An automatic identification, marking, and rejection system for surface defects in calcium silicate boards.

Abstract

This invention relates to the field of industrial automation inspection technology, specifically disclosing an automatic identification, marking, and removal system for surface defects in calcium silicate boards. The system includes an intelligent inspection collaborative control center, which is communicatively connected to the following modules: a polarization dust suppression and reconstruction module, used for morphological reconstruction optimized by polarization differential imaging fusion with an improved fruit fly algorithm. This module removes surface dust interference through polarization characteristics and uses the fruit fly algorithm to adaptively optimize structural elements, accurately reconstructing the microcrack contours. This invention utilizes polarization differential imaging technology and multi-angle polarized light sources to actively suppress the shadows of wet dust that are highly similar to the board's background color, effectively removing the masking interference of dynamic dust background on microcracks. This allows for high-contrast separation of hidden crack contours, fundamentally solving the problem of missed and false detections caused by dust artifacts in traditional visual inspection, and significantly improving the accuracy and reliability of defect identification.

Description

Technical Field

[0001] This invention relates to the field of industrial automation inspection technology, and in particular to an automatic identification, marking and rejection system for surface defects in calcium silicate boards. Background Technology

[0002] Calcium silicate board (fiber cement reinforced board) is a material widely used in the construction and decoration fields. It has good fire resistance, sound insulation and durability. However, during the production process, defects such as bubbles, cracks and scratches often appear on the surface of calcium silicate board, which will affect its performance and aesthetics. Therefore, quality control and defect detection of calcium silicate board are particularly important.

[0003] For example, the online defect detection method and system for calcium silicate decorative panels published in Chinese Patent Publication No. CN118443799A solves the technical problems of traditional defect detection methods, which usually use a single detection device, are prone to missed detection or false detection, and cannot comprehensively and accurately identify surface and internal defects, resulting in poor detection effect and low detection efficiency.

[0004] In existing technologies, because the wet dust on the surface of calcium silicate boards after production line cutting is very similar to the background color, the dust shadow can cover up the real microcracks or form artifacts, making it difficult to separate hidden cracks from the dynamic dust background. Moreover, the tendency of cracks to propagate under stress corrosion in a high-humidity curing environment cannot be predicted based on geometric morphology alone, and the actual degree of damage of defects lacks dynamic classification basis. In addition, when marking and removing defects based on the classification results, the boards will undergo slight elastic deformation and deviation during roller conveying, causing the marking coordinates to deviate from the actual defect position, and the removal accuracy is limited by deformation interference. To address these issues, an automatic identification and marking removal system for surface defects of calcium silicate boards is proposed. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides an automatic identification, marking and removal system for surface defects of calcium silicate boards, which can effectively solve the problems involved in the prior art.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides an automatic identification, marking, and removal system for surface defects in calcium silicate boards, including an intelligent inspection and collaborative control center, which is communicatively connected to the following modules: The polarization dust suppression and reconstruction module is used for morphological reconstruction optimized by polarization differential imaging fusion and improved fruit fly algorithm. It removes surface dust interference by polarization characteristics and uses fruit fly algorithm to adaptively optimize structural elements to accurately reconstruct microcrack contours, effectively removes wet dust interference, and significantly improves the integrity and accuracy of microcrack contour reconstruction. The stress field state classification module is used to accurately reconstruct the geometry of cracks. It uses a laser speckle interferometry fusion deep Q network to build an intelligent classification model, measures the strain field of the micro-region at the crack tip, and dynamically calibrates the crack propagation energy level through reinforcement learning. It classifies the defects of calcium silicate board in real time, realizing the dynamic calibration of crack propagation energy level and the real-time classification of defect severity. The edge pulse encoder sensing module is used to collect data on the displacement, deviation, and minute elastic deformation of the sheet material in real time during the conveying process through the deployed intelligent sensing system, generate a dynamic deformation vector field, provide real-time physical parameters for coordinate correction, and realize multi-dimensional real-time sensing of the sheet material conveying status. The coordinate correction module is used to integrate the twin neural network. Taking the plate image coordinates at the start of the transmission as the first frame reference, it compares the topological changes of the current image with the first frame in real time during the transmission process. Combined with the dynamic deformation vector field input by the edge pulse encoder, it dynamically corrects the defect mark coordinate system and completes the precise coordinate alignment under deformation compensation. The defective product positioning and rejection module is used to receive real-time corrected coordinate mappings and map them into the coordinate system of the rejection actuator. It also dynamically plans the optimal rejection path or timing to compensate for equipment response delays and drives the actuator to synchronously track the defective position of the board to reject defective boards. This ensures that the rejection action always accurately hits the actual defect point under conditions of continuous board conveying and dynamic deformation.

[0007] Preferably, the polarization dust suppression reconstruction module includes a polarization differential imaging unit and a morphological reconstruction unit; The polarization differential imaging unit is used to capture the differences in polarization characteristics of different materials (dust and cracks) on the surface of the board by using multi-angle polarization light source and differential imaging technology, actively suppress the shadow of wet dust similar to the background color, separate microcracks from the dynamic background with high contrast, and effectively eliminate artifact interference. The morphological reconstruction unit, based on the polarization-separated image, employs an improved fruit fly algorithm. Using crack contour integrity and background noise suppression as fitness, it adaptively optimizes the structural elements for morphological reconstruction. Through iterative optimization, it accurately reconstructs the fractured and blurred microcrack contours, avoiding over-detection or under-detection caused by traditional fixed morphological parameters, and achieving adaptive optimization of structural element parameters.

[0008] Preferably, the polarization differential imaging unit performs the following steps: Multi-angle polarized light sources deployed on the production line are controlled to irradiate the surface of calcium silicate board with different polarization angles in sequence, and the area array camera is triggered to collect the original image sequence under the corresponding polarization state. Through multi-angle time-series acquisition, complete polarization information is obtained in the same area, providing a data basis for subsequent differential analysis. Differential operations are performed on images with different polarization angles under the same field of view to extract the polarization degree feature parameters of dust and cracks caused by material differences, and a high-contrast polarization feature map is constructed. By utilizing the differences in material polarization characteristics, dust interference is effectively removed, and the contrast between cracks and background is significantly improved. By fusing polarization feature maps with regular grayscale images at the pixel level, the shadows of wet dust that are highly similar to the background color of the board are suppressed, making the outline of microcracks stand out from the dynamic dust background. The fusion enhancement makes the microcracks clearly visible in the dynamic dust environment and eliminates artifact interference.

