Intelligent detection method for strain gauge sticking quality

Abstract

The present application relates to a kind of intelligent detection methods of strain gauge sticking quality, the method is analyzed to strain gauge sticking after the specimen picture by picture processing technology, to judge its sticking quality.Specifically includes the following parts: P1) strain picture cutting, P2) regional feature extraction, P3) data and result.The beneficial effects of the present application are that the detection result can be obtained by inputting the sticking picture of strain gauge, which can be used as an auxiliary means for actual engineering detection.In addition, individualized features are set for different regions, improving the accuracy of detection while saving the computing power requirements of learning algorithm.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of engineering structure testing, and specifically designs an automatic and intelligent detection method for the quality of strain gauge pasting. BACKGROUND

[0002] The resistance strain gauge has the characteristics of simple structure, stable and reliable performance, high sensitivity and wide frequency range, and is widely used in the field of engineering structure testing.

[0003] On the other hand, as a contact measurement method, the strain gauge needs to be pasted to the surface of the object before measurement, and the quality of pasting often leads to the decline of its reliability and precision, so it is an important problem to ensure the quality of strain gauge pasting.

[0004] In actual engineering, the quality detection is mainly carried out by manual methods such as visual inspection and electrical measurement, which is very low in efficiency, and an intelligent automatic detection method needs to be developed to solve the above problems. SUMMARY

[0005] In view of the above technical problems, the present application constructs an intelligent detection process for judging the quality of strain gauge pasting by using picture processing means to assist the detection of strain gauge pasting quality.

[0006] Technical scheme:

[0007] In order to achieve the above application purpose, the intelligent detection method for the quality of strain gauge pasting can adopt the following technical scheme:

[0008] An intelligent detection method for the quality of strain gauge pasting, comprising the following steps:

[0009] S1) Establishing a sample database of strain gauge pasted specimen pictures, collecting a certain amount of strain gauge pasted specimen graphic data to form a database, and each picture sample is a complete strain gauge pasted specimen, which at least contains the above elements of pasted paper, welding point, strain grid and prestressed ring;

[0010] S2) According to the ASTM standard, the quality of strain gauge pasting in the sample database is artificially judged and recorded;

[0011] S3) dividing the test piece picture into different regions, and manually marking the two levels of region segmentation of the pictures in the sample database one by one, and training the machine learning model; the first level of picture segmentation of the pictures in the sample database is manually marked by using a rectangular frame, the first level of picture segmentation specifically refers to the soldering piece region, the strain gage region and the pre-stress ring region, a machine learning model suitable for picture segmentation is used, denoted as model I, and the model parameters of the first level of picture segmentation image region are obtained by using the sample database and the marking information for parameter training; the second level of picture segmentation of the pictures in the sample database is manually marked, the region segmented by the second level of picture segmentation refers to the soldering point region, a machine learning model suitable for picture segmentation is used, and the model parameters of the second level of picture segmentation image region are obtained by using the sample database and the marking information for parameter training;

[0012] S4) calculating the individualized features of the divided regions corresponding to the pictures in the sample database in sequence and training the classification model; the image is preprocessed by using the image preprocessing module M1, the features of the image in the sample are calculated by using the individualized feature module M2, and the image features and the marking information are trained for parameter training by using a machine learning model suitable for classification, so as to obtain the classification model parameters;

[0013] S5) using the machine learning model trained in step S3 to segment the regions of the given target picture, and returning the pictures of different regions after segmentation: the picture segmentation is divided into two levels: after the first level of picture segmentation, the second level of segmentation is further performed;

[0014] S6) using a first order differential operator and similarity analysis to judge the continuity of the strain gage region in the segmented image obtained in step S5, if the similarity condition is not met, it is determined that the pasting quality is unqualified, otherwise, it goes to step S7;

[0015] S7) using the image preprocessing module M1 and the individualized feature module M2 to process the segmented image obtained in step S5, obtaining the individualized features of the target picture TP, and using the classification model in step S4 and the model parameters thereof to determine the strain gage pasting quality in the target picture TP.

[0016] Step S2 first manually performs visual inspection according to the ASTM standard, and then confirms the overall working performance by instrument measurement, records the picture samples with normal performance as 1, and records the picture samples with unqualified performance as 0.

