Glass cutting on-line detection and self-adaptive deviation correction system based on machine vision

Abstract

The invention discloses a glass cutting online detection and self-adaptive correction system based on machine vision, and particularly relates to the field of intelligent manufacturing and industrial automatic detection, and the system comprises a high-dynamic imaging module, a stress-position feature decoupling module, an online deep learning decision module, a cooperative execution control module and a self-calibration module. An interference fringe image sequence of cutting stress and cutter position modulation information is generated and captured in real time through a polarized light interference technology, the image sequence is analyzed into independent stress distribution tensor and cutter path coordinates by utilizing a physical information neural network, joint feature description is output, and a recurrent neural network and an attention mechanism are combined to determine the cutting stress and the cutter position modulation information. According to the stress risk, the geometric deviation weight is dynamically adjusted, a three-dimensional self-adaptive compensation instruction is generated, closed-loop control of the cutting process is achieved through cooperation of a decoupled mechanical actuator and a decoupled geometric actuator, synchronous sensing and cooperative regulation and control of a mechanical state and a geometric path are achieved, and the precision and quality of glass cutting are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial automation inspection technology, and more specifically, to a machine vision-based online inspection and adaptive correction system for glass cutting. Background Technology

[0002] In the field of glass deep processing, glass cutting is a critical process with extremely high precision requirements. As high-end manufacturing industries such as display panels and optical devices place increasingly stringent demands on the dimensional accuracy, edge quality, and stress damage resistance of glass products, traditional offline and post-process inspection methods are no longer sufficient to meet the production needs of real-time quality control and process optimization. Therefore, intelligent glass cutting technology based on online inspection has become a current research hotspot and development direction.

[0003] Existing technologies mainly employ online detection methods based on ordinary machine vision. These methods involve acquiring images of the glass behind the cutting tool online using one or more industrial cameras, identifying the actual position of the cutting line using traditional image processing algorithms (such as edge detection and template matching), comparing it with a preset theoretical path, and calculating a single geometric position deviation. Subsequently, the system generates a control signal based on this deviation to drive the cutting platform to perform trajectory correction within the cutting plane (X and Y directions) to achieve the deviation correction function.

[0004] However, in practical use, it still has some shortcomings. For example, the detection dimension is singular, focusing only on geometric position deviations and failing to perceive the dynamic changes in internal stress of the glass during the cutting process. Stress concentration is a key factor leading to the propagation of microcracks and even fracture of the glass. Furthermore, the correction strategy is unidirectional and decoupled, that is, it only adjusts the path based on geometric errors and fails to consider the coupling relationship between process parameters such as cutting pressure and feed speed and stress state. The correction action itself may introduce new stress impacts. Finally, the vision system is susceptible to interference from ambient light, glass surface reflection, and cutting dust, lacks robustness, and lacks system-level online self-optimization capabilities. Its accuracy is prone to drift under long-term operation. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a machine vision-based online inspection and adaptive correction system for glass cutting, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine vision-based online detection and adaptive correction system for glass cutting, comprising a high dynamic imaging module including a polarized light interferometer unit and a high-speed camera. The polarized light interferometer unit generates a dynamic interference field in front of the cutting path, and the high-speed camera is coaxially configured to capture interference fringe image sequences in real time, wherein the interference fringe image sequences are formed by the joint modulation of cutting stress and tool position. Stress-position feature decoupling module: Receives the interference fringe image sequence, incorporates a physical information neural network, and analyzes the fringe distortion field into two sets of independent physical parameters: the cutting stress distribution tensor and the actual tool trajectory coordinates. Outputs a joint feature description including stress risk level and geometric deviation vector. The online deep learning decision module includes a recurrent neural network unit and an attention mechanism unit. The recurrent neural network unit establishes a cutting state evolution model based on the joint feature description on the time series. The attention mechanism unit dynamically allocates the weights of the geometric deviation vector according to the current stress risk level and generates three-dimensional adaptive compensation instructions including cutting pressure, feed rate and tool posture. The collaborative execution control module includes a decoupled mechanical actuator and a geometric actuator. The mechanical actuator dynamically adjusts the tool pressure according to the cutting pressure component in the three-dimensional adaptive compensation command, and the geometric actuator adjusts the tool spatial pose and motion parameters according to the attitude and velocity components in the three-dimensional adaptive compensation command, thereby realizing collaborative closed-loop control of the mechanical state and geometric path of the cutting process. Self-calibration module: Based on the batch cutting results, the boundary condition parameters of the physical information neural network and the initial phase of the polarization interference unit are optimized in reverse to achieve online self-optimization of detection accuracy and system robustness.

