characteristic sample data of large-size glass-ceramics produced under different batches and process conditions. These sample data have undergone rigorous screening and standardization to ensure the consistency and reliability of data quality. The database specifically includes stress concentration data, noise data, and surface undulation data of edge and center regions, which comprehensively reflects the characteristic differences of large-size glass-ceramics in different regions.

[0032] For example, edge regions typically exhibit high stress concentration (stress values ​​in the range of 1000-1500 MPa), high noise interference (signal-to-noise ratio below 20 dB), and large surface undulations (usually within ±30 μm); while the central region exhibits stable stress distribution (stress values ​​in the range of 200-500 MPa), low noise interference (signal-to-noise ratio above 40 dB), and small surface undulations (usually within ±5 μm). These detailed regional characteristic sample data provide a reliable basis for constructing accurate virtual models.

[0033] Geometric virtual models of samples are constructed based on historical data to map the external contours of large-size microcrystalline glass. The geometric virtual model accurately maps the outline of large-sized microcrystalline glass using 3D modeling technology (such as automatically generating the corresponding 3D mesh model through procedural modeling based on Unity-3D or similar 3D engines). This includes the circular outline (such as microcrystalline glass with a diameter of 600mm), thickness parameters (such as 5mm), and possible curved surface features. The geometric virtual model not only contains basic geometric dimension information, but also presets the boundary coordinates of the edge region (usually defined as the region within 100mm of the edge) and the center region (such as the central circular region with a diameter of 400mm) based on the regional division rules in historical data, providing a spatial reference framework for accurately locating the detection area.

[0034] The historical data of the region is fused with the geometric virtual model to generate an initial state of the regional characteristic task state collaborative tensor containing three dimensions: physical characteristic vector, task progress vector, and strategy effect vector. The physical characteristic vector is initialized with the regional characteristic data of similar samples in the historical data, the task progress vector is preset with the path coverage and theoretical execution time of the full detection process, and the strategy effect vector is initialized based on the historical detection strategy data. The physical property vector (in the physical property dimension) includes key parameter vectors such as stress concentration probability, noise level, and surface undulation. These vectors are initialized with regional characteristic data of similar samples (same material, same size, same process conditions) in historical data.

[0035] For example, for the edge region, the physical characteristic vector is initialized with a stress concentration probability of over 85%, a high noise level, and a surface undulation of ±30μm; for the center region, the physical characteristic vector is initialized with a stress concentration probability of less than 5%, a low noise level, and a surface undulation of ±5μm. This initialization method based on historical data enables the virtual model to quickly approximate the physical characteristics of the actual sample in the early stages of detection, reducing calibration time and errors.

[0036] The task progress vector (in the task progress dimension) includes task execution parameter vectors such as path coverage, theoretical execution time, and number of detected points. These vectors preset the path coverage of the entire detection process (usually above 98%), theoretical execution time, and the initial number of detected points (0). The task progress vector provides the system with a real-time reference for the detection progress, ensuring that the detection process proceeds as planned and avoiding missed detections or duplicate detections.

[0037] The strategy effect vector (in the strategy effect dimension) includes strategy evaluation parameter vectors such as parameter fit, accuracy achievement rate, and efficiency coefficient. These vectors are initialized based on historical detection strategy data and are usually set to parameter fit of 90%, accuracy achievement rate of 95%, and efficiency coefficient of 1.0. The strategy effect vector reflects the quality of the detection strategy and provides a quantitative basis for system optimization, enabling the detection process to achieve the best balance between accuracy and efficiency.

[0038] Through the above construction process, the regional characteristic digital twin model can accurately reflect the regional characteristic differences of large-size microcrystalline glass, providing a scientific basis for subsequent path planning and parameter configuration.

[0039] This model not only includes static geometric information but also integrates dynamic physical property data to form a complete virtual mapping. This enables the system to automatically generate highly adaptable and accurate detection strategies based on the characteristic differences of different regions, effectively solving the problems of missed detections and insufficient accuracy caused by regional characteristic differences in traditional detection methods.

[0040] The regional characteristic task state collaboration tensor, as the core data structure of the model, realizes the organic unity of physical characteristics, task progress and policy effect, and provides comprehensive and accurate information support for the decision-making of the reinforcement learning dual-agent module. It is a key foundation for realizing the closed-loop control process.

[0041] Furthermore, the multi-source sensing unit is integrated into the front end of the detection host and includes an industrial camera, a laser rangefinder, and a low-precision pre-scanning probe. The industrial camera is used to acquire high-definition images of the sample surface; the laser rangefinder is used to acquire surface undulation data of the sample surface along a preset scan line; and the low-precision pre-scanning probe is used to quickly cover the entire area and obtain overall trend data of stress distribution.

[0042] The data acquisition process of the multi-source sensing unit includes: Collect sample surface images, and extract actual region boundary coordinates using edge detection algorithms (such as the Canny edge detection algorithm or the Sobel operator algorithm, which can effectively identify the gray-scale change characteristics of the microcrystalline glass edge, accurately locate the edge position, and clarify the division range between the edge region and the center region) to clarify the division range between the edge and the center.