[0017] Step S3 specifically includes the following sub-steps:

[0018] S3.1) Artificial marking of the first level of image segmentation of the pictures in the sample database using a rectangular frame as the region shape, to obtain first marking information; the regions segmented by the first level of image segmentation specifically refer to the solder patch region, the strain grid region, and the pre-stress ring region, and the form of the record of this level of marking is the midpoint of the rectangle and the length and width of the rectangle;

[0019] S3.2) Parameter training using the sample database in step S1 and the first marking information in step S3.1 using a machine learning model suitable for image segmentation, denoted as model I, to obtain the model parameters of the image regions segmented by the first level of image segmentation, denoted as model parameters Ia;

[0020] S3.3) Artificial marking of the second level of image segmentation of the pictures in the sample database, to obtain second marking information, the regions segmented by the second level of image segmentation refer to the solder joint region, and the form of the record of this level of marking is the midpoint of the rectangle and the length and width of the rectangle;

[0021] S3.4) Parameter training using model I, the sample database in step S1, and the second marking information in step S3.3, to obtain the model parameters of the image regions segmented by the second level of image segmentation, denoted as model parameters Ib.

[0022] Step S4) specifically includes the following sub-steps:

[0023] S4.1) Perform subtraction operation using the strain grid region and the strain grid region of the adhesive bottom tape artificially marked in step S3.1 to obtain the strain grid region of the adhesive bottom;

[0024] S4.2) Perform picture preprocessing operation on the solder patch region, the strain grid region of the adhesive bottom obtained in step S4.1, one by one, using the image preprocessing module M1, and the picture preprocessing operation includes A) resampling with a sampling rate of 10; B) average sliding filter processing;

[0025] S4.3) Perform feature extraction on the solder patch region marked in step S3.1, the pre-stress ring region, the solder joint region marked in step S3.3, and the strain grid region of the adhesive bottom obtained in step S4.1 using the individualized feature module M2, and return the features;

[0026] S4.4) Train using the features extracted in step S4.3 and the artificial marking in step S3 using a classification model, denoted as model II, to obtain the corresponding classification parameters, denoted as model parameters II.

[0027] In step S4.1 above, the strain grid region of the adhesive bottom refers to a rectangle formed by three times the width of the strain grid and five-sixths the length of the strain grid.

[0028] Step S4.3 feature extraction is specifically: for the different regions segmented or labeled in the picture, different personalized features are calculated, wherein:

[0029] A) For the solder region, the following 14 features are calculated: image color moment features, 1, 2, and 3 low-order moments are extracted for each color channel, a total of 9 component features; 4 solder position and inertia moment of the surrounding region, a total of 5 features;

[0030] B) For the solder region, the following 48 features are calculated: 0, 45, 90, 135 direction second moment, contrast, correlation and entropy, a total of 16 features, four solder points a total of 48 features;

[0031] C) For the pre-stress ring region, after threshold processing and morphological feature extraction, the following 7 features are calculated: the Euler distance of the terminal point, a total of two component features; the included angle between the two ideal leads, a total of 1 component feature; the square sum of the offset of the lead path from the ideal lead, a total of 2 component features; the included angle between the ideal lead and the vertical edge, a total of 2 component features;

[0032] D) For the glue bottom strain gauge region, the following 10 features are calculated: the proportion of the lightest color appearing after color quartering, a total of 1 component feature; the first, second and third moments of the picture, a total of 9 component features.

[0033] Step S5) includes the following sub-steps:

[0034] 5.1) For the given target picture, use model I and model parameter Ia in step S3 to perform target recognition and first-level picture segmentation, obtaining the solder region, strain gauge region, pre-stress ring region, and glue bottom strain gauge region;

[0035] 5.2) Use the strain gauge region and glue bottom strain gauge region segmented in step S5.1 to perform reduction operation to obtain the glue bottom strain gauge region;

[0036] 5.3) For the solder region segmented in step S5.1, use model I and model parameter Ib in step S3 to perform target recognition and second-level picture segmentation to obtain the solder point region. It is assumed that the solder points must be located on the sticker, and since there are at most four solder points on the sticker, only the four detection boxes with the highest confidence are taken, and the relative coordinates of the picture and the center positions of the four solder points in the solder region are returned.

[0037] The process of step S6 is to check whether the strain grid has defects, to process the strain grid area by using a first-order differential operator, to determine whether there is an obvious discontinuity, if there is, it is considered that the target paste quality does not meet the requirements; and the right 12 / 13 of the grid image is intercepted, and after threshold processing, the grid image is reset to a matrix containing (0, 1), and compared with the standard case, if the different pixel points are greater than 10% of the overall pixel points, it is considered that the target paste quality does not meet the requirements; if the above two items are satisfied, step S7 is continued.