[0009] Preferably, the morphological reconstruction unit performs the following steps: The high-contrast polarization feature map output by the polarization differential imaging unit is received to initialize the population of structural elements for morphological reconstruction. Each individual characterizes the combination of shape and scale parameters of the structural element, effectively covering the geometric features of different crack morphologies and improving the matching accuracy of structural elements. Using crack contour integrity and background noise suppression rate as fitness functions, an improved fruit fly algorithm is used to iteratively update the structural element parameters. The global optimal structural element is adaptively optimized through the odor concentration determination mechanism, thereby avoiding over-detection or under-detection caused by manual settings. Based on the global optimal structuring element, morphological opening and closing reconstruction and connected component analysis are performed on the polarization feature map to accurately reconstruct the outline of fractured and blurred microcracks, realize the complete extraction of crack geometry, ensure the continuous reconstruction of fractured microcracks, and improve the integrity and reliability of defect identification.

[0010] Preferably, the stress field state classification module includes a strain field measurement unit and an energy level calibration unit; The strain field measurement unit is used to apply micro-perturbations to the identified crack region using laser speckle interferometry, and to measure the strain field distribution of the micro-region at the crack tip in a non-contact manner, thereby quantifying the stress concentration and potential propagation driving force around the crack and realizing high-precision non-contact measurement of the strain field at the crack tip. The energy level calibration unit constructs an intelligent classification model based on strain field measurement results and a deep Q-network. It takes the measured strain field data as the state input and crack propagation risk as the action value. Through a reinforcement learning mechanism, it dynamically calibrates the crack propagation energy level in a simulated high-humidity curing environment and outputs the real-time classification result of the defect hazard degree, thus realizing the adaptive dynamic calibration of the crack propagation energy level.

[0011] Preferably, the strain field measurement unit performs the following steps: A laser speckle interferometer is deployed and aligned above the identified crack area to apply controllable micro-perturbation excitation to the micro-region at the crack tip, thereby exciting the local strain response in the surrounding area of ​​the crack and realizing non-contact activation and capture of the strain state at the crack tip. By continuously acquiring speckle interferometric image sequences before and after disturbance, and calculating the displacement field distribution of the micro-region at the crack tip through a phase demodulation algorithm, a high-resolution strain field cloud map is generated, achieving high-precision visualization of the micron-level displacement field. The strain concentration factor and strain gradient characteristic parameters at the crack tip are extracted from the strain field cloud map to quantify the stress concentration and potential propagation driving force around the crack, thereby achieving quantitative characterization and assessment of crack propagation risk.

[0012] Preferably, the energy level calibration unit performs the following steps: The strain concentration coefficient and strain gradient characteristic parameter output by the strain field measurement unit are used as input to the state space. A smart classification model is constructed by combining it with a deep Q network. The crack propagation risk level is set as the action space, which realizes the quantitative characterization of crack propagation risk and lays the data foundation for classification decision-making. In a digital twin interactive platform simulating a high-humidity maintenance environment, the intelligent grading model is driven to perform actions and obtain reward and punishment feedback. The network parameters are iteratively trained until the strategy converges, enabling the model to have adaptive learning capabilities in complex maintenance environments and improving grading accuracy. The strain field data obtained in real time is input into the intelligent classification model that has been trained and converged, and the crack propagation energy level calibration results are dynamically output to complete the real-time classification of the degree of defect hazard. This enables the immediate determination of the defect hazard level and provides a reliable classification basis for subsequent removal.

[0013] Preferably, the edge pulse encoder sensing module performs the following steps: The intelligent sensing system includes an edge pulse encoder and a laser displacement sensor. The edge pulse encoder is symmetrically installed on both sides of the conveyor roller of the calcium silicate board. The wheel of the edge pulse encoder maintains a constant contact pressure with the edge of the board and collects the displacement pulse signal and lateral offset data of the edge of the board in the conveying direction in real time. The laser displacement sensor is arranged at four points at equal intervals along the lateral direction of the roller and collects the lateral deviation and elastic flexural deformation of the board in a non-contact manner. It covers the deformation characteristics of the entire cross section in the width direction of the board, effectively solving the problem of fusion between contact and multi-point measurement, and realizing the complete capture of the deformation characteristics of the entire cross section. Kalman filtering is used to fuse encoder pulse signals and laser displacement sensing signals to eliminate accumulated errors and measurement noise. Simultaneously, visual feedback from preset reference markers on the board surface is fused to correct the accumulated error of the edge pulse encoder. Through multi-source data fusion and visual calibration, the long-term stability and absolute accuracy of displacement measurement are significantly improved. An elastic deformation compensation algorithm is introduced to extract the minute elastic flexural deformation characteristics of the plate during transportation. The longitudinal displacement, lateral deviation and elastic flexural deformation are fused to generate a dynamic deformation vector field, which is transmitted to the intelligent inspection collaborative control center in real time through high-frequency data stream. This provides continuously updated physical parameters for the dynamic correction of the defect marking coordinate system, realizing the dynamic quantification and real-time transmission of deformation characteristics.

[0014] Preferably, the coordinate correction module performs the following steps: The first frame reference coordinate system is established using the plate image acquired at the start of the transport. The natural texture and edge contour of the plate surface are extracted as topological anchor point features to establish a unified absolute coordinate reference, providing a stable and reliable spatial reference for subsequent defect localization. During the transmission process, the current frame image and the first frame reference are input into the twin neural network to calculate the topological structure offset. The dynamic deformation vector field input from the edge pulse encoder sensing module is fused to achieve high-precision fusion of multi-source data and ensure that the offset tracks the dynamic changes of the board material in real time. A nonlinear deformation compensation model is constructed based on topological offset and dynamic deformation vector field. The defect marker coordinate system is dynamically corrected to achieve accurate alignment of defect coordinates under deformation conditions and accurate compensation for nonlinear deformation of the plate.

[0015] Preferably, the defective product location and rejection module performs the following steps: The system receives the real-time corrected defect coordinates, maps them to the base coordinate system of the removal actuator through coordinate transformation, and generates a defect motion trajectory prediction sequence to ensure that the defect coordinates are consistent with the spatial reference of the actuator, thus laying the foundation for accurate tracking. By combining the mechanical response delay parameters of the rejection actuator, the optimal rejection path and triggering sequence are dynamically planned to compensate for the fluctuation of conveying speed and the system hysteresis effect, so as to achieve precise synchronization between the rejection action and the arrival time of the defect and eliminate the influence of system hysteresis. The drive actuator synchronously tracks the defective positions of the sheet metal that undergoes elastic deformation and deviation during continuous conveying based on the real-time updated coordinate trajectory, and removes defective sheet metal. This ensures that the removal action always accurately hits the defect point under dynamic deformation conditions, thus improving the reliability of removal.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This automatic identification, marking, and removal system for surface defects in calcium silicate boards utilizes polarized light differential imaging technology. By employing multi-angle polarized light sources, it actively suppresses the shadows of wet dust that are highly similar to the background color of the board, effectively removing the interference of dynamic dust background on microcracks. This allows for high-contrast separation of the outlines of hidden cracks, fundamentally solving the problems of missed and false detections caused by dust artifacts in traditional visual inspection, and significantly improving the accuracy and reliability of defect identification.