[0043] Collect surface shape data of the sample surface along the preset scan line to obtain surface undulation data; During the scanning process, the sensor emits a laser beam to the sample surface. By measuring the time difference or phase difference of the reflected light, the height difference of each point on the sample surface relative to the reference plane is calculated, thereby obtaining surface undulation data. For large-size microcrystalline glass with a diameter of 600mm, the laser rangefinder usually scans along 3-5 diameter lines at different angles, collecting 50-100 data points on each scan line to comprehensively capture the surface features of the sample surface.

[0044] Based on boundary recognition and surface scanning, a low-precision pre-scanning probe begins operation. This probe has a scanning speed of 100 mm / s and a stress resolution of ±20 MPa. By emitting a low-power laser beam to cover the entire sample area, it quickly acquires overall stress distribution data. The pre-scanning process adopts a gridded scanning strategy to perform coarse-grained scanning on the surface of the microcrystalline glass, generating a stress distribution heat map. By analyzing the gradient changes in this heat map, the system identifies potential stress concentration coordinates, which typically represent areas where stress values ​​change drastically.

[0045] After completing the data acquisition at the three levels mentioned above, the multi-source sensing unit performs fusion processing on the acquired multi-source data. The fusion processing includes two key steps: removing outlier data and data normalization. In the outlier removal stage, the system uses statistical methods (such as the 3σ principle or box plot analysis) to identify and remove data points that deviate significantly from the normal range. These outliers may originate from surface contamination, measurement interference, or equipment errors. The data normalization operation transforms measurement data of different dimensions and ranges into a unified numerical range (usually 0-1), ensuring that various types of data are comparable in subsequent processing and eliminating the impact of dimensional differences on data analysis.

[0046] After fusion processing, the data forms an integrated data vector, which fully reflects the actual characteristics of the sample. The integrated data vector contains five core elements: region identifier (used to distinguish edge regions, central regions or other special regions), actual edge deviation (representing the offset between the measured edge position and the theoretical edge position), measured surface undulation (representing the height change of the sample surface relative to the ideal plane), measured stress concentration data (representing the distribution characteristics and concentration degree of stress values ​​within the region), and measured noise data (representing the degree of signal interference). These elements comprehensively reflect the true physical state of the glass-ceramic sample, providing an accurate basis for model calibration of the digital twin processing unit.

[0047] In practical applications, the workflow of the multi-source sensing unit is highly automated. After the sample is placed on the detection platform and the initial positioning is completed, the system automatically triggers the acquisition process of the multi-source sensing unit. The generated integrated data vector is transmitted to the digital twin processing unit in real time through the data interaction unit.

[0048] Furthermore, the virtual physical error calibration process is automatically triggered after the multi-source sensing unit completes surface data acquisition. At this time, the digital twin processing unit receives the real-time sensing data vector generated by the multi-source sensing unit. This vector contains key information such as region identifier, actual edge deviation, measured surface undulation, measured stress concentration data, and measured noise data. Simultaneously, the digital twin processing unit calls the constructed regional characteristic digital twin model to obtain the predicted value of the corresponding region. The calibration process first performs deviation calculation. The system calculates the deviation value between the measured data and the predicted value of the regional characteristic digital twin model through comparative analysis. The deviation value includes edge coordinate deviation, surface undulation deviation, and stress concentration probability deviation. Edge coordinate deviation refers to the difference in Euclidean distance between the actual boundary coordinates of the region and the preset boundary coordinates in the virtual model. This deviation reflects the degree of fit of the virtual model to the geometric shape of the sample and is an important indicator for judging the geometric accuracy of the model.

[0049] Surface undulation deviation refers to the root mean square error between the measured height data of the sample surface and the predicted height data of the virtual model. This deviation reflects the virtual model's ability to simulate the surface morphology of the sample and directly affects the determination of the reference plane for stress detection.

[0050] Stress concentration probability deviation refers to the degree of difference between the measured distribution of stress concentration areas and the distribution predicted by the virtual model. It is usually expressed as a percentage (%). This deviation reflects the accuracy of the virtual model in predicting the mechanical properties of the sample and is a key reference for optimizing the detection strategy.

[0051] When the deviation exceeds the preset threshold, the characteristic parameters of the regional characteristic digital twin model are adjusted. Optionally, the edge coordinate deviation threshold is set to ±5μm, the surface undulation deviation threshold is ±3μm, and the stress concentration probability deviation threshold is ±5%. When any type of deviation value exceeds the corresponding threshold, the system determines that there is a significant difference between the virtual model and the actual sample, and a global adjustment is required. At this time, the digital twin processing unit triggers a comprehensive adjustment process for the virtual model's characteristic parameters (such as using the least squares method or gradient descent method for optimization to ensure that the adjusted model approximates the measured data in multiple dimensions), and recalculates the core parameters of the model, including geometric parameters, material parameters, and boundary condition parameters, so that the entire virtual model converges to the state of the actual sample.

