An online defect detection method and system combining image analysis and processing technology

Abstract

This invention relates to the field of flaw detection technology, and more particularly to an online defect detection method and system combining image analysis and processing technology. It includes the following steps: S1: Acquiring and preprocessing detection image data, dividing the preprocessed image into a grid and extracting features; S2: Dividing the gridded image into even and odd grids, and calculating angles based on the division results; S3: Constructing a prediction model and performing real-time prediction analysis. This invention reduces computational complexity and improves processing efficiency by acquiring image data at equal intervals and performing preprocessing; it refines defect types and severity by dividing the grid into even and odd grids and calculating the included angles; it constructs an LSTM model for real-time prediction, combining historical data to improve prediction accuracy and reliability; and finally, it forms a closed-loop system to achieve real-time online monitoring, accurately determine defect types and severity, ensure the continuity and reliability of detection, and provide strong technical support for defect damage detection.

Description

Technical Field

[0001] This invention relates to the field of flaw detection technology, and in particular to an online defect detection method and system that combines image analysis and processing technology. Background Technology

[0002] The main directions of online flaw detection include ultrasonic testing, eddy current testing, magnetic particle testing, radiographic testing, and visual inspection. Existing online flaw detection methods offer advantages such as high sensitivity, non-destructive testing, real-time detection, high automation, wide applicability, and reliable results. However, these methods also have some drawbacks, such as high cost, demanding operator skills, limited detection speed, significant susceptibility to environmental factors, and limitations in the types of defects they can detect. Existing online flaw detection methods suffer from accuracy deficiencies, primarily due to insufficient precision in defect feature identification and classification, leading to frequent false positives and false negatives. Image analysis and processing technology can improve detection speed and accuracy, reduce reliance on operators, mitigate the impact of environmental factors, and expand the applicability of detection methods. In online flaw detection, image analysis and processing technology can significantly improve detection accuracy by effectively reducing false positives and false negatives through advanced algorithms and pattern recognition techniques, thereby improving the reliability and stability of the detection process. Summary of the Invention

[0003] To overcome the low accuracy of online defect detection in flaw detection, this invention provides a method and system for online defect detection that combines image analysis and processing technology.

[0004] The real-time technical solution of the present invention is: an online defect detection method combining image analysis and processing technology, comprising the following steps:

[0005] S1: Acquire detection image data and preprocess it; divide the preprocessed image into grids and extract features.

[0006] S2: Divide the gridded image into even and odd parts, and calculate the angle based on the division results;

[0007] S3: Build a prediction model and perform real-time prediction analysis to determine the degree of damage caused by the defect based on the prediction results;

[0008] S4: The entire system is designed as a closed loop to achieve real-time online detection of flaws.

[0009] Preferably, the step of acquiring and preprocessing the detected image data, and dividing the preprocessed image into grids and extracting features includes: acquiring image data at equal intervals t, cleaning and standardizing the image data; uniformly dividing the image data into n*n grids, using a CNN model to extract image defect features, calculating the mean and variance of the defect features in each grid, and recording the position of each defect feature in the grid.

[0010] Preferably, the step of dividing the gridded image into even and odd parts and calculating the angle based on the division result includes: marking the horizontal and vertical grids of n*n grids in odd and even order, with the horizontal grids marked in the order of 1, 2, 1, 2... and the vertical grids marked in the order of 1, 2, 1, 2...; obtaining the defect feature distribution, extracting the defect features 1 and 2 that are closest in the horizontal direction, and simultaneously extracting the defect feature 3 that is closest in the vertical direction; using defect features 1 and 2 as horizontal lines, connecting defect feature 3 to defect feature 1, and forming an angle between the lines connecting defect features 1, 2, and 3; calculating the angle; and classifying the angles.

[0011] Preferably, the classification of the included angles includes: acquiring image data before the formation of defect features and defining it as Class A image data, and acquiring image data in which defect features are not formed during the defect feature formation process and defining it as Class B image data; recording image change data during the process from Class A image to Class B image, classifying the included angles according to the image change data; if the change is horizontal during the process from Class A image to Class B image, then modifying defect features 1 and 2, and then forming a new included angle and defining it as a new horizontal included angle; if the change is vertical during the process from Class A image to Class B image, then modifying defect feature 3, and then forming a new included angle and defining it as a new vertical included angle; if the change is neither horizontal nor vertical, then modifying defect features 1, 2, and 3 according to the angle change, and then forming a new included angle and defining it as a diagonal included angle; and forming three included angle sequences based on the three new included angles.

