Abnormality monitoring method for intelligent flow production line of prefabricated slabs

Abstract

The invention discloses an abnormity monitoring method for an intelligent flow production line of prefabricated slabs, and belongs to the field of image processing. The method comprises the following steps: acquiring a demolded precast slab image, splitting the demolded precast slab image into R, G and B channel images, respectively positioning each channel precast slab area, and extracting contour pixel points to form a contour image; finding out the cross point of the contour image, segmenting to obtain an edge contour, calculating a pixel point edge missing risk coefficient according to the distance between the edge contour and a reference straight line, and constructing an edge missing risk image of each channel; and finally, processing the risk image by using an anomaly monitoring model to obtain an edge missing risk value. The problems that traditional prefabricated slab edge missing detection is low in efficiency, and accuracy and consistency are difficult to guarantee are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image monitoring, and particularly relates to an abnormal monitoring method for a precast slab intelligent assembly line. Background Art

[0002] As a commonly used building component, precast slabs are widely used in various construction projects. However, in the production process, due to the lack of strict quality control during the pouring process, the precast slabs after demoulding will have the phenomenon of edge chipping (i.e., there are depressions on the edges of the precast slabs). This defect not only damages the appearance of the precast slabs, but more importantly, it will significantly weaken the structural strength and load-bearing capacity of the precast slabs, resulting in abnormal stress distribution, causing potential safety hazards such as cracks, deformations, and even fractures, greatly threatening the stability and safety of the building structure. In addition, the edge chipping will also interfere with the installation accuracy of the precast slabs, causing splicing gaps, and then leading to corrosion due to problems such as water accumulation and dust accumulation, significantly shortening the service life of the building and increasing the maintenance cost. The existing detection of edge chipping of precast slabs after demoulding uses manual visual inspection, which is not only inefficient and difficult to meet the needs of large-scale production, but also greatly affected by the subjective factors of the inspectors, and it is difficult to guarantee the accuracy and consistency. Summary of the Invention

[0003] Aiming at the above deficiencies in the prior art, the abnormal monitoring method for a precast slab intelligent assembly line provided by the present invention solves the problems of low efficiency, difficult accuracy and consistency guarantee in the existing detection of edge chipping of precast slabs.

[0004] In order to achieve the above invention purpose, the technical solution adopted by the present invention is: an abnormal monitoring method for a precast slab intelligent assembly line, including the following steps:

[0005] Collect images of precast slabs after demoulding on the production line, and split the images into R, G, and B channel images;

[0006] Find the precast slab areas from the R, G, and B channel images respectively, to obtain the R-channel precast slab area, the G-channel precast slab area, and the B-channel precast slab area;

[0007] Extract the pixel points of the precast slab contours from the R-channel precast slab area, the G-channel precast slab area, and the B-channel precast slab area respectively, to obtain the R-channel precast slab contour image, the G-channel precast slab contour image, and the B-channel precast slab contour image;

[0008] Find the intersection points on each channel precast slab contour image, and cut according to the intersection points to obtain the edge contours;

[0009] According to the distance between each edge contour and the reference straight line, calculate the edge chipping risk coefficient of the pixel points on the edge contour, and construct the R-channel precast slab edge chipping risk image, the G-channel precast slab edge chipping risk image, and the B-channel precast slab edge chipping risk image;

[0010] The abnormal monitoring model is used to process the risk images of chamfer defects on precast slabs in the R channel, G channel, and B channel to obtain the risk values of chamfer defects.

[0011] Furthermore, the process of finding the precast slab area from each channel image includes:

[0012] In each channel image, calculate the distance between the channel value and the stored channel value of the precast slab to obtain the channel distance;

[0013] When the channel distance is less than the channel distance threshold, mark the pixel points corresponding to the channel value as valid pixel points;

[0014] Take the largest connected region formed by each valid pixel point as the precast slab area.

[0015] Furthermore, the process of obtaining the contour images of the precast slabs in the R channel, G channel, and B channel all includes: traversing each pixel point in the precast slab area. When there are pixel points in the non-precast slab area within the neighborhood range of the pixel point, this pixel point is a contour pixel point. Extract all contour pixel points to obtain the precast slab contour image. Among them, when the precast slab area is the precast slab area in the R channel, the precast slab contour image is the contour image of the precast slab in the R channel. When the precast slab area is the precast slab area in the G channel, the precast slab contour image is the contour image of the precast slab in the G channel. When the precast slab area is the precast slab area in the B channel, the precast slab contour image is the contour image of the precast slab in the B channel.