[0052] When all types of deviation values ​​are within the preset threshold range, the system determines that the virtual model is basically consistent with the actual sample and only local optimization is needed. At this time, the digital twin processing unit only updates the local characteristic data of the model, such as the stress distribution curve of a specific area or the surface fine-tuning parameters, while keeping the main structure of the model unchanged. This local update strategy not only ensures the accuracy of the model, but also avoids unnecessary computational overhead and improves the system response speed.

[0053] After calibration, the updated physical characteristic vector is written into the regional characteristic task state collaborative tensor. This physical characteristic vector is the core component of the collaborative tensor, containing the calibrated regional characteristic data, accurately reflecting the current physical state of the sample. The update process follows strict timing control to ensure data consistency. Subsequently, through the data interaction unit, the calibrated regional characteristic task state collaborative tensor is synchronized to the reinforcement learning dual agent module in real time.

[0054] Among them, the virtual physical error calibration process of the digital twin processing unit is the core link to ensure detection accuracy. This process dynamically adjusts the regional characteristic digital twin model by accurately quantifying the difference between the virtual model prediction value and the actual measurement value, so as to make it more accurately reflect the real physical state of the sample, thereby providing a reliable basis for subsequent path planning and parameter configuration.

[0055] Furthermore, the path planning agent automatically generates dynamic density path instructions adapted to different regional characteristics based on the physical characteristic vector in the task state coordination tensor of the regional characteristics.

[0056] The path planning agent is structurally composed of a feature recognition module, a path density decision module, a path generation module, and an instruction output module.

[0057] The feature recognition module is responsible for parsing the physical characteristic vector in the regional characteristic task state co-current tensor and extracting key parameters such as regional type, stress concentration probability, noise level, and surface undulation. The path density decision module determines the appropriate detection point density based on the recognition results. The path generation module plans the specific detection path according to the density decision results. The instruction output module writes the generated path instructions into the task progress vector of the regional characteristic task state co-current tensor. This modular design ensures the systematicness and accuracy of the path planning process.

[0058] The core of the dynamic density path generation method lies in the region-adaptive path density strategy. When the path planning agent identifies an edge region (typically characterized by a stress concentration probability greater than 80%, high noise level, and surface undulation greater than ±25μm), the system adopts a high-density path coverage scheme. Under this scheme, the spacing between adjacent detection points is set as the first spacing value, which is usually in the range of 0.4-0.5mm, preferably 0.45mm. The high density setting ensures that the stress gradient changes in the edge region can be fully captured. Especially for stress concentration points and stress change areas, this fine detection grid can effectively avoid the risk of missed detection. The high-density path usually adopts a spiral coverage mode, gradually covering from the edge inward to ensure that high-risk areas near the region boundary can be detected first.

[0059] When the path planning agent identifies the central region (typically characterized by a stress concentration probability of less than 10%, low noise level, and surface undulation of less than ±10μm), the system adopts a low-density path scheme. In this scheme, the spacing between adjacent detection points is set as a second spacing value, which is usually in the range of 0.7-0.8mm, preferably 0.75mm. The second spacing value is significantly larger than the first spacing value. This differentiated setting fully considers the relatively uniform stress distribution in the central region and avoids efficiency loss caused by over-detection. The low-density path usually adopts a grid-based coverage mode, detecting along the preset grid lines. This regular path planning is beneficial to improving movement efficiency and reducing non-detection time.

[0060] After path planning is completed, the system generates path instructions containing regional path sequences, movement speeds, and path node coordinates. The regional path sequences define the access order of detection points; a spiral sequence is typically used for edge regions, while a grid sequence is used for central regions. The movement speed is dynamically adjusted according to the regional characteristics; a low speed of 5 mm / s is typically set for edge regions to ensure detection accuracy, while the speed can be increased to 10 mm / s for central regions to improve detection efficiency. The path node coordinates are the specific locations of the detection points, based on the detection platform coordinate system. These path instructions are fully recorded in the task progress vector of the regional characteristic task state coordination tensor through data writing operations, providing precise execution basis for the motion control unit and data support for subsequent detection progress tracking and strategy optimization.

[0061] Furthermore, the parameter adjustment agent is responsible for automatically matching the optimal detection parameters according to the regional characteristics, ensuring that high-quality stress detection data can be obtained in different regions. This design fully considers the significant differences in noise level and stress distribution between the edge region and the center region of large-size microcrystalline glass, and realizes regional adaptive optimization of detection parameters.

[0062] The parameter regulation agent is structurally composed of a parameter identification module, a parameter decision-making module, a parameter fitness evaluation module, and an instruction generation module.