[0012] Preferably, the step of forming three angle sequences based on three new included angles includes: in the horizontal direction, if defect feature 3 is located to the left of defect feature 1 in the vertical direction or defect feature 3 is located within the vertical range of defect feature 1 and defect feature 2, then the new angle data caused by the change in the included angle due to the change in the defect feature forms a horizontal sequence; in the vertical direction, the new angle data formed by the change in the included angle in the vertical direction forms a vertical sequence; in the diagonal direction, the new angle data is sorted by the magnitude of the angle change so that it forms a diagonal sequence.

[0013] Preferably, the step of constructing a prediction model and performing real-time prediction analysis to determine the degree of damage of the defect based on the prediction results includes: acquiring a defect feature data training set, constructing an LSTM model, inputting the training set data into the LSTM model to obtain a trained LSTM model, using the LSTM model to predict defect features, obtaining the predicted defect features, and labeling them as the first defect feature; acquiring a historical defect feature dataset, and simultaneously acquiring a predicted defect feature dataset predicted by the LSTM model, using the historical defect feature dataset as the denominator and the predicted defect feature dataset as the numerator to construct a prediction factor, using the prediction factor to predict defect features, obtaining the predicted defect features, and labeling them as the second defect feature; analyzing the three included angle sequences using the first defect feature and the second defect feature; prediction factor = predicted defect feature dataset / historical defect feature dataset.

[0014] Preferably, the analysis of the three included angle sequences using the first defect feature and the second defect feature includes: matching the first defect feature, the second defect feature, and the three included angle sequences; if none of the three included angle sequences conform to the first defect feature and the second defect feature, then the defect feature is marked as a confirmed defect; if the three included angle sequences conform to the second defect feature but not the first defect feature, then the defect feature is marked as a defect to be determined; if all three included angles conform to the first defect feature and the second defect feature, then the parity change value of the angles in the three included angle sequences is obtained.

[0015] Preferably, obtaining the angle parity change values ​​in the three included angle sequences includes: obtaining a training set of angle parity change value data, constructing an SVM classification model, inputting the training set data into the SVM classification model to obtain a trained SVM classification model, and classifying the angle parity change using the SVM classification model; wherein, the classification criteria are: first, recording the parity ratio in the three included angle sequences, and second, using the relationship between the parity ratio and the angle change rate as the classification criterion; obtaining the intersection of the three included angle sequences, and if the determined defect, the undetermined defect, and the feature defect are located in different intersections, they are marked according to their respective classification rules; if the angle change rate in the three included angle sequences does not exceed the threshold (X), they are marked as non-defect areas.

[0016] Preferably, the step of constructing the entire system into a loop to achieve real-time online monitoring of flaw detection defects includes: according to the classification criteria and intersection, detecting all known defects, undetermined defects, and characteristic defects, and finally constructing the entire system into a loop system; the specific classification criteria are: if the rate of change of the angle in the three included angle sequences exceeds the threshold X, it is marked as a known defect.

[0017] An online defect detection system combining image analysis and processing technology includes:

[0018] Image acquisition module: used to acquire detection image data;

[0019] Image preprocessing module: used to preprocess the acquired image data;

[0020] Gridding and Feature Extraction Module: This module is used to uniformly divide the preprocessed image data into n*n grids and extract features.

[0021] Parity partitioning and angle calculation module: used to partition the gridded image into parity partitions and calculate the angles based on the partitioning results;

[0022] Predictive Model Building and Real-Time Prediction Module: Used to build predictive models and perform real-time predictive analysis;

[0023] Defect Feature Analysis and Result Output Module: This module is used to match the first defect feature with the second defect feature, determine the status of the defect feature, and output the final result.

