CN117274158A - A method for online monitoring of paving defects in ceramic photocuring additive manufacturing processes - Google Patents

Abstract

The invention discloses a method for online monitoring of paving defects in ceramic light-curing additive manufacturing processes, which includes correcting the side-view paving image collected by a surveillance camera into a front-view paving image based on the acquired image correction matrix; After image correction, the front-view paving image of the k-th layer and the front-view paving image of the k-N layer are combined with the k-N-th layer slice image to synthesize a three-channel color image, and divide the three-channel color image into several sub-images; build The deep learning model uses historical image data as a training set to train the deep learning model. After the training is completed, the defect detection model is obtained, and the sub-images are sequentially input into the defect detection model to achieve in-situ real-time detection of defects. The advantage is: integrating more information into the image that needs to be detected, as a supplement to the information of the image being inspected, while making full use of the feature extraction capabilities of the model, it gives the image more information and improves the detection accuracy.

Description

A method for online monitoring of paving defects in ceramic photocuring additive manufacturing processes

Technical field

The invention relates to the technical field of paving situation monitoring, and in particular to a method for online monitoring of paving defects in ceramic photocuring additive manufacturing processes.

Background technique

With the development of 3D printing technology in our country, printing technology has gradually matured, and the size of printed products has gradually developed towards large-scale. The problem is that the manufacturing time has doubled, and various defects are prone to occur during the manufacturing process.

As shown in Figure 1, there are three main types of defects in the ceramic additive manufacturing process. The first type of defects are crack defects, which are manifested as bubbles, pits and scratches. Although such defects are not serious during the molding process of the part body, It directly causes damage to the parts, but also leads to uneven distribution of materials inside the parts, which can easily lead to cracking of the parts during the subsequent degreasing and sintering processes. The second type of defects is fracture defects, which are fractures caused by lack of material, in the form of paving. There is no material in a certain part of the surface, forming an obvious edge between the area covered with raw materials and the area not covered. In the subsequent molding process, even if the missing parts are re-printed, due to the penetration of the ceramic slurry, Insufficient lightness makes it difficult to penetrate the slurry with two layers of paving thickness, making it difficult to fully solidify the raw materials. The lack of material defects seriously affects the interlayer bonding force of the green parts, causing the parts to break in layers; the third type of defects is damage defects, which manifest as Because of the collapse of the weak points of the parts, such defects will cause the molding process of the green parts to completely fail. Failure to identify the defects and terminate the printing task in time will lead to a huge waste of time and resources.

In current engineering applications, finding such errors generally relies on manual visual inspection. With the scale-up of ceramic additive manufacturing, the manufacturing time increases exponentially, making manual inspection methods increasingly unsuitable for this application scenario.

The digital image mask of the cured area in a ceramic additive molding process is shown in Figure 2(a), and a typical image of the paving surface under normal conditions is shown in Figure 2(b). A typical defective ( The image of the paving surface (lack of material) is shown in Figure 2(c). It can be intuitively seen from the figure that the information on the image includes the slurry solidification area formed by normal printing and the unexpected edges caused by defects (lack of material). The performance of the two edges on the two-dimensional image is very similar. For the target detection model It is quite confusing. For the problem of defect target detection in such complex surfaces, the existing target detection method is directly applied to the detection of such defects, but it cannot achieve good defect detection results. The model cannot accurately identify the printing edges and The edge of the material is missing, resulting in false detection. Taking the material shortage defect in the manufacturing process as an example, the edge caused by the paving material is used as the target to be identified in the defect, the data set image is annotated, and then the data set image and annotation information are input into the model for training to obtain the target. Detect the weight of the model, and then input the images of the verification set into the trained model to verify the training effect. The verification image is shown in Figure 3. It can be clearly seen that in the verification image, the model recognizes many normally solidified edges due to lack of The edges generated by the material cause false detections.

