A machine vision-based system for sorting appearance defects in calcium silicate board insulation boards.

Abstract

This invention relates to the field of industrial vision inspection technology, specifically disclosing a machine vision-based sorting system for appearance defects in calcium silicate insulation boards. The system includes an intelligent sorting control center, which is communicatively connected to the following modules: an anti-interference contour extraction module, used to optimize the high and low threshold parameters of Canny edge detection using an improved fruit fly algorithm, and integrating region growing and morphological skeleton extraction methods to locate candidate defect areas on the calcium silicate insulation board and refine the contours; and a damage weighting module, which, based on the geometric features of the refined contours, uses deep reinforcement learning to evaluate the damage weight of defects to insulation and structural performance. This invention, by integrating an improved fruit fly algorithm to optimize edge detection thresholds and combining region growing and morphological skeleton extraction, effectively overcomes the interference of low contrast and porous textures on the surface of calcium silicate insulation boards, achieving accurate and stable extraction of edge-peeling defect contours, and significantly improving the accuracy and robustness of defect detection.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection technology, and in particular to a sorting system for appearance defects of calcium silicate board insulation board based on machine vision. Background Technology

[0002] With the increasing market demand for building energy conservation, calcium silicate board, as an ideal thermal insulation material, is widely used in the construction and industrial fields, mainly to improve the energy efficiency and fire resistance of buildings. Appearance defect sorting technology plays a crucial role in the production of calcium silicate board, and timely and accurate identification and removal of defects can ensure the consistency and stability of the product.

[0003] For example, Chinese Patent Publication No. CN115456944A describes an intelligent identification method for appearance defects of ceramic substrates based on target detection. This method addresses the multi-form and multi-category problems of appearance defects on ceramic substrates, ensuring the detection cycle for multiple types of defects on ceramic substrates while improving the compatibility of product types and defect forms.

[0004] In existing technologies, calcium silicate insulation boards suffer from low contrast and porous texture interference on their surface, making it difficult to reliably separate the precise contours of edge-peeling defects from the background. Furthermore, the lack of dynamic grading criteria based on defect geometric features to assess the damage weight to the insulation performance and structural strength of the board results in single-size sorting standards failing to reflect actual quality impacts. Additionally, fluctuations in production line speed and changes in ambient lighting can lead to inconsistencies in the sorting grade of the same defect. Moreover, the real-time control logic of the actuator lacks adaptive adjustment capabilities to dynamic disturbances, further affecting sorting stability and consistency. Therefore, this paper proposes a machine vision-based calcium silicate insulation board appearance defect sorting system to address the aforementioned problems. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a machine vision-based sorting system for appearance defects of calcium silicate insulation 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 a machine vision-based sorting system for appearance defects in calcium silicate board insulation boards, including an intelligent sorting main control center, which is communicatively connected to the following modules: The anti-disturbance contour extraction module is used to optimize the high and low threshold parameters of Canny edge detection using an improved fruit fly algorithm, integrate region growth and morphological skeleton extraction methods, locate the defect candidate area on the calcium silicate board insulation board and refine the contour, realize the accurate contour extraction of edge peeling defects, and enhance the edge extraction stability under low contrast and porous texture. The damage weighting and grading module, based on the geometric features of the refined contour, uses deep reinforcement learning to evaluate the damage weight of defects on thermal insulation and structural performance. It also achieves dynamic grading of scrap, repair, and pass through three categories through fuzzy clustering, which solves the problem that a single size standard cannot reflect the actual quality impact. It establishes a nonlinear mapping between damage and grading to improve the rationality of grading. The dual-delay depth strategy module is used to train a strategy network based on a dual-delay depth deterministic strategy, using speed sensor and light sensor signals as states and thrust magnitude and rejection delay as actions. It outputs the optimal execution adjustment amount under non-stationary disturbances in real time to compensate for dynamic disturbances in the production line and ensure the consistency of sorting actions. The disturbance compensation execution module is used to dynamically correct the thrust curve of the pneumatic push rod and the rejection trigger time window according to the optimal execution adjustment amount output by the strategy network, so as to ensure that the same defect receives consistent sorting action under different production line conditions and eliminate the impact of speed and light fluctuations on rejection accuracy. The closed-loop incremental self-evolution module is used to collect misclassified samples from manual re-inspection or production line feedback. It adopts an elastic weight consolidation and memory replay mechanism to update the parameters of the deep Q network and fuzzy clustering without forgetting old knowledge. It also uses some parameters of the strategy network as a transferable base and makes rapid fine-tuning based on recent production line disturbance data. This enables the execution strategy to adaptively migrate and continuously optimize across different seasons and shifts, achieve cross-working-condition knowledge inheritance, and continuously improve sorting robustness.

[0007] Preferably, the anti-interference contour extraction module includes a fruit fly optimized threshold unit and a region skeleton refinement unit; The fruit fly optimized threshold unit is used to use the gradient amplitude distribution and local texture difference of the calcium silicate board insulation board surface image as the fitness function, and adopts the improved fruit fly algorithm to dynamically search the high and low threshold parameters of the Canny operator, outputting a binary edge map, avoiding edge breakage or false edges caused by manual setting, enhancing robustness to illumination and texture fluctuations, realizing threshold adaptive matching, and improving edge integrity under illumination fluctuations. The region skeleton refinement unit is used to fuse the region growth to lock the defect candidate area, and refine the contour along the main direction of the defect by combining the morphological skeleton extraction method. It eliminates false edge responses caused by the porous structure of the insulation board, outputs a closed and continuous real defect sub-pixel contour, effectively filters out porous texture interference, and outputs a high-precision closed defect boundary.

[0008] Preferably, the fruit fly optimized threshold unit performs the following steps: The gradient magnitude distribution of images on the surface of calcium silicate board insulation board was collected. The fitness function was constructed using the entropy value of the gradient histogram within the local window and the peak signal-to-noise ratio. The position of the fruit fly population was initialized as a candidate high and low threshold parameter pair, which effectively quantified the local texture complexity and edge contrast, laying the foundation for adaptive threshold search. An improved fruit fly algorithm is adopted, which introduces chaotic mapping and adaptive step size decay mechanism to guide individual fruit flies to sniff and search in the gradient magnitude space. The goal is to maximize the fitness function value and iteratively update the optimal position of the population, so as to avoid the threshold parameter from getting trapped in local extrema and ensure the stable convergence of the global optimal solution. After iterative convergence, the optimal position is decoded into the high and low thresholds of the Canny operator, which are used to suppress the pseudo-edge response caused by porous texture and retain the real edges with weak gradients, respectively. The output is a binary edge map with adaptive thresholds, which simultaneously realizes pseudo-edge suppression and weak edge preservation, and improves the completeness of edge extraction.