[0063] The parameter identification module is responsible for parsing the physical characteristic vector in the regional characteristic task state collaborative tensor and extracting key parameters such as noise data and stress concentration data; the parameter decision module determines the optimal combination of detection parameters for each region based on the identification results; the parameter fit evaluation module assesses the degree of matching between the parameter settings and the regional characteristics; and the instruction generation module writes the finally determined parameter instructions into the regional characteristic task state collaborative tensor. This hierarchical structure design ensures the accuracy and reliability of the parameter matching process.

[0064] The parameter matching process first optimizes regions where noise data exceeds a threshold (typically regions with high noise levels and a signal-to-noise ratio below 20dB, mainly located in edge regions). In these regions, the system sets the light source power to a first power value, which is usually in the range of 28-32mW, preferably 30mW; at the same time, it sets the integration time to a first duration value, which is usually in the range of 45-55μs, preferably 50μs. High light source power enhances signal strength, and long integration time increases signal acquisition duration. The combination of the two effectively improves the signal-to-noise ratio and reduces the impact of noise interference on the detection results. This parameter setting is particularly suitable for high-noise environments common in edge regions and can ensure clear capture of scattered light signals.

[0065] For regions where stress concentration data exceeds a threshold (typically regions with stress values ​​greater than 1000 MPa, mainly located in edge areas), the system sets the polarizer angle to a first angle value, which is usually in the range of 40-50°, preferably 45°. This specific angle can maximize the capture of the birefringence effect caused by stress, improve the sensitivity to stress gradient changes, and thus accurately measure the stress distribution in high stress concentration areas. Optimizing the polarizer angle is a key parameter adjustment to improve the accuracy of stress detection, especially for accurate measurement of edge stress concentration areas.

[0066] For the central region (typically a region with stress values ​​in the range of 200-500 MPa and low noise levels), the system employs different parameter configurations: the light source power is set to a second power value, typically in the range of 18-22 mW, preferably 20 mW; the integration time is set to a second duration value, typically in the range of 15-25 μs, preferably 20 μs; and the polarizer angle is set to a second angle value, typically in the range of -5° to +5°, preferably 0°. These parameters are significantly different from those set for the edge region, with the first power value being greater than the second power value and the first duration value being greater than the second duration value. This differentiated setting fully considers the characteristics of good signal quality and uniform stress distribution in the central region, improving detection efficiency while ensuring detection accuracy. The lower light source power and shorter integration time reduce single-point detection time, while the 0° polarizer angle optimizes the detection sensitivity to uniform stress fields.

[0067] After parameter matching is completed, the system generates parameter instructions that include parameter values, switching timing, and parameter suitability evaluation.

[0068] The parameter values ​​explicitly specify the light source power, integration time, and polarizer angle for each region; the switching timing is pre-calculated based on the path planning results to ensure that the parameter switching is completed 5-10ms before the detection host reaches the region boundary, avoiding interference with detection data acquisition during the parameter switching process; the parameter fit evaluation is achieved by calculating the matching value between the parameters and the region characteristics. This value is in the range of 0-100 and reflects the degree of fit between the parameter settings and the region characteristics. When the matching value is lower than the preset threshold (usually 90), the parameter adjustment agent writes a feedback flag to the region characteristic task state collaborative tensor, triggering the path planning agent to increase the path density of the corresponding region, forming a parameter-path collaborative optimization mechanism.

[0069] For example, in practical applications, for large-size microcrystalline glass with a diameter of 600mm, when the detection host moves from the central area to the edge area, the system automatically increases the light source power from 20mW to 30mW, extends the integration time from 20μs to 50μs, and adjusts the polarizer angle from 0° to 45° according to the pre-planned path and area boundary. This area-adaptive parameter matching mechanism, combined with dynamic density path planning, effectively solves the problems of insufficient detection accuracy in the edge area and low detection efficiency in the central area caused by traditional global unified parameter settings, and achieves dual optimization of accuracy and efficiency in the detection of surface stress of large-size microcrystalline glass.

[0070] Furthermore, the dual-agent interactive verification mechanism achieves the best balance between accuracy and efficiency in the detection of large-size microcrystalline glass. This mechanism breaks through the limitations of independent operation of path planning and parameter configuration in traditional detection systems, and establishes a dynamic feedback and collaborative optimization relationship between the two agents.

[0071] The interaction verification mechanism of the two agents is structurally composed of a parameter fit calculation unit, a decision criterion judgment unit, a feedback flag generation unit, a path density adjustment unit, and a data write-back unit.

[0072] The parameter fit calculation unit is responsible for calculating the parameter fit value; the decision criterion judgment unit compares the calculated value with the preset standard; the feedback flag generation unit generates a feedback flag when the fit is insufficient; the path density adjustment unit adjusts the path density in response to the feedback flag; and the data write-back unit writes the real-time detection data into the strategy effect vector.