There are two reasons for this misdetection: 1. In the ceramic additive manufacturing and molding process, since the raw material used is a mixture of ceramic solid powder and resin with a single component, no color information is generated during the curing process; 2. The additive manufacturing molding process is known for its strong flexibility. Each molding task makes the curing area on the surface of the paving material very flexible and changeable depending on the model, and the edges caused by the lack of material in the paving material are highly random. , the comprehensive result is that there is no obvious difference in geometric characteristics between the solidified edge and the missing edge of the paving surface, making it difficult to distinguish the two edges. The above two reasons can be summarized into one point, that is, the existing paving images do not adequately express the information of the additive manufacturing paving surface, making it impossible for the model to distinguish between printed edges and missing material edges.

Contents of the invention

The purpose of the present invention is to provide a method for online monitoring of paving defects in ceramic photocuring additive manufacturing processes, thereby solving the aforementioned problems existing in the prior art.

In order to achieve the above objects, the technical solutions adopted by the present invention are as follows:

A method for online monitoring of paving defects in ceramic photocuring additive manufacturing processes, including the following steps:

S1. Image correction:

The image correction matrix is obtained based on the image projected by the optical machine and the slice image. Based on the image correction matrix, the side-view paving image collected by the surveillance camera is corrected into a front-view paving image; the slice image is a manufacturing plane under the front viewing angle. Image of the slurry curing molding area;

S2, multi-layer image fusion:

After image correction, the front-view paving image of the k-th layer and the front-view paving image of the k-Nth layer are combined with the k-Nth layer slice image to synthesize a three-channel color image, and the three-channel color image is divided into several sub-images; where, N= 1,2,3,…,K-1;

S3. Defect detection:

Build a deep learning model and use historical image data as a training set to train the deep learning model. After the training is completed, the defect detection model is obtained, and the sub-images are sequentially input into the defect detection model to achieve in-situ online real-time detection of defects in ceramic additive manufacturing paving materials. .

Preferably, step S2 specifically includes the following content:

S21. Use optical machine projection to mark the outer contour of the molding area, install a narrow-band light-transmitting sheet on the front end of the lens of the surveillance camera, collect the image projected by the optical machine and calculate the four corner points of the molding area, and compare the four corner points of the molding area with The four corner points of the sliced image correspond to obtain the image correction matrix;

S22. Apply the image correction matrix to the side-view paving image collected by the surveillance camera to obtain a preliminary front-view paving image;

S23. Use the interpolation algorithm to fill in the blank pixels in the preliminary front-view paving image, and crop the preliminary front-view paving image to the same pixels as the sliced image to obtain the final front-view paving image, that is, the corrected image;

S24. Define the objective function and use the optimization algorithm to optimize the objective function to reduce the error between the corrected image and the original image.

Preferably, in step S24, the ratio of the first-order norm of the binarized slice image and the corrected image to the number of pixels of the true image is defined as the objective function, which is expressed as follows,

img _cali = img _src ∝H

(x′,y′)=(0,0),(0,H _truth ),(W _truth ,H _truth ),(W _truth ,0)

Among them, img _slic is the slice image, that is, the corrected true value; img _cali is the corrected and binarized image of the optical-mechanical projection image collected by the surveillance camera; img _src is the original image collected by the surveillance camera, and the original image is based on the image The correction matrix H completes the image correction; (x, y, 1) is the homogeneous coordinate point of the image in the original image; (x', y', 1) is the point in the corrected image corresponding to the original image, h ₁ ~ h ₈ are the eight independent elements of the correction matrix H; W _truth and H _truth are the maximum pixel values in the width and height directions of the true image respectively; x∈[0,w], y∈[0,h], w and h are respectively The pixel coordinate point in the original image; i, j are the pixels of the corresponding image.

Preferably, the optimization algorithm is an improved genetic algorithm,

The improved genetic algorithm uses the four corner point coordinates in S21 and the random points according to the normal distribution around the corner points as the initial population to provide a priori knowledge for the genetic algorithm; selection, crossover and mutation are performed in the initial population to Maximum efficiency approaches the optimal value of the objective function.

Preferably, in the process of optimizing the objective function using an improved genetic algorithm, the integer part and the decimal part of the correction point coordinates are encoded separately, and the decimal part is given a greater mutation probability to make the genetic algorithm's search ability stronger, and the integer part is Parts are given a smaller mutation probability to stabilize the evolutionary process; only the decimal part is crossed during the crossover process, and the integer part is not crossed to reduce the computational complexity;

Before each generation evolves, the fitness of each individual is sorted according to the objective function. The individual with the highest fitness directly enters the next generation. The purpose of the improved genetic algorithm is to find more accurate correction mark points between the integer pixels of the image.