[0017] 2. This automatic identification and marking system for defects on the surface of calcium silicate board adopts an improved fruit fly algorithm to adaptively optimize the structural elements of morphological reconstruction. The reconstruction parameters are dynamically adjusted according to the actual morphological characteristics of the cracks to avoid over-detection or under-detection of fractures and fuzzy cracks by fixed morphological operators. This achieves accurate reconstruction of the geometric contour of microcracks and provides high-quality morphological data for subsequent defect severity assessment.

[0018] 3. This automatic identification, marking, and removal system for surface defects in calcium silicate boards integrates laser speckle interferometry and deep Q-network to construct an intelligent grading model. By measuring the strain field distribution in the micro-region at the crack tip and combining reinforcement learning with dynamic interactive training in a digital twin platform under high humidity curing environment, it achieves real-time calibration of crack propagation energy levels. This overcomes the technical limitation that stress corrosion propagation tendency cannot be predicted solely based on geometric morphology, and provides a scientific dynamic grading basis for the degree of defect hazard.

[0019] 4. This automatic identification, marking and removal system for defects on the surface of calcium silicate board integrates an edge pulse encoder and a twin neural network to collect displacement, deviation and elastic deformation data of the board during the conveying process in real time. The system dynamically corrects the defect marking coordinate system based on the first frame image, effectively compensating for the coordinate offset caused by the slight elastic deformation and deviation of the board during roller conveying, and achieving accurate alignment of defect coordinates under deformation conditions. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the workflow of an automatic identification, marking and removal system for surface defects in calcium silicate board according to the present invention; Figure 2 This is a schematic diagram of the module structure of an automatic identification, marking and removal system for surface defects of calcium silicate board according to the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0022] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: an automatic identification, marking, and removal system for surface defects in calcium silicate boards, comprising an intelligent inspection and collaborative control center, which is connected to the following modules for communication: The polarization dust suppression and reconstruction module is used for morphological reconstruction optimized by polarization differential imaging fusion and improved fruit fly algorithm. It removes surface dust interference through polarization characteristics and uses fruit fly algorithm to adaptively optimize structural elements to accurately reconstruct microcrack contours. It effectively removes wet dust interference and significantly improves the integrity and accuracy of microcrack contour reconstruction. The polarization dust suppression and reconstruction module includes a polarization differential imaging unit and a morphological reconstruction unit. The polarization differential imaging unit utilizes multi-angle polarized light sources and differential imaging technology to capture the differences in polarization characteristics of different materials (dust and cracks) on the surface of the board. It actively suppresses the shadows of wet dust similar to the background color, separates microcracks from the dynamic background with high contrast, effectively eliminates artifact interference, and controls the multi-angle polarized light sources deployed on the production line to sequentially illuminate the surface of the calcium silicate board at different polarization angles. Simultaneously, it triggers the area array camera to acquire the original image sequence under the corresponding polarization state. Through multi-angle time-series acquisition, it ensures that complete polarization information is obtained for the same area, providing a basis for subsequent processing. The continuous difference provides the data foundation. Differential operation is performed on images with different polarization angles under the same field of view to extract the polarization degree feature parameters of dust and cracks caused by material differences. A high-contrast polarization feature map is constructed. By utilizing the difference in material polarization characteristics, dust interference is effectively removed, and the contrast between cracks and background is significantly improved. The polarization feature map is fused with the regular grayscale image at the pixel level to suppress the shadow of wet dust that is very similar to the background color of the board. This makes the outline of microcracks stand out and separate from the dynamic dust background. The fusion enhancement makes microcracks clearly distinguishable in the dynamic dust environment and eliminates artifact interference. It should be noted that four sets of linear LED polarized light sources are deployed above the production line conveyor rollers, corresponding to four polarization angles: 0°, 45°, 90°, and 135°. Each set of light sources has a uniform output power of 12W, and the illumination distance is maintained at 350mm above the surface of the board. Six 5-megapixel area array cameras are simultaneously triggered by the intelligent inspection and collaborative control center. The camera exposure time is set to 80μs, and the acquisition frequency is linked to the production line speed, supporting dynamic imaging at a maximum conveyor speed of 2.5m / s. Each set of polarized light sources is lit sequentially, and each time it is lit, the camera is simultaneously triggered to acquire a raw image of the corresponding polarization state, forming a sequence containing four polarized images. The light source switching interval is controlled within 20ms to ensure that the imaging position offset of the same board area under continuous polarization states is less than 0.1mm. For the four polarized images (0°, 45°, 90°, and 135°) acquired under the same field of view, the surface specular reflection component is eliminated by the difference between the 45° and 135° images, and then by the difference between the 0° and 90° images... Image differential enhancement enhances the polarization sensitivity of materials. Polarization feature parameters of dust and cracks are extracted separately, and pixel-level logarithmic transformation is used for differential operation. The lower limit of the differential threshold is set to 8 gray levels to suppress background noise, and the upper limit is set to 120 gray levels to retain high-contrast crack edges. The two sets of differential results are fused to generate a polarization feature map. In this feature map, the polarization value of the dust area is generally below 12°, and the polarization value of the crack area is between 25° and 45°, realizing the quantitative characterization of material differences. The polarization feature map is then fused with a weighted pixel-level image of the same field of view. The weight coefficient of the regular grayscale image is set to 0.3, and the weight coefficient of the polarization feature map is set to 0.7 to highlight the enhancing effect of polarization difference on microcracks. After fusion, the image is filtered by median (3×3 window) to remove isolated noise. The filtering intensity is set to a smoothness of 15%. While preserving the sharpness of crack edges, it effectively suppresses the shadow of wet dust similar to the background color of the board. After processing, the crack outline can be separated from the dynamic dust background with high stability. The morphological reconstruction unit, based on the polarization-separated image, employs an improved fruit fly algorithm. Using crack contour integrity and background noise suppression as fitness parameters, it adaptively optimizes the structural elements for morphological reconstruction. Through iterative optimization, it accurately reconstructs the