[0016] Furthermore, the process of obtaining the edge contour includes:

[0017] Take each contour pixel point on the channel precast slab contour image as the central pixel point, and take N contour pixel points in the neighborhood range to obtain a contour line segment, where N is a positive integer;

[0018] Calculate the curve length of the contour line segment and calculate the straight-line length from the starting pixel point to the ending pixel point in the contour line segment;

[0019] For the same contour line segment, take the ratio of the curve length to the straight-line length as the bending degree of the central pixel point;

[0020] Mark the central pixel points with a line segment bending degree greater than the bending degree threshold as suspected intersection pixel points;

[0021] Extract each continuous suspected intersection pixel point to obtain multiple suspected intersection contours;

[0022] Obtain the bending degree for the suspected intersection contours, and sort each suspected intersection contour from largest to smallest bending degree;

[0023] Take the pixel point corresponding to the maximum curvature among the first 4 cross-suspected contours as the intersection point;

[0024] Using each intersection point as a segmentation point, segment the contour image of the channel precast slab to obtain multiple edge contours.

[0025] Furthermore, the process of constructing the R-channel precast slab edge defect risk image, G-channel precast slab edge defect risk image, and B-channel precast slab edge defect risk image includes:

[0026] Connect the starting pixel point and the ending pixel point of each edge contour to obtain a reference straight line;

[0027] Take the distance from each pixel point on the edge contour to the reference straight line as the vertical offset;

[0028] According to the vertical offset, calculate the edge defect risk coefficient of the pixel point;

[0029] Take the edge defect risk coefficient as the new pixel value of this pixel point, and set the pixel values at other positions to 0 to obtain the precast slab edge defect risk image, where the precast slab edge defect risk image corresponding to the R-channel precast slab contour image is the R-channel precast slab edge defect risk image, the precast slab edge defect risk image corresponding to the G-channel precast slab contour image is the G-channel precast slab edge defect risk image, and the precast slab edge defect risk image corresponding to the B-channel precast slab contour image is the B-channel precast slab edge defect risk image.

[0030] Furthermore, the process of calculating the edge defect risk coefficient of the pixel point includes: taking the average of the vertical offsets of each pixel point belonging to the same edge contour to obtain the average vertical offset;

[0031] Subtract the average vertical offset from the vertical offset of each pixel point, and normalize the subtraction result to obtain the edge defect risk coefficient.

[0032] Furthermore, the anomaly monitoring model includes: a shallow feature extraction module, a first attention channel module, a second attention channel module, a first multi-scale feature extraction module, a second multi-scale feature extraction module, an adder A1, a CNN network, and a fully connected layer;

[0033] The input end of the shallow feature extraction module is used to input the R-channel precast slab edge defect risk image, G-channel precast slab edge defect risk image, and B-channel precast slab edge defect risk image, and its output end is respectively connected to the input end of the first attention channel module and the input end of the second attention channel module;

[0034] The output end of the first attention channel module is connected to the input end of the first multi-scale feature extraction module; the output end of the second attention channel module is connected to the input end of the second multi-scale feature extraction module; the input ends of adder A1 are respectively connected to the output ends of the first multi-scale feature extraction module and the second multi-scale feature extraction module, and its output end is connected to the input end of the CNN network; the output end of the CNN network is connected to the input end of the fully connected layer; the output end of the fully connected layer serves as the output end of the anomaly monitoring model.

[0035] Further, the shallow feature extraction module includes: a first pointwise convolutional layer, a second pointwise convolutional layer, a third pointwise convolutional layer, and a feature fusion layer;

[0036] The first pointwise convolutional layer is used to process the R-channel precast slab edge defect risk image; the second pointwise convolutional layer is used to process the G-channel precast slab edge defect risk image; the third pointwise convolutional layer is used to process the B-channel precast slab edge defect risk image; the feature fusion layer is used to fuse the outputs of the first pointwise convolutional layer, the second pointwise convolutional layer, and the third pointwise convolutional layer.