Preferably, step S3 specifically includes the following content:

S21. Collect side view paving images of the k-Nth layer and the kth layer respectively;

S22. Calibrate the side-view paving image to obtain the front-view paving image;

S23. Use the interpolation algorithm to fill in the blank pixels in the front-view paving image;

S24. Crop the front view paving image to the same pixel as the sliced image;

S25. Place the front-view paving image of the k-th layer after the above processing in the first channel, the front-view paving image of the k-Nth layer after processing in the second channel, and place the k-Nth layer slice image in the third channel, so as to Synthesize three-channel color images;

S26. Divide the three-channel color image into blocks, input the divided images into the defect detection model for detection, and output the identified defective target objects.

Preferably, the specific process of using surveillance cameras to collect paving images is as follows:

Install the surveillance camera side-axis diagonally above the manufacturing plane, and capture side-view paving images at each moment through the surveillance camera to achieve image collection;

When obtaining the image correction matrix, a detachable narrow-band light-transmitting sheet needs to be installed at the front end of the lens of the surveillance camera to obtain the corresponding image.

Preferably, the narrow-band light-transmitting sheet is a 405nm narrow-band light-transmitting sheet, and only light with a wavelength of 405n can pass through the narrow-band light-transmitting sheet.

The beneficial effects of the present invention are: 1. Using machine vision to automatically detect the entire ceramic manufacturing paving process, relying on image processing technology and deep learning technology to intelligently identify paving defects on the manufacturing plane, solving the problem of long-term processes The problem of detection efficiency of the process. 2. Propose a multi-layer image fusion method to incorporate more information into the images that need to be detected. As a supplement to the information of the inspected images, while making full use of the feature extraction capabilities of the model, it also gives the images more information, which helps Improve detection accuracy. 3. Correct the collected side view image to the front view image through the correction matrix, and solve the set objective function through the optimization algorithm to reduce the correction error. 4. Use an improved genetic algorithm to optimize the objective function. By using the initial corner point and the points around the corner point as the initial population, the genetic algorithm is provided with prior knowledge, so that it can approach the optimal value of the objective function with maximum efficiency and improve the algorithm. Calculation speed.

Description of the drawings

Figure 1 shows the typical defect manifestations during the laying process of ceramic additive manufacturing;

Figure 2 is a typical representation of normal paving images and defective paving images during the paving process of ceramic additive manufacturing;

Figure 3 is a schematic diagram of the solidified edge being mistakenly detected as a missing edge during the paving process of ceramic additive manufacturing;

Figure 4 is a schematic diagram of an additively manufactured silicon carbide reflector in an embodiment of the present invention;

Figure 5 is a side view image and a corrected image collected in the embodiment of the present invention;

Figure 6 is an effect diagram of the genetic algorithm solving the optimal correction point in the embodiment of the present invention;

Figure 7 is a schematic diagram of multi-layer image fusion in an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention.

Embodiment 1

In this embodiment, a method for online monitoring of paving defects in ceramic photocuring additive manufacturing processes is provided, which includes the following steps:

1. Image collection

In this embodiment, the surveillance camera is installed on the side axis obliquely above the manufacturing plane, and the surveillance camera captures side-view paving images at various times, thereby achieving image collection.

Install the narrow-band light-transmitting sheet when calibrating the parameters of the optical machine, and remove it after completing the calibration. That is, when obtaining the image correction matrix, a detachable narrow-band light-transmitting sheet needs to be installed at the front end of the lens of the surveillance camera to obtain the corresponding image. At other times, the narrow-band light-transmitting sheet is removed and the surveillance camera is used to obtain the image directly.

The narrow-band light-transmitting sheet is a 405nm narrow-band light-transmitting sheet, and only light with a wavelength of 405n can pass through the narrow-band light-transmitting sheet.

