A method for predicting and analyzing the causes of automotive paint bubble defects

Abstract

A method for predicting automobile paint bubble defects uses computer vision methods and deep neural network methods based on the shape of the gaps at the connection of the vehicle body steel plates to predict the occurrence of bubble defects. The method includes: setting a binary classification model based on gap contour data to obtain a classifier that can be used to predict whether bubbles exist; detecting the gap shape at the connection of the vehicle body steel plates to be painted and converting it into an image to obtain specific information about the gap; and using the classifier to predict the situation that causes the bubble defect based on the specific gap information. The present invention also discloses a method for analyzing the causes of automobile paint bubble defects. The present invention can effectively avoid the interference of human factors, improve the accuracy of bubble defect prediction in automobile painting production, and find the main causes of defects. It not only reduces production and maintenance time and labor costs, but also helps to improve product qualification rate and production efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobile coating, and relates to coating defect recognition and analysis, in particular to a method for recognition and analysis of coating bubble defects based on computer vision. Background Art

[0002] Painting is an important link in automobile production. For example, the work of the PVC sealing production line in the paint shop is to first fill the PVC glue into the gaps between the car body panels, and then dry it to harden the PVC glue to achieve the purpose of corrosion protection and stone impact resistance in the gaps. PVC glue is in a gel-like state at low temperatures, and its hardness gradually increases with increasing temperature. Due to insufficient filling of PVC glue at the folded edges of the car body panels, a certain amount of gaps remain inside after sealing. During the drying process, as the temperature of the car body rises rapidly, the gas in the gap expands rapidly, pushing the PVC glue up into bubbles, which are called glue bubbles, and are a type of production quality problem. If the above-mentioned production quality problem occurs, on the one hand, it will increase maintenance time and affect production efficiency; on the other hand, the maintenance process requires manual operation, which increases labor costs.

[0003] The existing solution to the above problem is to optimize the amount of hemming glue used to control the generation of gaps, thereby reducing the number of glue bubbles. On the other hand, this solution is difficult to implement and has little effect, mainly due to the following reasons:

[0004] 1. The gaps between the body panels are extremely small, and it is impossible to use equipment to penetrate the gaps for detection. The only way to predict whether there are defects is based on the experience of the specific operator, which has low accuracy;

[0005] 2. The location of bubble defects is uncertain and difficult to predict;

[0006] 3. The formation of bubble defects is mainly caused by gaps between the boards. However, due to the diverse shapes of the gaps, it is difficult to identify and find a solution to the problem manually. Summary of the Invention

[0007] In view of the above-mentioned shortcomings of the prior art, the purpose of the present invention is to provide a method for predicting automobile paint bubble defects, which is not affected by the specific operator's experience and improves accuracy.

[0008] Another object of the present invention is to provide a method for analyzing the causes of automobile coating foam defects.

[0009] The technical solution adopted by the present invention is: using computer vision method to predict automobile coating bubble defects and analyze the causes of the defects.

[0010] Furthermore, the present invention uses computer vision methods and deep neural network methods based on the shape of the gaps at the connection points of the vehicle body steel plates to predict the occurrence of blister defects and analyze the causes of the defects.

[0011] The computer vision method refers to the use of cameras, detectors and other technologies to replace the human eye to grasp, locate and measure objects in the physical world, and then based on artificial intelligence technology, imitate the human brain to recognize and judge the acquired images.

[0012] The following steps are involved:

[0013] Step 1: Set up a binary classification model based on the gap profile data to obtain a classifier that can be used to predict whether there are bubbles;

[0014] Step 2: Detect the shape of the gap at the connection of the car body steel panels to be painted and convert it into an image to obtain specific information about the gap;

[0015] Step three: Based on the specific information of the gap, the classifier is used to predict the situation that causes the bubble defect.

[0016] Specifically:

[0017] Step 1: Divide the glued area based on the gap angle. First, find and record the glued area with gap angle. Then, split the area with angle multiple times until all areas involving gaps have no angle.