[0009] Preferably, the region skeleton refinement unit performs the following steps: Using the edge points of connected regions in the binary edge map as the initial seed points for region growth, and based on the consistency criterion of local gray mean and gradient direction, the boundary of the candidate defect region is locked by expanding layer by layer in the eight neighborhoods, effectively suppressing the interference of porous texture and accurately locating the boundary of the defect region. Morphological skeleton extraction is performed on the locked defect candidate region. Boundary pixels are iteratively stripped along the main defect direction until a skeleton line with a single pixel width is obtained. At the same time, short branches and isolated skeleton segments with a length less than a preset threshold are removed, the main defect skeleton is retained, false branches are removed, and the purity of contour extraction is improved. The skeleton lines are back-projected onto the original gradient image, and a sub-pixel edge interpolation algorithm is used to fit the gradient maxima on both sides of the skeleton to output a closed and continuous real sub-pixel contour of the defect, thus achieving a sub-pixel level continuous closed contour and meeting the requirements of high-precision defect quantization.

[0010] Preferably, the damage weighting grading module includes a deep Q-network weighting unit and a fuzzy clustering grading unit; The deep Q-network weighting unit is used to train the deep Q-network to learn the nonlinear damage weights of different defect geometric features on the thermal insulation performance decay and structural strength degradation based on the porous matrix and layered structure of the calcium silicate board insulation board, and output the feature weight vector oriented towards the actual service performance of the board, quantifying the differential damage contribution of defects to service performance. The fuzzy clustering grading unit is used to input the feature weight vector output by the deep Q network into fuzzy C-means clustering. Combined with the scrap threshold standard and repair feasibility rules of calcium silicate board insulation board, the defects are dynamically divided into three defect levels: scrap, repair, and pass, so as to realize the flexible grading of boundary samples and reduce the misjudgment rate.

[0011] Preferably, the deep Q-network weighting unit performs the following steps: Extract the geometric feature set of the region surrounded by the sub-pixel contour of the real defect, including area, elongation, roundness, convex hull ratio, and the angle between the principal axis direction and the plate layer, and construct a 20-dimensional feature input vector to ensure that the geometry of the defect is fully quantified and to provide complete input for damage assessment. A dual-hidden-layer deep Q-network is constructed, with the feature input vector as the state and the combination of the defect-induced thermal insulation performance attenuation coefficient and the structural strength degradation coefficient as the action space. An empirical replay and target network soft update mechanism are used to train the Q-value approximator, realizing the nonlinear mapping between defect geometric features and damage weights, thereby improving the evaluation accuracy. After the network converges, the feature input vector of the defect to be classified is forward propagated, and the action combination that maximizes the cumulative damage Q value is output. That is, the thermal insulation performance attenuation weight and structural strength degradation weight of the defect are formed into a feature weight vector. The damage weight that reflects the actual service performance is output to support the subsequent dynamic classification decision.

[0012] Preferably, the fuzzy clustering hierarchical unit performs the following steps: The feature weight vector is used as the input sample for fuzzy C-means clustering. The number of clusters is initialized to 3, corresponding to scrap, repair and pass respectively. The membership matrix is ​​randomly initialized and the fuzzy weighting index is set to effectively establish the nonlinear mapping relationship between defect damage and quality level. Iteratively calculate the cluster centers and sample membership degrees of each category. When the change in the membership degree matrix between two adjacent iterations is less than the threshold, convergence is achieved. Output the membership degree values ​​of each defective sample to the three categories as the classification confidence level to ensure that the classification results are stable and have repeatable confidence expression. Combining the preset scrapping threshold standard and repair feasibility rules, the category with the highest membership degree is used as the initial grade. If the maximum membership degree is lower than the confidence level lower bound, it is marked as pending re-inspection. Finally, dynamic grading labels are output to achieve flexible adaptation of grading rules and a safe fallback for abnormal samples.

[0013] Preferably, the dual-delay depth strategy module performs the following steps: The instantaneous speed deviation of the production line collected by the speed sensor and the illuminance deviation collected by the light sensor are spliced ​​together to form a state vector. The thrust reference value of the pneumatic push rod and the rejection trigger delay are used as two-dimensional continuous actions to construct a Markov decision process, so that the system can perceive and quantify the impact of dynamic disturbances on rejection accuracy. A dual-delay deep deterministic policy gradient algorithm is adopted to construct the current policy network, the target policy network, and the dual-Q network respectively. The target policy smoothing noise is added to the action output, and the Q-value estimation is stabilized by soft update, suppressing Q-value overestimation and improving the stability and convergence reliability of policy learning. After the policy network training converges, it receives the current state vector forward propagation output of the optimal execution adjustment amount in real time, including the thrust correction coefficient and delay compensation value, and sends it to the intelligent sorting main control center to realize millisecond-level adaptive adjustment of execution parameters under dynamic working conditions.

[0014] Preferably, the disturbance compensation execution module performs the following steps: The thrust correction coefficient and delay compensation value are extracted from the intelligent sorting main control center. The thrust reference value is multiplied by the thrust correction coefficient and mapped to the nonlinear curve of the PWM duty cycle of the pneumatic push rod. This enables smooth adjustment of the thrust soft start and soft stop, effectively suppresses mechanical shock of the push rod, and extends the service life of the solenoid valve. The opening time and width of the rejection trigger time window are dynamically adjusted based on the delay compensation value, so that the center of the time window is aligned with the predicted time when the defect target arrives at the push rod action area, compensating for the positional offset caused by speed fluctuations, eliminating rejection position deviations caused by speed fluctuations, and improving hit accuracy. The corrected thrust curve parameters and time window parameters are sent to the programmable logic controller to drive the pneumatic push rod to perform the rejection action, so that the same defect can be sorted in a consistent manner under different production line conditions, ensuring the consistency and repeatability of the sorting action and stabilizing the rejection effect.

[0015] Preferably, the closed-loop incremental self-evolution module performs the following steps: Collect misclassified samples labeled by the manual re-inspection platform and samples with deviations in sorting results reported from the end of the production line, and store them in a timestamped misclassified sample buffer. Each sample contains the original image, feature weight vector and correct grading label, and continuously learns new misclassification patterns while maintaining the stability of the original grading ability. The elastic weight consolidation algorithm is used to calculate the important weights of the deep Q network and fuzzy clustering parameters in the current task. Combined with the memory replay mechanism, misclassified samples are uniformly sampled from the buffer for offline incremental training to suppress catastrophic forgetting, correct misjudgment boundaries, and maintain the stability of the original clustering topology. The parameters of the first two layers of the dual-delay deep strategy network are frozen as a transferable base. The output layer parameters are quickly fine-tuned using production line disturbance data within the last 24 hours. This enables the execution strategy to adapt to different seasons and shifts, quickly adapt to environmental and operational changes, and ensure consistent removal.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This machine vision-based calcium silicate insulation board appearance defect sorting system optimizes the edge detection threshold by integrating an improved fruit fly algorithm and combining region growth and morphological skeleton extraction. It can effectively overcome the interference of low contrast and porous texture on the surface of calcium silicate insulation board, achieve accurate and stable extraction of edge peeling defect contours, and significantly improve the accuracy and robustness of defect detection.