[0038] The specific steps of step S7 are:

[0039] S7.1) using the image preprocessing module M1 to preprocess the given middle solder piece area and the glue bottom area of the strain grid area;

[0040] S7.2) using the individualized feature module M2 to extract features from the solder piece area obtained in step S5.1), the pre-stressed ring area, the glue bottom area of the strain grid area obtained in step 5.2), and the solder joint area obtained in step S5.3),

[0041] S7.3) using model II, model parameters para-II and the features extracted in S7.2) to classify, and returning result 0, i.e. unqualified or result 1, i.e. qualified.

[0042] In step S7.3, when the picture result to be detected is in the middle region of the two support vectors, in order to meet the engineering application, for this kind of result, it is judged as 0, i.e. unqualified.

[0043] Beneficial effects:

[0044] The application applies intelligent technology, and only inputs the paste picture of the strain piece, and the result can be output, which can be used as an auxiliary means for detection in actual engineering. Secondly, the application sets individualized features for different areas of the strain piece picture, significantly physical meaning, and saves the algorithm power requirement of learning algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a framework diagram of the strain piece paste quality intelligent detection method;

[0046] Figure 2 is a picture of a sample containing a complete strain piece after pasting;

[0047] Figure 3 is a sample picture with a first level of segmentation label, wherein the red frame marks the solder piece area, the pre-stressed ring area, and the strain grid area from left to right;

[0048] Figure 4is a sample picture with second level segmentation and labeling, in which 4 red boxes mark 4 solder joints respectively;

[0049] Figure 5 The middle red boxes mark the strain gate region and the strain gate region with glue bottom respectively;

[0050] Figure 6 is the picture effect of the solder region before and after resampling;

[0051] Figure 7 is the solder region;

[0052] Figure 8 is the color threshold processing for the pre-stress ring region. DETAILED DESCRIPTION

[0053] The technical solutions of the present application will be further described in detail below in combination with the drawings of the specification and specific embodiments.

[0054] An individualized feature extraction means for different physical meanings of picture regions, which adopts a picture preprocessing image preprocessing module M1 and an individualized feature module M2;

[0055] M1: Preprocess the picture, including A) resampling, set the sampling rate to 10; B) average sliding filter processing

[0056] M2: Calculate different individualized features for different regions after picture segmentation or labeling: A) For the solder region, calculate the following 14 features: color moment features, extract 1, 2, and 3 low-order moments for each color channel, a total of 9 components; 4 solder joint positions and the inertia moments of the surrounding regions, a total of 5 features; B) For the solder joint region, calculate the following 48 features: 0, 45, 90, and 135 direction second-order moments, contrast, correlation, and entropy, a total of 16 features, and four solder joints have a total of 48 features; C) For the pre-stress ring region, after threshold processing and morphological feature extraction of the specific pre-stress ring path, calculate the following 7 features: Euler distance of the terminal point, a total of two components; the included angle between the two ideal leads, a total of 1 component; the square sum of the offset of the lead path from the ideal lead, a total of 2 components; the included angle between the ideal lead and the vertical edge, a total of 2 components; D) For the glue bottom strain gate region, calculate the following 10 features: the proportion of the lightest color after color quartering, a total of 1 component; picture first, second, and third moments, a total of 9 components.

[0057] The intelligent detection process includes the following steps:

[0058] S1) Establish a sample database of pictures of test pieces after strain gauge pasting: collect a certain amount of sample database of test piece pictures after strain gauge pasting, generally more than 100 samples, to form a database. A single picture sample should be a complete test piece after strain gauge pasting, including pasted paper, welding points, strain grids, and pre-stress rings.

[0059] S2) According to the ASTM standard, manually judge the strain gauge pasting quality in the sample database and record it.

[0060] S3) Divide the test piece pictures into different areas, manually mark the areas in the sample database one by one at two levels, and train the machine learning model: manually mark the first level of picture segmentation in the sample database using a rectangular frame, and the areas segmented by the first level of picture segmentation are specifically the welding piece area, the strain grid area, and the pre-stress ring area. Use a machine learning model suitable for picture segmentation (denoted as model I), and use the sample database and the marking information to train the parameters to obtain the model parameters of the first level of picture segmentation image area; manually mark the second level of picture segmentation in the sample database, and the areas segmented by the second level of picture segmentation are specifically the welding point area. Use a machine learning model suitable for picture segmentation, and use the sample database and the marking information to train the parameters to obtain the model parameters of the second level of picture segmentation image area.