contours of fractured and blurred microcracks, avoiding over-detection or under-detection caused by traditional fixed morphological parameters. This achieves adaptive optimization of structural element parameters. It receives a high-contrast polarization feature map output from the polarization differential imaging unit and initializes a population of structural elements for morphological reconstruction. Each individual represents the combination of shape and scale parameters of the structural element, effectively covering the geometric features of different crack morphologies and improving the matching accuracy of structural elements. Using crack contour integrity and background noise suppression as fitness functions, the improved fruit fly algorithm iteratively updates the structural element parameters. An odor concentration determination mechanism adaptively optimizes the globally optimal structural element, achieving adaptive optimization of structural element parameters and avoiding over-detection or under-detection caused by manual settings. Based on the globally optimal structural element, morphological opening and closing reconstruction and connected component analysis are performed on the polarization feature map to accurately reconstruct the contours of fractured and blurred microcracks, achieving complete extraction of crack geometry and ensuring continuous reconstruction of fractured microcracks, thus improving the completeness and reliability of defect identification. It should be noted that after receiving the high-contrast polarization feature map output by the polarization differential imaging unit, gray-level statistics and edge energy distribution analysis are first performed on the polarization feature map to determine the initial search space of the structuring element. The structuring element population adopts a real-number encoding method, and each individual is composed of a combination of shape type parameters and scale parameters. The shape type includes three types: linear, rhomboid, and disk. The scale parameters correspond to the principal axis length and radius of the structuring element, respectively, with a value range of 3 to 21 pixels and a step size of 2 pixels. The population size is fixed at 30, and the maximum number of iterations is 50 generations. The initial population is uniformly and randomly generated within the parameter space to ensure coverage of different scale and shape combinations. The crack contour integrity is quantified by extracting the area of ​​the connected components of the crack region in the morphologically reconstructed image and the coverage of the standard reference contour, with a weight of 0.6. The background noise suppression rate is quantified by calculating the decrease ratio of the density of isolated points with gray values ​​higher than the local threshold in the non-crack region, with a weight of 0.4. The improved fruit fly algorithm is improved in the iterative process. In this study, morphological opening and closing reconstruction is performed on each individual in the population through an odor concentration determination mechanism. The reconstruction results are input into the fitness function to calculate the comprehensive score, and the population position and scale parameters are updated according to the elite retention strategy. The convergence condition is set to the global optimal fitness change rate being less than 0.5% for five consecutive generations. Finally, the global optimal structural element parameter combination is output. Based on the global optimal structural element obtained by the improved fruit fly algorithm, morphological opening and closing reconstruction operations are performed on the polarization feature map in sequence. Open reconstruction is used to eliminate the small artifacts formed by dust shadow residue, and closing reconstruction is used to connect the fractured crack segments. The shape of the structural element adopts the rhombus type of the final optimized output, with a principal axis length of 11 pixels. After reconstruction, connected component analysis is performed on the image. The connected component area threshold is set to 15 pixels. Isolated noise areas are filtered out, and narrow regions with an area greater than the threshold and an eccentricity greater than 0.75 are retained as effective crack contours. After this processing, the fractured and blurred microcrack contours are accurately reconstructed, and the complete crack geometry data is output to the stress field state classification module. The stress field grading module is used to accurately reconstruct the geometry of cracks. It uses a laser speckle interferometry fusion deep Q network to build an intelligent grading model, measures the strain field of the micro-region at the crack tip, and dynamically calibrates the crack propagation energy level through reinforcement learning. It then grades the defects of the calcium silicate board in real time, realizing the dynamic calibration of the crack propagation energy level and the real-time grading of the degree of defect hazard. The stress field grading module includes a strain field measurement unit and an energy level calibration unit. The strain field measurement unit employs laser speckle interferometry to apply micro-perturbations to the identified crack region, non-contactly measuring the strain field distribution in the micro-region at the crack tip. This quantifies the stress concentration and potential propagation driving force around the crack, achieving high-precision non-contact measurement of the crack tip strain field. A deployed laser speckle interferometer is aligned above the identified crack region, applying controllable micro-perturbations to the crack tip micro-region to excite the local strain response in the surrounding area. This enables non-contact activation and capture of the crack tip strain state. The unit continuously acquires speckle interferometry image sequences before and after the perturbation, calculates the displacement field distribution in the crack tip micro-region using a phase demodulation algorithm, and generates a high-resolution strain field cloud map. This achieves high-precision visualization of the micrometer-level displacement field. The strain concentration coefficient and strain gradient characteristic parameters at the crack tip are extracted from the strain field cloud map, quantifying the stress concentration and potential propagation driving force around the crack, thus achieving quantitative characterization and assessment of crack propagation risk. It should be noted that after the crack identification area is located, the automatically controlled robotic arm moves the measuring head of the laser speckle interferometer directly above the crack, with a fixed working distance of 500mm. This ensures that the measurement field of view covers the crack tip and its surrounding 20mm × 20mm micro-area. A 532nm wavelength solid-state laser is used as the illumination source, with a stable output power of 50mW. After beam expansion, a uniform speckle field is formed. A sinusoidal micro-perturbation excitation with a frequency of 2Hz and an amplitude of 0.5μm is applied to the crack tip micro-area through a piezoelectric ceramic driving lens. The excitation direction is perpendicular to the main crack propagation direction, and the duration is 0.5 seconds. These excitation parameters have been pre-tested and can excite sufficient speckle without causing crack propagation. To measure the strain response, during the excitation process, an industrial camera is simultaneously triggered to continuously acquire speckle interferometric image sequences before and after the disturbance at a frame rate of 30 frames per second. Each sequence contains 15 frames before the disturbance and 15 frames after the disturbance, which are used for subsequent phase demodulation calculations. The acquired speckle interferometric image sequences are processed using a spatial phase demodulation algorithm. First, the phase information of the speckle pattern is extracted through Fourier transform. Then, pixel-by-pixel phase difference operations are performed on the image sequences before and after the disturbance to eliminate the static background phase, resulting in a displacement field