2. Image correction:

The image correction matrix is obtained based on the image projected by the optical machine and the slice image. Based on the image correction matrix, the side-view paving image collected by the surveillance camera is corrected into a front-view paving image; the slice image is a manufacturing plane under the front viewing angle. Image of the slurry cured molding area.

Specifically include the following content,

1. Use optical machine projection to mark the outer contour of the molding area, install a narrow-band light-transmitting sheet on the front end of the lens of the surveillance camera, collect the image projected by the optical machine and calculate the four corner points of the molding area, and compare the four corner points of the molding area with The four corner points of the sliced image correspond to obtain the image correction matrix;

2. Apply the image correction matrix to the side-view paving image collected by the surveillance camera to obtain a preliminary front-view paving image;

3. Use the interpolation algorithm to fill in the blank pixels in the preliminary front-view paving image, and crop the preliminary front-view paving image to the same pixels as the sliced image to obtain the final front-view paving image, which is the corrected image;

4. Define the objective function and use the optimization algorithm to optimize the objective function to reduce the error between the corrected image and the original image. The optimization algorithm is an improved genetic algorithm,

The improved genetic algorithm uses the four corner point coordinates in S21 and the random points according to the normal distribution around the corner points as the initial population to provide a priori knowledge for the genetic algorithm; selection, crossover and mutation are performed in the initial population to Maximum efficiency approaches the optimal value of the objective function.

In this embodiment, the ratio of the first-order norm of the binarized slice image and the corrected image to the number of pixels in the true image is defined as the objective function, which is expressed as follows:

img _cali = img _src ∝H

(x′,y′)=(0,0),(0,H _truth ),(W _truth ,H _truth ),(W _truth ,0)

Among them, img _slic is the slice image, that is, the corrected true value; img _cali is the corrected and binarized image of the optical-mechanical projection image collected by the surveillance camera; img _src is the original image collected by the surveillance camera, and the original image is based on the image The correction matrix H completes the image correction; (x, y, 1) is the homogeneous coordinate point of the image in the original image; (x', y', 1) is the point in the corrected image corresponding to the original image, h ₁ ~ h ₈ are the eight independent elements of the correction matrix H; W _truth and H _truth are the maximum pixel values in the width and height directions of the true image respectively; x∈[0,w], y∈[0,h], w and h are respectively The pixel coordinate point in the original image; i, j are the pixels of the corresponding image.

In this embodiment, according to the properties of plane projective mapping, a plane projective transformation is a linear transformation of its homogeneous vector in 3 dimensions. This linear transformation relationship can be represented by a non-singular 3x3 matrix, that is

After installing the 405nm narrow-band light-transmitting sheet on the front end of the lens of the surveillance camera, the camera only collects the 405nm light signal. At this time, the optical system is used to calibrate the four corner points of the manufacturing plane. The above formula is obtained based on the four sets of corresponding points between the two planes. Using the eight independent elements of the transformation matrix in , the transformation matrix between the two planes is obtained, and then the matrix is applied to the entire plane, which can eliminate the distortion on the plane and restore the image from the side view to the image from the front view.

In this embodiment, since the slice image of the k-Nth layer is needed as an auxiliary information source in multi-layer image fusion to provide the curing area information of the k-Nth layer's manufacturing surface, it should be noted that the slice image and paving material are required in this process. Only the images that can strictly correspond to the sliced images can accurately provide information on the cured area of the paving surface. However, in order to ensure the efficient utilization and uniform distribution of the optical machine energy in the manufacturing surface, the position directly above the manufacturing surface is occupied by the optical machine, and the industrial camera used for monitoring can only be placed on the side of the optical machine to collect paving images, as shown in the figure As shown in 5(a). Therefore, the side-view paving image needs to be corrected to the front-view perspective to complete the correspondence between the slice image and the manufacturing surface. According to the image correction theory, image correction requires the calculation of a unique 3x3 correction matrix between two images. Under the homogeneous coordinate system, only eight independent ratios in the matrix need to be calculated. Calculating these eight independent ratios requires eight equations. The system of equations is solved. The horizontal and vertical coordinates of a set of corresponding points on the two images can provide two equations, and four sets of corresponding points can provide the required eight equations.