[0037] Further, the first attention channel module includes: an average pooling layer, a first Sigmoid layer, and a multiplier M1;

[0038] The first input end of multiplier M1 is connected to the input end of the average pooling layer and serves as the input end of the first attention channel module; the output end of the average pooling layer is connected to the input end of the first Sigmoid layer; the second input end of multiplier M1 is connected to the output end of the first Sigmoid layer, and its output end serves as the output end of the first attention channel module;

[0039] The second attention channel module includes: a max pooling layer, a second Sigmoid layer, and a multiplier M2;

[0040] The first input end of multiplier M2 is connected to the input end of the max pooling layer and serves as the input end of the second attention channel module; the output end of the max pooling layer is connected to the input end of the second Sigmoid layer; the second input end of multiplier M2 is connected to the output end of the second Sigmoid layer, and its output end serves as the output end of the second attention channel module.

[0041] Further, both the first multi-scale feature extraction module and the second multi-scale feature extraction module include: a first convolutional block, a second convolutional block, a third convolutional block, and a Concat layer;

[0042] The input end of the first convolutional block serves as the input end of the first multi-scale feature extraction module or the second multi-scale feature extraction module, and its output end is respectively connected to the input end of the second convolutional block and the input end of the Concat layer; the output end of the second convolutional block is respectively connected to the input end of the third convolutional block and the input end of the Concat layer; the output end of the third convolutional block is connected to the input end of the Concat layer; the output end of the Concat layer serves as the output end of the first multi-scale feature extraction module or the second multi-scale feature extraction module.

[0043] The beneficial effects of the present invention are as follows:

[0044] The present invention collects images of precast slabs after demoulding on a production line, splits the images into R, G, and B channel images, processes the R, G, and B channels separately, obtains the precast slab regions respectively, further obtains the pixel points of the precast slab contour, constructs the contour images of the R, G, and B channels respectively, then finds the intersection points from the precast slab contour images of each channel, extracts the edge contours, so as to facilitate comparing the edge contours with the reference straight line, obtaining the lack-of-edge risk coefficients of the pixel points on the edge contours, constructing the lack-of-edge risk images of the precast slabs in the R, G, and B channels respectively, highlighting the lack-of-edge conditions of the pixel points on each edge contour, and then using an anomaly monitoring model to process the lack-of-edge risk images of the precast slabs in the R, G, and B channels, comprehensively considering the lack-of-edge conditions of the pixel points in the three channels, and improving the accuracy of anomaly monitoring. Compared with the manual detection method, the present invention has a higher detection efficiency for lack of edges in precast slabs, avoids errors caused by subjective judgment differences in manual detection, makes the detection results more objective and accurate, ensures the consistency of detection for different batches and at different times, and improves the reliability of quality detection of precast slabs. Description of the Drawings

[0045] Figure 1 It is a flowchart of an anomaly monitoring method for an intelligent precast slab production line;

[0046] Figure 2 It is a structural schematic diagram of an anomaly monitoring model;

[0047] Figure 3 It is a structural schematic diagram of a shallow feature extraction module;

[0048] Figure 4 It is a structural schematic diagram of a first attention channel module;

[0049] Figure 5 It is a structural schematic diagram of a second attention channel module;

[0050] Figure 6 It is a structural schematic diagram of a first multi-scale feature extraction module and a second multi-scale feature extraction module. Detailed Embodiments

[0051] The specific embodiments of the present invention will be described below to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0052] As Figure 1 shown, an abnormal monitoring method for a precast slab intelligent production line includes the following steps:

[0053] Collect images of precast slabs after demoulding on the production line, and split the images into R, G, and B channel images;

[0054] Find the precast slab areas from the R, G, and B channel images respectively to obtain the R-channel precast slab area, G-channel precast slab area, and B-channel precast slab area;

[0055] Extract the pixel points of the precast slab contour from the R-channel precast slab area, G-channel precast slab area, and B-channel precast slab area respectively to obtain the R-channel precast slab contour image, G-channel precast slab contour image, and B-channel precast slab contour image;

[0056] Find the intersection points on each channel precast slab contour image, and cut according to the intersection points to obtain the edge contour;

[0057] According to the distance between each edge contour and the reference straight line, calculate the missing edge risk coefficient of the pixel points on the edge contour, and construct the R-channel precast slab missing edge risk image, G-channel precast slab missing edge risk image, and B-channel precast slab missing edge risk image;

[0058] Use the abnormal monitoring model to process the R-channel precast slab missing edge risk image, G-channel precast slab missing edge risk image, and B-channel precast slab missing edge risk image to obtain the missing edge risk value.