[0024] Beneficial Effects: This invention proposes an online defect detection method combining image analysis and processing technologies. By acquiring image data at equal intervals *t* and preprocessing it, the temporal continuity and consistency of the data are ensured, avoiding data deviations caused by inconsistent acquisition time intervals. Next, the preprocessed image is gridded and features are extracted. The image is uniformly divided into n*n grids, and a CNN model is used to extract image defect features. The mean and variance of defect features in each grid are calculated, and the position of each defect feature in the grid is recorded. This not only reduces computational complexity but also improves processing efficiency, while local processing better captures subtle defect features. Based on this, the gridded image is divided into odd and even sections, and angles are calculated according to the division results. By marking the horizontal and vertical grids with odd / even order, it is easier to identify and track the distribution and changes of defects in different directions. The nearest defect features in the horizontal and vertical directions are extracted, forming angles, and the angle is calculated, further refining the type and severity of defects and providing more detailed information for subsequent defect analysis and processing. Subsequently, a predictive model is constructed and real-time predictive analysis is performed. By acquiring a training set of defect feature data, an LSTM model is built. The characteristics of the LSTM model are used to predict defect features in the current detection image in real time. Predictive factors are constructed by combining historical and predicted data to quantify the difference between the predicted results and historical data, further improving the accuracy and reliability of the prediction. By combining the first and second defect features, the spatial relationships and evolution patterns of defect features are analyzed more comprehensively, improving the accuracy of diagnosis. Finally, the entire system is constructed as a loop to achieve real-time online monitoring of flaw detection defects. By analyzing the odd / even change values ​​of angles in the intersection, the type and severity of defects are more accurately determined, avoiding misjudgments. The specific classification standard is: if the rate of change of angles in the three included angle sequences exceeds a threshold X, it is marked as a confirmed defect, ensuring the continuity and real-time nature of defect detection and providing reliable technical support for assessing the degree of defect damage. Attached Figure Description

[0025] Figure 1 This is a flowchart of an online defect detection method and system combining image analysis and processing technology according to the present invention;

[0026] Figure 2 This is a schematic diagram of the structure of an online defect detection system that combines image analysis and processing technology according to the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1: An online defect detection method combining image analysis and processing technology, such as... Figure 1 As shown, it includes the following steps:

[0029] S1: Acquire detection image data and preprocess it; divide the preprocessed image into grids and extract features.

[0030] S2: Divide the gridded image into even and odd parts, and calculate the angle based on the division results;

[0031] S3: Build a prediction model and perform real-time prediction analysis to determine the degree of damage caused by the defect based on the prediction results;

[0032] S4: The entire system is designed as a closed loop to achieve real-time online detection of flaws.

[0033] The process involves acquiring and preprocessing detection image data, dividing the preprocessed image into grids, and extracting features. This includes: acquiring image data at equal intervals t, cleaning and standardizing the image data, dividing the image data into n*n grids, using a CNN model to extract image defect features, calculating the mean and variance of the defect features in each grid, and recording the position of each defect feature in the grid.

[0034] Further explanation is as follows: The image data is acquired at equal intervals t: This ensures the temporal continuity and consistency of the data, avoiding data deviations caused by inconsistent acquisition time intervals. The image preprocessing involves cleaning and standardizing the acquired image data, using image processing algorithms to remove noise, and adjusting brightness and contrast to make defect features in the image more prominent. The gridding divides the image evenly into n*n grids; for example, a 1024x1024 image is divided into 64 small 16x16 grids. This division reduces computational complexity, improves processing efficiency, and allows for better capture of subtle defect features through localized processing. The location recording records the position of each defect feature within the grid, including its center coordinates and range. For example, the center coordinates of a crack feature might be (10, 10), with a range of 16x16 grids. Accurately recording the defect location information helps in locating and tracking defect development, providing guidance for repair and maintenance.

[0035] The image is divided into even and odd grids, and the angles are calculated based on the division results. This includes: marking the horizontal and vertical grids of the n*n grids in odd and even order, with the horizontal grids marked in the order of 1, 2, 1, 2... and the vertical grids marked in the order of 1, 2, 1, 2...; obtaining the defect feature distribution, extracting the horizontally closest defect features 1 and 2, and simultaneously extracting the vertically closest defect feature 3; using defect features 1 and 2 as horizontal lines, connecting defect feature 3 to defect feature 1, and forming an angle between the lines connecting defect features 1, 2, and 3; calculating the angle and classifying the angles.

[0036] Further explanation is that the parity division of the gridded image makes it easier to identify and track the distribution and changes of defects in different directions. The defect feature distribution is obtained, and the horizontally closest defect features 1 and 2 are extracted, while the vertically closest defect feature 3 is also extracted. By extracting the closest defect features, the relative positional relationship between defects can be determined more accurately, which helps in subsequent angle calculation and classification. Using defect features 1 and 2 as a horizontal line, and connecting defect feature 3 to defect feature 1, the angle between defect features 1 and 2, 3 is calculated. Calculating the angle helps the system more accurately describe the spatial relationship between defects, thereby better understanding the morphology and distribution of defects. The angles are then classified: by classifying the angles, the type and severity of defects are further refined, providing more detailed information for subsequent defect analysis and processing. Assuming it is used to detect surface defects in metal sheets, the gridded image is divided into parity parts: the 16x16 image grid is divided into parity parts, with the horizontal and vertical grids labeled in parity order.