[0073] The interactive verification mechanism begins with the parameter adjustment agent calculating the parameter fit value. This calculation is based on the degree of matching between parameter settings and regional characteristics. By quantitatively analyzing the adaptability of parameter configuration to specific regional characteristics (such as noise level and stress distribution), a parameter fit value in the range of 0-100 is generated. The preset standard is usually set to 90. When the parameter fit value reaches or exceeds this standard, it indicates that the current parameter configuration matches the regional characteristics well, and the parameter adjustment agent directly outputs the path and parameter instructions without additional adjustment. When the parameter fit value is lower than the preset standard, the parameter adjustment agent writes a feedback flag to the regional characteristic task state collaborative tensor. This feedback flag contains the regional identifier and information on the degree of insufficient fit, providing a basis for adjustment for the path planning agent.

[0074] The path planning agent continuously monitors the regional characteristics, task state, and collaborative tensor. Once a feedback flag is detected, the path density adjustment process is immediately initiated. This process calculates the required increase in path density based on the degree of inadequacy contained in the feedback flag.

[0075] For example, when the fit value is 85 (below the standard of 90), the system increases the path density by 10%; when the fit value is 80, the system increases the path density by 20%. This proportional adjustment ensures that the increase in path density matches the degree of parameter misfit, neither excessively increasing the number of detection points leading to decreased efficiency nor affecting detection accuracy due to insufficient adjustment.

[0076] After the path command and parameter command are generated, they are transmitted to the motion control unit and parameter adjustment unit respectively through the data interaction unit for execution. The data interaction unit adopts the high-speed Ethernet communication protocol to ensure the real-time performance and reliability of command transmission.

[0077] During execution, the motion control unit drives the detection host to move precisely according to the path instructions, with a positioning accuracy of ±5μm; the parameter adjustment unit completes parameter switching 5ms before the detection host reaches the target area to ensure that the parameter settings are synchronized with the detection position.

[0078] Real-time data during the detection process includes detection point coordinates, stress measurement values, measurement accuracy, actual movement speed, and actual parameter values. This data is written back in real time to the strategy effect vector of the regional characteristic task state coordinating tensor through the data write-back unit. The strategy effect vector records the actual performance of each detection point, providing a data basis for subsequent strategy optimization. When the detection data reflects insufficient accuracy or low efficiency, the system can dynamically adjust the detection strategy of subsequent regions based on the historical data of the strategy effect vector, forming a closed-loop optimization mechanism.

[0079] For example, when detecting the edge region of a large-size microcrystalline glass with a diameter of 600mm, when the parameter adjustment agent calculates the parameter fit value of a certain edge sub-region to be 85 (below the preset standard of 90), the system writes a feedback flag into the regional characteristic task state collaborative tensor; after the path planning agent detects the flag, it increases the path density of the sub-region from 5 points / mm² to 5.5 points / mm²; the adjusted path and parameter instructions are transmitted to the execution unit, and after the detection is completed, the detection data is written back to the policy effect vector.

[0080] This interactive verification mechanism effectively solves the detection accuracy problem caused by insufficient parameter adaptation. At the same time, by dynamically increasing the path density rather than increasing the density globally, it avoids excessive sacrifice of detection efficiency and achieves a dynamic balance between accuracy and efficiency.

[0081] In summary, the closed-loop control process includes: The detection host transmits the detection data of a single area to the digital twin processing unit; The digital twin processing unit analyzes the accuracy of the area, compares the measured stress value with the reference value of the standard sample, and counts the proportion of detection points that meet the accuracy requirements. When the accuracy compliance rate reaches the preset standard, the current strategy effect data will be written into the regional characteristic task state collaborative tensor. When the accuracy compliance rate is lower than the standard, the area is marked as a high uncertainty area, and the multi-source sensing unit is instructed to increase the sampling frequency of the area. Re-perform virtual physical error calibration and update the physical property vector of the region; Based on the calibrated physical characteristic vector, the dual-agent module reconstructs the path and parameter instructions for the region and re-executes the detection.

[0082] Furthermore, the update mechanism of the regional characteristic task state collaborative tensor is a core technical link in achieving accurate detection and decision-making. The update formula of the regional characteristic task state collaborative tensor is as follows: ; Among them, T t+1 The regional characteristic task state coordination tensor at time t+1 serves as the core data carrier for system decision-making, and its updates must maintain consistency with the state T at the previous time step. t The continuity of decision-making should be maintained to avoid decision-making shocks caused by sudden changes.

[0083] The ⊕ symbol represents a tensor element-level fusion operator, which is implemented by weighted summation of the corresponding elements of the tensor. This design ensures the organic integration of historical states and new information, preserving historical decision-making experience while incorporating the latest perceived data in a timely manner.

[0084] α is the credibility weight of multi-source sensing data, with a baseline value set at 0.3; β is the credibility weight of digital twin calibration data, with a baseline value set at 0.4; γ is the credibility weight of agent decision data, with a baseline value set at 0.3. This initial allocation ratio reflects that in typical detection scenarios, the reliability of historical model data is usually higher than that of real-time sensing data, while decision effect data plays a balancing role.