[0007] The technical effects and advantages of this invention are as follows: This invention utilizes polarized light interferometry imaging technology to simultaneously decouple the cutting stress distribution from the actual tool trajectory from a single interference fringe image sequence, overcoming the limitation of existing technologies that can only detect geometric deviations but cannot perceive stress states. Furthermore, the system employs a deep learning decision model with an integrated attention mechanism, using stress risk as a dynamic weight to collaboratively generate three-dimensional compensation commands that include both mechanical parameters (pressure) and geometric parameters (pose and velocity). This achieves a leap from "single deviation correction" to "mechanical-geometric collaborative control," fundamentally avoiding secondary stress damage that may be caused by simple path correction.

[0008] This invention employs a physical information neural network for feature decoupling, embedding the photoelasticity equation as a physical constraint into the model training. This ensures that the feature extraction process conforms to fundamental physical laws, significantly improving the accuracy and physical reliability of stress field inversion and trajectory recognition. Simultaneously, the system's built-in self-calibration module optimizes key model parameters and initial imaging system settings based on batch processing results, achieving online self-evolution and long-term maintenance of detection accuracy and system robustness. This effectively overcomes the performance drift problem caused by environmental changes and device aging in traditional systems.

[0009] This invention dynamically adjusts the focus of the control strategy (such as ensuring safety in high-risk situations and ensuring accuracy in low-risk situations) based on the real-time perceived stress risk level, and performs forward-looking compensation based on time-series prediction, enabling the system to cope with normal working conditions and intelligently handle complex situations such as material property fluctuations and abnormal working conditions. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall structure of the present invention; Figure 2 This is a schematic diagram of image acquisition and preprocessing in the high dynamic range imaging module of the present invention; Figure 3 This is a schematic diagram of the stress-position feature decoupling module of the present invention; Figure 4 This is a schematic diagram illustrating the generation of compensation instructions for the online deep learning decision module of the present invention; Figure 5 This is a schematic diagram illustrating the parameter optimization of the self-calibration module of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] As attached Figure 1 The machine vision-based online inspection and adaptive correction system for glass cutting shown includes a high dynamic imaging module, comprising a polarized light interferometer and a high-speed camera. The polarized light interferometer generates a dynamic interference field in front of the cutting path, and the high-speed camera is coaxially configured to capture interference fringe image sequences in real time. The interference fringe image sequences are formed by the modulation of cutting stress and tool position.

[0013] It should be specifically noted that the high dynamic range imaging module includes a polarization interference unit, a high-speed camera, and a synchronization triggering unit; the polarization interference unit includes a high extinction ratio polarizer, an electrically adjustable liquid crystal phase delayer, and a spherical interference lens group, which is installed at a set distance in front of the cutting tool to generate a dynamic interference field; the high-speed camera is coaxially configured with the cutting spindle to capture in real time the interference fringe image sequence formed by the joint modulation of cutting stress and tool position; the synchronization triggering unit is used to realize the synchronous triggering of spindle speed, camera frame rate, and phase delayer modulation frequency.

[0014] As attached Figure 2As shown, the acquisition frame rate and exposure time of the high-speed camera, as well as the modulation frequency of the liquid crystal phase delay, are adaptively and dynamically configured according to the physical characteristics of the material being cut.

[0015] It should be further explained that the polarization interference unit is composed of a high extinction ratio Glan-Taylor polarizer (extinction ratio ≥10000:1), an electrically adjustable liquid crystal phase retarder (response time ≤50μs), and a spherical interference lens group. It is installed at 50-80mm from the front end of the cutting tool and is set in the same plane as the cutting path. The liquid crystal phase retarder is controlled by a high-precision drive power supply and dynamically adjusts the phase of the interference field according to the characteristics of the cutting material (adjustment range 0-2π) to generate a dynamic interference field that adapts to the stress response characteristics of different materials.