distribution that only reflects the micro-disturbance. The phase demodulation accuracy is controlled within ±0.05 rad, corresponding to a displacement resolution better than 0.02 μm. Based on the displacement field data, the in-plane strain components are calculated using the finite difference method. , and shear strain A high-resolution strain field cloud map with a resolution of 512×512 pixels is generated. Each pixel in the cloud map corresponds to an actual physical size of 0.05mm×0.05mm. The strain field cloud map is presented in pseudo-color, and the strain concentration zone at the crack tip is automatically marked using contour lines. From the generated high-resolution strain field cloud map, the point with the maximum strain is automatically identified as the crack tip location, and the strain concentration factor is extracted along the leading edge of the crack extension direction. The value is defined as the ratio of the maximum strain at the crack tip to the average strain in the far field. Simultaneously, the strain gradient distribution is extracted at intervals of 0.1 mm perpendicular to the crack direction, and the maximum gradient change rate is calculated as the strain gradient characteristic parameter. The above characteristic parameter extraction process is repeated three times, and the average value is taken as the final output value to eliminate random errors in single measurements. Based on the empirical model established by the calibration experiment, when the strain concentration factor is greater than 3.5 and the strain gradient exceeds 120 με / mm, the crack is determined to have a significant propagation driving force. The quantification results are output to the energy level calibration unit in numerical form as the input basis for crack propagation risk assessment. The energy level calibration unit constructs an intelligent grading model based on strain field measurement results and a deep Q-network. It uses the measured strain field data as state input and crack propagation risk as action value. Through reinforcement learning, it dynamically calibrates the crack propagation energy level in a simulated high-humidity curing environment, outputting real-time grading results of defect severity. This achieves adaptive dynamic calibration of the crack propagation energy level. The strain concentration coefficient and strain gradient characteristic parameters output by the strain field measurement unit are used as state space input. Combined with a deep Q-network, an intelligent grading model is constructed, setting the crack propagation risk level as the action space. This achieves a quantitative representation of crack propagation risk, laying a data foundation for grading decisions. In a digital twin interactive platform simulating a high-humidity curing environment, the intelligent grading model is driven to execute actions and obtain reward / penalty feedback. The network parameters are iteratively trained until the strategy converges, enabling the model to have adaptive learning capabilities in complex curing environments and improving grading accuracy. Real-time measured strain field data is input into the trained and converged intelligent grading model, dynamically outputting crack propagation energy level calibration results to complete real-time grading of defect severity. This achieves immediate determination of defect severity levels, providing a reliable grading basis for subsequent removal. It should be noted that after the strain field measurement unit completes data acquisition, the extracted strain concentration coefficient and strain gradient feature parameters are used as inputs to the intelligent hierarchical model constructed based on a deep Q-network. The state space dimension is set to a two-dimensional continuous space, the strain concentration coefficient ranges from 1.0 to 8.0, and the strain gradient ranges from 0 to 300 με / mm. The deep Q-network adopts a three-layer fully connected neural network structure. The input layer contains 2 neurons corresponding to the state parameters, the hidden layer contains 64 neurons using the ReLU activation function, and the output layer contains 4 neurons corresponding to the crack propagation risk level. Four discrete actions were assigned: low risk, medium risk, low-high risk, and high risk. The network learning rate was set to 0.001, and the experience replay pool capacity was 10,000 sets. 128 sets of samples were randomly selected for batch gradient descent in each training session. An ε-greedy strategy was used for action selection, with an initial exploration rate ε set to 0.9, which linearly decayed to 0.05 with each training step. A virtual simulation scenario of a high-humidity curing environment was constructed in a digital twin interactive platform. The relative humidity was set to 95%, the temperature was kept constant at 45℃, and the curing cycle was simulated to be 28 days. The platform incorporated crack propagation dynamics based on the finite element method. The training model calculates the crack propagation rate and path in a simulated environment based on the input strain field data. Each time the intelligent grading model executes an action, the digital twin platform simulates and calculates the crack propagation state at the next moment based on the current crack state and the intervention strategy corresponding to the selected action. It then returns a reward value according to a preset reward and penalty function. The reward and penalty function is designed as follows: correctly identifying a high-risk crack awards a +10 reward; misjudging a low-risk crack as high-risk awards a -5 penalty; missing a high-risk crack awards a -15 penalty; and correctly identifying a low-risk crack awards a +2 reward. Each training round contains 200 decision steps. The training process is repeated for 5000 rounds until the strategy converges. The convergence criterion is that the average reward value is stable above +8 for 500 consecutive rounds, and the action selection strategy converges to the utilization stage with ε=0.05. After the training converges, the strain concentration coefficient and strain gradient characteristic parameters obtained in real time are input into the intelligent classification model, and the crack propagation energy level calibration results are output. The calibration results are presented on the operation interface in the form of four-color indicator lights. At the same time, a defect report containing the defect location, energy level, and recommended treatment method is generated and transmitted to the production management system via industrial Ethernet to provide a classification basis for rejection decisions. The edge pulse encoder sensing module is used to collect data on the displacement, deviation, and minute elastic deformation of the sheet material in real time during the conveying process through the deployed intelligent sensing system, generate a dynamic deformation vector field, provide real-time physical parameters for coordinate correction, and realize multi-dimensional real-time sensing of the sheet material conveying status. The coordinate correction module is used to integrate the twin neural network. Taking the plate image coordinates at the start of the transmission as the first frame reference, it compares the topological changes of the current image with the first frame in real time during the transmission process. Combined with the dynamic deformation vector field input by the edge pulse encoder, it dynamically corrects the defect mark coordinate system and completes the precise coordinate alignment under deformation compensation. The defective product positioning and rejection module is used to receive real-time corrected coordinate mappings and map them into the coordinate system of the rejection actuator. It also dynamically plans the optimal rejection path or timing to compensate for equipment response delays. This drives the actuator to synchronously track the defect location of the sheet metal to reject defective sheets. It ensures that under conditions of continuous sheet metal conveying and dynamic deformation, the rejection action always accurately hits the actual defect point, achieving synchronous tracking and precise rejection of defect locations under dynamic deformation conditions.