[0059] In this embodiment, the process of finding the precast slab area from each channel image includes:

[0060] In each channel image, calculate the distance between the channel value and the stored precast slab channel value to obtain the channel distance;

[0061] When the channel distance is less than the channel distance threshold, mark the pixel points corresponding to the channel value as valid pixel points;

[0062] Take the largest connected area formed by each valid pixel point as the precast slab area.

[0063] In this embodiment, calculate the distance between the R-channel value and the R-channel value of the stored precast slab, the distance between the G-channel value and the G-channel value of the stored precast slab, and the distance between the B-channel value and the B-channel value of the stored precast slab. The distance is the absolute value of the difference between the two. The R-channel value of the stored precast slab, the G-channel value of the stored precast slab, and the B-channel value of the stored precast slab are the R, G, and B channel values corresponding to the precast slab image stored in advance.

[0064] In this embodiment, the processes of obtaining the R-channel precast slab contour image, the G-channel precast slab contour image, and the B-channel precast slab contour image all include: traversing each pixel point in the precast slab area. When there are pixel points in the non-precast slab area within the neighborhood range of the pixel point, this pixel point is a contour pixel point. Extract all the contour pixel points to obtain the precast slab contour image. Among them, when the precast slab area is the R-channel precast slab area, the precast slab contour image is the R-channel precast slab contour image; when the precast slab area is the G-channel precast slab area, the precast slab contour image is the G-channel precast slab contour image; when the precast slab area is the B-channel precast slab area, the precast slab contour image is the B-channel precast slab contour image.

[0065] In the present invention, by traversing each pixel point in the precast slab area, when there are pixel points in the non-precast slab area within the 3×3 neighborhood range of the pixel point, it indicates that this pixel point is on the edge. Therefore, this pixel point is a pixel point on the contour.

[0066] In this embodiment, the process of obtaining the edge contour includes:

[0067] Taking each contour pixel point on the channel precast slab contour image as the central pixel point, and taking N contour pixel points in the neighborhood range to obtain a contour line segment, where N is a positive integer;

[0068] Calculating the curve length of the contour line segment and calculating the straight-line length from the starting pixel point to the ending pixel point in the contour line segment;

[0069] For the same contour line segment, taking the ratio of the curve length to the straight-line length as the bending degree of the central pixel point;

[0070] Marking the central pixel points with a line segment bending degree greater than the bending degree threshold as suspected intersection pixel points;

[0071] Extracting each continuous suspected intersection pixel point to obtain multiple suspected intersection contours;

[0072] Obtaining the bending degree for the suspected intersection contour (here the bending degree refers to the ratio of the curve length of the suspected intersection contour to the straight-line length from the starting pixel point to the ending pixel point in the suspected intersection contour), and sorting each suspected intersection contour from largest to smallest bending degree;

[0073] Taking the pixel point corresponding to the maximum curvature in the first 4 suspected intersection contours as the intersection point;

[0074] Taking each intersection point as a segmentation point, segment the contour image of the channel precast slab to obtain multiple edge contours.

[0075] In this embodiment, the channel distance threshold and the bending degree threshold are specifically set according to experiments or experience.

[0076] In this embodiment, N can be values such as 4, 6, 8, etc. When it is 4, 2 pixel points are taken on each side of the central pixel point. When it is 6, 3 pixel points are taken on each side of the central pixel point. When it is 8, 4 pixel points are taken on each side of the central pixel point. It is equivalent to taking a contour line segment with a length of N + 1 centered on each pixel point.

[0077] The present invention identifies suspected intersection pixel points by calculating the ratio of the curve length to the straight-line length (bending degree) of the contour line segment. This method can effectively quantify the bending degree of the contour, so as to accurately locate the turning positions on the contour.

[0078] Since the actual shape of the precast slab itself is a cuboid, when photographing the demolded precast slab from a top view, its imaging shape is a rectangle. Extracting each continuous suspected intersection pixel point to obtain multiple cross-suspected contours, the 4 corners of the precast slab have the largest bending degree. Therefore, select the pixel point corresponding to the maximum curvature among the first 4 cross-suspected contours as the intersection point, and segment the contour image of the channel precast slab according to the intersection point to obtain 4 edge contours.

