A multimodal fusion edge detection method for cutting glass

Abstract

This invention relates to the field of glass inspection technology, and more particularly to a multimodal fusion edge detection method for cut glass. The method includes assessing the impact of operating condition fluctuations, constructing a multimodal fusion assessment network based on a historical database, and calculating an initial edge defect risk coefficient; dynamically allocating weights for each detection mode and generating an edge quality index, thereby adjusting the laser ultrasonic scanning range and the sampling frequency of the line scan camera; using the spatial consistency coefficient of each mode's results to determine diagnostic reliability, calculating the diagnostic deviation value when it fails to meet the standard, and optimizing the monitoring cycle and model update based on the cumulative deviation. This invention solves the technical problem that existing detection methods cannot fuse visible light images, polarized light stress distribution, and laser ultrasonic echo information online, making it difficult to perform real-time spatial localization and risk quantification assessment of subsurface hidden microcracks.

Description

Technical Field

[0001] This invention relates to the field of glass inspection technology, and in particular to a multimodal fusion edge detection method for cutting glass. Background Technology

[0002] In the glass substrate cutting process of LCD panels, the quality of the cut edge directly affects the yield of subsequent processes such as polarizer attachment and defoaming. During mechanical cutting, the glass substrate is subjected to shear stress, which not only produces surface defects such as edge chipping and serrated cracks at the cut edge, but may also form hidden microcracks on the subsurface that have not yet extended to the surface. These subsurface defects are difficult to detect by conventional visual inspection methods after cutting, but they will gradually expand under the thermal and mechanical stress of subsequent processes, eventually leading to a decrease in panel strength or even delayed fracture, becoming a bottleneck problem for improving the yield of narrow bezel cutting in high-generation LCDs.

[0003] Currently, production lines commonly use line-scan cameras for online imaging and inspection of cutting edges, identifying surface chipping and obvious cracks through grayscale thresholding or deep learning algorithms. However, visible light imaging is limited by physical principles and cannot detect subsurface microcracks that have not yet formed morphological changes on the surface. While laser ultrasonic technology can invert crack depth and location through acoustic echoes, the traditional C-scan point-by-point inspection mode is extremely slow and cannot match the continuous operation cycle of the production line. Polarized stress imaging can characterize the residual stress distribution at the cutting edge, but traditional photoelastic methods are only used as offline laboratory tools. These three detection methods operate independently, and the data are not correlated, lacking a technical solution for online fusion and evaluation of apparent geometric features, subsurface crack information, and residual stress distribution.

[0004] Therefore, there is an urgent need for a detection method that can integrate multiple physical field information online and achieve a comprehensive evaluation of both apparent and hidden defects at the cutting edge, in order to meet the high standards of cutting edge quality required by high-generation LCD panels.

[0005] Chinese Patent Publication No. CN121837139A discloses a method and apparatus for detecting defects in substrate glass, relating to the field of glass inspection technology. The method includes: Step S1, a line scan imaging mechanism acquires a first image of the substrate glass to be inspected, the first image being an overall image including the edge region and the middle region of the substrate glass; Step S2, a detection module identifies defect regions of the substrate glass to be inspected based on the first image and marks the defect coordinates of the defect regions; Step S3, based on the defect coordinates, a planar array imaging mechanism is triggered to magnify the defect regions n times to form a second image; Step S4, a processing module classifies the defect types of the defect regions based on the second image; Step S5, a grade determination is made based on the inspection results of the substrate glass to be inspected.

[0006] However, the aforementioned method and apparatus for detecting defects in substrate glass have the following problems:

[0007] This scheme uses a visible light detection route that combines line scanning imaging and area array secondary imaging. It relies on only a single physical field mode, which inherently has a blind spot for detecting hidden microcracks on the subsurface. Furthermore, the detection parameters are fixed and lack the ability to perform multimodal fusion evaluation and adaptive optimization based on diagnostic bias. As a result, it is difficult to effectively detect and quantify the risk of subsurface microcracks on the cutting edge. Summary of the Invention

[0008] To address this issue, the present invention provides a multimodal fusion edge detection method for cutting glass, which overcomes the problem in the prior art that it is impossible to fuse visible light images, polarized light stress distribution and laser ultrasonic echo information online, making it difficult to perform real-time spatial positioning and risk quantification assessment of subsurface hidden microcracks.

[0009] To achieve the above objectives, this invention provides a multimodal fusion edge detection method for cutting glass. It includes:

[0010] The impact of operating condition interference characteristics on the detection mode of edge detection is determined based on the operating conditions of the preset monitoring cycle.

[0011] A multimodal fusion evaluation network was constructed based on a historical database of glass cutting edge defects.

[0012] The edge defect risk characteristics are determined by the initial edge defect risk characteristics output by the feature recognition output of several sets of feature data of the detection modality based on the multimodal fusion evaluation network, and the edge risk status of the cut glass is determined.

[0013] Based on the comprehensive decision coefficient of fusion working condition interference and defect risk, the weight allocation of each detection mode is determined in order to assess the edge quality of the cut glass edge to be tested.

[0014] Based on the edge quality assessment results, the area range for directional scanning of the glass cutting edge by the laser ultrasonic probe is determined, and the line trigger sampling frequency of the line scan camera is adjusted simultaneously.

[0015] The diagnostic deviation characteristics of a single test are determined based on the spatial consistency between the test results of each modality.

[0016] Based on the cumulative deviation characteristics of the preset sliding window, the monitoring period for the adjustment condition interference characteristics and the sample sampling ratio when the multimodal fusion evaluation network is updated are determined.

[0017] Furthermore, the aforementioned operating condition interference characteristics include a tool wheel wear interference component, a thickness fluctuation interference component, and a light source attenuation interference component, wherein,

[0018] The wear interference component of the cutter wheel is determined based on the ratio of the cumulative cutting mileage of the cutter wheel to the rated life mileage of the cutter wheel;

[0019] The thickness fluctuation interference component is determined based on the ratio of the thickness value at the current cutting position of the glass to the standard thickness value.

[0020] The light source attenuation interference component is determined based on the ratio of the current cumulative lighting time of the light source of the line scan camera to the rated lifespan of the light source.

[0021] Furthermore, the plurality of sets of feature data include a visible light feature data set, a polarized light stress feature data set, and an ultrasonic acoustic feature data set;

[0022] The initial edge defect risk characteristic is characterized by the initial edge defect risk characteristic coefficient, wherein...

[0023] The initial edge defect risk characteristic coefficient is determined based on the sum of the probabilities of four types of defects, and is used to determine the edge risk status of the cut glass.