[0016] The high-speed camera uses a single-photon-level CMOS high-speed camera with a pixel resolution of 1024×1024 and a frame rate range of 1000-10000fps. It is equipped with a 50mm fixed-focus industrial lens with an aperture of F1.4 and an adjustable focal length. The camera and the cutting spindle adopt a coaxial optical design, and the interference light and imaging light are transmitted coaxially through a semi-transparent and semi-reflective mirror to ensure that the captured interference fringe image is strictly aligned with the tool position and avoid parallax error.

[0017] The synchronous triggering unit uses an FPGA as the core control chip and is connected to the high-speed camera, the cutting spindle encoder, and the LCD phase delay drive power supply respectively to achieve three-in-one synchronization of "spindle speed - camera frame rate - phase delay modulation frequency" with a synchronization error ≤1μs.

[0018] The acquired data consists of a "dynamic interference fringe image sequence", supplemented by "camera frame synchronization trigger signal", "spindle encoder position signal" and "liquid crystal phase delay modulation parameters".

[0019] Selection criteria: The degree of distortion of interference fringes is directly related to the cutting stress. Based on the photoelastic effect, the greater the stress, the more significant the fringe distortion. The spatial position offset of the fringes is directly related to the actual tool trajectory. By carrying two sets of physical information simultaneously in the same image sequence, the spatiotemporal registration error of multi-sensor data can be avoided. The synchronous trigger signal and modulation parameters are used as auxiliary data to eliminate the phase noise caused by the dynamic modulation of the interference field in the subsequent decoupling module, thereby improving the decoupling accuracy.

[0020] The acquisition parameters are dynamically adjusted based on the stress wave propagation velocity (1500-3000 m / s) and cutting speed (0.5-5 m / s) of the material being cut. For high-hardness materials (Vickers hardness HV≥300, elastic modulus E≥200 GPa, stress wave propagation velocity ≥2500 m / s, and stress response exhibiting brittle characteristics), the camera frame rate is set to 8000-10000 fps, pixel exposure time to 50-100 ns, and liquid crystal phase delay modulator modulation frequency to 1000 Hz. For flexible materials (Vickers hardness HV<150, elastic modulus E<50 GPa, stress wave propagation velocity ≤2000 m / s, and stress response exhibiting plastic deformation characteristics), the frame rate is set to 3000-5000 fps, exposure time to 100-200 ns, and modulation frequency to 500 Hz. The single acquisition cycle is 10 ms, and each group acquires 100-500 frames of image sequences to form a continuous time-series data stream.

[0021] The acquired raw image sequence is preprocessed to eliminate noise and systematic errors, providing high-quality data for subsequent decoupling. This preprocessing includes dark field correction: acquiring dark field images without interference light illumination, and performing pixel-level subtraction between the original image and the dark field image. Specifically, if the grayscale value of a pixel in the original image is G1(x,y) and the corresponding grayscale value of the pixel in the dark field image is G2(x,y), for each pixel with identical coordinates in both images, the processed pixel grayscale value G(x,y) = G1(x,y) - G2(x,y) is calculated. This removes fixed background noise from the original image and retains the effective signal of interference fringes generated only by the cutting stress and tool position modulation, improving the accuracy of subsequent image feature extraction.

[0022] Eliminate camera CMOS dark current noise; Phase noise elimination: Based on the acquired liquid crystal phase delay modulator modulation parameters, Fourier filtering algorithm is used to extract the fundamental frequency component of interference fringes and filter out high-frequency phase noise generated by dynamic modulation; Image registration: Using the spindle encoder position signal as a reference, multiple frames of images corresponding to the same cutting position are translated and rotated to eliminate image offset caused by cutting vibration; Dimensionality reduction and enhancement: Wavelet transform algorithm is used to reduce the dimension of the image, retaining the low-frequency components of fringe distortion, while contrast-limited adaptive histogram equalization (CLAHE) is used to enhance the fringe edge details, outputting a preprocessed image sequence with a dimension of 256×256.