[0079] In this embodiment, the specific process of calculating the curve length of the contour line segment includes: starting from the starting pixel point of the contour line segment, traversing each pixel point in turn. For two adjacent pixel points, calculate the Euclidean distance between them, and add up the distances between all adjacent pixel points. The sum obtained is the curve length of the contour line segment.

[0080] The straight-line length is the Euclidean distance from the starting pixel point to the ending pixel point.

[0081] In this embodiment, the process of constructing the R-channel precast slab edge-defect risk image, the G-channel precast slab edge-defect risk image, and the B-channel precast slab edge-defect risk image includes:

[0082] Connect the starting pixel point and the ending pixel point of each edge contour to obtain a reference straight line;

[0083] Take the distance from each pixel point on the edge contour to the reference straight line as the vertical offset;

[0084] Calculate the edge-defect risk coefficient of the pixel point according to the vertical offset;

[0085] Taking the risk coefficient of edge defect as the new pixel value of this pixel point and setting the pixel values at other positions to 0, a precast slab edge defect risk image is obtained. Among them, the precast slab edge defect risk image corresponding to the precast slab contour image in the R channel is the R channel precast slab edge defect risk image, the precast slab edge defect risk image corresponding to the precast slab contour image in the G channel is the G channel precast slab edge defect risk image, and the precast slab edge defect risk image corresponding to the precast slab contour image in the B channel is the B channel precast slab edge defect risk image.

[0086] The expression of the reference straight line is: Ax + By + C = 0, where A = y 2 -y 1 , B = x 1 -x 2 , C = x 2 y 1 -x 1 y 2 , A is the vertical change coefficient, B is the horizontal change coefficient, C is the position offset constant, (x 1 , y 1 ) is the coordinate of the starting pixel point, (x 2 , y 2 ) is the coordinate of the ending pixel point. For any pixel point (x i , y i ) on the edge contour, the vertical offset to the reference straight line is: , where d i is the vertical offset of the i-th pixel point, | | is the absolute value operation, and i is the number of the pixel point on the edge contour.

[0087] In this embodiment, the process of calculating the edge defect risk coefficient of a pixel point includes: taking the mean of the vertical offsets of each pixel point belonging to the same edge contour to obtain the average vertical offset;

[0088] Subtracting the average vertical offset from the vertical offset of each pixel point and normalizing the subtraction result to obtain the edge defect risk coefficient: , where r i is the edge defect risk coefficient of the i-th pixel point, d avg is the average vertical offset, d max is the maximum vertical offset, and d min is the minimum vertical offset.

[0089] In the present invention, a reference line is obtained by connecting the starting and ending pixel points of the edge contour, providing a standard for measuring the deviation of the edge contour. The distance from each pixel point on the edge contour to the reference line is used as the vertical offset, which can intuitively reflect the position deviation of the pixel point relative to the ideal line. The average vertical offset is obtained by taking the average of the vertical offsets of each pixel point on the same edge contour, and this is used as a reference. By subtracting the average vertical offset from the vertical offset of each pixel point, the overall deviation caused by inaccurate selection of the intersection point can be eliminated, thereby highlighting the relative lack of edge in the local area.

[0090] As Figure 2 shown, the anomaly monitoring model includes: a shallow feature extraction module, a first attention channel module, a second attention channel module, a first multi-scale feature extraction module, a second multi-scale feature extraction module, adder A1, a CNN network, and a fully connected layer;

[0091] The input end of the shallow feature extraction module is used to input the R-channel precast slab edge lack risk image, the G-channel precast slab edge lack risk image, and the B-channel precast slab edge lack risk image, and its output end is respectively connected to the input end of the first attention channel module and the input end of the second attention channel module;

[0092] The output end of the first attention channel module is connected to the input end of the first multi-scale feature extraction module; the output end of the second attention channel module is connected to the input end of the second multi-scale feature extraction module; the input end of adder A1 is respectively connected to the output end of the first multi-scale feature extraction module and the output end of the second multi-scale feature extraction module, and its output end is connected to the input end of the CNN network; the output end of the CNN network is connected to the input end of the fully connected layer; the output end of the fully connected layer is used as the output end of the anomaly monitoring model.