[0024] Furthermore, the comprehensive decision-making risk coefficient is determined based on the product of operating condition disturbance and defect risk, wherein,

[0025] The operating condition interference is the operating condition interference coefficient, which is the degree of interference of the current operating condition on the detection mode.

[0026] The working condition interference coefficient is determined based on the tool wheel wear interference component, the thickness fluctuation interference component, and the light source attenuation interference component.

[0027] The defect risk is the initial edge defect risk characteristic coefficient.

[0028] Furthermore, the weight allocation includes,

[0029] The first weight allocation is an ultrasound-priority weight allocation, determined based on the comprehensive decision coefficient being less than or equal to the first preset comprehensive decision coefficient.

[0030] The second weight allocation is a balanced and collaborative weight allocation, which is determined based on the comprehensive decision coefficient being greater than the first preset comprehensive decision coefficient and less than or equal to the second preset comprehensive decision coefficient.

[0031] The third weight allocation is a visible light priority weight allocation, which is determined based on the comprehensive decision coefficient being greater than the second preset comprehensive decision coefficient.

[0032] Furthermore, the edge quality assessment result includes an edge quality index, wherein,

[0033] The edge quality index, determined based on the sum of the probability of edge chipping defects and the probability of subsurface hidden microcracks, is used to determine the integrity of the edge of the current detection area in order to execute the corresponding detection strategy. The process of determining the area range for directional scanning by the laser ultrasonic probe along the cutting edge based on the edge quality assessment index includes...

[0034] Based on the fact that the weighted edge quality index is less than or equal to the first preset quality index, it is determined that the directional scanning area of ​​the laser ultrasound probe will be narrowed down to cover only the core area of ​​the polarization stress gradient anomaly region.

[0035] Furthermore, the detection strategy includes,

[0036] The first detection strategy involves narrowing the directional scanning area of ​​the laser ultrasonic probe to cover only the core area of ​​the polarized light stress gradient anomaly region, and reducing the line-triggered sampling frequency of the line scan camera, in order to eliminate the risk of subsurface cracks.

[0037] Furthermore, the detection strategy also includes,

[0038] The second detection strategy is to expand the directional scanning area of ​​the laser ultrasonic probe to cover the entire polarized light stress gradient anomaly area and its outer boundary, and to keep the line trigger sampling frequency of the line scan camera at the standard frequency, so as to accurately identify the correlation between the chipping size and the subsurface crack, and accurately distinguish between sawtooth cracks and hidden microcracks.

[0039] The third detection strategy involves further expanding the directional scanning area of ​​the laser ultrasound probe to cover the entire polarized light stress gradient anomaly area and extending it forward and backward along the cutting edge, while increasing the line-triggered sampling frequency of the line scan camera.

[0040] Furthermore, the spatial location consistency includes a spatial location consistency coefficient, wherein,

[0041] The spatial position consistency coefficient is determined based on the set of visible light profile anomaly points, the set of polarized light stress anomaly regions, and the set of ultrasonic crack location points.

[0042] The diagnostic deviation feature includes a diagnostic deviation value, which is determined based on the accuracy deviation of defect type classification, the accuracy deviation of defect boundary positioning, and the crack depth estimation error, and is used to quantify the degree of deviation between a single detection result and the actual defect state.

[0043] Further, the cumulative deviation feature includes a cumulative diagnostic deviation value, wherein,

[0044] The cumulative deviation value is determined based on the arithmetic mean of all diagnostic deviation values ​​within a preset sliding window, and is used to measure the overall degradation trend of the diagnostic performance of the detection system during continuous operation, so as to implement corresponding optimization strategies.

[0045] The optimization strategy includes,

[0046] The first optimization strategy is to shorten the duration of the preset monitoring cycle;

[0047] The second optimization strategy is to adjust the sample sampling ratio when updating the multimodal fusion evaluation network.

[0048] Compared with existing technologies, the advantages of this invention lie in its ability to quantify the wear of the cutting wheel, substrate thickness fluctuations, and light source attenuation into interference components, thereby establishing a quantitative mapping relationship between operating condition drift and detection signal quality. Specifically, cutting wheel wear alters the initiation depth and extension direction of subsurface cracks, thickness fluctuations increase the dispersion of crack morphology, and light source attenuation weakens the ability of visible light images to identify minute defects. By synthesizing these three types of interference components in real time, the fuzzy changes in production line conditions are transformed into precise interference level indicators. This provides a quantitative basis for the dynamic adjustment of detection mode weights, enabling the detection system to adaptively adjust its dependence on each mode according to operating conditions, thus maintaining the consistency and reliability of detection results under fluctuating operating conditions.

[0049] Furthermore, this invention constructs a multimodal fusion evaluation network and introduces a cross-modal attention fusion module into the architecture, enabling the network to autonomously learn the correlation and verification patterns between different modal features during training. This significantly improves the ability to identify hidden microcracks and to finely distinguish between chipped edges and sawtooth cracks, solving the technical problem of the complete blind zone of subsurface defects in traditional single visual detection methods.

[0050] Furthermore, this invention, through real-time high-precision processing of three modal signals, explicitly outputs the spatial coordinate sequence and quantization error parameters required for subsequent diagnosis during the feature extraction stage, providing an accurate numerical basis for subsequent steps. At the same time, by utilizing the comprehensive reasoning ability of deep learning networks for multi-dimensional features, the original physical signal is transformed into a quantized initial risk probability, resulting in high detection sensitivity for extremely weak early subsurface cracks.

[0051] Furthermore, this invention quantifies and couples operating condition interference and defect risk through comprehensive decision coefficients, constructing a dynamic allocation mechanism for detection mode weights. Under low-risk, favorable operating conditions, visible light defect characteristics are weak, and the ultrasonic mode has high specificity for subsurface cracks, thus increasing its weight ensures zero missed detection of hidden defects. Under high-risk, harsh operating conditions, wear of the cutting wheel and attenuation of the light source lead to a decrease in the ultrasonic signal-to-noise ratio, while the visible light channel is less affected by interference, thus increasing its weight ensures the measurement accuracy of edge chipping and sawtooth cracks, thereby achieving an adaptive balance between detection sensitivity and anti-interference robustness.

[0052] Furthermore, this invention quantifies the integrity of the cutting edge through an edge quality index and adaptively adjusts the laser ultrasonic scanning range and the sampling frequency of the line scan camera accordingly. When the edge quality is excellent, the scanning area is reduced and the sampling frequency is lowered to save detection resources; when the risk is moderate, the scanning is expanded to the outer boundary of the stress anomaly zone, while maintaining the standard frequency to cover the area associated with edge chipping and microcracks; when the quality is poor, the scanning range is further expanded and the sampling frequency is increased to capture the edge chipping morphology and crack tip position with ultra-high resolution, providing accurate defect location and type differentiation basis for process adjustment, and achieving seamless scheduling of detection resources from macroscopic screening to microscopic positioning.