[0017] 2. This machine vision-based calcium silicate board insulation board appearance defect sorting system uses deep reinforcement learning to evaluate the damage weight of defect geometric features and combines fuzzy clustering to achieve dynamic grading. It can reflect the actual nonlinear damage impact of defects on insulation performance and structural strength, overcome the shortcomings of traditional single-size sorting standards that cannot reflect quality differences, and make the sorting results more in line with the product service performance requirements.

[0018] 3. This machine vision-based calcium silicate board insulation board appearance defect sorting system constructs an execution adjustment mechanism based on a dual-delay depth deterministic strategy gradient, which can perceive production line speed fluctuations and ambient light changes in real time, adaptively adjust the rejection thrust and action delay, and ensure that the same defect receives consistent sorting action under different dynamic disturbance conditions, significantly improving the consistency and stability of sorting. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the workflow of a machine vision-based calcium silicate board insulation board appearance defect sorting system of the present invention. Figure 2 This is a schematic diagram of the module structure of a machine vision-based calcium silicate board insulation board appearance defect sorting system according to the present invention. Detailed Implementation

[0020] 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.

[0021] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a machine vision-based sorting system for appearance defects in calcium silicate insulation boards, comprising an intelligent sorting control center, which is connected to the following modules for communication: The anti-disturbance contour extraction module is used to optimize the high and low threshold parameters of Canny edge detection using an improved fruit fly algorithm, integrate region growth and morphological skeleton extraction methods, locate the defect candidate area on the calcium silicate board insulation board and refine the contour, realize the accurate contour extraction of edge peeling defects, and enhance the edge extraction stability under low contrast and porous texture. The anti-disturbance contour extraction module includes a fruit fly optimized threshold unit and a region skeleton refinement unit. The Drosophila optimization thresholding unit uses the gradient amplitude distribution and local texture differences of the calcium silicate board insulation surface image as the fitness function. It employs an improved Drosophila algorithm to dynamically search for high and low threshold parameters of the Canny operator, outputting a binary edge map. This avoids edge breaks or false edges caused by manual setting, enhances robustness to illumination and texture fluctuations, achieves adaptive threshold matching, and improves edge integrity under illumination fluctuations. The unit collects the gradient amplitude distribution of the calcium silicate board insulation surface image, constructs a fitness function using the entropy value of the gradient histogram within a local window and the peak signal-to-noise ratio, and initializes the Drosophila population position as candidate high and low threshold parameter pairs, effectively quantifying local texture differences. To improve the complexity and edge contrast, an adaptive threshold search is established. This paper adopts an improved fruit fly algorithm that introduces chaotic mapping and adaptive step size decay mechanism to guide individual fruit flies to sniff and search in the gradient magnitude space. The goal is to maximize the fitness function value and iteratively update the optimal position of the population, avoiding the threshold parameter from getting trapped in local extrema and ensuring the stable convergence of the global optimal solution. The optimal position after iterative convergence is decoded into the high threshold and low threshold of the Canny operator, which are used to suppress the pseudo-edge response caused by porous texture and retain the real edge of weak gradient, respectively. The output is a threshold-adaptive binary edge map, which simultaneously realizes pseudo-edge suppression and weak edge preservation, and improves the completeness of edge extraction. It should be noted that, firstly, 8-bit grayscale images of the calcium silicate board insulation board surface were acquired using a linear scan camera, and the gradient magnitude of each pixel was calculated using the Sobel operator. The global gradient magnitude distribution was statistically obtained. Using a local sliding window of 16×16 pixels as the unit, the entropy value (quantifying the uncertainty of the gradient direction) and peak signal-to-noise ratio (PSNR, reflecting the contrast between defect edges and background texture) of the gradient histogram within the window were calculated. The fitness function was constructed by weighting the two values, with the entropy value weight set to 0.6 and the PSNR weight set to 0.4. The initial population size of fruit flies was 20, and the two-dimensional position of each individual was... Corresponding to the Canny high threshold (search interval [30, 150]) and low threshold (search interval [10, 80]), and with a mandatory constraint that the low threshold is at least 15 gray levels smaller than the high threshold, the initial position of the population is generated using a Logistic chaotic mapping (r=3.9, initial value x0=0.32) to ensure uniform coverage of the search space; at the start of the iteration, each fruit fly individual generates a random direction vector using the chaotic mapping, searches for candidate threshold pairs in the neighborhood along the current sniffing radius (initial radius set to 12 gray levels) in the gradient magnitude space, calculates the fitness function value of the new position, and if the candidate position If the fitness is higher than the original position, the individual's optimal position is updated. After each round of sniffing and updating of all individuals, the global optimal position of the population is found, and an adaptive step-size decay mechanism is used to adjust the sniffing radius. The decay coefficient is 0.95, and the minimum radius is limited to 0.5 gray levels. The iteration process continues until the rate of change of the global optimal fitness value is less than 0.1% for 15 consecutive generations. The number of iterations is 40 to 60 generations. During this process, the introduction of chaotic mapping avoids the threshold parameter from getting trapped in local extrema, ensuring a balance between the ability of the high threshold to suppress porous textures and the ability of the low threshold to preserve weak gradient true edges. The iteration results... After convergence, the global optimal position of the population is decoded into a high threshold and a low threshold of the Canny operator. The high threshold obtained by decoding (typical convergence value between 60 and 90) is used in the gradient non-maximum suppression stage to eliminate pseudo-edge responses with low amplitude caused by the porous structure of the insulation board. The low threshold (typical convergence value between 25 and 45) is used in the hysteresis tracking stage of the dual threshold connection to retain weak gradient segments with gentle gray-level changes in the defect edges. Then, the edge connection operation is performed to output a threshold-adaptive binary edge map, in which each pixel value is 0 or 255, where 255 indicates that it is determined to be an edge point. The region skeleton refinement unit is used to fuse region growth to lock defect candidate areas. Combined with morphological skeleton extraction methods, it refines the contour along the main defect direction, eliminates false edge responses caused by the porous structure of the insulation board, and outputs closed and continuous real defect sub-pixel contours. It effectively filters out porous texture interference and outputs high-precision closed defect boundaries. Using the edge points of connected components in the binary edge map as the initial seed points for region growth, it expands