[0061] S4) Calculate the individualized features of the divided areas corresponding to the pictures in the sample database and train the classification model: use image preprocessing modules M1 and M2 to preprocess and calculate the features of the pictures in the sample, and use a machine learning model suitable for classification to train the parameters of the picture features and the marking information to obtain the classification model parameters.

[0062] S5) Use the machine learning model trained in step S3 to segment the areas of the given target picture and return the segmented pictures of different areas: picture segmentation is divided into two levels: after the first level of picture segmentation, further second level of segmentation is performed.

[0063] S6) Use a first-order differential operator and similarity analysis to judge the continuity of the strain grid area in the segmented image obtained in step S5, and if the similarity condition is not met, it is determined that the pasting quality is unqualified, otherwise go to step S7.

[0064] S7) Use M1 module and M2 module to process the segmented image obtained in step S5 to obtain the individualized features of the target picture TP, and use the classification model and its model parameters in step S4 to determine the strain gauge pasting quality in the TP picture.

[0065] The application discloses an intelligent detection method for strain gauge sticking quality.

[0066] S1) Prepare and collect a certain number of strain gauge sticking pictures of test pieces to form a sample database, denoted as SDB.

[0067] S2) Manually distinguish the strain gauge sticking quality in the sample database SDB, and record the judgment results one by one.

[0068] S3) Manually mark the two-level region segmentation of the pictures in the sample database SDB one by one, train a machine learning model to generate the two-level model parameters, denoted as Para-Ia and Para-Ib.

[0069] S4) Input the segmented region pictures in the sample database SDB into an image preprocessing module M1 one by one, calculate the preferred features of the strain gauge sticking quality, train a classification model using the above features, and generate model parameters, denoted as Para-II.

[0070] S5) Input a target picture to be distinguished, denoted as TP, perform region segmentation on the target picture using the model and the model parameters in step S3, and obtain segmented images of different regions.

[0071] S6) Judge the continuity of the strain gauge region in the segmented image obtained in step S5 using a first-order differential operator, and the judgment process is to check whether the strain gauge is flawed. The strain gauge is generally uniformly manufactured in a factory, and the quality is guaranteed. The common flaws in a laboratory mainly include a fracture at the strain gauge. In order to search for the defect, the strain gauge region is processed using the first-order differential operator to judge whether there is an obvious discontinuity. If there is, it is considered that the target sticking quality does not meet the requirements. The right 12 / 13 of the strain gauge picture is intercepted, threshold processing is performed, the strain gauge picture is reset to a matrix containing (0, 1), and the strain gauge picture is compared with a standard case. If the different pixel points are greater than 10% of the overall pixel points, it is considered that the target sticking quality does not meet the requirements. If the above two items are met, step S7 is performed.

[0072] If the sticking quality is not qualified, the process proceeds to step S7.

[0073] S7) Process the segmented image obtained in step S5 using the M1 module and the M2 module to obtain the individualized features of the target picture TP, and judge the strain gauge sticking quality in the TP picture using the classification model and the model parameters in step S4.

[0074] The specific steps of step S1) are as follows:

[0075] S1.1) Collect a certain amount of strain gauge after the sample data sample database, generally more than 100 samples, constitute a database. When collecting, the collection equipment should be as parallel as possible to the paste plane of the strain gauge, and the picture edge should be as parallel as possible to the sample edge.

[0076] S1.2) In order to determine the quality of the model, a complete strain gauge after the sample should be included in a single picture sample, including the elements such as the paste after the tape, the welding point, the strain grid and the pre-stress ring. As shown in Figure 1 , and keep a certain degree of clarity.

[0077] In step S2), for the manual determination method of strain gauge paste quality, there is no relevant standard issued in China, which can be based on the standard ASTM E 1237-93 of American Society for Testing and Materials. First, visual inspection is carried out, and then the overall performance is confirmed through instrument measurement, and the sample with normal performance is recorded as 1, and the picture sample with unqualified performance is recorded as 0.

[0078] The specific steps of step S3 are:

[0079] S3.1) The first level of picture segmentation is manually marked by using a rectangular frame to the picture in the sample database. The area segmented by the first level of picture segmentation is specifically the welding sheet area, the strain grid area and the pre-stress ring area. As shown in Figure 2 , the form of this layer of marking record is the midpoint of the rectangle and the length and width of the rectangle.