[0053] Furthermore, this invention establishes a multimodal cross-validation mechanism by calculating the spatial overlap between visible light profile anomalies, polarized light stress anomalies, and ultrasonic crack location points. When the three spatially overlap significantly, the diagnostic result is considered reliable; when their spatial distribution deviates significantly, the diagnostic result is considered unreliable. This mechanism automatically identifies diagnostic inconsistencies caused by sensor noise or random factors, selects high-confidence results for process improvement, and quantifies the deviation of low-confidence results as deviation values, providing error feedback data for optimizing the detection model. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the steps of a multimodal fusion edge detection method for cutting glass according to an embodiment of the present invention;

[0055] Figure 2 This is a logic block diagram of how the corresponding detection strategy is determined based on the edge quality index in an embodiment of the present invention.

[0056] Figure 3 This is a logic block diagram illustrating how the reliability of a single fusion diagnostic result is determined based on the spatial consistency coefficient in an embodiment of the present invention.

[0057] Figure 4 This is a logic block diagram of how the corresponding optimization strategy is determined based on the cumulative diagnostic deviation value in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0059] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0060] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0061] Please see Figure 1 As shown, it is a schematic diagram of the steps of the multimodal fusion edge detection method for cutting glass according to an embodiment of the present invention;

[0062] The multimodal fusion edge detection method for cutting glass according to embodiments of the present invention includes:

[0063] Step S1: Determine the impact of operating condition interference characteristics on the detection mode of edge detection based on the operating conditions of the preset monitoring cycle;

[0064] Step S2: Construct a multimodal fusion evaluation network based on a historical database of glass cutting edge defects.

[0065] Step S3: Based on the initial edge defect risk characteristics output by the feature recognition of several sets of feature data of the detection mode by the multimodal fusion evaluation network, determine the edge risk status of the cut glass;

[0066] Step S4: Based on the comprehensive decision coefficient of the fusion working condition interference and defect risk, determine the weight allocation of each detection mode in order to assess the edge quality of the cut glass edge to be tested.

[0067] Step S5: Based on the edge quality assessment results, determine the area range for the laser ultrasonic probe to perform directional scanning on the edge of the cut glass, and simultaneously adjust the line trigger sampling frequency of the line scan camera;

[0068] Step S6: Determine the diagnostic deviation characteristics of a single detection based on the spatial positional consistency among the detection results of each modality;

[0069] Step S7: Based on the cumulative deviation characteristics of the preset sliding window, determine the monitoring period of the adjustment condition interference characteristics and the sample sampling ratio when the multimodal fusion evaluation network is updated.

[0070] In this embodiment of the invention, the working condition interference characteristics include a tool wheel wear interference component, a thickness fluctuation interference component, and a light source attenuation interference component, wherein,

[0071] The wear interference component of the cutter wheel is determined based on the ratio of the cumulative cutting mileage of the cutter wheel to the rated life mileage of the cutter wheel. It is used to determine the degree of interference of the actual cutting force attenuation caused by wear of the cutter wheel on the detection mode.

[0072] The rated life mileage of the cutter wheel is obtained by testing under standard cutting conditions based on the cutter wheel material, cutter wheel angle and glass substrate material, and the value ranges from 5000m to 10000m. In this embodiment, 8000m is preferred.

[0073] The thickness fluctuation interference component is determined based on the ratio of the thickness value at the current cutting position of the cut glass to the standard thickness value. It is used to determine the degree of interference of the crack propagation energy threshold change caused by the deviation of the cut glass thickness from the standard value on the detection mode.

[0074] The standard thickness value is the design thickness for cutting glass.

[0075] The light source attenuation interference component is determined based on the ratio of the current cumulative illumination time of the light source of the line scan camera to the rated lifespan of the light source. It is used to determine the degree of interference of light flux attenuation caused by light source aging on the visible light imaging quality.

[0076] The rated lifespan of the light source is determined according to the test method of the LED light source industry standard IES LM-80.

[0077] The preset monitoring cycle is the time taken to complete a preset cutting length in a continuous glass cutting production line. In this embodiment, the preset cutting length is preferably 100m of glass.

[0078] Specifically, this invention establishes a quantitative mapping relationship between operating condition drift and detection signal quality by quantifying tool wear, substrate thickness fluctuation, and light source attenuation into interference components. Tool wear alters the initiation depth and propagation direction of subsurface cracks, thickness fluctuation increases the dispersion of crack morphology, and light source attenuation weakens the ability of visible light images to identify minute defects. By synthesizing these three types of interference components in real time, the ambiguous changes in production line conditions are transformed into precise interference level indicators. This provides a quantitative basis for the dynamic adjustment of detection mode weights, enabling the detection system to adaptively adjust its dependence on each mode according to operating conditions, thereby maintaining the consistency and reliability of detection results under fluctuating operating conditions.

[0079] In this embodiment of the invention, the operating conditions, including the cumulative cutting mileage of the cutter wheel, the online thickness measurement of the glass substrate, and the cumulative illumination time of the light source of the line scan camera, are acquired within a preset monitoring period. The interference components of cutter wheel wear, thickness fluctuation, and light source attenuation are calculated respectively to determine the interference characteristics of the operating conditions, so as to determine the impact of the current operating condition fluctuation on the three detection modes of visible light scanning imaging, polarized light stress imaging, and laser ultrasonic echo signal.

[0080] Optionally in this embodiment, the influence of working condition interference features on the detection mode of edge detection can be determined by the weighted sum of the above-mentioned tool wheel wear interference component, thickness fluctuation interference component and light source attenuation interference component. This embodiment does not make specific limitations.

[0081] In this embodiment of the invention, the historical glass cutting edge defect sample database includes,

[0082] The image includes a visible light grayscale scanned image of the glass cutting edge under known defect conditions, a transmitted polarized light interference phase image, a laser-ultrasonic time-domain echo waveform, and the corresponding actual defect state label.

[0083] The actual defect status labels include four categories: edge chipping defects, serrated crack defects, surface microcrack defects, and subsurface hidden microcrack defects.

[0084] In this embodiment of the invention, the multimodal fusion evaluation network includes multiple feature encoders and a cross-modal attention fusion module, wherein,

[0085] The plurality of feature encoders include,

[0086] A visible light feature encoder, built on a convolutional neural network, is used to extract edge contour curvature and cross-sectional roughness features from visible light images of cut edges.