and locks the boundary of the defect candidate area layer by layer in the eight neighborhoods according to the consistency criterion of local gray-level mean and gradient direction. It effectively suppresses porous texture interference and accurately locates the defect region boundary. Morphological skeleton extraction is performed on the locked defect candidate area. It iteratively peels off the boundary pixels along the main defect direction until a skeleton line with a single pixel width is obtained. At the same time, short branches and isolated skeleton segments with a length less than a preset threshold are removed. The main defect skeleton is retained, false branches are removed, and the purity of contour extraction is improved. The skeleton line is back-projected onto the original gradient image. The sub-pixel edge interpolation algorithm is used to perform curve fitting on the gradient maxima points on both sides of the skeleton to output closed and continuous real defect sub-pixel contours. It achieves sub-pixel level continuous closed contours and meets the requirements of high-precision defect quantization. It should be noted that in the binary edge map, all 8-connected edge points are first extracted as the initial seed point set for region growing. For each seed point, consistency is determined based on the local gray-level mean and gradient principal direction angle within its 3×3 neighborhood: when the difference between the gray-level value of a neighboring pixel and the gray-level value of the seed point is less than a preset gray-level tolerance of 8 gray levels, and the gradient direction angle does not exceed 15°, the pixel is included in the candidate region boundary. This process is repeated layer by layer until no new pixels are added in two consecutive rounds, thus locking the defect candidate region. During the expansion process, if the candidate region area is less than 20 pixels, it is judged as a noise point and directly removed, not entering the subsequent skeleton extraction process. For the locked defect candidate region, the Hilditch thinning algorithm is used to iteratively peel off the boundary pixels along the defect principal direction. In each iteration, only boundary points that satisfy the topological connectivity preservation condition are deleted, and the deletion does not destroy the 8-connectivity of the skeleton. The iteration continues until all boundary pixels are no longer deletable, thus obtaining the desired result. A single-pixel-width skeleton line is generated. Then, each branch node of the skeleton line is traversed, and the path length of each branch is calculated. Short branches with a length of less than 8 pixels and isolated skeleton segments not connected to the main skeleton are removed, leaving only the main skeleton trunk with a length of not less than 12 pixels to ensure that the skeleton line truly reflects the geometric extension direction of the defect. The refined skeleton line is back-projected onto the original gradient magnitude image, and local maxima of the gradient magnitude are searched within a range of 2 pixels on both sides of the skeleton line. For each pair of symmetrically distributed maxima, a cubic spline interpolation function is used to perform curve fitting along the normal of the skeleton line with a fitting step size of 0.2 pixels and a minimum of 5 interpolation points. All fitting points are connected to generate a closed sub-pixel contour of the real defect. The contour curvature is required to be continuous and without self-intersection. Before output, the contour is length normalized to ensure that the contour point sequence is evenly arranged in a clockwise direction and the spacing between adjacent contour points is controlled between 0.3 and 0.8 pixels. The damage weighting and grading module, based on the geometric features of the refined contour, uses deep reinforcement learning to evaluate the damage weight of defects on thermal insulation and structural performance. It also achieves dynamic grading of three categories—scrap, repair, and pass—through fuzzy clustering, which solves the problem that a single size standard cannot reflect the actual quality impact. It establishes a nonlinear mapping between damage and grading to improve the rationality of grading. The damage weighting and grading module includes a deep Q-network weighting unit and a fuzzy clustering grading unit. The deep Q-network weighting unit is used to train the deep Q-network to learn the nonlinear damage weights of different defect geometric features on the attenuation of thermal insulation performance and the degradation of structural strength, based on the porous matrix and layered structure of the calcium silicate board insulation. It outputs a feature weight vector oriented towards the actual service performance of the board, quantifying the differentiated damage contribution of defects to service performance. It extracts the geometric feature set of the region enclosed by the sub-pixel contour of the real defect, including area, elongation, roundness, convex hull ratio, and the angle between the principal axis direction and the board's layering, constructing a 20-dimensional feature input vector to ensure that the defect geometry is fully quantified, providing complete input for damage assessment. A dual-hidden-layer deep Q-network is constructed, with the feature input vector as the state and the combination of the thermal insulation performance attenuation coefficient and the structural strength degradation coefficient of the defect as the action space. An empirical replay and target network soft update mechanism are used to train the Q-value approximator to achieve a nonlinear mapping between defect geometric features and damage weights, thereby improving the evaluation accuracy. After the network converges, the feature input vector of the defect to be classified is forward propagated, and the action combination that maximizes the cumulative damage Q-value is output, which is the thermal insulation performance attenuation weight and the structural strength degradation weight of the defect. This forms a feature weight vector, and the output damage weights that reflect the actual service performance support subsequent dynamic classification decisions. It should be noted that after extracting the sub-pixel contour of the real defect, the geometric feature set of the region it encloses is first calculated. Specifically, Green's formula is used to calculate the pixel-level area within the contour and convert it into the actual physical area; the elongation is obtained by the aspect ratio of the minimum bounding rectangle of the contour; the roundness is calculated by the ratio of the area to the square of the perimeter, where the perimeter is obtained by accumulating the Euclidean distances of the contour points; the convex hull ratio is constructed using the Andrew monotonic chain algorithm, and the ratio of the original contour area to the convex hull area is calculated; the principal axis direction is obtained by calculating the second moment of the contour and is compared with the thermal insulation... The included angle of the precast layer direction is used as an independent feature. In addition, a 20-dimensional feature input vector is constructed by combining the compactness, rectangularity, eccentricity of defects, and the first eight low-frequency components of the Fourier descriptor. All features are normalized to zero mean and unit variance to eliminate the influence of dimensions. The number of input layer nodes of the double hidden layer deep Q network is 20, the number of first hidden layer nodes is 128, and the number of second hidden layer nodes is 64. The activation function is ReLU, and the output layer is a 5×5 discrete action space, corresponding to thermal insulation performance attenuation coefficient and structural strength degradation coefficient both being within 0. Within the interval up to 1, five levels are selected with a step size of 0.25. During training, the experience replay pool capacity is set to 20,000 sets, and 32 sets of samples are randomly sampled for batch updates each time. The target network is soft-updated every 100 steps, with a soft update coefficient τ of 0.01. The Q-value loss function uses mean squared error, the