[0080] S3.2) A machine learning model suitable for picture segmentation is used, which is recorded as model I, such as artificial neural network yolo model. The sample database in step S1 and the label information in step S3.1 are used for parameter training to obtain the model parameters of the first level of picture segmentation image area, which is recorded as model parameter Ia.

[0081] S3.3) The second level of picture segmentation is manually marked. The area segmented by the second level of picture segmentation is the welding point area, as shown in Figure 3 , the form of this layer of marking record is the midpoint of the rectangle and the length and width of the rectangle.

[0082] S3.4) Model I is used, the sample database in step S1 and the label information in step S3.3 are used for parameter training to obtain the model parameters of the second level of picture segmentation image area, which is recorded as model parameter Ib.

[0083] The specific steps of step S4 are,

[0084] S4.1) The strain grid area and the adhesive tape strain grid area marked by manual marking in step S3.1 are used for reduction operation, as shown in Figure 4The glue bottom area corresponding to the strain gauge region is obtained.

[0085] S4.2) The image preprocessing module M1 is used to perform preprocessing operations on the solder patch region and the glue bottom area corresponding to the strain gauge region in the sample one by one.

[0086] S4.3) The individualized feature module M2 is used to extract features from the solder patch region marked in step S3.1, the pre-stress ring region, the solder joint region marked in step S3.3, and the glue bottom area corresponding to the strain gauge region obtained in step S4.1, and return the features.

[0087] S4.4) The classification model, denoted as model II, is used to train the features extracted in step S4.2 with the manual marking in step S2, to obtain corresponding classification parameters, denoted as parameters II.

[0088] In the above step S4.1, the glue bottom area corresponding to the strain gauge region refers to a rectangle formed by three times the width of the strain gauge and five-sixths the length of the strain gauge.

[0089] Further, the specific steps of step S5 are as follows:

[0090] 5.1) For the given target picture, the model I and model parameters Ia in step S3 are used to perform target recognition and first-level picture segmentation, to obtain the solder patch region, the strain gauge region, the pre-stress ring region, and the glue bottom area corresponding to the strain gauge region.

[0091] 5.2) The strain gauge region and the glue bottom area corresponding to the strain gauge region segmented in S5.1 are used to perform reduction operations to obtain the glue bottom area corresponding to the strain gauge region.

[0092] 5.3) For the solder patch region segmented in S5.1, the model I and model parameters Ib in step S3 are used to perform target recognition and second-level picture segmentation to obtain the solder joint region. It is assumed that the solder joint must be located on the sticker, and since there are at most four solder joints on the sticker, only the four detection boxes with the highest confidence are taken, such as Figure 5 to identify the solder joint region and the returned confidence. The return is the picture and the relative coordinates of the center position of the four solder joints in the solder patch region.

[0093] Further, the process of step S6) is to check whether the strain grid has defects. The strain grid is generally uniformly manufactured by the factory, and its quality is guaranteed. The common defects in the laboratory are mainly grid fracture. In order to detect this defect, a first-order differential operator is used to process the strain grid area to determine whether there is an obvious discontinuity. If there is, it is considered that the target pasting quality does not meet the requirements. The right 12 / 13 of the grid picture is intercepted, and after threshold processing, the grid picture is reset to a matrix containing (0, 1). Compared with the standard case, if the different pixel points are greater than 10% of the overall pixel points, it is considered that the target pasting quality does not meet the requirements. If the above two items are met, step S7) is continued.

[0094] Further, the specific steps of step S7) are:

[0095] S7.1) using the image preprocessing module M1 to preprocess the given middle solder area and the glue bottom strain grid area.

[0096] S7.2) using the personalized feature module M2 to extract features from the solder area obtained in step S5.1, the pre-stressed ring area, the glue bottom strain grid area obtained in step S5.2, and the solder joint area obtained in step S5.3.

[0097] S7.3) using a classification algorithm to classify and calculate the features extracted in S7.2, the model II, and the model parameters para-II. The result is 0, which means unqualified, and the result is 1, which means qualified.

[0098] Further, in step S7.3, when the result of the picture to be detected is in the middle region of the two support vectors, in order to meet the engineering application, the result of this type is strictly judged as 0, that is, unqualified.

[0099] The present application comprises a general image preprocessing module M1, which preprocesses the picture. The image preprocessing module M1 is suitable for the set formed by the regions marked by the samples in the database or the regions segmented from the target picture TP to be judged. This module can optimize the picture and reduce the computational load in the later stage.