This invention uses optical machine projection to mark the outer contour of the molding area, installs a narrow-band light-transmitting sheet at the front end of the lens of the surveillance camera, collects the image projected by the optical machine and calculates the four corner points of the molding area, and compares the four corner points of the molding area with The correction matrix of the image can be obtained by corresponding the four corner points of the sliced image. Then the correction matrix is applied to the original image. After the image is cropped, the front view image shown in Figure 5(b) is obtained. In this application method, the detected positions of the four corner points of the molding area are inaccurate because the resolution of the surveillance camera used is close to the opto-mechanical resolution used for projection. In engineering applications, the resolution of the camera is required to be much greater than the projection resolution of the light machine in order to use images to accurately describe corner points. However, sunken light-curing ceramic additive manufacturing is known for its high precision. The single pixel in the manufacturing area The size is only about 60 microns. It is undoubtedly impossible to use a camera with a resolution much larger than that of the optical machine when the resolution of the projection machine itself is already very large. In order to solve the problem of image correction, the present invention defines the ratio of the first-order norm of the binarized slice image and the corrected image to the number of pixels of the true image as the loss function. The smaller the function value, the smaller the correction error.

The process of calculating the corrected image from the original image by the correction matrix requires corresponding position calculation, interpolation, binarization, and cropping. Therefore, the relationship between the original image and the corrected image is not a simple equation. For this reason, the target The function value of the function cannot obtain an accurate analytical solution. Instead, a feasible method is to exhaustively enumerate all the point combinations on the original image through the exhaustive method, calculate the correction matrix separately and then calculate the objective function value. When there is approximate information about the corner point position, In the case of prior knowledge, the calculation cost of possible optimal value combinations is controllable.

However, the default precondition of the exhaustive method is that the value range of the optimal solution is discrete values (because the selected points are all pixels in the original image), but this is not the actual situation. The optimal correction point may be somewhere between two pixels. In this case heuristic algorithm can get better results. Commonly used heuristic algorithms include neural networks, simulated annealing, ant colony algorithm and genetic algorithms. Combining various algorithms with the characteristics of this topic, the present invention uses genetic algorithms to optimize the objective function.

Genetic algorithms find optimal solutions by simulating the evolution process of natural populations. However, one of its major disadvantages is that it is difficult to achieve good results when there are too many decision variables, because different decision variables often need to mutate toward the optimal solution at the same time to gain adaptability. Stronger populations, but the inheritance and variation of different decision variables are independent of each other, so the more decision variables there are, the more difficult it is to achieve satisfactory results. A feasible method to solve such problems is to provide a priori knowledge for the genetic algorithm to guide the optimization process. In the present invention, the projection range of the optical machine in Figure 5(a) is the range of the manufacturing area, the four corner points of the boundary and The random points according to the normal distribution around the boundary are used as the population of the genetic algorithm, providing a priori knowledge for the genetic algorithm to calculate. In the present invention, the set genetic algorithm parameters are shown in Table 2.

Table 2 Genetic algorithm parameters

The reason why a smaller population size and a larger number of iterations are selected in this invention is that a small population size can effectively improve the calculation speed of the algorithm. Under the premise of having prior knowledge, the initial population is already an "elite individual". Selection, crossover and mutation among elite individuals can approach the optimal value of the objective function with maximum efficiency.

In the process of optimizing the objective function using the improved genetic algorithm, the integer part and the decimal part of the correction point coordinates are encoded separately, and the decimal part is given a higher mutation probability to make the algorithm's search ability stronger, and the integer part is given a smaller The mutation probability is to make the evolutionary process stable; only the decimal part is crossed during the crossover process, and the integer part is not crossed. The reason for this is that after the initial population is defined, the integer part similarity of individuals is high and there is no need for crossover, so this process is omitted to reduce the complexity of calculations.

In addition, in order to ensure the stability of the algorithm, the fitness of each individual is sorted according to the objective function before each generation evolves, and the individual with the highest fitness directly enters the next generation. The ultimate goal of the algorithm is to find more accurate correction mark points between the integer pixels of the image.

After a large number of experiments, it was found that the method used can stably approximate the optimal solution. After about 500 generations of iterations, the average fitness of the population has been very close to the fitness of the optimal individual. It has almost completely coincided with the fitness of the optimal individual by the 1000th generation. As shown in Figure 6.