optimizer is Adam, the learning rate is set to 0.001, the discount factor γ is set to 0.95, and the exploration strategy uses the ε-greedy method, with an initial ε of 0.9, which decays by 0.01 every 500 steps to a minimum of 0.05. The 20-dimensional features of the defects to be classified are... The input vector is forward propagated to calculate the Q value corresponding to all 25 action combinations. The action corresponding to the largest Q value is selected as the output, which is the thermal insulation performance attenuation weight and structural strength degradation weight of the defect. The two are combined to form a 2D feature weight vector. This feature weight vector retains the nonlinear damage mapping relationship between the defect geometry and the service performance of the insulation board, and serves as the input basis for subsequent fuzzy clustering grading units. In actual operation, multiple defects on each insulation board independently output their feature weight vectors and participate in the grading judgment separately, without interfering with each other. The fuzzy clustering grading unit is used to input the feature weight vector output by the deep Q-network into fuzzy C-means clustering. Combined with the scrap threshold standard and repair feasibility rules for calcium silicate board insulation, defects are dynamically classified into three defect levels: scrap, repair, and pass. This achieves flexible grading of boundary samples, reducing the false positive rate. The feature weight vector is used as the input sample for fuzzy C-means clustering, with an initial cluster number of 3 corresponding to the scrap, repair, and pass categories. The membership matrix is ​​randomly initialized, and a fuzzy weighting index is set, effectively establishing a nonlinear relationship between defect damage and quality level. The mapping relationship is used to iteratively calculate the cluster centers and sample membership degrees of each category. When the change in the membership degree matrix between two adjacent iterations is less than the threshold, the system converges. The membership degree values ​​of each defective sample to the three categories are output as the classification confidence level to ensure that the classification results are stable and have repeatable confidence expression. Combined with the preset scrapping threshold standard and repair feasibility rules, the category with the highest membership degree is used as the initial classification level. If the maximum membership degree is lower than the lower bound of the confidence level, it is marked as pending re-inspection. Finally, dynamic classification labels are output to realize flexible adaptation of classification rules and safe fallback for abnormal samples. It should be noted that the feature weight vector output by the deep Q-network is used as the input sample for fuzzy C-means clustering. Each sample corresponds to the two-dimensional damage weight of an independent defect on the insulation board. The initial number of clusters is fixed at 3, representing three quality levels: scrap, repair, and pass. The fuzzy weighting index is set to 2.0 based on extensive experiments to balance the fuzziness and convergence stability of the clustering results. The membership matrix is ​​initialized with random numbers in the interval [0, 1], and the constraint that the sum of the membership degrees of each sample to all categories is 1 is enforced. The initial value of the cluster center is obtained by randomly selecting the feature weight vectors of three samples. During the iteration process, the coordinates of the cluster center and the membership values ​​of each sample are updated sequentially. When the membership values ​​of two adjacent iterations are... Convergence is determined when the maximum change in the membership matrix is ​​less than 0.001. The number of iterations is between 15 and 30. After convergence, each defect sample obtains three membership values, representing the confidence level of its classification into the scrap, repair, and pass categories, respectively. After completing the membership calculation, an initial grade determination is performed based on the preset scrap threshold standard and repair feasibility rules. The scrap threshold standard is set as follows: when the membership degree of a defect sample to the scrap category is greater than or equal to 0.70, it is directly classified as scrap, without considering the membership degree of other categories. If the membership degree of the scrap category is lower than 0.70 but higher than 0.40, it enters the repair feasibility rule evaluation stage. This rule comprehensively considers whether the defect area is less than 50 square millimeters and whether the elongation is... The following three conditions must be met: a score greater than 0.85, and whether the sample is located within 10 mm of the edge of the insulation board. Meeting any two of these conditions qualifies the sample as a repair grade; otherwise, the pass grade is evaluated. The pass grade requires a membership degree of ≥0.65 to the pass category, a membership degree of <0.30 to the scrap category, and a membership degree of <0.40 to the repair category. If the maximum membership degree for all three categories is below the lower confidence threshold of 0.60, the sample is marked as pending re-inspection, and no automatic grading label is output. When finally outputting the dynamic grading label, a structured label containing a grade code and a confidence value is generated for each defect sample. The grade code is represented by an integer value, where 0 represents pass, 1 represents repair, and 2 represents scrap. 3 represents samples to be scrapped and 3 represents samples to be re-inspected. The confidence level value is directly taken as the membership value of the sample to the initially determined grade category, and three significant figures are retained. For samples marked as to be re-inspected, their original images and feature weight vectors are stored in the re-inspection queue, waiting for the intervention of the manual re-inspection platform for confirmation. Each insulation board may contain multiple defect samples. Each sample outputs a grade label independently. The highest grade among all defect samples is taken as the final sorting grade of the whole board. That is, if any defect is judged to be scrapped, the whole board is scrapped. If there is repair and no scrapping, the whole board needs to be repaired. The board can only pass when all defects are at the pass grade. The grade label is pushed to the intelligent sorting main control center in real time as the decision basis for triggering the subsequent actions of the execution mechanism. The dual-delay depth strategy module is used to train a strategy network based on a dual-delay depth deterministic strategy, using speed sensor and light sensor signals as states and thrust magnitude and rejection delay as actions. It outputs the optimal execution adjustment amount under non-stationary disturbances in real time to compensate for dynamic disturbances in the production line and ensure the consistency of sorting actions. The disturbance compensation execution module is used to dynamically correct the thrust curve of the pneumatic push rod and the rejection trigger time window according to the optimal execution adjustment amount output by the strategy network, so as to ensure that the same defect receives consistent sorting action under different production line conditions and eliminate the impact of speed and light fluctuations on rejection accuracy. The closed-loop incremental self-evolution module is used to collect misclassified samples from manual re-inspection or production line feedback. It adopts an elastic weight consolidation and memory replay mechanism to update the parameters of the deep Q network and fuzzy clustering without forgetting old knowledge. It also uses some parameters of the strategy network as a transferable base and makes rapid fine-tuning based on recent production line disturbance data. This enables the execution strategy to adaptively migrate and continuously optimize across different seasons and shifts, achieve cross-working-condition knowledge inheritance, and continuously improve sorting robustness.