[0037] The included angles are classified as follows: image data before defect features are formed is acquired and defined as Class A image data, and image data without defect features formed during the defect feature formation process is acquired and defined as Class B image data; image change data during the process from Class A image to Class B image is recorded, and the included angles are classified according to the image change data. If the change during the process from Class A image to Class B image is in the horizontal direction, defect features 1 and 2 are modified, and a new included angle is formed and defined as a new horizontal included angle; if the change during the process from Class A image to Class B image is in the vertical direction, defect feature 3 is modified, and a new included angle is formed and defined as a new vertical included angle; if the change is neither in the horizontal nor the vertical direction, defect features 1, 2, and 3 are modified according to the angle change, and a new included angle is formed and defined as an oblique included angle; based on the three new included angles, three included angle sequences are formed.

[0038] Further explanation involves: acquiring image data before defect features form and defining it as Class A image data: This establishes a baseline for comparing and analyzing the formation process of defect features. Class A image data provides a reference in a defect-free state, helping to identify and quantify the occurrence and development of defects. Acquiring image data during the defect feature formation process but before defect features are formed and defining it as Class B image data: This records the early stages of defect feature formation, helping to analyze the initial morphology and evolution of defects. Class B image data provides the transition stage from defect-free to defective, helping to more precisely identify early signs of defects. Recording image change data during the transition from Class A to Class B images: By recording image change data, the formation process of defect features is tracked, providing detailed information for subsequent analysis. This change data helps to understand the development trend and speed of defects. Classifying the included angles based on image change data: By classifying the included angles, the spatial relationships and evolution patterns of defect features are described more accurately, helping to identify different types of defects. If the change from image A to image B is horizontal, then defect feature 1 and defect feature 2 are modified to form a new included angle, which is defined as the new horizontal included angle. By modifying the position of the defect feature, the change in the defect in the horizontal direction is more accurately reflected, which helps to identify and analyze defect expansion in the horizontal direction. If the change in the horizontal direction is vertical, then defect feature 3 is modified to form a new included angle, which is defined as the new vertical included angle. By modifying the position of the defect feature, the change in the vertical direction is more accurately reflected, which helps to identify and analyze defect expansion in the vertical direction. If the change is neither horizontal nor vertical, then defect features 1, 2, and 3 are modified according to the angle change, and a new included angle is formed, which is defined as the oblique included angle. By comprehensively considering the changes in defect features in multiple directions, the complex morphology of the defect is more comprehensively described, which helps to identify and analyze oblique defect expansion. Based on the three new included angles, three included angle sequences are formed. By forming included angle sequences, the change process of defect features is systematically recorded and analyzed, providing data support for subsequent prediction and diagnosis. Assuming the detection of surface defects in metal sheets is used, the image data before defect features form is acquired and defined as Class A image data: An image of a defect-free metal sheet is acquired and defined as Class A image data. Image data during the defect feature formation process, but before defect features are formed, is acquired and defined as Class B image data: During the production process, several images are acquired of the metal sheet surface where micro-cracks have begun to appear but have not yet fully formed into obvious defects; these are defined as Class B image data. Image change data is recorded during the transition from Class A to Class B images: The propagation process of cracks on the metal sheet surface is recorded during the transition from Class A to Class B images, including changes in crack length, width, and distribution.The included angles are classified based on image change data: Analyze the recorded image change data, calculate the included angle changes between defect features 1, 2, and 3, and classify them as horizontal, vertical, or oblique angles. If the change occurs horizontally from image A to image B: Assuming the horizontal distance between defect features 1 and 2 increases during the transition, recalculate the new included angle, defining it as the new horizontal included angle. If the change occurs vertically from image A to image B: Assuming defect feature 3 moves vertically during the transition, recalculate the new included angle, defining it as the new vertical included angle. If the change occurs neither horizontally nor vertically: Assuming the change occurs obliquely during the transition from image A to image B, recalculate the new included angle, defining it as the oblique included angle. Based on the three new included angles, three included angle sequences are formed: according to the new horizontal included angle, the new vertical included angle, and the diagonal included angle, three included angle sequences are formed, namely the horizontal sequence, the vertical sequence, and the diagonal sequence.