The optimization results are shown in Table 3. It can be seen that the fitness of the correction points has been improved by an order of magnitude after optimization by the genetic algorithm, and the discrete values of the coordinate points before optimization have been extended to continuous values, thereby more accurately locating the correction points. point to improve the optimization effect. After image correction, the sliced image of the part and the paving image can be more accurately matched, which provides a basis for subsequent work.

Table 3 Comparison of optimization results

3. Multi-layer image fusion:

After image correction, the front-view paving image of the k-th layer and the front-view paving image of the k-Nth layer are combined with the k-Nth layer slice image to synthesize a three-channel color image, and the three-channel color image is divided into several sub-images. The value of N can be selected according to the actual situation to better meet actual needs.

In this embodiment, the slice image of the k-1th layer is used as an auxiliary information source to provide the manufacturing surface solidification area information of the k-1th layer, thereby achieving multi-layer image fusion, specifically including the following:

1. Collect the side view paving images of the k-1th layer and the kth layer respectively;

2. Correct the side-view paving image to obtain the front-view paving image;

3. Use the interpolation algorithm to fill in the blank pixels in the front-view paving image;

4. Crop the front view paving image to the same pixels as the sliced image;

5. Place the front-view paving image of the k-th layer after the above processing in the first channel, the front-view paving image of the k-1th layer after processing in the second channel, and the k-1th layer slice image in the second channel. three-channel to synthesize a three-channel color image;

6. Divide the three-channel color image into blocks, input the divided images into the defect detection model for detection, and output the identified defect targets. In this step, the purpose of blocking is to realize paving and testing at the same time. The parts of the platform that have been paved with materials can be tested first instead of waiting for the entire platform to be tested together. This can improve the work efficiency of the equipment. .

In this embodiment, in order to solve the problem of target detection on complex surfaces, the present invention uses a multi-layer image fusion method to incorporate more information into the image that needs to be detected as an information supplement to the image being detected. Specifically, several other image information is integrated into the inspected image to provide information reference for the detection model. In the traditional target detection model, after the inspected image is input into the detection framework, it is down-sampled several times for feature extraction to generate feature maps of different scales, and then the target information is decoded in the corresponding feature maps to complete the prediction of target objects at different scales.

The present invention proposes a new detection method, which uses several auxiliary images as information supplements of the inspected image, extracts features of the main inspected image and the auxiliary information image at the same time in the detection model, and combines the feature layer of the auxiliary image with the main inspected image. The feature layers of the image are stacked and fused to jointly output the prediction results. During the blank forming process of ceramic additive manufacturing parts, the parts are formed layer by layer in a time sequence. The currently formed working layer is defined as the k-th layer, and the layer formed before the current layer is defined as the k-Nth layer. In this embodiment, two auxiliary images are selected as the information supplement of the inspected image, namely the k-Nth layer paving image of the main inspected image and the k-Nth layer slice image. The reason for selecting these two images is that the paving The manifestation of material defects on the image is the edge mutation of the paving image of the kth layer relative to the paving image of the k-Nth layer. In practical applications, the exposure process is generally performed after the paving process is normal. Therefore, the paving images of the k-Nth layer are generally normal paving images. The image information of the normal paving images of the k-Nth layer is integrated into the current k-th layer. In the inspected image of the layer, it helps the model distinguish the difference between normal edges and defective edges, and plays a very important guiding role in the detection of paving defects in the current layer.

The reason for choosing the slice image of the k-Nth layer as another auxiliary image is that the edge mutation on the surface of the pavement is sometimes not caused by defects in the paving material, such as the silicon carbide mirror shown in Figure 4(a). The thickness of this model is 3mm. During the molding process with a layer thickness of 50um, the structure changes from mirror to back between layer 41 and layer 42. During this process, the paving surface mutates from a circular solidified area to a grid row solidified area, creating an image. The mutation information of the edge is shown in Figure 4(b)-Figure 4(c). In this case, the normal edge mutation information needs to be input into the detection model. The physical meaning of the slice image is the slurry solidification area on the manufacturing plane, which can provide the surface solidification area information of the upper layer to assist the model in accurately identifying the edges of the normal solidification area.