[0087] A polarization stress feature encoder, built on a convolutional neural network, is used to extract stress phase delay and stress gradient distribution features from stress phase delay images.

[0088] An ultrasonic crack feature encoder, built on a one-dimensional convolutional neural network, is used to extract crack depth and crack orientation features from ultrasonic echo waveforms.

[0089] The cross-modal attention fusion module connects to each feature encoder to calculate the attention weights between different modal features, and performs weighted fusion of each modal feature based on the attention weights to output a fused feature vector.

[0090] Specifically, this invention constructs a multimodal fusion evaluation network and introduces a cross-modal attention fusion module into the architecture, enabling the network to autonomously learn the correlation and verification patterns between different modal features during training. This significantly improves the ability to identify hidden microcracks and to finely distinguish between chipped edges and sawtooth cracks, solving the technical problem of the complete blind zone of subsurface defects in traditional single visual detection methods.

[0091] In this embodiment of the invention, a training set and a validation set are extracted from a historical glass cutting edge defect sample database. The visible light image, polarized light image, and ultrasonic echo waveform of each sample in the training set are input into the corresponding feature encoder. The defect probability distribution obtained by classifying the fused feature vector output by the cross-modal attention weighted fusion module is compared with the actual defect state label. The classification loss is calculated using a classification loss function. The network parameters are iteratively updated using an optimizer. Training is terminated when the loss function value on the validation set meets the preset termination condition. The network weight parameters at this time are saved as the initial evaluation network for the multimodal fusion.

[0092] Optionally in this embodiment, the visible light feature encoder adopts a convolutional neural network structure with ResNet-50 as the backbone. The input is a visible light grayscale image block with a size of 64×64 pixels along the cutting edge, and the output is a 256-dimensional visible light feature vector containing edge contour curvature and cross-sectional roughness.

[0093] Optionally in this embodiment, the polarization stress feature encoder adopts a convolutional neural network structure with ResNet-50 as the backbone. The input is a stress phase delay distribution image block corresponding to the spatial location, and the output is a 256-dimensional stress feature vector containing the stress phase delay and stress gradient distribution.

[0094] Optionally in this embodiment, the laser ultrasonic crack feature encoder adopts a structure of one-dimensional convolutional neural network combined with self-attention mechanism. The input is an A-scan waveform sequence with a time domain length of 1024 sampling points, and the output is a 128-dimensional ultrasonic feature vector containing crack depth and crack orientation.

[0095] In this embodiment of the invention, the cross-modal attention weighted fusion module receives the three feature vectors mentioned above, maps the three feature vectors to the same 256-dimensional embedding space through linear transformation, then calculates the scaling dot product attention weight between the two modal features to generate a 3×3 attention weight matrix, and finally uses this matrix to perform weighted fusion of each modal feature to output a 256-dimensional fused feature vector. This fused feature vector is passed through a fully connected layer and a Softmax classifier to output the probability distribution corresponding to the four defect state categories.

[0096] In this embodiment of the invention, the plurality of sets of feature data include,

[0097] The visible light feature data set is obtained by sub-pixel edge detection of visible light images of the cut glass edge, and is used to provide the contour coordinate sequence and local geometric properties of the cut edge.

[0098] The polarized light stress feature data set is obtained by stress calculation based on the polarized light interference phase image of the cut glass edge, and is used to provide the boundary coordinates and stress distribution characteristics of the stress gradient anomaly region.

[0099] The ultrasonic acoustic feature data set, which is obtained by inverting the acoustic path of the ultrasonic echo waveform at the edge of the cut glass, is used to provide the spatial coordinates of the subsurface crack, the crack depth, the crack orientation, and the depth estimation error.

[0100] In this embodiment of the invention, the initial edge defect risk characteristic is characterized by an initial edge defect risk characteristic coefficient, wherein,

[0101] The initial edge defect risk characteristic coefficient is determined based on the sum of the probabilities of four types of defects, and is used to determine the edge risk status of the cut glass.

[0102] Specifically, this invention performs real-time high-precision processing of three modal signals, and explicitly outputs the spatial coordinate sequence and quantization error parameters required for subsequent diagnosis during the feature extraction stage, providing an accurate numerical basis for subsequent steps. At the same time, it utilizes the comprehensive reasoning ability of deep learning networks on multi-dimensional features to transform the original physical signal into a quantized initial risk probability, thereby enabling high detection sensitivity for extremely weak early subsurface cracks.

[0103] In this embodiment of the invention, after constructing a multimodal fusion initial evaluation network, visible light images, polarized light interference phase images, and ultrasonic echo waveforms of the cut glass edge are acquired synchronously in real time. The visible light feature data set, the polarized light stress feature data set, and the ultrasonic acoustic feature data set are processed to obtain the three sets of feature data. The three sets of feature data are input into the multimodal fusion initial evaluation network. The multimodal fusion initial evaluation network outputs the probability distribution of four categories: edge chipping defects, sawtooth cracks, surface microcracks, and subsurface hidden microcracks. The sum of the probabilities of the four types of defects is used as the initial edge defect risk coefficient to determine the edge risk state of the cut glass.

[0104] Optionally, in this embodiment, after preprocessing the visible light image, a subpixel edge detection algorithm based on cubic spline interpolation is used to extract the cutting edge contour. The local curvature is calculated point by point along the contour, and the gray-level variance in the contour normal direction is statistically analyzed as the cross-sectional roughness. At the same time, the pixel coordinates of each point on the contour are recorded to form a subpixel edge geometric feature vector containing the subpixel edge contour coordinate sequence and local geometric attributes.

[0105] Optionally in this embodiment, for the polarized light interference phase image, the stress phase delay of each pixel is first calculated using a digital phase shift algorithm to generate a full-field stress distribution map. Then, the local stress gradient is calculated using the Sobel operator, and connected regions with stress gradient values ​​exceeding a preset gradient value are marked as stress gradient anomaly regions. The boundary pixel coordinates of the anomaly regions are extracted to form a full-field stress gradient anomaly feature map containing the boundary coordinates of the stress gradient anomaly regions. The preset gradient value is preferably set to 0.15 MPa / μm.

[0106] Optionally, in this embodiment, the ultrasonic echo waveform is bandpass filtered for noise reduction. Then, the arrival time difference between the surface wave and the defect reflection wave is extracted. Combined with the sound velocity constant of the glass substrate, the crack spatial coordinates and crack depth are calculated by sound path inversion. At the same time, the crack orientation is estimated by analyzing the echo amplitude difference under different receiving angles. The crack depth value is compared with the statistical average crack depth corresponding to the same defect type in the historical glass cutting edge defect sample database. The absolute value of the difference between the two is calculated to obtain the depth estimation error. Finally, a subsurface crack acoustic characteristic parameter set is formed, which includes the crack spatial coordinates, crack depth and orientation parameters obtained by sound path inversion, and depth estimation error.