[0023] Stress-position feature decoupling module: Receives the interference fringe image sequence, incorporates a physical information neural network, and analyzes the fringe distortion field into two sets of independent physical parameters: the cutting stress distribution tensor and the actual tool trajectory coordinates. Outputs a joint feature description including stress risk level and geometric deviation vector.

[0024] It should be specifically noted that the physical information neural network includes a physically constrained convolutional layer, an attention decoupling layer, and a dual-output layer; wherein the physically constrained convolutional layer embeds the photoelasticity equation as a constraint in the convolution operation; the attention decoupling layer integrates channel attention and spatial attention mechanisms to separate stress-related features and position-related features; the dual-output layer is used to output the cutting stress distribution tensor and the actual trajectory coordinates of the tool, respectively.

[0025] As attached Figure 3 As shown, it should be further explained that the stress-location feature decoupling module incorporates a Physical Information Neural Network (PINN), which includes an input layer, a physically constrained convolutional layer, an attention decoupling layer, and a dual output layer, with the specific structure as follows: Input layer: Receives the preprocessed interference fringe image sequence (dimension 256×256×T, where T is the time step, set to 10), and simultaneously inputs auxiliary data (modulation parameters of liquid crystal phase delay device, physical parameters of cutting material). Physically constrained convolutional layers: 3×3 convolutional kernels (64 in total) are used. During convolution, photoelasticity equations (physical constraint terms) are embedded. This means that the output features are constrained to meet the physical relationship of "stress-optical path difference" through the loss function, preventing the model output from violating fundamental physical laws. The specific photoelasticity equations are as follows: Combined with the optical path difference formula A quantitative relationship between stress and stripe distortion was constructed, in which... The change in the refractive index of the medium. The photoelastic coefficient of the material. These are two principal stresses in the plane. The thickness is the distance through which light propagates in the material.

[0026] Attention Decoupling Layer: Introduces a fusion module of channel attention mechanism and spatial attention mechanism. Channel attention is used to distinguish between "stress-related feature channels" and "position-related feature channels", while spatial attention is used to locate "stress-induced diffusion distortion" and "position-induced translation distortion" in stripe distortion, achieving dual separation of spatial and channel features. Dual output layers: output the cutting stress distribution tensor (dimension: H×W×3, where H and W are the image height and width, and 3 corresponds to the stress components in the x, y, and z directions) and the actual tool trajectory coordinates (dimension: 3×1, corresponding to the three-dimensional coordinates of x, y, and z).

[0027] Selected data: The training data includes "preprocessed interference fringe image sequence under multiple materials and working conditions", "synchronously acquired measured stress data", "measured tool trajectory data", and "material physical parameter library".

[0028] The 100,000 sets of training data were divided into training, validation, and test sets in a 7:2:1 ratio. The stress data were normalized (mapped to the [0,1] interval), and the trajectory coordinate data were standardized (to eliminate the influence of dimensions). Data augmentation techniques were used to expand the amount of training data by rotating (±5°), scaling (0.9-1.1 times), and adding Gaussian noise (signal-to-noise ratio ≥30dB) the interference fringe image.

[0029] The preprocessed interference fringe image sequence and auxiliary data are input into the PINN model. A physically constrained convolutional layer extracts image features while satisfying the constraints of the photoelasticity equations, outputting a feature map fused with physical information. A fusion attention module separates the stress feature map and position feature map, which are then input into their respective fully connected layers. The fully connected layers output the cutting stress distribution tensor and the actual tool trajectory coordinates. Based on the stress distribution tensor, the maximum principal stress value is calculated, specifically by obtaining a three-dimensional cutting stress distribution tensor from decoupling, including normal stresses in the x, y, and z directions. and shear stress The tensor form is The principal stresses σ1, σ2, and σ3 are obtained by solving the principal stress formula, and σ1 ≥ σ2 ≥ σ3. Among them, σ1, which has the largest value, is the maximum principal stress value.