[0093] The shallow feature extraction module is used to preliminarily process the R, G, and B-channel precast slab edge lack risk images, fuse the features of the three channels, and enhance the features at the same pixel point position. The first and second attention channel modules act respectively, automatically screening out the features important for edge lack risk assessment, suppressing the interference of irrelevant information, enhancing the sensitivity to subtle edge lack features, and improving the pertinence and effectiveness of feature extraction. The first and second multi-scale feature extraction modules extract features from different scales. The data after preliminary feature extraction is input into the CNN network, and the powerful convolution operation ability of the CNN can further mine deep features and learn the complex associations between features. The fully connected layer integrates the features obtained from the previous processing and maps the features to the final edge lack risk value.

[0094] As Figure 3 shown, the shallow feature extraction module includes: a first pointwise convolution layer, a second pointwise convolution layer, a third pointwise convolution layer, and a feature fusion layer;

[0095] The first pointwise convolutional layer is used to process the R-channel precast slab edge defect risk image; the second pointwise convolutional layer is used to process the G-channel precast slab edge defect risk image; the third pointwise convolutional layer is used to process the B-channel precast slab edge defect risk image; the feature fusion layer is used to fuse the outputs of the first pointwise convolutional layer, the second pointwise convolutional layer, and the third pointwise convolutional layer.

[0096] In this embodiment, the feature fusion layer is an adder, and the pointwise convolutional layer is a 1×1 convolutional layer, which ensures that the size of the precast slab edge defect risk image of each channel remains unchanged, and then the images of the three channels are added element-wise, so that the features at the same position are added to achieve feature enhancement.

[0097] As Figure 4 shown, the first attention channel module includes: an average pooling layer, a first Sigmoid layer, and a multiplier M1;

[0098] The first input end of the multiplier M1 is connected to the input end of the average pooling layer and serves as the input end of the first attention channel module; the output end of the average pooling layer is connected to the input end of the first Sigmoid layer; the second input end of the multiplier M1 is connected to the output end of the first Sigmoid layer, and its output end serves as the output end of the first attention channel module.

[0099] As Figure 5 shown, the second attention channel module includes: a max pooling layer, a second Sigmoid layer, and a multiplier M2;

[0100] The first input end of the multiplier M2 is connected to the input end of the max pooling layer and serves as the input end of the second attention channel module; the output end of the max pooling layer is connected to the input end of the second Sigmoid layer; the second input end of the multiplier M2 is connected to the output end of the second Sigmoid layer, and its output end serves as the output end of the second attention channel module.

[0101] Average pooling can highlight the overall feature distribution by calculating the average value of the pixels in the region; max pooling selects the maximum value in the region to emphasize the prominent features. The pooled features are input to the Sigmoid layer to generate attention weights. The multiplier multiplies the attention weights by the original input features, enabling the model to adaptively focus on the key features related to the precast slab edge defect risk, suppressing irrelevant or secondary information, and enhancing the pertinence and effectiveness of feature extraction.

[0102] The multiplier M1 and the multiplier M2 perform element-wise multiplication.

[0103] As Figure 6 shown, both the first multi-scale feature extraction module and the second multi-scale feature extraction module include: a first convolutional block, a second convolutional block, a third convolutional block, and a Concat layer;

[0104] The input end of the first convolutional block serves as the input end of the first multi-scale feature extraction module or the second multi-scale feature extraction module, and its output end is respectively connected to the input end of the second convolutional block and the input end of the Concat layer; the output end of the second convolutional block is respectively connected to the input end of the third convolutional block and the input end of the Concat layer; the output end of the third convolutional block is connected to the input end of the Concat layer; the output end of the Concat layer serves as the output end of the first multi-scale feature extraction module or the second multi-scale feature extraction module.

[0105] In the present invention, the convolutional kernel size of the first convolutional block is 3×3, the convolutional kernel size of the second convolutional block is 5×5, and the convolutional kernel size of the third convolutional block is 7×7. Different convolutional kernel sizes have different receptive fields for images, and the extracted features are diverse, realizing the mutual complementation of features at multiple scales.

[0106] The convolutional block includes: a convolutional layer, an activation function layer, and a normalization layer.