[0100] The steps of the picture preprocessing module M1 are:

[0101] I) resample the picture, set the sampling ratio to 10 to simplify the subsequent operation.

[0102] Taking the color moment of the solder area as an example, the picture effects before and after resampling are as follows Figure 6 The color moments of the pictures before and after resampling are as follows:

[0103]

[0104]

[0105] It is shown that the resampling has little effect on the features, and no important information of the picture or the specific region in the picture is lost. The purpose of the picture resampling is to reduce the computational complexity.

[0106] II) Average filtering is performed on the picture, including but not limited to the nine-pixel sliding average method. For the pre-stress ring and the strain grid region, this step is not performed.

[0107] In order to reduce the influence of the speckle in the picture, the average filtering is used to process the intercepted region. It is worth noting that, since the pre-stress ring and the strain grid region have no obvious boundary, the original and processed pictures of the region are retained.

[0108] The present application comprises a general personalized feature module M2, which optimizes the features for each region after the picture is segmented, and for regions with different physical meanings. The input of the personalized feature module M2 is the picture of each region after the segmentation of the sample database SDB or the picture of each region after the segmentation of the target picture TP to be identified by step S5, and the output is a group of vectors indicating the features of the sample or the target picture to be identified. The personalized feature module M2 comprises the following contents:

[0109] 1) For the solder pad region, the following 14 features are calculated: A) Calculate the color moment features of the picture, extract 1, 2, and 3 low-order moments for each color channel, a total of 9 components; B) Calculate the inertia moments of 4 solder joint positions and the regions surrounded by them, a total of 5 features.

[0110] For the solder pad region, color moments are used to represent the characteristics of data distribution. In statistics, the first moment represents the mean of the data, the second moment represents the variance of the data, and the third moment represents the skewness of the data distribution. In mathematical representation, there are also fourth-order and higher-order moments to represent the density of color. However, it is generally believed that the information is mainly concentrated in the low-order moments, so the first, second, and third moments are used as a feature. For a color picture, the color moments of the picture have a total of 9 components, and each color channel has 3 low-order moments.

[0111] In order to represent the feature that the solder joint region is distributed in the solder pad region, the personalized feature module M2 should input the centroid of the recognition of the 4 solder joint positions and the inertia moment as its feature at the same time as the input of the solder joint picture data.

[0112] 2) For the solder joint region, the following 48 features are calculated: 0, 45, 90, 135 direction second moment, contrast, correlation, and entropy, a total of 16 features, and a total of 48 features for four solder joints.

[0113] The gray level co-occurrence matrix can extract the comprehensive information of the welding defect image, such as the direction, local texture arrangement rule, and texture change rule. Four statistical features, including energy (second moment), contrast, correlation, and entropy, are adopted at four common angles of 0, 45, 90, and 135 degrees.

[0114] 3) For the prestressed ring area, after the threshold processing and morphological feature extraction of the specific prestressed ring path, the following seven features are calculated: A) the Euler distance of the terminal point, a total of two components; B) the included angle between the two ideal conductive lines, a total of one component; C) the square sum of the offset of the conductive line path from the ideal conductive line, a total of two components; D) the included angle between the ideal conductive line and the vertical edge, a total of two components.

[0115] Since the edge distribution of the prestressed ring is not obvious, some preprocessing is needed to reflect its true trajectory. First, color thresholding is performed, as shown in Figure 7 .

[0116] It can be seen that the edges on the left (soldering area) and right (raster area) are relatively obvious. By using a first-order differential operator, the original image is divided into three regions (matrix). And due to the large difference between the left and right backgrounds, the shape of the conductive line can be directly detected using the threshold method.

[0117] For the middle part, it can be seen that the color of the conductive line is similar to that of the background. If a traditional edge detection operator is used, it will be difficult to segment, and additional morphological assistance is required. The following rules are specified: the upper and lower conductive lines do not overlap and are not too unstable, i.e., if the conductive line is a function, its ideal state should be a straight line, but in actual work, the difference between the maximum and minimum values should not be too large. Therefore, based on edge detection, we select the shape according to the terminal point and the difference between the maximum and minimum values and the variance, both of which are as small as possible and the terminal point meets the requirements.

[0118] The following seven features are used: the Euler distance of the terminal point, the included angle between the two ideal conductive lines, the square sum of the offset of the conductive line path from the ideal conductive line, the included angle between the ideal conductive line and the vertical edge, and the like.