[0085] When the system detects that the accuracy rate is below 95%, it indicates that the accuracy of the perceived data has a significant impact on decision-making. At this time, the α value increases by 0.1-0.2, and the β value decreases accordingly. When the detection efficiency is not up to standard (efficiency coefficient is greater than 1.2), it indicates that the decision-making strategy needs to be optimized. At this time, the γ value increases by 0.1-0.2, and the β value decreases accordingly. This dynamic adjustment mechanism enables the system to adapt to different detection scenarios and ensure that the collaborative tensor always maintains high reliability.

[0086] Multi-source sensing data vector S sense It includes five key dimensions: region identification (integer value, 0 represents the edge region, 1 represents the center region), actual edge deviation (in micrometers, representing the distance difference between the measured edge and the theoretical edge), measured surface undulation (in micrometers, representing the change in surface height), measured stress concentration probability (percentage, representing the possibility of stress concentration in the region), and measured noise level (integer value, 1 represents low noise, 2 represents medium noise, and 3 represents high noise).

[0087] Digital twin calibration data vector S twin It includes three key dimensions: the error value between the virtual model prediction and the measured value (percentage, indicating the degree of deviation between the model prediction and the actual measurement), the model confidence value (0-1 range, indicating the degree of certainty of the model in the current prediction), and the model's fit with historical samples (0-1 range, indicating the similarity between the current sample and historical samples).

[0088] Agent decision data vector S agent It includes three key dimensions: path density adjustment value (percentage, representing the adjustment ratio of path density relative to the baseline value), parameter and regional characteristic adaptation value (0-100 range, representing the degree of matching between parameter configuration and regional characteristics), and current decision and historical best decision matching value (0-1 range, representing the similarity between current decision and historical successful decision).

[0089] The engineering implementation of this formula can use the tensor operation function of the Python-NumPy library. During continuous detection, the system performs a collaborative tensor update every 100ms to ensure the real-time performance and accuracy of the decision-making basis.

[0090] The design of this collaborative tensor update formula fully considers the special requirements of large-size microcrystalline glass detection. Through dynamic weighting mechanism and multi-source data fusion, it effectively solves the limitations of traditional single data source decision-making, enabling the system to adapt to the characteristics of different regions and achieve high-precision and high-efficiency stress detection.

[0091] Furthermore, the evaluation process of the strategy effect vector adopts a joint decision-making evaluation mechanism. This evaluation mechanism is the core link in the decision optimization of the closed-loop control process, providing the system with objective and quantitative decision-making basis, and ensuring that the detection process achieves the best balance between accuracy, efficiency and resource consumption. The evaluation process of the strategy effect vector adopts the following calculation method: ; Among them, E joint The value represents the joint decision evaluation, which is a comprehensive indicator for measuring the overall performance of the detection strategy. The larger the value, the better the strategy performance.

[0092] λ1 is the weighting coefficient for the accuracy achievement rate, λ2 is the weighting coefficient for the efficiency achievement rate, and λ3 is the weighting coefficient for the resource consumption value. The sum of the three is equal to 1. Optionally, in the initial case, λ1=0.5, λ2=0.3, λ3=0.2. This allocation ratio reflects that in most detection scenarios, detection accuracy is the primary consideration, followed by efficiency, and resource consumption is relatively minor.

[0093] When the object being tested is microcrystalline glass for high-end optical instruments, the accuracy requirement is extremely high. In this case, adjusting λ1=0.7, λ2=0.2, and λ3=0.1 significantly increases the accuracy weight. When the object being tested is microcrystalline glass for mass-produced consumer electronics products, the efficiency requirement is relatively high. In this case, adjusting λ1=0.3, λ2=0.6, and λ3=0.1 significantly increases the efficiency weight.

[0094] P acc This indicates the percentage of tests that meet accuracy standards, reflecting the reliability of the test results; P eff Indicates the percentage of efficiency achieved, reflecting the time efficiency of the testing process; C res This represents the value of resource consumption, reflecting the energy and equipment wear and tear costs of the testing process.

[0095] Accuracy compliance rate P acc The calculation method is to count the number of detection points in the test data whose errors meet the accuracy requirements (e.g., ±8MPa), and divide this number by the total number of detection points. For example, in the edge region, out of 500 detection points, 475 points have errors within the ±8MPa range, then P... acc =475 / 500=0.95. This indicator directly reflects the reliability of the test results and is the core parameter for evaluating the quality of the test.

[0096] Efficiency compliance rate P eff The calculation method is to divide the theoretical optimal detection time (calculated based on ideal path planning and parameter settings) by the actual execution time. For example, if the theoretical optimal time is 120 seconds and the actual execution time is 150 seconds, then P... eff =120 / 150=0.8. This indicator reflects the time efficiency of the detection process. The closer the value is to 1, the higher the efficiency.

[0097] Resource consumption value C res The calculation method involves multiplying the laser power (mW) by the detection duration (s). For example, using 30mW power for 150 seconds, C... res =30×150=4500mW·s, and then converted into a dimensionless value in the range of 0.5-1.5 through normalization, which is convenient for mathematical calculation with other indicators.