[0022] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the dual-delay deep strategy module performs the following steps: the instantaneous speed deviation of the production line collected by the speed sensor and the illuminance deviation collected by the light sensor are spliced ​​together as the state vector; the thrust reference value of the pneumatic push rod and the rejection trigger delay amount are used as two-dimensional continuous actions to construct a Markov decision process, so that the system can perceive and quantify the impact of dynamic disturbances on rejection accuracy; the dual-delay deep deterministic strategy gradient algorithm is adopted to construct the current strategy network, the target strategy network and the dual-Q network respectively; the target strategy smoothing noise is added to the action output; the Q value estimation is stabilized by soft update method, the Q value overestimation is suppressed, and the stability and convergence reliability of strategy learning are improved; after the strategy network training converges, the current state vector is received in real time and the forward propagation output of the optimal execution adjustment amount, including the thrust correction coefficient and the delay compensation value, is sent to the intelligent sorting main control center to realize the millisecond-level adaptive adjustment of the execution parameters under dynamic working conditions; It should be noted that in actual production line deployment, the speed sensor and illumination sensor are first installed coaxially with the line scan camera and at the front end of the pneumatic actuator, respectively. Real-time instantaneous speed deviation and illumination deviation values ​​are acquired with a sampling period of 20 milliseconds. These two values ​​are then concatenated into a two-dimensional state vector, which serves as the input to the Markov decision process. The speed deviation range is -0.5 m / s to +0.5 m / s, and the illumination deviation range is -200 lux to +200 lux. When these ranges are exceeded, sensor self-calibration is triggered. The thrust reference value of the pneumatic actuator is set to a normalized unit from 0 to 100, corresponding to the PWM duty cycle of the proportional solenoid valve. The rejection trigger delay is set from -50 milliseconds to +50 milliseconds, ranging from 0% to 100%. Negative values ​​indicate early triggering, and positive values ​​indicate delayed triggering. This state-action space definition ensures that the sorting system can sense dynamic disturbances on the production line and output continuous adjustment amounts. The network structure design of the dual-delay deep deterministic policy gradient algorithm is as follows: Both the current policy network and the target policy network adopt a three-layer fully connected structure. The number of nodes in the input layer is 2 (velocity deviation, illumination deviation), and the number of nodes in the two hidden layers are 128 and 64, respectively. The activation function is ReLU, and the number of nodes in the output layer is 2 (thrust correction coefficient, delay compensation value). The dual-Q network structure is used for the input layer. A 2-node, hidden layer (128 and 64 nodes), and output layer (1 node) are used to evaluate the Q-value of state-action pairs. Gaussian noise with a mean of 0 and a standard deviation of 0.2 is added to the action output and clipped to the range of -0.5 to +0.5 as a smoothing regularization term for the target policy. The learning rate of the policy network is set to 0.0001, the learning rate of the Q network is set to 0.0003, the soft update coefficient τ is 0.005, the experience replay pool capacity is 50,000 sets, and 64 sets are sampled in each batch for updating. The policy network is updated every 2 steps, and the Q network is updated every step. After the policy network completes offline training and convergence, the network weights are frozen and loaded onto the production line's industrial control computer for real-time operation. During this phase, the industrial control computer reads the state vector at a 20-millisecond cycle, calculates the thrust correction coefficient and delay compensation value through forward propagation of the strategy network, and maps the thrust correction coefficient to the PWM duty cycle of the proportional solenoid valve after multiplying it by the preset thrust reference value. The delay compensation value is superimposed on the reference trigger time to generate a dynamic time window. The output adjustment amount is sent to the programmable logic controller through the EtherCAT bus at a 5-millisecond cycle to drive the pneumatic push rod to perform the rejection action. When the thrust correction coefficient exceeds the range of 0.85 to 1.15 for 10 consecutive sampling cycles, the incremental update request of the strategy network is automatically triggered, and the closed-loop incremental self-evolution module performs rapid fine-tuning. The disturbance compensation execution module performs the following steps: It extracts the thrust correction coefficient and delay compensation value from the intelligent sorting main control center, multiplies the thrust reference value by the thrust correction coefficient, and maps it to a nonlinear curve of the PWM duty cycle of the pneumatic push rod. This enables smooth adjustment of the thrust soft start and soft stop, effectively suppresses mechanical impact of the push rod, extends the service life of the solenoid valve, and dynamically adjusts the opening time and window width of the rejection trigger time window according to the delay compensation value. This ensures that the center of the time window is aligned with the predicted time when the defect target arrives at the push rod's action area, compensates for the positional offset caused by speed fluctuations, eliminates rejection position deviations caused by speed fluctuations, and improves the hit accuracy. The corrected thrust curve parameters and time window parameters are sent to the programmable logic controller to drive the pneumatic push rod to perform rejection actions. This ensures that the same defect receives consistent sorting action execution under different production line conditions, guaranteeing the consistency and repeatability of the sorting action and stabilizing the rejection effect. It should be noted that the intelligent sorting main control center first multiplies the thrust correction coefficient output by the strategy network with the preset thrust reference value to obtain the dynamic thrust target value. This target value is converted into the PWM duty cycle of the pneumatic push rod proportional solenoid valve through a pre-calibrated nonlinear mapping function. The mapping curve adopts a quadratic function form to compensate for the saturation and dead zone characteristics of the solenoid valve. The lower limit of the duty cycle is set to 15% to correspond to the minimum sustainable thrust, and the upper limit is set to 92% to avoid overheating of the solenoid valve. To achieve soft start and soft stop of thrust, a trapezoidal transition segment is automatically inserted when the PWM duty cycle change rate exceeds 5% per millisecond. The transition time of the rising edge and falling edge is set to 8 milliseconds and 12 milliseconds, respectively, to effectively suppress the mechanical impact of the push rod action and avoid secondary damage to the plate caused by sudden thrust changes. The center time of the time window is calculated in real time based on the defect detection time, the fixed distance from the defect center to the push rod action area, and the current instantaneous speed of the production line. The delay compensation value is added to the center time as a deviation correction term. The initial time window width is set to 40 milliseconds. When the absolute value of the speed deviation exceeds 0.2 meters per second, the window width is adaptively expanded to 60 milliseconds to cover the position uncertainty caused by speed fluctuations. The window opening and closing times correspond to the center time minus half the width and plus half the width, respectively, to ensure that the push rod action always covers the margin range before and after the actual arrival time of the defect and to compensate for the position offset caused by speed fluctuations. The programmable logic controller generates a PWM waveform with a corresponding duty cycle based on the thrust curve parameters, drives the proportional solenoid valve to control the pneumatic push rod to perform the rejection action, and enables the push rod action window according to the time window parameters. Rejection is only allowed when the current time falls within the time window range. The thrust curve parameters include the current cycle thrust target value and the transition segment flag. The time window parameters include the absolute timestamps of the opening and closing times. The two work together to ensure that the same defect receives consistent sorting action execution under different production line speeds and lighting conditions, eliminating the impact of dynamic disturbances on rejection accuracy. The closed-loop incremental self-evolution module performs the following steps: It collects misclassified samples labeled by the manual re-inspection platform and samples with sorting result deviations fed back from the end of the production line, and stores them in a timestamped misclassified sample buffer. Each sample contains the original image, feature weight vector, and correct classification label. It continuously learns new misclassification patterns while maintaining the stability of the original classification ability. It uses the elastic weight consolidation algorithm to calculate the important weights of the deep Q network and fuzzy clustering parameters in the current task. It combines the memory replay mechanism to uniformly sample misclassified samples from the buffer for offline incremental training, suppressing catastrophic forgetting, correcting misjudgment boundaries, and maintaining the stability of the original clustering topology. It freezes the parameters of the first two layers of the dual-delay deep strategy network as a transferable base and uses the production line disturbance data in the last 24 hours to quickly fine-tune the output layer parameters, realize the adaptive migration of the execution strategy between different seasons and shifts, quickly adapt to environmental and operational changes, and ensure the consistency of elimination. It should be noted that the automatic grading results are reviewed by a manual re-inspection platform. Samples deemed misclassified, along with sorting deviation samples reported from the end of the production line, are stored in a misclassified sample buffer. Each stored sample contains the original 8-bit grayscale image of the defect, a 20-dimensional feature weight vector output by a deep Q-network, and a manually labeled correct grading label, along with a millisecond-level timestamp for tracing production line conditions. The buffer capacity is set to 2000 sets, using a first-in, first-out queue management system. When the number of stored sets reaches 500, an incremental training process is automatically triggered. To suppress catastrophic forgetting, the Fisher information matrix of each connection weight in the deep Q-network under the current task is calculated as a quantification of parameter importance. By combining a memory replay mechanism, 32 sets of misclassified samples are uniformly sampled from the buffer for offline training. The elastic weight consolidation coefficient λ is set to 0.75. When updating network parameters, a secondary constraint is applied to important weights to keep the performance degradation of old tasks within 5% during the learning of new knowledge. For the parameter update of fuzzy clustering hierarchical units, the elastic weight consolidation algorithm is also used to handle the incremental optimization of cluster centers and membership matrices. Specifically, the feature weight vector of the misclassified sample is used as the new input to calculate the importance weight of the current three cluster centers (corresponding to scrapped, repaired, and passed) in the feature space, which is the reciprocal of the sample density around each cluster center. During incremental training, the memory replay mechanism samples from the buffer... Sixty-four misclassified samples were sampled and mixed with the current batch data at a 1:3 ratio before undergoing fuzzy C-means iteration. The fuzzy weighting exponent remained constant at 2.0, and the convergence threshold was set to 0.001. When updating cluster centers, the old centers were constrained by the elastic weight coefficient λ=0.75, limiting the pulling effect of new samples on cluster centers. This approach corrects misclassification boundaries while maintaining the original cluster topology and preventing cluster center drift caused by a small number of abnormal samples. The adaptive transfer of the dual-delay deep policy network employs a strategy combining layer freezing and rapid fine-tuning. The weight parameters of the first two fully connected layers of the offline-trained policy network are frozen, preserving its ability to extract higher-order features of speed and illumination biases. As a transferable base, the weights of the output layer (third layer) are rapidly fine-tuned using production line disturbance data collected in the last 24 hours. The fine-tuning sample set includes state-action-reward triplets under dynamic conditions such as step changes in speed and sinusoidal fluctuations in illumination, with a sample size of no less than 200 sets. During the fine-tuning process, the learning rate of the policy network is increased to 0.001 (three times that of the initial training), the training rounds are limited to 50 rounds, 64 sets are sampled in each batch, and only the 8 connection weights of the output layer (128-dimensional input and 2-dimensional output) are updated. The time taken for a single fine-tuning is controlled within 2 seconds, which can be completed during production line shift changes, realizing the adaptive transfer of the execution policy to changes in ambient light in different seasons and differences in shift operating habits.