[0018] Step 2: Based on the gap profile detection method of the plate density, first determine the different plate densities in different areas; then, use the flaw detection instrument to use the double-sided detection method to calculate the double-sided plate thickness and total thickness at a certain point of the plate, and thus calculate the gap width; finally, continuously adjust the angle and position of the detector until all sample points are sampled.

[0019] Step 3: A method for predicting the quality of foam defects based on gap contours. First, a method for abstracting contour data is set to obtain a digital expression of the contour. Then, the gap contour data is converted into RGB image format and combined with actual production quality information to generate a data set. Finally, a binary classification model based on gap contour data is set to obtain a classifier that can be used to predict whether foam exists.

[0020] When the generation of a bubble defect is predicted according to the above method, a method for analyzing the cause of the automobile coating bubble defect further includes:

[0021] Step 4: Root cause analysis of blister defects based on the blister defect prediction model. First, obtain the weights of the feature layers in the model of step 3, then calculate the heat map, and finally find the main causes of blister defects from the heat map.

[0022] In the third step of the method for predicting the quality of blister defects based on gap contours, a method for abstracting contour data is first set to obtain a digital representation of the contour; the gap contour data is then converted into an RGB image format and combined with actual production quality information to generate a data set; finally, a binary classification model based on the gap contour data is set to obtain a classifier that can be used to predict whether there are bubbles. The specific steps are as follows:

[0023] 3.1) Gap profile data abstraction method: First, a two-dimensional coordinate system is constructed for the position to be detected in step 2 to determine the coordinates of each point; then, the three values ​​detected in step 2 are matched with the coordinate points to generate historical data.

[0024] 3.2) Gap contour data conversion and calibration: First, the multidimensional data corresponding to each coordinate is preprocessed; then the three columns of data for each coordinate point are mapped to RGB respectively to generate the corresponding image; finally, a labeled dataset is generated corresponding to the actual production quality.

[0025] 3.3) Model setting: First, the algorithm task is set as a binary classification task based on gap contour data and drying temperature. A classifier that can be used to predict the quality of the rubber bubble should be obtained. The predicted rubber bubble defect quality can be obtained based on the input gap contour, that is, whether there is a rubber bubble; then the training data is determined to be the gap contour RGB image, the label is whether there is a quality problem, and the model selects the convolutional neural network; finally, the corresponding CNN parameter calculation method is applied to obtain the classifier.

[0026] The fourth step is to analyze the root causes of the foam defects based on the foam defect prediction model. First, the weights of the feature layers in the model of step three are obtained, then a heat map is calculated, and finally the main causes of the foam defects are found from the heat map. The specific steps are as follows:

[0027] 4.1) Select the CNN model from step 3 and apply the Grad-CAM algorithm to generate a heatmap of the gap contours, identifying the correlation between the cause of the bubble defect and the gap contours. CNN feature extraction is a good choice because it preserves spatial features. Based on the prediction results from step 3, the final layer of the convolutional layer is obtained, represented as feature layer A.

[0028] 4.2) Calculate the weight of each channel of feature layer A.

[0029]

[0030] in Indicates whether there is a bubble quality problem in the classification problem, the weight of the problem item in the kth channel; y c The score of the problematic item in the final prediction result is represented by Represents the data of (i, j) in the kth channel in the feature layer A; Z represents the number of all pixels in the channel.

[0031] 4.3) Calculate L according to the following formula to obtain the thermal diagram.

[0032]

[0033] in A heat map showing problematic items; k Represents the value of feature layer A in the kth channel.

[0034] 4.4) From the heat map, we can see that the hottest area is the key part for the computer to predict whether the area will appear. The hottest area is the key factor that causes the bubble defect.