[0030] Based on the preset stress risk level threshold (set according to the material fracture strength), the stress risk level (low risk, medium risk, high risk) is output. Low risk means that the maximum principal stress value during the cutting process is ≤ 50% of the material's allowable stress, and there is no risk of plastic deformation of the material; medium risk means that the maximum principal stress value is between 50% and 80% of the material's allowable stress, and the material may undergo slight plastic deformation; high risk means that the maximum principal stress value is > 80% of the material's allowable stress, and the material has the risk of fracture or serious damage. Based on the difference between the actual and theoretical trajectory coordinates of the tool, the geometric deviation vectors Δx, Δy, and Δz are calculated, and finally, a joint feature description of "stress risk level - geometric deviation vector" is output.

[0031] The online deep learning decision module includes a recurrent neural network unit and an attention mechanism unit. The recurrent neural network unit establishes a cutting state evolution model based on the joint feature description on the time series, and the attention mechanism unit dynamically allocates the weights of the geometric deviation vector according to the current stress risk level to generate three-dimensional adaptive compensation instructions including cutting pressure, feed rate and tool posture.

[0032] It should be specifically noted that, based on the stress risk level, the attention mechanism unit dynamically adjusts the weight coefficients of each directional component of the geometric deviation vector using a predefined adaptive weight allocation rule, and, combined with the cutting state prediction value output by the recurrent neural network unit, generates the three-dimensional adaptive compensation instruction through a multi-objective optimization algorithm.

[0033] As attached Figure 4 As shown, it is necessary to further explain that the recurrent neural network (RNN) unit adopts a long short-term memory network (LSTM) structure. The input is a joint feature description on the time series (20 time steps in length, each time step including stress risk level and geometric deviation vector). Through the gating mechanism of LSTM (input gate, forget gate, output gate), it captures the temporal evolution law of the cutting state and establishes a cutting state evolution model. The model output is the cutting state prediction value (stress risk level prediction, geometric deviation vector prediction) for the next 5 time steps, which is used to predict the trend of cutting state changes in advance.

[0034] An adaptive weight allocation mechanism is adopted, using the current stress risk level as the trigger condition to dynamically adjust the weights of the deviations in the x, y, and z directions of the geometric deviation vector. Specific rules are as follows: When the stress risk level is "low risk," the focus is on geometric accuracy, with a weight of 0.4 for the x and y directions (trajectory plane) and 0.2 for the z direction (cutting depth); when the stress risk level is "medium risk," a balance is struck between mechanical state and geometric accuracy, with a weight of 0.3 for the x and y directions and 0.3 for the z direction, while also introducing a stress change rate weight of 0.1; when the stress risk level is "high risk," mechanical safety is prioritized, with a stress change rate weight of 0.5 and x, y, and z direction deviation weights of 0.17 each.

[0035] The selected data are time-series joint feature descriptions, historical cutting compensation instructions, and historical cutting quality feedback data.

[0036] Sliding window sampling with a window size of 20 is used to sample the temporal joint feature description to form temporal samples; historical compensation instructions are normalized and mapped to [0,1] according to the maximum value range of cutting parameters; cutting quality feedback data is graded into levels 1-5, which are used as inputs to the reward function. Level 1 (optimal) is cutting surface roughness Ra≤0.8μm and dimensional error≤±0.01mm, Level 2 is Ra≤1.6μm and dimensional error≤±0.02mm, Level 3 is Ra≤3.2μm and dimensional error≤±0.03mm, Level 4 is Ra≤6.3μm and dimensional error≤±0.05mm, and Level 5 (worst) is Ra>6.3μm or dimensional error>±0.05mm.