[0107] The present invention collects images of precast slabs after demoulding on the production line, splits the images into R, G, and B channel images, processes the R, G, and B channels separately, obtains the precast slab regions respectively, further obtains the pixel points of the precast slab contour, constructs the contour images of the R, G, and B channels respectively, then finds the intersection points from the precast slab contour images of each channel, and extracts the edge contours, so as to facilitate the comparison of the edge contours with the reference straight line, obtain the chamfer risk coefficients of the pixel points on the edge contours, construct the chamfer risk images of the precast slabs in the R, G, and B channels respectively, highlight the chamfer conditions of the pixel points on each edge contour, and then use the anomaly monitoring model to process the chamfer risk images of the precast slabs in the R, G, and B channels, comprehensively considering the chamfer conditions of the pixel points in the three channels, and improving the accuracy of anomaly monitoring. Compared with the manual detection method, the present invention has a higher detection efficiency for chamfers of precast slabs, avoids the errors caused by subjective judgment differences in manual detection, makes the detection results more objective and accurate, ensures the consistency of detection for different batches and at different times, and improves the reliability of quality detection of precast slabs.

[0108] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for abnormal monitoring of a prefabricated panel intelligent production line, characterized in that: The following steps are involved: Collect images of prefabricated panels after demoulding on the production line, and split the images into R, G, and B channel images; Find the prefabricated board area from the R, G, and B channel images respectively, and obtain the R channel prefabricated board area, the G channel prefabricated board area, and the B channel prefabricated board area; Extracting pixel points of the prefabricated board contour from the R channel prefabricated board area, the G channel prefabricated board area, and the B channel prefabricated board area, respectively, to obtain the R channel prefabricated board contour image, the G channel prefabricated board contour image, and the B channel prefabricated board contour image; Find the intersection points on the contour image of each channel precast panel, and segment according to the intersection points to obtain the edge contour; According to the distance between each edge contour and the reference straight line, the edge defect risk coefficient of the pixel point on the edge contour is calculated, and the R channel prefabricated board edge defect risk image, the G channel prefabricated board edge defect risk image and the B channel prefabricated board edge defect risk image are constructed; The abnormal monitoring model is used to process the R channel precast slab missing edge risk image, the G channel precast slab missing edge risk image and the B channel precast slab missing edge risk image to obtain the missing edge risk value.

2. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 1 is characterized in that: The process of finding the precast panel area from each channel image includes: In each channel image, the distance between the channel value and the stored prefabricated board channel value is calculated to obtain the channel distance; When the channel distance is less than the channel distance threshold, the pixel corresponding to the channel value is marked as a valid pixel; The largest connected area formed by each valid pixel point is taken as the prefabricated board area.

3. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 1 is characterized in that: The process of obtaining the R channel prefabricated board contour image, the G channel prefabricated board contour image and the B channel prefabricated board contour image all include: traversing each pixel point in the prefabricated board area, when there is a pixel point in the non-prefabricated board area within the neighborhood range of the pixel point, the pixel point is a contour pixel point, and all contour pixels are extracted to obtain the prefabricated board contour image, wherein, when the prefabricated board area is the R channel prefabricated board area, the prefabricated board contour image is the R channel prefabricated board contour image, when the prefabricated board area is the G channel prefabricated board area, the prefabricated board contour image is the G channel prefabricated board contour image, and when the prefabricated board area is the B channel prefabricated board area, the prefabricated board contour image is the B channel prefabricated board contour image.

4. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 1 is characterized in that: The process of obtaining the edge contour includes: Taking each contour pixel point on the channel prefabricated plate contour image as the central pixel point, taking N contour pixel points in the neighborhood range to obtain the contour line segment, where N is a positive integer; Calculate the curve length of the contour segment and the straight line length from the starting pixel point to the ending pixel point in the contour segment; For the same contour segment, the ratio of the curve length to the straight line length is taken as the curvature of the central pixel point; The central pixel point where the curvature of the line segment is greater than the curvature threshold is marked as a suspected crossing pixel point; Extract each continuous suspected intersection pixel point to obtain multiple suspected intersection contours; Obtain the curvature of the cross suspected contours, and sort the cross suspected contours from large to small according to the curvature; The pixel point corresponding to the maximum curvature in the first four intersecting suspected contours is taken as the intersection point; Taking each intersection as a segmentation point, the channel precast panel contour image is segmented to obtain multiple edge contours.

5. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 1 is characterized in that: The process of constructing the R channel precast slab edge defect risk image, the G channel precast slab edge defect risk image, and the B channel precast slab edge defect risk image includes: Connect the starting pixel point and the ending pixel point of each edge contour to obtain a reference straight line; The distance from each pixel point on the edge contour to the reference straight line is taken as the vertical offset; According to the vertical offset, the edge defect risk coefficient of the pixel point is calculated; The chipping edge risk coefficient is used as the new pixel value of the pixel point, and the pixel values ​​at other positions are set to 0 to obtain the prefabricated board chipping edge risk image, wherein the prefabricated board chipping edge risk image corresponding to the R channel prefabricated board contour image is the R channel prefabricated board chipping edge risk image, the prefabricated board chipping edge risk image corresponding to the G channel prefabricated board contour image is the G channel prefabricated board chipping edge risk image, and the prefabricated board chipping edge risk image corresponding to the B channel prefabricated board contour image is the B channel prefabricated board chipping edge risk image.

6. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 5 is characterized in that: The process of calculating the edge defect risk coefficient of a pixel point includes: taking the average of the vertical offsets of each pixel point belonging to the same edge contour to obtain an average vertical offset; The vertical offset of each pixel is subtracted from the average vertical offset, and the subtraction result is normalized to obtain the edge defect risk coefficient.

7. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 1 is characterized in that: The anomaly monitoring model includes: a shallow feature extraction module, a first attention channel module, a second attention channel module, a first multi-scale feature extraction module, a second multi-scale feature extraction module, an adder A1, a CNN network and a fully connected layer; The input end of the shallow feature extraction module is used to input the R channel prefabricated board missing edge risk image, the G channel prefabricated board missing edge risk image and the B channel prefabricated board missing edge risk image, and its output end is connected to the input end of the first attention channel module and the input end of the second attention channel module respectively; The output end of the first attention channel module is connected to the input end of the first multi-scale feature extraction module; the output end of the second attention channel module is connected to the input end of the second multi-scale feature extraction module; the input end of the adder A1 is respectively connected to the output end of the first multi-scale feature extraction module and the output end of the second multi-scale feature extraction module, and its output end is connected to the input end of the CNN network; the output end of the CNN network is connected to the input end of the fully connected layer; the output end of the fully connected layer serves as the output end of the abnormality monitoring model.

8. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 7 is characterized in that: The shallow feature extraction module includes: a first point-by-point convolution layer, a second point-by-point convolution layer, a third point-by-point convolution layer and a feature fusion layer; The first point-by-point convolution layer is used to process the R channel prefabricated board chipping risk image; the second point-by-point convolution layer is used to process the G channel prefabricated board chipping risk image; the third point-by-point convolution layer is used to process the B channel prefabricated board chipping risk image; the feature fusion layer is used to fuse the output of the first point-by-point convolution layer, the output of the second point-by-point convolution layer and the output of the third point-by-point convolution layer.

9. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 7 is characterized in that: The first attention channel module includes: an average pooling layer, a first sigmoid layer, and a multiplier M1; The first input end of the multiplier M1 is connected to the input end of the average pooling layer and serves as the input end of the first attention channel module; the output end of the average pooling layer is connected to the input end of the first Sigmoid layer; the second input end of the multiplier M1 is connected to the output end of the first Sigmoid layer, and its output end serves as the output end of the first attention channel module; The second attention channel module includes: a maximum pooling layer, a second sigmoid layer and a multiplier M2; The first input end of the multiplier M2 is connected to the input end of the maximum pooling layer and serves as the input end of the second attention channel module; the output end of the maximum pooling layer is connected to the input end of the second Sigmoid layer; the second input end of the multiplier M2 is connected to the output end of the second Sigmoid layer, and its output end serves as the output end of the second attention channel module.

10. The abnormality monitoring method for the prefabricated panel intelligent production line according to claim 7 is characterized in that: The first multi-scale feature extraction module and the second multi-scale feature extraction module both include: a first convolution block, a second convolution block, a third convolution block and a Concat layer; The input end of the first convolution block serves as the input end of the first multi-scale feature extraction module or the second multi-scale feature extraction module, and its output end is respectively connected to the input end of the second convolution block and the input end of the Concat layer; the output end of the second convolution block is respectively connected to the input end of the third convolution block and the input end of the Concat layer; the output end of the third convolution block is connected to the input end of the Concat layer; the output end of the Concat layer serves as the output end of the first multi-scale feature extraction module or the second multi-scale feature extraction module.

Images