In this embodiment, red (R), green (G), and blue (B) are the three optical primary colors. The three colors can be combined to produce any color of the image. Most true-color digital images are composed of these three primary colors. Channel 24-bit deep bitmap, similarly, the input images of almost all image recognition deep learning models are three-channel images.

In the present invention, since the surface defects of the paving materials have no color information, and the signal-to-noise ratio of the black-and-white camera is better than that of the color camera, the images collected are always 8-bit deep grayscale images, which are copied into two other images when inputting the deep learning model. channel is converted into a three-channel image, which greatly wastes the feature extraction capability of the model in this process. In order to solve this problem, as shown in Figure 7, the present invention proposes multi-layer paving image fusion technology for the first time, combining different images into a 24-bit deep bitmap, replacing the previous method of directly copying a single channel, and making full use of the model. At the same time, the feature extraction capability gives the image more information, which helps to improve the detection accuracy.

4. Defect detection:

Build a deep learning model and use historical image data as a training set to train the deep learning model. After the training is completed, the defect detection model is obtained, and each sub-image is input into the defect detection model in turn to realize in-situ online real-time performance of ceramic additive manufacturing paving defects. detection.

5. Method application

The method of the present invention is put into practical application, and the detection results are fed back to the control terminal of the paving equipment. If there is a defect, the paving equipment operates normally. When a defect is detected, the detection results are stored and corresponding processing actions are performed.

By adopting the above technical solutions disclosed in the present invention, the following beneficial effects are obtained:

The present invention provides a method for online monitoring of paving defects in ceramic light-curing additive manufacturing processes. It uses machine vision to automatically detect the entire ceramic manufacturing paving process, and relies on image processing technology and deep learning technology to intelligently identify The paving defects of the manufacturing plane solve the problem of detection efficiency in the long-term process. A multi-layer image fusion method is proposed to incorporate more information into the image that needs to be detected. As a supplement to the information of the inspected image, it makes full use of the feature extraction capabilities of the model and gives the image more information, which helps to improve detection. Accuracy. The collected side view image is corrected to the front view image through the correction matrix, and the set objective function is solved through the optimization algorithm to reduce the correction error. An improved genetic algorithm is used to optimize the objective function. By using the initial corner point and the points around the corner point as the initial population, the genetic algorithm is provided with prior knowledge, allowing it to approach the optimal value of the objective function with maximum efficiency and improve the calculation speed of the algorithm. .

The above are only preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention. These improvements and modifications can also be made. The scope of protection of the present invention should be considered.

Claims (8)