[0035] in:

[0036] 1. Improve the gap detection method. For locations that cannot be directly detected manually, the corresponding flaw detector can be used to measure the defect shape and convert it into an image to obtain specific information about the gap that causes the foam defect;

[0037] 2. Optimize the ability to identify complex gap information. For complex gap types, the application of artificial intelligence-based classification algorithms can effectively identify the characteristics of each category and predict whether bubble defects will occur;

[0038] 3. Identify the key factors causing defects based on the gap contours. When performing prediction tasks, we can analyze the last convolutional layer of the convolutional neural network to obtain the weights of different pixels affecting the classification, and then identify the main causes of the defects.

[0039] Due to the adoption of the above technical solution, the present invention can effectively avoid the interference of human factors and improve the accuracy of the prediction of bubble defects in automobile painting production; it can further find out the main causes of the defects and provide decision support for production management quality optimization, which not only reduces production maintenance time and labor costs, but also helps to improve product qualification rate and production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The figure is a flow chart of a method for predicting and analyzing defects of PVC foam in automobile coatings based on gap profiles in an embodiment of the present invention.

[0041] Figure 2 Schematic diagram of component area division in an embodiment of the present invention.

[0042] Figure 3 Schematic diagram of the gap profile and gap width in the component in the embodiment of the present invention.

[0043] Figure 4Schematic diagram of data acquisition and detection angles in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] This paper proposes a method for predicting and analyzing the causes of automotive paint bubble defects. Based on the concept of computer vision and the application of deep neural network technology, the method predicts the occurrence of bubble defects based on the shape of the gap at the connection between the car body steel plates, while keeping other influencing factors unchanged, and analyzes the main causes of the defects.

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] like Figure 1 As shown, the flowchart of the method for predicting and analyzing the causes of automobile coating PVC foam defects based on gap contours of the present invention includes: PVC coating area division based on gap angles; gap contour detection method based on board density; foam quality prediction method based on gap contours; foam defect root cause analysis based on foam defect prediction model.

[0047] A method for predicting and analyzing defects in PVC foam for automotive coatings based on gap profiles is proposed. The specific implementation steps are as follows:

[0048] Step 1: Divide the PVC glue coating area based on the gap angle. The location, shape and size of the gap are different in different car areas. Divide different analysis areas according to the shape of the gap location. If there is a gap angle at a certain location, it is necessary to divide the two sides into different areas based on the angle. Figure 2 As shown, the position of a component with a folded angle divides the component into area 1 and area 2. It should be noted that the "folded angle" can be any angle except a straight angle, and is not limited to Figure 2 The right angle shown in ; the boundary between "Area 1" and "Area 2" is whether this shape can be differentiated, and the corners are not differentiable.

[0049] Step 2: Crack profile detection based on sheet material density. This step involves obtaining information about the crack profile. As shown in the figure, density values ​​for different areas and both sides of the crack must first be determined. These values ​​can be used by the flaw detector to calculate the thickness of the sheet material on both sides. Next, considering the extremely small crack, it is impossible to penetrate the crack location to collect information. Therefore, a double-sided inspection method is used. Given the large number of samples, high frequency, and limited time, an automated detector is used to collect relevant data. The detection angle is set perpendicular to the tangent of each detection point. The obtained double-sided sheet material thickness is then used to calculate the crack width using the following formula.

[0050] Gap width = overall width – width on one side – width on the other side

[0051] like Figure 3As shown, the detection angle is perpendicular to the tangent of each detection point for collection.

[0052] Finally, according to the actual situation of the plate outside the gap, the list is traversed and the detection is performed uniformly with a step size smaller than the standard size of the foam quality. For example, the detection is performed uniformly with a step size of one sixteenth of the standard size of the foam quality until all points are detected. Figure 4 As shown, there is a schematic diagram of the gap interface data acquisition content, which includes a schematic diagram of the component cross section and the gap measurement range.