[0107] In this embodiment of the invention, the comprehensive decision coefficient is determined based on the product of operating condition disturbance and defect risk, wherein,

[0108] The operating condition interference is the operating condition interference coefficient, which is the degree of interference of the current operating condition on the detection mode.

[0109] The operating condition interference coefficient is determined based on the interference components of tool wheel wear, thickness fluctuation, and light source attenuation. It is used to quantify the impact of tool wheel wear, glass thickness deviation, and light source attenuation on the quality of the detection signal during production line operation.

[0110] The defect risk is the initial edge defect risk characteristic coefficient.

[0111] In this embodiment of the invention, the weight allocation includes,

[0112] The first weight allocation is an ultrasound-priority weight allocation.

[0113] The second weight allocation is a balanced and collaborative weight allocation.

[0114] The third weight allocation is a visible light priority weight allocation;

[0115] Specifically, this invention quantifies and couples operating condition interference and defect risk through comprehensive decision coefficients, constructing a dynamic allocation mechanism for detection mode weights. Under low-risk, favorable operating conditions, visible light defect characteristics are weak, and the ultrasonic mode has high specificity for subsurface cracks, so its weight is increased to ensure zero missed detection of hidden defects. Under high-risk, harsh operating conditions, wear of the cutting wheel and attenuation of the light source lead to a decrease in the ultrasonic signal-to-noise ratio, while the visible light channel is less affected by interference, so its weight is increased to ensure the measurement accuracy of edge chipping and sawtooth cracks, thereby achieving an adaptive balance between detection sensitivity and anti-interference robustness.

[0116] In this embodiment of the invention, the comprehensive decision coefficient is calculated based on the product of the operating condition interference and the initial edge defect risk coefficient. The comprehensive decision coefficient is compared with the first preset comprehensive decision coefficient and the second preset comprehensive decision coefficient. The corresponding weight allocation is determined according to the comparison result. The weights of each detection mode are adjusted according to the determined weight allocation. The response values ​​of each mode are fused according to the adjusted weights to generate an edge quality assessment index that characterizes the edge integrity of the current detection area.

[0117] In this embodiment, the ultrasonic priority weight allocation is optional and is suitable for situations where the initial defect risk is low and the working conditions are good. At this stage, the surface quality of the cutting edge is usually good and the visible light image features are not obvious. The main hidden danger is whether there are early micro-cracks on the subsurface. Therefore, the weight of the laser ultrasonic channel is significantly increased to 0.5 to conduct in-depth investigation of subsurface cracks with high specificity and ensure zero missed detection of hidden defects.

[0118] Optionally in this embodiment, the balanced collaborative allocation is suitable for situations where a certain level of risk is detected or the operating condition is at a moderate level. In such cases, both surface chipping and subsurface cracks may exist. Therefore, by assigning relatively balanced weights to the three features, and through information complementarity and cross-modal cross-validation, a comprehensive assessment of edge quality and accurate differentiation of defect types are achieved, balancing the sensitivity and specificity of detection.

[0119] In this embodiment, the optional visible light priority weight allocation is suitable for situations with high initial risk, such as when obvious edge chipping is detected and the working conditions are poor, with severe wear of the cutting wheel and attenuation of the light source. At this stage, the macroscopic defect characteristics of the visible light channel are most significant and relatively less affected by working conditions, while the signal-to-noise ratio of the laser ultrasonic signal decreases due to broadband noise interference caused by severe wear of the cutting wheel. Therefore, the visible light feature weight is increased to 0.6 to ensure high-precision measurement and positioning of macroscopic edge chipping and sawtooth cracks, while simultaneously using the stress channel to monitor crack propagation trends.

[0120] In a preferred embodiment, if the comprehensive decision coefficient is less than or equal to the first preset comprehensive decision coefficient, then an ultrasound-priority weight allocation is determined; if the comprehensive decision coefficient is greater than the first preset comprehensive decision coefficient and less than or equal to the second preset comprehensive decision coefficient, then a balanced and coordinated weight allocation is determined; if the comprehensive decision coefficient is greater than the second preset comprehensive decision coefficient, then a visible light-priority weight allocation is determined.

[0121] In a preferred embodiment of the present invention, the working condition interference coefficient can be determined by weighted summation of the tool wheel wear interference component, the thickness fluctuation interference component, and the light source attenuation interference component, without specific limitations here.

[0122] In an optional embodiment of the present invention, the first preset comprehensive decision coefficient is set to 0.3. During the system debugging phase, 100 sets of glass cutting edge detection samples under different working conditions were collected. Statistical analysis revealed that when the comprehensive decision coefficient is less than or equal to 0.3, the cutting edge is in a low-risk and good working condition. At this time, using an ultrasonic priority weight allocation strategy, the detection rate of subsurface hidden microcracks can reach over 92%, and the misjudgment rate of edge chipping defects is controlled within 3%. When the comprehensive decision coefficient exceeds 0.3, if the ultrasonic priority strategy is still used, the ultrasonic signal-to-noise ratio decreases significantly due to the wear noise of the cutting wheel and the attenuation of the light source. The detection rate of subsurface microcracks drops to below 78%, and the misjudgment rate rises to over 8%. Therefore, 0.3 is used as the first preset comprehensive decision coefficient to distinguish between low-risk and medium-risk working conditions, ensuring that the high sensitivity advantage of ultrasonic modes for hidden defects is prioritized in the low-risk stage.

[0123] In an optional embodiment of the present invention, the second preset comprehensive decision coefficient is set to 0.6. Statistical analysis of the above 100 sets of samples revealed that when the comprehensive decision coefficient is greater than 0.6, the probability of a clear edge chipping at the cutting edge exceeds 75%, and the average values ​​of the blade wear interference component and the light source attenuation interference component reach 0.55 and 0.40 or higher, respectively. At this point, the effective echo recognition rate of the ultrasonic signal under broadband noise interference drops below 60%. If a visible light priority weight allocation strategy is adopted, increasing the visible light feature weight to 0.6, the detection rate of edge chipping defects can be restored to over 90%, and the recognition accuracy of sawtooth cracks reaches over 85%. However, if a balanced collaborative strategy is continued, the boundary positioning accuracy deviation of edge chipping defects exceeds 0.15mm, making it difficult to meet the edge accuracy requirements of subsequent polarizer attachment. Therefore, 0.6 is used as the second preset comprehensive decision coefficient to distinguish between medium-risk and high-risk working conditions, ensuring priority is given to ensuring the measurement accuracy of macroscopic defects under high-risk and severe working conditions.