1. A method for online monitoring of paving defects in ceramic photocuring additive manufacturing processes, which is characterized by: including the following steps: S1. Image correction: The image correction matrix is obtained based on the image projected by the optical machine and the slice image. Based on the image correction matrix, the side-view paving image collected by the surveillance camera is corrected into a front-view paving image; the slice image is a manufacturing plane under the front viewing angle. Image of the slurry curing molding area; S2, multi-layer image fusion: After image correction, the front-view paving image of the k-th layer and the front-view paving image of the k-Nth layer are combined with the k-Nth layer slice image to synthesize a three-channel color image, and the three-channel color image is divided into several sub-images; where, N= 1,2,3,…,K-1; S3. Defect detection: Build a deep learning model and use historical image data as a training set to train the deep learning model. After the training is completed, the defect detection model is obtained, and the sub-images are sequentially input into the defect detection model to achieve in-situ online real-time detection of defects in ceramic additive manufacturing paving materials. . 2. The method for online monitoring of paving defects in ceramic light-curing additive manufacturing process according to claim 1, characterized in that step S2 specifically includes the following content: S21. Use optical machine projection to mark the outer contour of the molding area, install a narrow-band light-transmitting sheet on the front end of the lens of the surveillance camera, collect the image projected by the optical machine and calculate the four corner points of the molding area, and compare the four corner points of the molding area with The four corner points of the sliced image correspond to obtain the image correction matrix; S22. Apply the image correction matrix to the side-view paving image collected by the surveillance camera to obtain a preliminary front-view paving image; S23. Use the interpolation algorithm to fill in the blank pixels in the preliminary front-view paving image, and crop the preliminary front-view paving image to the same pixels as the sliced image to obtain the final front-view paving image, that is, the corrected image; S24. Define the objective function and use the optimization algorithm to optimize the objective function to reduce the error between the corrected image and the original image. 3. The method for online monitoring of paving defects in ceramic photocuring additive manufacturing process according to claim 2, characterized in that: in step S24, the first-order norm of the binarized slice image and the corrected image is The ratio to the number of pixels in the true image is defined as the objective function, expressed as follows, img _cali = img _src ∝H (x′, y′) = (0, 0), (0, H _truth ), (W _truth , H _truth ), (W _truth , 0) Among them, img _slic is the slice image, that is, the corrected true value; img _cali is the corrected and binarized image of the optical-mechanical projection image collected by the surveillance camera; img _src is the original image collected by the surveillance camera, and the original image is based on the image The correction matrix H completes the image correction; (x, y, 1) is the homogeneous coordinate point of the image in the original image; (x', y', 1) is the point in the corrected image corresponding to the original image, h ₁ ~ h ₈ are the eight independent elements of the correction matrix H; W _truth and H _truth are the maximum pixel values in the width and height directions of the true image respectively; x∈[0, w], y∈[0, h], w and h are respectively The pixel coordinate point in the original image; i, j are the pixels of the corresponding image. 4. The method for online monitoring of paving defects in ceramic light-curing additive manufacturing process according to claim 2, characterized in that: the optimization algorithm is an improved genetic algorithm, The improved genetic algorithm uses the four corner point coordinates in S21 and the random points according to the normal distribution around the corner points as the initial population to provide a priori knowledge for the genetic algorithm; selection, crossover and mutation are performed in the initial population to Maximum efficiency approaches the optimal value of the objective function. 5. The method for online monitoring of paving defects in ceramic light-curing additive manufacturing processes according to claim 4, characterized in that: in the process of optimizing the objective function using an improved genetic algorithm, the integer part of the correction point coordinates is combined with The decimal part is encoded separately, and the decimal part is given a greater mutation probability to make the genetic algorithm's search ability stronger, and the integer part is given a smaller mutation probability to make the evolution process stable; only the decimal part is crossed during the crossover process, and the integer part is Some parts do not intersect to reduce computational complexity; Before each generation evolves, the fitness of each individual is sorted according to the objective function. The individual with the highest fitness directly enters the next generation. The purpose of the improved genetic algorithm is to find more accurate correction mark points between the integer pixels of the image. 6. The method for online monitoring of paving defects in ceramic light-curing additive manufacturing process according to claim 2, characterized in that step S3 specifically includes the following content: S21. Collect side view paving images of the k-Nth layer and the kth layer respectively; S22. Calibrate the side-view paving image to obtain the front-view paving image; S23. Use the interpolation algorithm to fill in the blank pixels in the front-view paving image; S24. Crop the front view paving image to the same pixel as the sliced image; S25. Place the front-view paving image of the k-th layer after the above processing in the first channel, the front-view paving image of the k-Nth layer after processing in the second channel, and place the k-Nth layer slice image in the third channel, so as to Synthesize three-channel color images; S26. Divide the three-channel color image into blocks, input the divided images into the defect detection model for detection, and output the identified defective target objects. 7. The method for online monitoring of paving defects in ceramic light-curing additive manufacturing process according to claim 1, characterized in that: the specific process of using a monitoring camera to collect paving images is: Install the surveillance camera side-axis diagonally above the manufacturing plane, and capture side-view paving images at each moment through the surveillance camera to achieve image collection; When obtaining the image correction matrix, a detachable narrow-band light-transmitting sheet needs to be installed at the front end of the lens of the surveillance camera to obtain the corresponding image. 8. The method for online monitoring of paving defects in ceramic light-curing additive manufacturing process according to claim 6, characterized in that: the narrow-band light-transmitting sheet is a 405nm narrow-band light-transmitting sheet, and only light of 405n wavelength can pass through it. Narrowband light-transmitting sheet.