[0037] The temporal joint feature description after sliding window sampling is input into the LSTM unit; the LSTM unit outputs the cutting state prediction values ​​for the next 5 time steps; according to the current stress risk level, the attention mechanism unit assigns weights to each direction of the geometric deviation vector and stress change rate weights. For low risk, geometric accuracy is emphasized, with higher weights in the x and y directions; for medium risk, mechanical and geometric performance are balanced, with weights evenly distributed across directions and stress change rate weights introduced; for high risk, mechanical safety is prioritized, with the stress change rate weight being the highest; based on the state prediction values ​​and dynamic weights, a multi-objective optimization algorithm (NSGA-Ⅲ) is used, with the optimization objectives being "minimize geometric deviation" and "minimize stress risk level," and the constraints being the physical limits of the cutting parameters (e.g., cutting pressure range 0-500N, feed rate range 0.5-5m / s). The algorithm outputs three-dimensional adaptive compensation commands for cutting pressure, feed rate, and tool attitude (pitch angle, yaw angle); a first-order low-pass filtering algorithm is used to smooth the compensation commands, avoiding secondary stress impacts caused by sudden changes in cutting parameters. The specific formula is as follows:

[0038] in, This is the current filter output value. Enter the value for the current compensation instruction. This is the filtered output value from the previous moment. These are the filter coefficients, with values ​​ranging from 0 to... <1.

[0039] The collaborative execution control module includes a decoupled mechanical actuator and a geometric actuator. The mechanical actuator dynamically adjusts the tool pressure according to the cutting pressure component in the three-dimensional adaptive compensation command, while the geometric actuator adjusts the tool spatial pose and motion parameters according to the attitude and velocity components in the three-dimensional adaptive compensation command, thereby realizing collaborative closed-loop control of the mechanical state and geometric path of the cutting process.

[0040] It should be specifically noted that the mechanical actuator adopts a piezoelectric ceramic micro-displacement actuator and integrates a force sensor; the geometric actuator adopts a six-degree-of-freedom parallel robot and integrates a displacement sensor; the mechanical actuator and the geometric actuator are decoupled and synchronized through a real-time control unit.

[0041] It should be further explained that the mechanical actuator is a piezoelectric ceramic micro-displacement actuator, equipped with a force sensor with a measurement range of 0-500N and an accuracy of ±0.1N. It is installed at the tool holder of the cutting tool and is responsible for receiving the cutting pressure component in the three-dimensional adaptive compensation command. The dynamic adjustment of the tool pressure is achieved through the micro-displacement drive of the piezoelectric ceramic (response time ≤10μs).

[0042] The geometric actuator employs a six-degree-of-freedom parallel robot equipped with a grating ruler displacement sensor with a measurement accuracy of ±0.01μm. It is responsible for receiving the feed rate and tool attitude components in the compensation command, and adjusting the spatial pose of the tool through six-degree-of-freedom linkage, with a pitch angle range of ±5° and a deflection angle range of ±5°, along with motion parameters, to achieve precise correction of the cutting path (response time ≤1ms).

[0043] The real-time control unit uses a DSP as the core controller and is connected to two actuators and a high dynamic imaging module to achieve real-time closed-loop control of "compensation command - actuator action - status feedback" with a control cycle of ≤100μs.

[0044] The selected data include three-dimensional adaptive compensation commands, force feedback data from mechanical actuators, displacement feedback data from geometric actuators, and interference fringe images acquired in real time by a high-dynamic imaging module.

[0045] The force feedback data and displacement feedback data are filtered using the Kalman filter algorithm with a filter coefficient of 0.05 to eliminate noise caused by actuator vibration. Feature values ​​(fringe distortion degree) are extracted from the real-time interference fringe image and compared with the feature values ​​before compensation. The compensation effect evaluation index (distortion elimination rate, supplemented by a comprehensive judgment of force error and displacement error) is calculated.

[0046] The real-time control unit decomposes the three-dimensional adaptive compensation command into mechanical control sub-commands (cutting pressure) and geometric control sub-commands (feed speed, tool posture), which are sent to the mechanical actuator and the geometric actuator respectively. The mechanical actuator adjusts the piezoelectric ceramic driving voltage according to the mechanical control sub-command to achieve dynamic adjustment of the cutting pressure. The geometric actuator drives the six-degree-of-freedom parallel robot according to the geometric control sub-command to adjust the tool posture and feed speed. The force feedback data of the mechanical actuator and the displacement feedback data of the geometric actuator are collected in real time, and the compensated interference fringe image is also collected. The feedback data is compared with the command target value to calculate the execution error. If the error exceeds the threshold (force error > 0.5N, displacement error > 0.01mm), a correction command is generated to adjust the actuator action. At the same time, the compensation effect is fed back to the online deep learning decision module according to the compensation effect evaluation index of the interference fringe image for subsequent decision optimization. Through the clock synchronization mechanism of the real-time control unit, the actions of the mechanical actuator and the geometric actuator are strictly synchronized (synchronization error ≤ 1μs) to avoid the decrease in cutting quality caused by the timing deviation of the action.