[0023] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the edge pulse encoder sensing module performs the following steps: the intelligent sensing system includes an edge pulse encoder and a laser displacement sensor. The edge pulse encoder is symmetrically installed on both sides of the conveyor roller of the calcium silicate board. The wheel of the edge pulse encoder maintains a constant contact pressure with the edge of the board, and collects the displacement pulse signal and lateral offset data of the edge of the board in the conveying direction in real time. The laser displacement sensor is arranged at four points at equal intervals along the lateral direction of the roller, and collects the lateral deviation and elastic flexural deformation of the board in a non-contact manner, covering the deformation characteristics of the entire cross-section in the width direction of the board, effectively solving the problem of fusion between contact and multi-point measurement, and realizing the complete capture of the deformation characteristics of the entire cross-section. The system employs Kalman filtering to fuse encoder pulse signals and laser displacement sensing signals, eliminating accumulated errors and measurement noise. Simultaneously, it integrates visual feedback from pre-set reference markers on the plate surface to correct accumulated errors in the edge pulse encoder. Through multi-source data fusion and visual calibration, the system significantly improves the long-term stability and absolute accuracy of displacement measurement. Furthermore, it introduces an elastic deformation compensation algorithm to extract the minute elastic flexural deformation characteristics of the plate during transport. The system fuses longitudinal displacement, lateral deviation, and elastic flexural deformation to generate a dynamic deformation vector field, which is transmitted in real-time to the intelligent inspection collaborative control center via high-frequency data stream. This provides continuously updated physical parameters for the dynamic correction of the defect marker coordinate system, enabling dynamic quantification and real-time transmission of deformation characteristics. It should be noted that two sets of high-precision edge pulse encoders are symmetrically installed on both sides of the calcium silicate board conveyor rollers. The encoder wheels are made of polyurethane-coated structure, maintaining a constant contact pressure of 0.5N with the edge of the board to ensure that the wheels rotate synchronously with the board without slippage. The encoders output 5000 pulse signals per revolution, corresponding to a displacement resolution of 0.05mm, and collect the longitudinal displacement and lateral offset of the board in the conveying direction in real time. Four laser displacement sensors are evenly spaced along the transverse direction of the rollers, with an adjacent spacing of 300mm, covering the entire 1200mm width of the board. The cross-section sensor employs a diffuse reflection laser ranging principle, with a sampling frequency set to 200Hz. It non-contactly collects the lateral deviation of the sheet metal during transport and the elastic flexural deformation caused by its own weight and unevenness of the roller conveyor. The encoder pulse signal and the laser displacement sensor signal are input to the intelligent inspection and collaborative control center. A Kalman filter algorithm is used for multi-source data fusion, with the process noise covariance matrix set to 0.001 times the identity matrix. The measurement noise covariance is set to 0.02mm for the encoder and 0.01mm for the laser sensor based on sensor calibration results, effectively eliminating wheel noise. To address the cumulative error of slippage and the environmental noise of laser measurement, a visual feedback calibration mechanism is introduced simultaneously. Three circular reference marks with a diameter of 5mm are pre-made on the surface of the plate at the initial conveying position. The position of the marks is photographed every 500mm by an industrial camera above, and compared with the cumulative displacement of the encoder. When the deviation between the two exceeds 0.2mm, an error correction command is triggered to clear the cumulative error of the encoder, ensuring the long-term stability and absolute accuracy of displacement measurement under long-term continuous conveying conditions. Based on the longitudinal displacement data after Kalman filtering fusion, the lateral deviation collected by the laser sensor, and the elastic flexural deformation measured synchronously by the four-point laser array, a dynamic deformation vector field with three degrees of freedom is constructed. The elastic flexural deformation uses a cubic spline interpolation algorithm to fit the cross-section of the four-point measurement values ​​to obtain the flexural amount at any position in the width direction of the plate. The flexural deformation compensation accuracy is controlled within ±0.05mm. The dynamic deformation vector field is generated at a frequency of 200Hz. Each vector contains a longitudinal displacement value, a lateral offset value, and a corresponding flexural deformation value. It is transmitted to the intelligent inspection and collaborative control center in the form of a data stream via EtherCAT real-time industrial Ethernet. The coordinate correction module performs the following steps: First, a reference coordinate system is established using the board image acquired at the start of transport. Natural textures and edge contours on the board surface are extracted as topological anchor point features. A unified absolute coordinate reference is established to provide a stable and reliable spatial reference for subsequent defect localization. During transport, the current frame image and the first frame reference are input into a twin neural network to calculate the topological offset. The dynamic deformation vector field input from the edge pulse encoder perception module is fused to achieve high-precision fusion of multi-source data, ensuring that the offset tracks the dynamic changes of the board in real time. A nonlinear deformation compensation model is constructed based on the topological offset and the dynamic deformation vector field to dynamically correct the defect marker coordinate system, achieving accurate alignment of defect coordinates under deformation conditions and accurately compensating for the nonlinear deformation of the board. It should be noted that when the front end of the board touches the deployed light-emitting sensor, a 5-megapixel area array camera is simultaneously triggered to acquire the initial image. The image resolution is set to 2448×2048 pixels, and the field of view covers 1200mm of the board width. The SIFT algorithm is used to extract no less than 200 natural texture feature points from the image. Combined with the Canny edge detection operator, the four-dimensional outline of the board is extracted to construct the initial topological anchor point set. The first frame reference coordinate system takes the intersection of the left edge of the board and the conveying direction as the origin, with the positive X-axis along the conveying direction and the positive Y-axis perpendicular to the conveying direction as the positive Y-axis, establishing an absolute coordinate system. All defect mark coordinates are uniformly represented in this coordinate system. The Siamese neural network uses VGG16 as the backbone network, with the input size set to 224×224 pixels. The feature extraction layer maps the current frame and the first frame image to a 256-dimensional feature space, calculates the Euclidean distance between feature vectors, and determines the topological offset based on the feature point matching degree. The output shows the board's position on the conveying line. When fusing the dynamic deformation vector field for longitudinal and lateral offsets in the conveying direction, the overall offset output by the twin neural network and the local deformation data measured by the encoder are weighted and fused using the timestamp as the synchronization reference. The weight coefficients are dynamically adjusted according to the real-time confidence of the sensors. After fusion, the offset data update frequency reaches 200Hz, keeping it synchronized with the conveying speed of the sheet metal. The nonlinear deformation compensation model adopts the thin plate spline interpolation algorithm, using the anchor point coordinates in the first frame reference coordinate system as the source point set and the corresponding point coordinates in the current frame after offset compensation as the target point set. The global deformation mapping function is calculated, which describes the nonlinear deformation law of the sheet metal caused by elastic deflection, lateral deviation and longitudinal stretching during the conveying process. After the defect mark coordinates are input into the nonlinear deformation compensation model, the corrected actual coordinates are obtained through mapping transformation. The corrected coordinate system is updated in real time at a frequency of 100Hz and transmitted to the rejection actuator via Ethernet to ensure that the subsequent rejection action can accurately hit the defect position in motion. The defective product positioning and rejection module performs the following steps: It receives the defect coordinates after real-time correction, maps them to the base coordinate system of the rejection actuator after coordinate transformation, generates a defect motion trajectory prediction sequence, ensures that the defect coordinates are consistent with the spatial reference of the actuator, lays the foundation for accurate tracking, and dynamically plans the optimal rejection path and triggering sequence based on the mechanical response delay parameters of the rejection actuator to compensate for the fluctuation of the conveying speed and the system hysteresis effect, achieves accurate synchronization between the rejection action and the arrival time of the defect, eliminates the influence of system hysteresis, drives the actuator to synchronously track the position of the plate defect that has undergone elastic deformation and deviation during continuous conveying according to the real-time updated coordinate trajectory, and performs rejection operation on the defective plate, ensuring that the rejection action always accurately hits the defect point under dynamic deformation conditions, and improves the reliability of rejection. It should be noted that after receiving the defect correction coordinates output in real time from the coordinate correction module, the defect position is mapped from the plate reference coordinate system to the basic coordinate system of the rejection actuator through coordinate transformation. The mapping process comprehensively considers the spatial relative relationship between the installation position of the rejection actuator and the conveyor roller. After calibration, the rotation matrix and translation vector are obtained, and the transformation accuracy is controlled within ±0.1mm. Based on the corrected defect coordinate sequence, combined with the current plate conveying speed and the displacement pulse signal fed back by the encoder in real time, a Kalman filter prediction algorithm is used to generate a defect motion trajectory prediction sequence. The prediction time domain is set to 200ms, covering the action response window of the rejection actuator, to ensure that the trajectory prediction is synchronized with the actual movement state of the plate. The rejection actuator adopts a high-speed pneumatic injection device. The response delay of the solenoid valve is determined to be 12ms after calibration. The full opening time of the jet valve core is 8ms, and the closing time is 6ms. Based on the defect motion trajectory prediction sequence and the mechanical response delay parameters, the optimal rejection trigger sequence is calculated, based on the defect movement to rejection. A trigger command is issued 15ms before the center position of the workstation to compensate for the inherent hysteresis of the actuator. At the same time, the conveyor speed fluctuation is monitored in real time. When the speed deviation exceeds ±0.05m / s, the trigger advance is dynamically adjusted to ensure that the spatiotemporal matching accuracy between the rejection action and the defect position is better than ±2mm, effectively compensating for the positioning error caused by the stability fluctuation of the conveyor system. The drive actuator synchronously tracks and rejects the defect positions of the sheet metal that have undergone elastic deformation and deviation during continuous conveying based on the real-time updated coordinate trajectory and trigger sequence. The air nozzle array of the high-speed pneumatic jet device is arranged laterally along the roller conveyor with a spacing of 25mm, covering the entire 1200mm cross section of the sheet metal in the width direction. The controller selects the corresponding nozzle group to execute the jetting action according to the component of the defect correction coordinate in the width direction of the sheet metal. The jetting duration is set to 50ms to ensure that the defective sheet metal is reliably blown away from the conveyor line. After rejection, the rejection status is confirmed by photoelectric sensors and the rejection result is fed back to the production management system, forming a complete closed-loop control of defect identification, classification, positioning and rejection.