[0119] Additionally, if the image quality is poor and the conductive line shape cannot be obtained, the square sum of the offset of the conductive line path from the ideal conductive line cannot be calculated, and the value of this item is taken as 0.5.

[0120] 4) For the glue bottom strain gauge area, the following ten features are calculated: A) the proportion of the lightest color after four equal divisions, a total of one component; B) the first, second, and third moments of the image, a total of nine components.

[0121] In the ideal condition of picture definition, the strain gage region of the glue bottom should reflect a small bubble, but due to the lack of actual shooting definition, it often reflects the texture characteristics. Since it is believed that the strain gage region of the glue bottom has little effect on the results, the gray level co-occurrence matrix is not used here, but the strain gage region of the glue bottom is first grayed, and color is divided into two and four.

[0122] The proportion of the appearance of lighter colors, the first, second and third order moments are characterized.

[0123] Overall, for a fixed picture of a strain gauge, this module always returns a set of features.

Claims

1. A method for intelligently detecting the quality of strain gauge adhesion, characterized by, The method comprises the following steps: S1) Establishing a sample database of strain gauge pasted specimen pictures, collecting a certain amount of strain gauge pasted specimen picture data to form a database, and each picture sample being a complete strain gauge pasted specimen, at least including a pasted sticker, a welding spot, a strain grid and a pre-stress ring; S2) According to the ASTM standard, the strain gauge pasting quality in the sample database is artificially judged and recorded; S3) The specimen picture is divided into different regions, and the picture in the sample database is artificially marked for two levels of region segmentation one by one, and a machine learning model is trained; a rectangular frame is used to artificially mark the first level of picture segmentation of the picture in the sample database, the first level of picture segmentation specifically refers to the welding piece region, the strain grid region and the pre-stress ring region, a machine learning model suitable for picture segmentation is used, denoted as model I, and the sample database and the marking information are used for parameter training to obtain the model parameters of the first level of picture segmentation image region; the second level of picture segmentation of the picture in the sample database is artificially marked, the region segmented by the second level of picture segmentation is the welding spot region, a machine learning model suitable for picture segmentation is used, and the sample database and the marking information are used for parameter training to obtain the model parameters of the second level of picture segmentation image region; S4) The personalized features of the divided regions corresponding to the pictures in the sample database are calculated and the classification model is trained; an image preprocessing module M1 is used to preprocess the image, a personalized feature module M2 is used to calculate the features of the image in the sample, and a machine learning model suitable for classification is used to train the image features and the marking information to obtain the classification model parameters; S5) The machine learning model trained in step S3 is used to segment the regions of the given target picture, and the segmented pictures of different regions are returned: the picture segmentation is divided into two levels: after the first level of picture segmentation, the second level of segmentation is further performed; S6) The strain grid region in the segmented image obtained in step S5 is subjected to continuity judgment by using a first-order differential operator and similarity analysis, if the similarity condition is not met, it is determined that the pasting quality is unqualified, otherwise, it goes to step S7; S7) The segmented image obtained in step S5 is processed by using the image preprocessing module M1 and the personalized feature module M2 to obtain the personalized features of the target picture TP, and the strain gauge pasting quality in the target picture TP is determined by using the classification model in step S4 and the model parameters thereof.

2. The method of claim 1, wherein the method comprises: In step S2, according to the ASTM standard, the performance of the whole is first measured by artificial visual inspection and then confirmed by instrument measurement, the picture sample with normal performance is recorded as 1, and the picture sample with unqualified performance is recorded as 0.

3. The method of claim 1, wherein the method further comprises: Step S3 specifically comprises the following sub-steps: S3.1) performing first-level image segmentation manual labeling of the pictures in the sample database using a rectangular frame as the region shape to obtain first labeling information; the regions segmented by the first-level image segmentation specifically refer to the solder patch region, the strain grid region, and the pre-stress ring region; the first-level labeling record is in the form of a rectangular midpoint and the length and width of the rectangle; S3.2) performing parameter training using a machine learning model suitable for image segmentation, denoted as model I, using the sample database in step S1 and the first labeling information in step S3.1 to obtain the model parameters of the first-level image segmentation image region, denoted as model parameter Ia; S3.3) performing second-level image segmentation manual labeling of the pictures in the sample database to obtain second labeling information; the regions segmented by the second-level image segmentation refer to the solder joint region; the second-level labeling record is in the form of a rectangular midpoint and the length and width of the rectangle; S3.4) performing parameter training using model I, the sample database in step S1, and the second labeling information in step S3.3 to obtain the model parameters of the second-level image segmentation image region, denoted as model parameter Ib.