[0047] Self-calibration module: Based on the batch cutting results, the boundary condition parameters of the physical information neural network and the initial phase of the polarization interference unit are optimized in reverse to achieve online self-optimization of detection accuracy and system robustness.

[0048] It should be specifically noted that the self-calibration module connects the high dynamic imaging module and the stress-position feature decoupling module, and is used to dynamically adjust the boundary condition parameters of the physical information neural network and the initial phase of the polarization interference unit based on the batch cutting result data through a reverse optimization algorithm.

[0049] The self-calibration module aims to minimize the average dimensional error and stress detection error of batch cutting, and uses a particle swarm optimization algorithm to jointly optimize the boundary condition parameters and the initial phase.

[0050] It should be further explained that a multi-parameter inverse optimization model is constructed, with the optimization objectives being "minimizing the average size error of batch cutting" and "minimizing the stress detection error", and the optimization variables being the boundary condition parameters (constraint coefficients of the photoelasticity equation) of the physical information neural network and the initial phase of the polarization interference unit.

[0051] As attached Figure 5 As shown, the selected data includes batch cutting results, detection data during the batch cutting process, and system operating environment data.

[0052] Statistical analysis was performed on the batch cutting results data to calculate the average dimensional error, maximum dimensional error, and stress detection error; environmental data was normalized and used as an interference term in the optimization model; correlation analysis was used to link the process detection data and the result data to locate error-sensitive links.

[0053] After completing a batch of cutting (batch quantity ≥ 50 pieces), collect cutting result data, process detection data, and environmental data; calculate the batch average dimensional error and stress detection error. If the error exceeds the preset threshold (dimensional error > 0.05 mm, stress detection error > 5 MPa), initiate the self-calibration process; input the error data into the inverse optimization model, and use the particle swarm optimization algorithm (PSO) to optimize the boundary condition parameters of the physical information neural network (adjusting the constraint coefficients of the photoelasticity equation) and the initial phase of the polarization interference unit (achieved by adjusting the initial driving voltage of the liquid crystal phase delay unit); write the optimized parameters into the corresponding modules (physical information neural network, polarization interference unit driving power supply) to complete the parameter update; select 3-5 pieces Cutting tests are conducted on similar parts to collect test data and verify the calibration effect. If the error still does not meet the standard, steps 3-4 are repeated. If the error still does not meet the standard after repeating steps 3-4 (parameter optimization-parameter update) 3 times, the system automatically triggers an anomaly warning mechanism and performs the following operations: calling the historical optimal parameter library, matching the current cutting material and working conditions, loading historically verified and effective calibration parameters as a temporary solution to ensure the stable operation of subsequent cutting operations; outputting an anomaly diagnosis report, clearly indicating the core reasons that may have caused the optimization failure, guiding manual troubleshooting and repair, and restarting the self-calibration process after the fault is eliminated; associating and storing the optimized parameters with the corresponding environmental data and cutting condition data to establish a parameter library for rapid calibration under similar working conditions in the future.