[0023] The following describes the workflow of a machine vision-based calcium silicate board insulation board appearance defect sorting system.

[0024] After the system starts up, it first acquires images of the insulation board surface using a linear scan camera. The anti-interference contour extraction module uses an improved fruit fly algorithm to dynamically optimize the high and low thresholds of Canny edge detection, and integrates region growth and morphological skeleton extraction methods to refine the defect edges while suppressing the interference of porous textures, and outputs closed sub-pixel level defect contours. Subsequently, the damage weight grading module extracts the geometric feature set of the defect contours, and calculates the insulation performance decay weight and structural strength degradation weight through forward propagation of a deep Q network to form a two-dimensional feature weight vector. This vector is then input into fuzzy C-means clustering, and combined with the scrap threshold standard and repair feasibility rules, each defect is dynamically divided into scrap, repair, or pass levels, and the highest level of all defects on the board is taken as the sorting result of the whole board. Based on the grading results and real-time collected production line speed and illumination deviations, the intelligent sorting control center outputs thrust correction coefficients and delay compensation values ​​through a dual-delay deep strategy network. The disturbance compensation execution module dynamically adjusts the thrust curve of the pneumatic push rod and the rejection trigger time window accordingly, driving the programmable logic controller to execute precise rejection actions. The closed-loop incremental self-evolution module continuously collects misclassified samples from manual re-inspection and deviation data from production line feedback. It uses an elastic weight consolidation and memory replay mechanism to periodically update the deep Q-network and fuzzy clustering parameters, and rapidly fine-tunes the output layer of the strategy network, ensuring that the system maintains sorting stability and consistency under different production line conditions.

[0025] 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. A machine vision-based sorting system for appearance defects in calcium silicate board insulation panels, comprising an intelligent sorting control center, characterized in that, The intelligent sorting main control center has the following communication connection modules: The anti-disturbance contour extraction module is used to optimize the high and low threshold parameters of Canny edge detection using an improved fruit fly algorithm, and integrates region growth and morphological skeleton extraction methods to locate and refine the contours of defect candidate areas on calcium silicate board insulation boards. The damage weighting and grading module, based on the geometric features of the refined contour, uses deep reinforcement learning to evaluate the damage weight of defects on thermal insulation and structural performance, and achieves dynamic grading of three categories: scrap, repair, and pass through fuzzy clustering. The dual-delay deep strategy module is used to train a policy network based on a dual-delay deep deterministic strategy, using the signals from the velocity sensor and the illumination sensor as the state and the thrust magnitude and culling delay as the action, and outputs the optimal execution adjustment amount under non-stationary perturbations in real time. The disturbance compensation execution module is used to dynamically correct the thrust curve of the pneumatic push rod and the elimination trigger time window based on the optimal execution adjustment amount output by the strategy network. The closed-loop incremental self-evolution module is used to collect misclassified samples from manual re-inspection or production line feedback. It adopts an elastic weight consolidation and memory replay mechanism to update the parameters of the deep Q network and fuzzy clustering, and uses some parameters of the strategy network as a transferable base to make rapid fine-tuning in combination with recent production line disturbance data.

2. The machine vision-based calcium silicate board insulation board appearance defect sorting system according to claim 1, characterized in that: The anti-interference contour extraction module includes a fruit fly optimized threshold unit and a region skeleton refinement unit. The fruit fly optimization threshold unit is used to dynamically search for the high and low threshold parameters of the Canny operator using the gradient amplitude distribution and local texture difference of the calcium silicate board insulation board surface image as the fitness function, and outputs a binary edge map by using the improved fruit fly algorithm. The region skeleton refinement unit is used to fuse region growth to lock defect candidate areas, and refine the contour along the main defect direction by combining morphological skeleton extraction method, eliminating false edge responses caused by the porous structure of the insulation board, and outputting a closed and continuous real defect sub-pixel contour.