4. The method of claim 3, wherein the method further comprises: Step S4 specifically comprises the following sub-steps: S4.1) performing reduction operation using the strain grid region and the glue bottom strain grid region labeled manually in step S3.1 to obtain the glue bottom strain grid region; S4.2) using the image preprocessing module M1 to perform picture preprocessing operation on the solder patch region and the glue bottom strain grid region in the sample one by one; the picture preprocessing operation includes A) resampling with a sampling rate of 10; B) average sliding filter processing; S4.3) using the individualized feature module M2 to extract features from the solder patch region, the pre-stress ring region labeled in step S3.1, the solder joint region labeled in step S3.3, and the glue bottom strain grid region obtained in step S4.1, and returning the features; S4.4) using a classification model, denoted as model II, to train the features extracted in step S4.3 and the manual labeling in step S3 to obtain corresponding classification parameters, denoted as model parameter II.

5. The method of claim 4, wherein the method further comprises: In the above step S4.1, the glue bottom strain grid region refers to a rectangle formed by three times the width of the strain grid and five-sixths the length of the strain grid.

6. The method of claim 4, wherein the method further comprises: Step S4.3 feature extraction specifically refers to calculating different individualized features for different regions segmented by the image or labeled, wherein: A) for the solder patch region, the following 14 features are calculated: image color moment features, one, two, and three low-order moments are extracted for each color channel, a total of 9 component features; 4 solder joint positions and the inertia moment of the region surrounded thereby, a total of 5 features; B) for the solder joint region, the following 48 features are calculated: second-order moments, contrast, correlation, and entropy in 0, 45, 90, and 135 directions, a total of 16 features, and a total of 48 features for four solder joints; C) For the pre-stress ring area, after the path of the specific pre-stress ring is extracted by threshold processing and morphological feature extraction, the following 7 features are calculated: the Euler distance of the terminal point, a total of two component features; the included angle between the two ideal guide lines, a total of 1 component feature; the square sum of the offset of the guide line path from the ideal guide line, a total of 2 component features; the included angle between the ideal guide line and the vertical edge, a total of 2 component features; D) For the glue bottom strain grid area, the following 10 features are calculated: the proportion of the lightest color after color quartering, a total of 1 component feature; the first, second and third moments of the picture, a total of 9 component features. Step S5) includes the following sub-steps:

7. The method of claim 4, wherein the method further comprises: 5.1) For the given target picture, target recognition is performed using model I and model parameter Ia in step S3, and a first-level picture segmentation is performed to obtain the solder patch area, the strain grid area, the pre-stress ring area, and the glue bottom strain grid area; 5.2) The strain grid area and the glue bottom strain grid area segmented in step S5.1 are used to perform reduction operation to obtain the glue bottom strain grid area; 5.3) For the solder patch area segmented in step S5.1, target recognition is performed using model I and model parameter Ib in step S3, and a second-level picture segmentation is performed to obtain the solder point area. It is assumed that the solder point must be located on the sticker, and since there are at most four solder points on the sticker, only the four detection boxes with the highest confidence are taken, and the relative coordinates of the picture and the center positions of the four solder points in the solder patch area are returned. The process of step S6 is to check whether the strain grid has defects. A first-order differential operator is used to process the strain grid area to determine whether there is a significant discontinuity. If there is, it is considered that the target pasting quality does not meet the requirements. The right 12 / 13 of the grid image is intercepted, threshold processing is performed, and the grid image is reset to a matrix containing (0, 1). If the number of different pixels is greater than 10% of the total number of pixels, it is considered that the target pasting quality does not meet the requirements. If both of the above conditions are met, step S7 is continued. 8.The method of claim 1, wherein The specific steps of step S7 are: S7.1) using the image preprocessing module M1 to preprocess the given solder patch area and glue bottom strain grid area; 9.The method of claim 1, wherein S7.2) using the individualized feature module M2 to extract features from the solder patch area obtained in step S5.1, the pre-stress ring area, the glue bottom strain grid area obtained in step 5.2, and the solder point area obtained in step S5.3; S7.3) using model II, model parameter para-II, and the features extracted in S7.2 to classify, and returning result 0, i.e. unqualified, or result 1, i.e. qualified. In step S7.3, when the picture result to be detected is in the middle region of the two support vectors, in order to meet the engineering application, result 0, i.e. unqualified, is returned.

10. The method of claim 9, wherein the method further comprises: ​

Images