[0124] Please see Figure 2 As shown, it is a logic block diagram of an embodiment of the present invention for determining the corresponding detection strategy based on the edge quality index.

[0125] In this embodiment of the invention, the edge quality assessment result can be characterized by an edge quality index, wherein,

[0126] The edge quality index, which is determined based on the sum of the probability of edge chipping defects and the probability of subsurface hidden microcrack defects, is used to determine the integrity of the edge of the current detection area in order to execute the corresponding detection strategy.

[0127] The detection strategy includes,

[0128] The first detection strategy involves narrowing the directional scanning area of ​​the laser ultrasonic probe to cover only the core area of ​​the polarized light stress gradient anomaly region, and reducing the line-triggered sampling frequency of the line scan camera to 0.5 times the standard frequency, in order to eliminate the risk of subsurface cracks.

[0129] The core region is a sub-region within the stress gradient anomaly zone where the stress gradient value exceeds 1.2 times the preset gradient value.

[0130] The second detection strategy is to expand the directional scanning area of ​​the laser ultrasonic probe to cover the entire polarized light stress gradient anomaly area and its outer boundary, and to keep the line trigger sampling frequency of the line scan camera at the standard frequency, so as to accurately identify the correlation between the chipping size and the subsurface crack, and accurately distinguish between sawtooth cracks and hidden microcracks.

[0131] The outward expansion distance is 0.5mm.

[0132] The third detection strategy involves further expanding the directional scanning area of ​​the laser ultrasonic probe to cover the entire polarized light stress gradient anomaly area and extending it 5mm forward and backward along the cutting edge. At the same time, the line-triggered sampling frequency of the line scan camera is increased to 1.5 times the standard frequency to capture the micro-morphology of the chipped edge with ultra-high resolution, so as to accurately distinguish and precisely locate the chipped edge defect, sawtooth crack defect and hidden micro-crack defect.

[0133] Specifically, this invention quantifies the integrity of the cutting edge through an edge quality index and adaptively adjusts the laser ultrasonic scanning range and the sampling frequency of the line scan camera accordingly. When the edge quality is excellent, the scanning area is reduced and the sampling frequency is lowered to save detection resources; when the risk is moderate, the scanning is expanded to the outer boundary of the stress anomaly zone, while maintaining the standard frequency to cover the area associated with edge chipping and microcracks; when the quality is poor, the scanning range is further expanded and the sampling frequency is increased to capture the edge chipping morphology and crack tip position with ultra-high resolution, providing accurate defect location and type differentiation basis for process adjustment, and achieving seamless scheduling of detection resources from macroscopic screening to microscopic positioning.

[0134] In this embodiment of the invention, after obtaining the edge quality index, it is compared with the first preset quality index and the second preset quality index. Based on the comparison result, the corresponding detection strategy is determined, and the laser ultrasonic probe and the line scan camera are controlled to perform corresponding scanning area adjustments and sampling frequency switching to complete the fine detection of the edge of the currently cut glass.

[0135] In a preferred embodiment, if the edge quality index is less than or equal to a first preset quality index, then a first detection strategy is determined to be executed; if the edge quality index is greater than the first preset quality index and less than or equal to a second preset quality index, then a second detection strategy is determined to be executed; if the edge quality index is greater than the second preset quality index, then a third detection strategy is determined to be executed.

[0136] In an optional embodiment of the present invention, the first preset quality index is set to 0.3. During the system debugging phase, 80 sets of cut glass edge samples were collected. Statistical analysis revealed that when the edge quality index is less than or equal to 0.3, the actual probability of subsurface hidden microcracks is less than 5%, and the coverage area of ​​the polarized light stress gradient anomaly zone is less than 10% of the total area of ​​the cut edge. At this time, a first detection strategy of reducing the scanning range to the core area and reducing the sampling frequency is adopted. Under the premise of ensuring zero missed detection of subsurface cracks, the average detection time per test is shortened to 35% of the standard detection time. When the edge quality index exceeds 0.3, the probability of subsurface microcracks increases to more than 18%. If the first detection strategy is still used, 12% of hidden microcracks will be missed due to insufficient scanning range. Therefore, 0.3 is used as the first preset quality index to distinguish between high-quality edges and risky edges that require expanded scanning.

[0137] In an optional embodiment of the present invention, the second preset quality index is set to 0.7. Statistical analysis of the above samples revealed that when the edge quality index is greater than 0.7, the proportion of severe edge chipping or large-scale stress concentration at the cutting edge exceeds 80%, and the proportion of high-risk defects with subsurface crack depths exceeding 50 μm reaches 45%. If the second detection strategy with standard frequency and limited outward expansion distance is continued, the edge chipping boundary positioning accuracy deviation will exceed 0.12 mm, which cannot meet the requirements of the precision bonding process. After adopting the third detection strategy, through high-density ultrasonic scanning and ultra-high resolution line scanning imaging, the edge chipping depth measurement accuracy is improved to ±5 μm, and the subsurface crack tip positioning accuracy reaches ±0.1 mm. Therefore, 0.7 is used as the second preset quality index to distinguish between medium-risk and high-risk edges, ensuring that the highest precision detection resource allocation is initiated in the high-risk stage.

[0138] In an optional embodiment of the present invention, the standard frequency is determined based on the ratio of the production line conveying speed to the target space sampling interval. In this embodiment of the present invention, the production line conveying speed is 200 mm / s, the target space sampling interval is 0.05 mm, and the standard frequency is 4000 Hz.

[0139] The production line conveying speed and the target space sampling interval can be adjusted according to the actual detection accuracy requirements and production line operating parameters, and are not specifically limited here.

[0140] Please see Figure 3 As shown, it is a logic block diagram of determining the reliability of a single fusion diagnostic result based on the spatial consistency coefficient in an embodiment of the present invention.

[0141] In this embodiment of the invention, the spatial location consistency can be characterized by a spatial location consistency index, wherein...

[0142] The spatial location consistency index is determined based on the set of visible light profile anomalies, the set of polarized light stress anomalies, and the set of ultrasonic crack location points. It is used to evaluate the degree of consistency in the spatial location judgment of the same defect by different detection modes.

[0143] The set of visible light contour anomalies is determined based on curvature abrupt change points or roughness anomalies selected from the visible light subpixel edge contour coordinate sequence.

[0144] The set of polarized light stress anomaly regions is determined based on the coordinates of all pixels within the region enclosed by the boundary coordinates of the polarized light stress gradient anomaly region.