3. The machine vision-based calcium silicate board insulation board appearance defect sorting system according to claim 2, characterized in that: The fruit fly optimized threshold unit performs the following steps: Gradient amplitude distribution of images on the surface of calcium silicate board insulation board is collected. Fitness function is constructed using the entropy value of the gradient histogram within a local window and the peak signal-to-noise ratio. The position of the fruit fly population is initialized as a candidate high and low threshold parameter pair. An improved fruit fly algorithm is adopted, which introduces chaotic mapping and adaptive step size decay mechanism to guide individual fruit flies to sniff and search in the gradient magnitude space, and iteratively updates the optimal position of the population with the goal of maximizing the fitness function value. The optimal position after iterative convergence is decoded into the high and low thresholds of the Canny operator, which are used to suppress the pseudo-edge response caused by porous texture and retain the real edges with weak gradients, respectively, and output a binary edge map with adaptive thresholds.

4. The machine vision-based calcium silicate board insulation board appearance defect sorting system according to claim 2, characterized in that: The region skeleton refinement unit performs the following steps: Using the edge points of connected regions in the binary edge map as the initial seed points for region growth, and based on the consistency criterion of local gray mean and gradient direction, the boundary of the defect candidate region is locked by expanding layer by layer in the eight neighborhoods. Morphological skeleton extraction is performed on the locked defect candidate region. Boundary pixels are iteratively stripped along the main defect direction until a skeleton line with a single pixel width is obtained. At the same time, short branches and isolated skeleton segments with a length less than a preset threshold are removed. The skeleton lines are back-projected onto the original gradient image, and a subpixel edge interpolation algorithm is used to fit the gradient maxima on both sides of the skeleton to output a closed and continuous real defect subpixel contour.

5. A machine vision-based sorting system for appearance defects in calcium silicate board insulation boards according to claim 2, characterized in that: The damage weighting module includes a deep Q-network weighting unit and a fuzzy clustering grading unit; The deep Q-network weighting unit is used to train the deep Q-network to learn the nonlinear damage weights of different defect geometric features on the thermal insulation performance decay and structural strength degradation based on the porous matrix and layered structure of the calcium silicate insulation board, and output the feature weight vector oriented towards the actual service performance of the board. The fuzzy clustering hierarchical unit is used to input the feature weight vector output by the deep Q network into fuzzy C-means clustering, and dynamically classify defects into three defect levels: scrap, repair, and pass, in combination with the scrap threshold standard and repair feasibility rules of calcium silicate board insulation board.

6. The machine vision-based calcium silicate board insulation board appearance defect sorting system according to claim 5, characterized in that: The deep Q-network weighting unit performs the following steps: Extract the geometric feature set of the region surrounded by the sub-pixel contour of the real defect, including area, elongation, roundness, convex hull ratio, and the angle between the principal axis direction and the board layering, and construct a 20-dimensional feature input vector; A dual-hidden-layer deep Q-network is constructed, with the feature input vector as the state and the combination of the thermal insulation performance attenuation coefficient and the structural strength degradation coefficient as the action space. The Q-value approximator is trained using empirical replay and target network soft update mechanism. After the network converges, the feature input vector of the defect to be classified is forward propagated, and the action combination that maximizes the cumulative damage Q value is output. That is, the thermal insulation performance attenuation weight and structural strength degradation weight of the defect, forming the feature weight vector.

7. A machine vision-based sorting system for appearance defects in calcium silicate board insulation boards according to claim 5, characterized in that: The fuzzy clustering hierarchical unit performs the following steps: The feature weight vector is used as the input sample for fuzzy C-means clustering. The number of clusters is initialized to 3, corresponding to the three classes of scrap, repair, and pass respectively. The membership matrix is ​​randomly initialized and the fuzzy weighting index is set. Iteratively calculate the cluster centers and sample membership degrees of each category. When the change in the membership degree matrix between two adjacent iterations is less than the threshold, convergence is achieved. Output the membership degree values ​​of each defect sample to the three categories as the hierarchical confidence level. Combining the preset scrapping threshold standard and repair feasibility rules, the category with the highest membership degree is used as the initial grade. If the maximum membership degree is lower than the confidence level lower bound, it is marked as pending re-inspection, and finally, a dynamic grade label is output.

8. A machine vision-based sorting system for appearance defects in calcium silicate board insulation boards according to claim 5, characterized in that: The dual-delay depth strategy module performs the following steps: The instantaneous speed deviation of the production line collected by the speed sensor and the illuminance deviation collected by the light sensor are spliced ​​together to form a state vector. The thrust reference value of the pneumatic push rod and the rejection trigger delay are used as two-dimensional continuous actions to construct a Markov decision process. A dual-delay deep deterministic policy gradient algorithm is adopted to construct the current policy network, the target policy network, and the dual-Q network respectively. Target policy smoothing noise is added to the action output, and Q-value estimation is stabilized through soft update. After the policy network training converges, it receives the current state vector forward propagation output of the optimal execution adjustment amount, including the thrust correction coefficient and the delay compensation value, and sends it to the intelligent sorting main control center in real time.

9. A machine vision-based sorting system for appearance defects in calcium silicate board insulation boards according to claim 8, characterized in that: The disturbance compensation execution module performs the following steps: The thrust correction coefficient and delay compensation value are extracted from the intelligent sorting main control center. The thrust reference value is multiplied by the thrust correction coefficient and mapped to the nonlinear curve of the PWM duty cycle of the pneumatic push rod, so as to realize the smooth adjustment of thrust soft start and soft stop. The opening time and width of the elimination trigger time window are dynamically adjusted based on the delay compensation value, so that the center of the time window is aligned with the predicted time when the defect target arrives at the push rod action area, compensating for the positional offset caused by speed fluctuations. The corrected thrust curve parameters and time window parameters are sent to the programmable logic controller to drive the pneumatic push rod to perform the rejection action, so that the same defect can be sorted in a consistent manner under different production line conditions.

10. A machine vision-based sorting system for appearance defects in calcium silicate board insulation panels according to claim 9, characterized in that: The closed-loop incremental self-evolution module performs the following steps: Collect misclassified samples labeled by the manual re-inspection platform and samples with deviations in sorting results reported from the end of the production line, and store them in a timestamped misclassified sample buffer. Each sample contains the original image, feature weight vector, and correct grading label. The elastic weight consolidation algorithm is used to calculate the important weights of the deep Q network and fuzzy clustering parameters in the current task. Combined with the memory replay mechanism, misclassified samples are uniformly sampled from the buffer for offline incremental training. By freezing the parameters of the first two layers of the dual-delay deep strategy network as a transferable base, and using production line disturbance data from the past 24 hours to quickly fine-tune the output layer parameters, the execution strategy can be adaptively migrated between different seasons and shifts.

Images