[0053] Step 3: A method for predicting bubble defects based on gap contours. First, a method for abstracting contour data is set to obtain a digital representation of the contour. Then, the gap contour data is converted into RGB images and combined with actual production quality information to generate a dataset. Finally, a binary classification model based on the gap contour data is set to obtain a classifier that can be used to predict the presence of bubbles. The specific steps are as follows:

[0054] 3.1) Gap profile data abstraction method: First, all the detection points in step 2 are determined as coordinate points. A two-dimensional coordinate system is constructed based on the region. The starting detection point is determined as the origin, and all coordinate points are detected. Then, the detected data are matched with the detection coordinates.

[0055] 3.2) Gap profile data conversion and calibration: First, the historical data is standardized. Based on the thickness of the double-sided board and the gap thickness at each coordinate point, three values ​​are determined. These three values ​​are normalized to a value between 0 and 255. Using the RGB color generation method, the color of each coordinate point is generated. All coordinate points are integrated to generate a two-dimensional color image of the area. The normalization formula is as follows:

[0056]

[0057] 3.3) Model Setup: For the CNN model, choose VGG16. Use the data from step 3, point 2, containing temperature and gap contours to train the model parameters. The model has 16 layers, including 13 convolutional layers and 3 fully connected layers. The input image size is 224*224*3, and the activation function is ReLU. The convolution kernel is 3*3 and the pooling kernel is 2*2. The training and test sets can be divided into 8-2 categories. The training result is a classifier that can predict production quality results based on gap contours.

[0058] Step 4: Root cause analysis of blister defects based on the blister defect prediction model. First, obtain the weights of the feature layers in the model in step 3, then calculate the heat map, and finally find the main causes of blister defects from the heat map. The specific steps are as follows:

[0059] 4.1) Select the CNN model from step 3 and apply the Gradient-CAM algorithm to generate a heat map of the gap contour. This algorithm is used because CNN feature extraction preserves spatial features. Based on the prediction results from step 3, the final layer of data from the convolutional layer is obtained, represented as feature layer A.

[0060] 4.2) Calculate the weight of each channel of feature layer A.

[0061]

[0062] in Indicates whether there is a bubble quality problem in the classification problem, the weight of the problem item in the kth channel; y c It represents the score of the problematic item in the final prediction result; Represents the data of (i, j) in the kth channel in the feature layer A; Z represents the number of all pixels in the channel.

[0063] 4.3) Calculate L according to the following formula to obtain the thermal diagram.

[0064]

[0065] in A heat map showing problematic items; k Represents the value of feature layer A in the kth channel.

[0066] 4.4) The heat map shows that the hottest area is the key location for the computer to predict whether the area will appear. This hottest area is the key factor causing the blister defect. Furthermore, the key defect location is identified and described using a 3D color image generated by combining the 2D coordinates and the crack outline. This allows the crack shape at the key location causing the blister defect to be determined.

[0067] In summary, the present invention proposes a method for predicting and analyzing the quality of PVC foam in automotive coatings. First, the vehicle's various areas are divided into regions based on whether the gaps have corners. Each region is then gridded and the coordinates of each point within the region are determined. A double-sided inspection method is then applied using an ultrasonic flaw detector to determine the gap profile. Three sets of data are then generated based on the thickness of the panels on either side of the gap and the gap width. The gap profiles are then converted into RGB images and compared with the actual quality inspection results to form a dataset. Based on this dataset, a convolutional neural network is constructed and trained to generate a gap image classifier, which is then used to predict the presence of foam defects. Finally, based on the prediction results, the classification basis of the prediction model is analyzed to identify the key factors that lead to foam defects.

[0068] The above descriptions and embodiments are provided to facilitate understanding and application of the present invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to these contents and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above descriptions and embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for predicting automobile paint bubble defects, characterized by: Based on the shape of the gap at the connection between the body steel plates, computer vision methods and deep neural network methods are used to predict the occurrence of blister defects. The following steps are included: Step 1: Setting a binary classification model based on the gap contour data to obtain a classifier that can be used to predict whether there are glue bubbles; specifically, the following steps are included: (1) dividing the glue coating area based on the gap angle; (2) a gap contour detection method based on the density of the board material; (3) obtaining the classifier based on the area division and gap contour detection; Step 2: Detect the shape of the gap at the connection of the car body steel panels to be painted and convert it into an image to obtain specific information about the gap; Step three: Based on the specific information of the gap, the classifier is used to predict the situation that causes the bubble defect.