[0098] For example, for the detection of the edge region of a large-sized microcrystalline glass with a diameter of 600mm, the system performs joint decision evaluation according to the following steps: First, the detection host completes the detection task of the edge area and transmits the detection data to the digital twin processing unit; The digital twin processing unit calculates the accuracy rate P of the region. acc =0.92 (92% of the detection point errors are within ±8MPa). Meanwhile, the motion control unit recorded the actual detection time of 180 seconds, and the system calculated the theoretically optimal detection time of 150 seconds, thus obtaining the efficiency compliance rate P. eff =150 / 180=0.833; The parameter adjustment unit calculates the laser power as 30mW, and based on the detection time of 180 seconds, the resource consumption value C is calculated. res After normalization, it becomes 1.2; In the scenario of testing microcrystalline glass for high-end optical instruments, the system is set with evaluation weight coefficients λ1=0.7, λ2=0.2, and λ3=0.1. Substitute into the formula to calculate: E joint =(0.7×0.92+0.2×0.833) / (0.1×1.2)=6.755; The system's preset threshold is 0.8. Since 6.755 > 0.8, the detection strategy is deemed qualified and no adjustment is required.

[0099] In another scenario, for the detection of the same area, when there is strong interference in the detection environment, the accuracy compliance rate drops to P. acc =0.85, the efficiency compliance rate dropped to P eff =0.75, the resource consumption value increases to C due to the extended integration time. res=1.4; Calculated using the same weighting coefficient: E joint =(0.7×0.85+0.2×0.75) / (0.1×1.4)=5.321; Although this value is still greater than 0.8, the system recorded that this value has decreased significantly compared to the previous detection, predicting that subsequent detections may not meet the standards, thus triggering a mild optimization mechanism in advance.

[0100] In the third scenario, when detecting the central area, the accuracy compliance rate P acc =0.98, efficiency compliance rate P eff =0.7 (due to the use of high-density paths), resource consumption value C res =1.3; In a mass production scenario, the system is set to λ1=0.3, λ2=0.6, and λ3=0.1; E is calculated. joint =(0.3×0.98+0.6×0.7) / (0.1×1.3)=5.492>0.8, the strategy is deemed qualified.

[0101] When E joint When the value falls below the preset threshold of 0.8, the system determines that the current detection strategy does not meet the requirements and initiates the optimization process. First, the parameter adjustment agent improves the parameter accuracy of the relevant area, such as increasing the light source power and extending the integration time; Secondly, the path planning agent increases the path density value of the corresponding area and reduces the distance between adjacent detection points; Finally, the two agents reconstruct the path and parameter instructions for the region, and the detection host re-executes the detection.

[0102] For example, in a certain edge region detection, the accuracy rate dropped to P due to sample surface contamination. acc =0.75, efficiency compliance rate P eff =0.65, resource consumption value C res =1.5; In the scenario of testing microcrystalline glass for high-end optical instruments, the system sets the evaluation weight coefficients λ1=0.7, λ2=0.2, and λ3=0.1; the calculated E joint =(0.7×0.75+0.2×0.65) / (0.1×1.5)=0.437<0.8; The system increases the path density from 5 points / mm² to 6 points / mm², the light source power from 30mW to 32mW, and the integration time from 50μs to 60μs; After retesting, the accuracy compliance rate increased to P. acc =0.94, efficiency compliance rate P eff =0.78, resource consumption value C res =1.6; Recalculate E joint=(0.7×0.94+0.2×0.78) / (0.1×1.6)=0.509. Although 0.509 is still less than 0.8, it is a significant improvement over the previous value. The system continues to iterate and optimize, eventually reaching the value shown in E. joint =0.85 > 0.8, which meets the requirements.

[0103] The design of this evaluation formula fully considers the special requirements of large-size microcrystalline glass detection. Through comprehensive evaluation of quantitative indicators, it avoids system performance imbalance caused by optimizing a single indicator. The numerator of the formula (λ1×P) acc +λ2×P eff This reflects positive returns and emphasizes the importance of accuracy and efficiency; the denominator (λ3×C) res This reflects negative costs and highlights the constraints on resource consumption.

[0104] This structural design conforms to the basic principles of engineering optimization, enabling the system to maximize detection performance under resource constraints.

[0105] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-size glass-ceramic surface stress detection device, characterized in that, The system comprises a detection host, a multi-source perception unit, a digital twin processing unit, a motion control unit, a parameter adjustment unit, a data interaction unit, and a reinforcement learning double-agent module. The digital twin processing unit is configured to construct a regional characteristic digital twin model of the large-size glass-ceramic and generate a regional characteristic task state collaborative tensor. The multi-source perception unit is configured to collect surface data of the glass-ceramic sample and generate a perception data vector. The digital twin processing unit is configured to calibrate the regional characteristic digital twin model according to the perception data vector and update the regional characteristic task state collaborative tensor. The reinforcement learning double-agent module comprises a path planning agent and a parameter adjustment agent, the path planning agent generates a path instruction based on the updated regional characteristic task state collaborative tensor, and the parameter adjustment agent generates a parameter instruction based on the updated regional characteristic task state collaborative tensor. The data interaction unit transmits the path instruction to the motion control unit and transmits the parameter instruction to the parameter adjustment unit. The motion control unit drives the detection host to move according to the path instruction, and the parameter adjustment unit adjusts the detection parameters according to the parameter instruction. The detection host performs stress detection and feeds back detection data to the digital twin processing unit.