[0145] The set of ultrasonic crack location points is determined based on the inversion coordinates of the acoustic path of all detected cracks in the subsurface crack acoustic characteristic parameter set.

[0146] The diagnostic deviation feature is characterized by a diagnostic deviation value, wherein...

[0147] The diagnostic deviation value is determined based on the defect type classification accuracy deviation, defect boundary positioning accuracy deviation, and crack depth estimation error, and is used to quantify the degree of deviation between a single detection result and the actual defect state.

[0148] The accuracy deviation of the defect type classification is determined based on the ratio of the total number of units to the number of correctly diagnosed units.

[0149] The defect boundary positioning accuracy deviation is determined based on the average Euclidean distance of all correctly diagnosed units.

[0150] The crack depth estimation error is determined based on the average of the depth estimation errors of all crack elements.

[0151] Specifically, this invention establishes a multimodal cross-validation mechanism by calculating the spatial overlap between visible light profile anomalies, polarized light stress anomalies, and ultrasonic crack location points. When the three spatially overlap significantly, the diagnostic result is considered reliable; when their spatial distribution deviates significantly, the diagnostic result is considered unreliable. This mechanism automatically identifies diagnostic inconsistencies caused by sensor noise or random factors, selects high-confidence results for process improvement, and quantifies the deviation of low-confidence results as deviation values, providing error feedback data for optimizing the detection model.

[0152] In this embodiment of the invention, after obtaining the detection results of each modality, the set of visible light contour anomaly points, the set of polarized light stress anomaly areas, and the set of ultrasonic crack location points are extracted respectively. The spatial position consistency index of the three is calculated. The reliability of this diagnosis is determined based on the comparison result of the spatial position consistency index and the preset consistency coefficient. If the reliability does not meet the standard, the error of defect type classification accuracy, defect boundary positioning accuracy, and crack depth estimation error are further calculated, and the diagnosis deviation value is calculated.

[0153] In a preferred embodiment of the present invention, the spatial location consistency index can be determined by the ratio of the area of ​​the intersection region to the area of ​​the union region of the three sets, without specific limitation here.

[0154] In a preferred embodiment, if the spatial position consistency coefficient is less than a preset consistency coefficient, the single fusion diagnosis result is determined to be unreliable, and the diagnosis deviation value is calculated; if the spatial position consistency coefficient is greater than or equal to the preset consistency coefficient, the single fusion diagnosis result is determined to be reliable, and the edge detection result of the cut glass is output.

[0155] In a preferred embodiment, the diagnostic deviation value can be determined by weighted summation of the defect type classification accuracy deviation, the defect boundary positioning accuracy deviation, and the crack depth estimation error; no specific limitation is made here.

[0156] In an optional embodiment of the present invention, the preset consistency coefficient is set to 0.7. During the system debugging phase, 60 sets of cut glass edge defect samples confirmed by offline re-inspection were collected. Statistical analysis revealed that when the three-modal spatial position consistency index is greater than or equal to 0.7, the consistency between the multimodal diagnostic results and the offline re-inspection results reaches over 93%, and the missed detection rate of edge chipping defects and subsurface microcracks is less than 3%. When the consistency index is less than 0.7, the consistency between the diagnostic results and the offline re-inspection results drops to below 72%, and the missed detection rate of subsurface microcracks rises to over 15%. Therefore, 0.7 is used as the preset consistency coefficient to distinguish between high-confidence diagnoses and low-confidence diagnoses requiring deviation assessment.

[0157] Please see Figure 4 As shown, it is a logic block diagram of an embodiment of the present invention for determining the execution of the corresponding optimization strategy based on the cumulative diagnostic deviation value.

[0158] In this embodiment of the invention, the cumulative deviation feature is characterized by a cumulative diagnostic deviation value, wherein,

[0159] The cumulative diagnostic deviation value is determined based on the arithmetic mean of all diagnostic deviation values ​​within a preset sliding window. It is used to measure the overall degradation trend of the diagnostic performance of the detection system during continuous operation, so as to implement corresponding optimization strategies.

[0160] The optimization strategy includes,

[0161] The first optimization strategy is to shorten the duration of the preset monitoring cycle, so as to increase the monitoring density of the detection process by increasing the data acquisition frequency when the system performance gradually drifts.

[0162] The second optimization strategy is to adjust the sample sampling ratio when the multimodal fusion evaluation network is updated, so that when the system performance degrades significantly, the model can learn more fully the new feature patterns introduced by recent changes in operating conditions by increasing the sampling ratio of new samples.

[0163] Specifically, this invention addresses the issue where frequent unreliable diagnostics indicate performance degradation. Small deviations suggest gradual performance drift, requiring a slight reduction in the monitoring cycle to capture changing operating conditions. Larger deviations indicate significant performance degradation, where recent data is rich in new operating condition characteristics, necessitating an increase in the proportion of new samples to allow the model to fully learn current feature patterns and accelerate accuracy recovery. This two-tiered progressive optimization mechanism strikes a balance between optimization effectiveness and computational cost, extending the effective technical lifespan of the detection system.

[0164] In this embodiment of the invention, after obtaining the current diagnostic deviation value, the arithmetic mean of all diagnostic deviation values ​​within a preset sliding window is calculated as the cumulative diagnostic deviation value. The cumulative diagnostic deviation value is compared with the preset cumulative deviation value, and the corresponding optimization strategy is determined based on the comparison result. The duration parameter of the preset monitoring period and the sample sampling ratio when the multimodal fusion evaluation network is updated are adjusted synchronously.

[0165] In a preferred embodiment, if the cumulative diagnostic deviation value is less than or equal to a preset cumulative deviation value, then the first optimization strategy is determined to be executed; if the cumulative diagnostic deviation value is greater than the preset cumulative deviation value, then the second optimization strategy is determined to be executed.

[0166] In a preferred embodiment of the present invention, the length of the preset sliding window can be the detection cycle corresponding to the most recent 30 consecutive diagnoses that were determined to be unreliable in step S6, and is not specifically limited here.

[0167] In a preferred embodiment of the present invention, the preset cumulative deviation value is 0.20. In the initial stage of the detection system's deployment, at least 50 fusion diagnostic deviation values ​​from unreliable diagnoses are collected. The average and standard deviation of these deviation values ​​are calculated. Statistical analysis reveals that when the cumulative diagnostic deviation value is less than or equal to the average plus one standard deviation, fluctuations in the system's diagnostic performance are mainly caused by gradual factors such as normal wear of the cutting wheel or slow decay of the light source. In this case, the first optimization strategy is adopted, appropriately shortening the monitoring cycle to promptly capture the changing trend of the operating conditions. When the cumulative diagnostic deviation value exceeds this threshold, it is usually accompanied by abrupt factors such as the cutting wheel approaching the end of its lifespan, a significant replacement of glass batches, or the light source approaching its rated lifespan. Recent diagnostic data contains a large amount of information reflecting the characteristics of new operating conditions. In this case, the second optimization strategy is adopted, increasing the sampling ratio of new samples to allow the model to more fully learn the characteristics of the current operating conditions, thereby accelerating the recovery of diagnostic accuracy. Therefore, 0.20 is used as the preset cumulative deviation value to distinguish between gradual performance drift and significant performance degradation.