2. The method for predicting automobile paint bubble defects according to claim 1, characterized in that: In the step (1), first, the glued areas with gap corners are found and recorded; then, the areas with gap corners are split multiple times until all areas involving gaps have no corners.

3. The method for predicting automobile paint bubble defects according to claim 2, characterized in that: The angle is an angle other than a straight angle, and the angle divides the two sides into different areas. The angle at the boundary between different areas is not differentiable.

4. The method for predicting automobile paint bubble defects according to claim 1, characterized in that: In the step (2): first, different plate densities in different areas are determined; then, the thickness of the double-sided plate at a certain point and the total thickness of the plate are calculated, thereby calculating the gap width; finally, the angle and position of the detector are continuously adjusted until all sample points are sampled.

5. The method for predicting automobile paint bubble defects according to claim 4, characterized in that: The double-sided plate thickness at a certain point and the total thickness of the plate are calculated by using a flaw detection instrument using a double-sided detection method.

6. The method for predicting automobile paint bubble defects according to claim 1, characterized in that: In the step (3): first, a method for abstracting the contour data is set to obtain a digital expression of the contour; then, the gap contour data is converted into an RGB image format and combined with actual production quality information to generate a data set; finally, a binary classification model based on the gap contour data is set to obtain a classifier that can be used to predict whether there are bubbles.

7. The method for predicting automobile paint bubble defects according to claim 6, characterized in that: The following operations are included: (3-1) Gap profile data abstraction method: First, a two-dimensional coordinate system is constructed for the location to be detected, and the coordinates of each point are determined; then, the three detected values ​​are matched with the coordinate points to generate historical data; (3-2) Gap profile data conversion and calibration: First, pre-process the multidimensional data corresponding to each coordinate; then, correspond the three columns of data of each coordinate point to RGB respectively to generate the corresponding image; finally, correspond to the actual production quality to generate a labeled data set; (3-3) Model setting: First, the algorithm task is set as a binary classification task based on gap contour data and drying temperature. A classifier that can be used to predict the quality of the rubber bubble should be obtained. The predicted rubber bubble defect quality can be obtained based on the input gap contour, that is, whether there is a rubber bubble; then the training data is determined to be the gap contour RGB image, the label is whether there is a quality problem, and the model selects the convolutional neural network; finally, the corresponding CNN parameter calculation method is applied to obtain the classifier.

8. The method for predicting automobile paint bubble defects according to any one of claims 1 to 7, characterized in that: The method further includes calculating a thermal map and determining the key position causing the bubble defect from the thermal map.

9. The method for predicting automobile paint bubble defects according to claim 8, characterized in that: The main causes of bubble defects are found from the heat map by obtaining the weights of the feature layers in the model.

10. The method for predicting automobile paint bubble defects according to claim 9, characterized in that: The following steps are involved: (a) Select the CNN model and apply the Grad-CAM algorithm to generate a heat map of the gap contour; the last layer of convolutional layer data is obtained, which is represented as feature layer A; (b) Calculate the weight of each channel of feature layer A: in Indicates whether there is a bubble quality problem in the classification problem, the weight of the problem item in the kth channel; y C The score of the problematic item in the final prediction result is represented by Indicates the data of (i, j) in the kth channel in the feature layer A; Z represents the number of all pixels in the channel; (c) Calculate L according to the following formula to obtain the thermal diagram; in A heat map showing problematic items; K Represents the value of feature layer A in the kth channel; (d) From the thermal diagram, it can be seen that the hottest part is the key factor leading to bubble defects.

Images