[0054] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 machine vision-based online inspection and adaptive correction system for glass cutting, characterized in that, include: High dynamic imaging module: includes a polarized light interferometer unit and a high-speed camera. The polarized light interferometer unit generates a dynamic interference field in front of the cutting path. The high-speed camera is coaxially configured to capture interference fringe image sequences in real time. The interference fringe image sequences are formed by the co-modulation of cutting stress and tool position. Stress-position feature decoupling module: Receives the interference fringe image sequence, incorporates a physical information neural network, and analyzes the fringe distortion field into two sets of independent physical parameters: the cutting stress distribution tensor and the actual tool trajectory coordinates. Outputs a joint feature description including stress risk level and geometric deviation vector. The online deep learning decision module includes a recurrent neural network unit and an attention mechanism unit. The recurrent neural network unit establishes a cutting state evolution model based on the joint feature description on the time series. The attention mechanism unit dynamically allocates the weights of the geometric deviation vector according to the current stress risk level and generates three-dimensional adaptive compensation instructions including cutting pressure, feed rate and tool posture. The collaborative execution control module includes a decoupled mechanical actuator and a geometric actuator. The mechanical actuator dynamically adjusts the tool pressure according to the cutting pressure component in the three-dimensional adaptive compensation command, and the geometric actuator adjusts the tool spatial pose and motion parameters according to the attitude and velocity components in the three-dimensional adaptive compensation command, thereby realizing collaborative closed-loop control of the mechanical state and geometric path of the cutting process. Self-calibration module: Based on the batch cutting results, the boundary condition parameters of the physical information neural network and the initial phase of the polarization interference unit are optimized in reverse to achieve online self-optimization of detection accuracy and system robustness.

2. The machine vision-based online inspection and adaptive correction system for glass cutting according to claim 1, characterized in that: The high dynamic imaging module includes a polarization interference unit, a high-speed camera, and a synchronization triggering unit. The polarization interference unit includes a high extinction ratio polarizer, an electrically adjustable liquid crystal phase delayer, and a spherical interference lens group, which is installed at a set distance in front of the cutting tool to generate a dynamic interference field. The high-speed camera is coaxially configured with the cutting spindle to capture in real time the interference fringe image sequence formed by the joint modulation of cutting stress and tool position. The synchronization triggering unit is used to realize the synchronous triggering of spindle speed, camera frame rate, and phase delayer modulation frequency.

3. The machine vision-based online inspection and adaptive correction system for glass cutting according to claim 2, characterized in that: The frame rate and exposure time of the high-speed camera, as well as the modulation frequency of the liquid crystal phase delay, are adaptively and dynamically configured according to the physical characteristics of the material being cut.

4. The machine vision-based online inspection and adaptive correction system for glass cutting according to claim 1, characterized in that: The physical information neural network includes a physically constrained convolutional layer, an attention decoupling layer, and a dual-output layer; wherein the physically constrained convolutional layer embeds the photoelasticity equation as a constraint in the convolution operation. An attention decoupling layer integrates channel attention and spatial attention mechanisms to separate stress-related features from location-related features. The dual output layers are used to output the cutting stress distribution tensor and the actual tool trajectory coordinates, respectively.

5. The machine vision-based online inspection and adaptive correction system for glass cutting according to claim 1, characterized in that: Based on the stress risk level, the attention mechanism unit dynamically adjusts the weight coefficients of each directional component of the geometric deviation vector using a predefined adaptive weight allocation rule. Combined with the cutting state prediction value output by the recurrent neural network unit, the unit generates the three-dimensional adaptive compensation instruction through a multi-objective optimization algorithm.

6. The machine vision-based online inspection and adaptive correction system for glass cutting according to claim 1, characterized in that: The mechanical actuator is a piezoelectric ceramic micro-displacement actuator with an integrated force sensor; the geometric actuator is a six-degree-of-freedom parallel robot with an integrated displacement sensor; the mechanical actuator and the geometric actuator are decoupled and synchronized through a real-time control unit.

7. The machine vision-based online inspection and adaptive correction system for glass cutting according to claim 1, characterized in that: The self-calibration module connects the high dynamic imaging module and the stress-position feature decoupling module. It is used to dynamically adjust the boundary condition parameters of the physical information neural network and the initial phase of the polarization interference unit based on the batch cutting result data through a reverse optimization algorithm.

8. The machine vision-based online inspection and adaptive correction system for glass cutting according to claim 7, characterized in that: The self-calibration module aims to minimize the average dimensional error and stress detection error of batch cutting, and uses a particle swarm optimization algorithm to jointly optimize the boundary condition parameters and the initial phase.

Images