2. The device for detecting surface stress of large-size microcrystalline glass according to claim 1, characterized in that, The construction process of the regional characteristic digital twin model comprises: calling a preset historical detection database, the historical detection database stores regional historical data of different batches of large-size glass-ceramics, including stress concentration data, noise data, and surface fluctuation data of edge regions and center regions; constructing a geometric virtual model of the sample based on the historical data, mapping the contour of the large-size glass-ceramic; fusing the regional historical data with the geometric virtual model to generate an initial state of the regional characteristic task state collaborative tensor including a physical characteristic vector, a task progress vector, and a strategy effect vector.

3. The device for detecting surface stress of a large-size microcrystalline glass according to claim 2, characterized in that, The data collection process of the multi-source perception unit comprises: collecting sample surface images, extracting actual regional boundary coordinates, and determining the division range of the edge and the center; collecting surface shape data of the sample along the preset scanning line to obtain surface fluctuation data; covering the entire region to obtain stress distribution data and identify stress concentration coordinates; The multi-source perception unit fuses the multi-source data to generate an integrated data vector reflecting the actual characteristics of the sample, which includes regional identification, actual edge deviation, measured surface fluctuation, measured stress concentration data, and measured noise data.

4. The device for detecting surface stress of large-size microcrystalline glass according to claim 3, characterized in that, The virtual physical error calibration process of the digital twin processing unit comprises: calculating the deviation value of the measured data and the predicted value of the regional characteristic digital twin model, the deviation value including edge coordinate deviation, surface fluctuation deviation, and stress concentration probability deviation; when the deviation exceeds the preset threshold, adjusting the characteristic parameters of the regional characteristic digital twin model; when the deviation is within the preset range, updating the local characteristic data of the regional characteristic digital twin model.

5. The apparatus for detecting surface stress of a large-size microcrystalline glass according to claim 2, wherein The dynamic density path generation method of the path planning agent comprises: when identifying the edge region, a high-density path coverage scheme is adopted, and the distance between adjacent detection points is a first distance value; When the central region is identified, a low-density path scheme is adopted, and the spacing between adjacent detection points is a second spacing value, which is greater than the first spacing value; After the path planning is completed, path instructions containing the sequence of regional paths, moving speed and path node coordinates are generated.

6. The apparatus for detecting surface stress of a large-size microcrystalline glass according to claim 5, wherein, The parameter matching process of the parameter adjustment agent includes: For the region where the noise data exceeds the threshold value, the light source power is set to a first power value, and the integration time is set to a first time length value; For the region where the stress concentration data exceeds the threshold value, the polarizer angle is set to a first angle value; for the central region, the light source power is set to a second power value, the integration time is set to a second time length value, and the polarizer angle is set to a second angle value, the first power value is greater than the second power value, and the first time length value is greater than the second time length value; After the parameter matching is completed, parameter instructions containing parameter values, switching opportunities and parameter adaptation degree evaluations are generated; The parameter adaptation degree evaluation is calculated by calculating the matching value of the parameter and the region characteristic, and when the matching value is lower than the preset threshold value, a feedback flag is written to the region characteristic task state collaborative tensor.

7. The apparatus for detecting surface stress of a large-size microcrystalline glass according to claim 2, wherein The interaction verification mechanism of the double-agent module includes: The parameter adjustment agent calculates the parameter adaptation degree value, and when the adaptation degree value reaches the preset standard, the path instructions and the parameter instructions are outputted; When the adaptation degree value is lower than the standard value, the parameter adjustment agent writes a feedback flag to the region characteristic task state collaborative tensor; After the path planning agent detects the feedback flag, the path density value of the corresponding region is increased.

8. The apparatus for detecting surface stress of a large-size microcrystalline glass according to claim 2, wherein, The closed-loop control process includes: The detection host transmits the detection data of a single region to the digital twin processing unit; The digital twin processing unit analyzes the accuracy of the region, compares the measured stress value with the reference value of the standard sample, and calculates the proportion of detection points that meet the accuracy requirements; When the accuracy meets the preset standard, the current strategy effect data is written to the region characteristic task state collaborative tensor; When the accuracy is lower than the standard, the region is marked as a high-uncertainty region, and the multi-source perception unit is instructed to increase the sampling frequency of the region; The virtual physical error calibration is re-executed, and the physical characteristic vector of the region is updated; Based on the calibrated physical characteristic vector, the double-agent module reconstructs the path and parameter instructions of the region and re-executes the detection.

Images