[0024] The following describes the workflow of the automatic identification, marking, and removal system for surface defects in calcium silicate boards.

[0025] After the system is started, the intelligent inspection and collaborative control center simultaneously triggers the multi-angle polarization light source and area array camera to sequentially collect the original image sequence at four polarization angles of 0°, 45°, 90° and 135° along the conveyor roller of the production line. Through differential operation and weighted fusion, a high-contrast polarization feature map is generated, which effectively suppresses the shadow of wet dust similar to the background color of the board and enables the microcrack outline to be separated from the dynamic background with high stability. Subsequently, the morphological reconstruction unit adopts the improved fruit fly algorithm to adaptively optimize the structural elements and performs morphological opening and closing reconstruction and connected domain analysis on the polarization feature map to accurately reconstruct the geometric morphology of fractured and blurred microcracks. After completing the crack geometry reconstruction, the stress field state classification module controls the robotic arm to move the laser speckle interferometer device directly above the crack, applies micro-perturbation excitation and acquires a speckle interferometer image sequence, generates a high-resolution strain field cloud map through phase demodulation and finite difference method, and extracts the strain concentration coefficient and strain gradient feature parameters; the energy level calibration unit, based on the intelligent classification model constructed by the deep Q network, inputs the above feature parameters into the trained and converged model, dynamically outputs the crack propagation energy level calibration results, and realizes real-time classification of the degree of defect hazard; During the defect marking and rejection stage, the edge pulse encoder sensing module collects the longitudinal displacement, lateral deviation, and elastic flexural deformation of the board in real time through the encoder and laser displacement sensor. After Kalman filtering, a dynamic deformation vector field is generated. The coordinate correction module establishes the first frame reference coordinate system based on the initial image. By comparing the topological structure offset between the current image and the first frame through a twin neural network, a nonlinear deformation compensation model is constructed by fusing the dynamic deformation vector field to dynamically correct the defect mark coordinates. The defect location and rejection module maps the corrected coordinates to the rejection actuator coordinate system. Combined with the equipment response delay and speed fluctuation, the trigger timing is dynamically planned to drive the high-speed pneumatic injection device to synchronously track and reject the moving defect position, forming a complete closed-loop control of defect identification, classification, location, and rejection.

[0026] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An automatic identification, marking, and rejection system for surface defects in calcium silicate boards, comprising an intelligent inspection and collaborative control center, characterized in that, The intelligent inspection collaborative control center has the following communication connections: The polarization dust suppression and reconstruction module is used for morphological reconstruction optimized by polarization differential imaging fusion and improved fruit fly algorithm. It removes surface dust interference through polarization characteristics and uses fruit fly algorithm to adaptively optimize structural elements to accurately reconstruct the microcrack contour. The stress field state classification module is used to accurately reconstruct the crack geometry. It uses a laser speckle interferometry fusion deep Q network to build an intelligent classification model, measures the strain field of the micro-region at the crack tip, and dynamically calibrates the crack propagation energy level through reinforcement learning to classify the defects of the calcium silicate board in real time. The edge pulse encoder sensing module is used to collect data on the displacement, deviation, and minute elastic deformation of the sheet material in real time during the conveying process through a deployed intelligent sensing system, and generate a dynamic deformation vector field. The coordinate correction module is used to integrate the twin neural network. Taking the plate image coordinates at the start of the transmission as the first frame reference, it compares the topological changes of the current image with the first frame in real time during the transmission process. Combined with the dynamic deformation vector field input from the edge pulse encoder, it dynamically corrects the defect mark coordinate system. The defective product positioning and rejection module is used to receive real-time corrected coordinate mappings and map them into the coordinate system of the rejection actuator. It also dynamically plans the optimal rejection path or timing to compensate for equipment response delays and drives the actuator to synchronously track the defective positions of the sheet metal to reject defective sheets.

2. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 1, characterized in that: The polarization dust suppression reconstruction module includes a polarization differential imaging unit and a morphological reconstruction unit; The polarization differential imaging unit is used to capture the differences in polarization characteristics of different materials on the surface of the board by using multi-angle polarization light source and differential imaging technology, actively suppress the shadow of wet dust similar to the background color, and separate microcracks from the dynamic background with high contrast. The morphological reconstruction unit, based on the polarization-separated image, employs an improved fruit fly algorithm, using crack contour integrity and background noise suppression as fitness to adaptively optimize the structural elements of morphological reconstruction. Through iterative optimization, it accurately reconstructs the fractured and blurred microcrack contours.

3. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 2, characterized in that: The polarization differential imaging unit performs the following steps: The multi-angle polarized light source deployed on the production line is controlled to irradiate the surface of the calcium silicate board with different polarization angles in sequence, and the area array camera is simultaneously triggered to acquire the original image sequence under the corresponding polarization state. Differential operations are performed on images with different polarization angles under the same field of view to extract the polarization degree feature parameters of dust and cracks caused by material differences, and a high-contrast polarization feature map is constructed. By fusing polarization feature maps with regular grayscale images at the pixel level, the shadows of wet dust that are highly similar to the background color of the board are suppressed, making the microcrack outline stand out from the dynamic dust background.

4. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 2, characterized in that: The morphological reconstruction unit performs the following steps: The high-contrast polarization feature map output by the polarization differential imaging unit is received, and the population of structural elements for morphological reconstruction is initialized. Each individual represents the combination of shape and scale parameters of the structural element. Using crack profile integrity and background noise suppression rate as fitness functions, an improved fruit fly algorithm is used to iteratively update the structural element parameters, and the global optimal structural element is adaptively optimized through the odor concentration determination mechanism. Based on the global optimal structuring element, morphological opening and closing reconstruction and connected domain analysis are performed on the polarization feature map to accurately reconstruct the fractured and blurred microcrack contours.

5. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 2, characterized in that: The stress field state classification module includes a strain field measurement unit and an energy level calibration unit; The strain field measurement unit is used to apply micro-perturbations to the identified crack region using laser speckle interferometry to non-contactly measure the strain field distribution in the micro-region at the crack tip, and to quantify the stress concentration and potential propagation driving force around the crack. The energy level calibration unit constructs an intelligent classification model based on strain field measurement results and a deep Q-network. It takes the measured strain field data as the state input and crack propagation risk as the action value. Through a reinforcement learning mechanism, it dynamically calibrates the crack propagation energy level in a simulated high-humidity curing environment and outputs the real-time classification result of the degree of defect hazard.

6. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 5, characterized in that: The strain field measurement unit performs the following steps: A laser speckle interferometer is deployed and aligned above the identified crack area to apply a controllable micro-perturbation excitation to the micro-region at the crack tip, thereby exciting the local strain response in the surrounding area of ​​the crack. A series of speckle interferometric images were continuously acquired before and after the disturbance. The displacement field distribution of the micro-region at the crack tip was calculated by the phase demodulation algorithm to generate a high-resolution strain field cloud map. The strain concentration factor and strain gradient characteristic parameters at the crack tip are extracted from the strain field cloud map to quantify the stress concentration and potential propagation driving force around the crack.

7. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 5, characterized in that: The energy level calibration unit performs the following steps: The strain concentration factor and strain gradient characteristic parameter output by the strain field measurement unit are used as inputs to the state space. A smart hierarchical model is constructed by combining it with a deep Q-network, and the crack propagation risk level is set as the action space. In a digital twin interactive platform simulating a high-humidity maintenance environment, the intelligent hierarchical model is driven to perform actions and obtain reward and punishment feedback, and the network parameters are iteratively trained until the strategy converges. The strain field data obtained from real-time measurement is input into the intelligent classification model that has been trained and converged, and the crack propagation energy level calibration results are dynamically output to complete the real-time classification of the degree of defect hazard.

8. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 5, characterized in that: The edge pulse encoder sensing module performs the following steps: The intelligent sensing system includes an edge pulse encoder and a laser displacement sensor. The edge pulse encoder is symmetrically installed on both sides of the conveyor roller of the calcium silicate board. The wheel of the edge pulse encoder maintains a constant contact pressure with the edge of the board and collects the displacement pulse signal and lateral offset data of the edge of the board in the conveying direction in real time. The laser displacement sensor is arranged at four points at equal intervals along the lateral direction of the roller and collects the lateral deviation and elastic flexural deformation of the board in a non-contact manner, covering the deformation characteristics of the entire cross section in the width direction of the board. Kalman filtering is used to fuse the encoder pulse signal and the laser displacement sensing signal to eliminate accumulated error and measurement noise. Simultaneously, visual feedback from the reference markers on the board surface is fused to correct the accumulated error of the edge pulse encoder. An elastic deformation compensation algorithm is introduced to extract the minute elastic flexural deformation characteristics of the plate during the conveying process. The longitudinal displacement, lateral deviation and elastic flexural deformation are fused to generate a dynamic deformation vector field, which is transmitted to the intelligent inspection and collaborative control center in real time through high-frequency data stream.

9. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 8, characterized in that: The coordinate correction module performs the following steps: A reference coordinate system for the first frame is established using the board image acquired at the start of the transmission, and the natural texture and edge contour of the board surface are extracted as topological anchor point features. During the transmission process, the current frame image and the first frame reference are input into the Siamese neural network to calculate the topological offset and fuse the dynamic deformation vector field input from the edge pulse encoder perception module. A nonlinear deformation compensation model is constructed based on topological offset and dynamic deformation vector field to dynamically correct the defect marker coordinate system.

10. The automatic identification, marking, and rejection system for surface defects of calcium silicate board according to claim 9, characterized in that: The defective product location and rejection module performs the following steps: The defect coordinates are received in real time and mapped to the base coordinate system of the removal actuator after coordinate transformation to generate a defect motion trajectory prediction sequence. By combining the mechanical response delay parameters of the rejection actuator, the optimal rejection path and triggering sequence are dynamically planned to compensate for the fluctuation of conveying speed and the system hysteresis effect. The drive actuator synchronously tracks the location of defects in the sheet metal that undergoes elastic deformation and deviation during continuous conveying based on the real-time updated coordinate trajectory, and removes defective sheet metal.

Images