[0168] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A multimodal fusion edge detection method for cutting glass, characterized in that, include, The impact of operating condition interference characteristics on the detection mode of edge detection is determined based on the operating conditions of the preset monitoring cycle. A multimodal fusion evaluation network was constructed based on a historical database of glass cutting edge defects. The edge defect risk characteristics are determined by the initial edge defect risk characteristics output by the feature recognition output of several sets of feature data of the detection modality based on the multimodal fusion evaluation network, and the edge risk status of the cut glass is determined. Based on the comprehensive decision coefficient of fusion working condition interference and defect risk, the weight allocation of each detection mode is determined in order to assess the edge quality of the cut glass edge to be tested. Based on the edge quality assessment results, the area range for directional scanning of the glass cutting edge by the laser ultrasonic probe is determined, and the line trigger sampling frequency of the line scan camera is adjusted simultaneously. The diagnostic deviation characteristics of a single test are determined based on the spatial consistency among the test results of each modality. Based on the cumulative deviation characteristics of the preset sliding window, the monitoring period for the adjustment condition interference characteristics and the sample sampling ratio when the multimodal fusion evaluation network is updated are determined.

2. The multimodal fusion edge detection method for cutting glass according to claim 1, characterized in that, The aforementioned operating condition interference characteristics include a tool wheel wear interference component, a thickness fluctuation interference component, and a light source attenuation interference component, wherein... The wear interference component of the cutter wheel is determined based on the ratio of the cumulative cutting mileage of the cutter wheel to the rated life mileage of the cutter wheel; The thickness fluctuation interference component is determined based on the ratio of the thickness value at the current cutting position of the glass to the standard thickness value. The light source attenuation interference component is determined based on the ratio of the current cumulative lighting time of the light source of the line scan camera to the rated lifespan of the light source.

3. The multimodal fusion edge detection method for cutting glass according to claim 2, characterized in that, The plurality of sets of feature data include visible light feature data set, polarized light stress feature data set, and ultrasonic acoustic feature data set; The initial edge defect risk characteristic is characterized by the initial edge defect risk characteristic coefficient, wherein... The initial edge defect risk characteristic coefficient is determined based on the sum of the probabilities of four types of defects, and is used to determine the edge risk status of the cut glass. The four types of defects include edge chipping defects, serrated crack defects, surface microcrack defects, and subsurface hidden microcrack defects.

4. The multimodal fusion edge detection method for cutting glass according to claim 3, characterized in that, The comprehensive decision-making risk coefficient is determined based on operating condition disturbances and defect risks, wherein... The operating condition interference is characterized by the operating condition interference coefficient, which represents the degree of interference of the detected modal under the current operating conditions. The defect risk is characterized by the initial edge defect risk characteristic coefficient; The working condition interference coefficient is determined based on the interference components of tool wheel wear, thickness fluctuation, and light source attenuation.

5. The multimodal fusion edge detection method for cutting glass according to claim 4, characterized in that, The weight allocation includes, The first weight allocation is an ultrasound-priority weight allocation, determined based on the comprehensive decision coefficient being less than or equal to the first preset comprehensive decision coefficient. The second weight allocation is a balanced and collaborative weight allocation, which is determined based on the comprehensive decision coefficient being greater than the first preset comprehensive decision coefficient and less than or equal to the second preset comprehensive decision coefficient. The third weight allocation is a visible light priority weight allocation, which is determined based on the comprehensive decision coefficient being greater than the second preset comprehensive decision coefficient.

6. The multimodal fusion edge detection method for cutting glass according to claim 5, characterized in that, The edge quality assessment results, This includes the edge quality index, among which, The edge quality index, which is determined based on the sum of the probability of edge chipping defects and the probability of subsurface hidden microcrack defects, is used to determine the integrity of the edge of the current detection area in order to execute the corresponding detection strategy.

7. The multimodal fusion edge detection method for cutting glass according to claim 6, characterized in that, The detection strategy includes, The first detection strategy involves narrowing the directional scanning area of ​​the laser ultrasonic probe to cover only the core area of ​​the polarized light stress gradient anomaly region, and reducing the line-triggered sampling frequency of the line scan camera, in order to eliminate the risk of subsurface cracks.

8. The multimodal fusion edge detection method for cutting glass according to claim 7, characterized in that, The detection strategy also includes, The second detection strategy is to expand the directional scanning area of ​​the laser ultrasonic probe to cover the entire polarized light stress gradient anomaly area and its outer boundary, and to keep the line trigger sampling frequency of the line scan camera at the standard frequency, so as to accurately identify the correlation between the chipping size and the subsurface crack, and accurately distinguish between sawtooth cracks and hidden microcracks. The third detection strategy involves further expanding the directional scanning area of ​​the laser ultrasound probe to cover the entire polarized light stress gradient anomaly area and extending it forward and backward along the cutting edge, while increasing the line-triggered sampling frequency of the line scan camera.

9. The multimodal fusion edge detection method for cutting glass according to claim 8, characterized in that, The spatial location consistency includes a spatial location consistency coefficient, wherein... The spatial position consistency coefficient is determined based on the set of visible light profile anomaly points, the set of polarized light stress anomaly regions, and the set of ultrasonic crack location points. The diagnostic deviation feature includes a diagnostic deviation value, which is determined based on the defect type classification accuracy deviation, defect boundary positioning accuracy deviation, and crack depth estimation error, and is used to quantify the degree of deviation between a single detection result and the actual defect state.

10. The multimodal fusion edge detection method for cutting glass according to claim 9, characterized in that, The cumulative deviation characteristic, This includes the cumulative diagnostic deviation value, where, The cumulative deviation value is determined based on the arithmetic mean of all diagnostic deviation values ​​within a preset sliding window, and is used to measure the overall degradation trend of the diagnostic performance of the detection system during continuous operation, so as to implement corresponding optimization strategies. The optimization strategy includes, The first optimization strategy is to shorten the duration of the preset monitoring cycle; The second optimization strategy is to adjust the sample sampling ratio when updating the multimodal fusion evaluation network.

Images