[0039] Based on three new included angles, three included angle sequences are formed, including: in the horizontal direction, if defect feature 3 is located to the left of defect feature 1 in the vertical direction or defect feature 3 is located within the vertical range of defect feature 1 and defect feature 2, then the new angle data caused by the change in the included angle due to the change in the defect feature forms a horizontal sequence; in the vertical direction, the new angle data formed by the change in the included angle in the vertical direction forms a vertical sequence; in the diagonal direction, the new angle data is sorted by the magnitude of the angle change so that it forms a diagonal sequence.

[0040] To further explain, recording changes in the included angle in the horizontal direction allows for a more precise description of the expansion and evolution of defects in that direction, aiding in the identification of defect features in the horizontal direction. Similarly, recording changes in the included angle in the vertical direction provides a more precise description of the expansion and evolution of defects in that direction, helping to identify defect features in the vertical direction. Finally, recording changes in the included angle in the diagonal direction provides a more comprehensive description of the expansion and evolution of defects in that direction, aiding in the identification of defect features in the diagonal direction.

[0041] Construct a prediction model and perform real-time predictive analysis to determine the degree of damage based on the prediction results. This includes: acquiring a training set of defect feature data, constructing an LSTM model, inputting the training set data into the LSTM model to obtain a trained LSTM model, using the LSTM model to predict defect features, obtaining the predicted defect features, and labeling them as the first defect feature; obtaining a historical defect feature dataset and a dataset of predicted defect features from the LSTM model, using the historical defect feature dataset as the denominator and the predicted defect feature dataset as the numerator to construct a prediction factor, using the prediction factor to predict defect features, obtaining the predicted defect features, and labeling them as the second defect feature; and analyzing the three included angle sequences using the first and second defect features. The prediction factor is calculated as: predicted defect feature dataset / historical defect feature dataset.

[0042] To further explain, LSTM is a special type of recurrent neural network that excels at processing time-series data, capturing temporal dependencies, and is suitable for predicting defect features. A trained LSTM model can predict defect features in the currently detected image in real time, providing preliminary results for subsequent analysis. By comparing historical and predicted data, the accuracy and stability of the prediction model are evaluated, allowing for further model optimization. Constructing predictive factors quantifies the differences between the predicted results and historical data, further improving prediction accuracy. Correcting the predictive factors further enhances the accuracy and reliability of the predictions. Combining the first and second defect features allows for a more comprehensive analysis of the spatial relationships and evolution patterns of defect features, improving diagnostic accuracy.

[0043] The analysis of the three included angle sequences is performed using the first and second defect features, including: matching the first and second defect features with the three included angle sequences; if none of the three included angle sequences meet the first and second defect features, the defect feature is marked as a confirmed defect; if the three included angle sequences meet the second defect feature but not the first defect feature, the defect feature is marked as a defect to be determined; if all three included angle sequences meet the first and second defect features, the odd / even change value of the angle in the three included angle sequences is obtained.

[0044] To further explain, by matching the first and second defect features with three included angle sequences, the spatial relationships and evolution patterns of the defect features can be analyzed more comprehensively, improving diagnostic accuracy. If none of the three included angle sequences match the first or second defect features, it indicates that these features are completely different from known defect patterns and are marked as confirmed defects requiring further detailed examination and processing. If the three included angle sequences match the second defect feature but not the first, it indicates that these features require further verification and are marked as defects to be determined for further examination and analysis. If all three included angle sequences match both the first and second defect features, it indicates that these features are highly consistent with known defect patterns, and the characteristics of the defect can be more accurately described by further analyzing the odd / even variation values ​​of the angles.

[0045] Obtaining the odd / even change values ​​of the angles in three included angle sequences includes: obtaining a training set of angle odd / even change value data, constructing an SVM classification model, inputting the training set data into the SVM classification model to obtain a trained SVM classification model, and classifying the angle odd / even changes using the SVM classification model. The classification criteria are as follows: first, record the odd / even ratio in the three included angle sequences; second, use the relationship between the odd / even ratio and the angle change rate as the classification criterion. Obtain the intersection of the three included angle sequences. If the identified defect, the unidentified defect, and the feature defect are located in different intersections, they are marked according to their respective classification rules. If the angle change rate in the three included angle sequences does not exceed the threshold (X), they are marked as non-defect areas.

[0046] To further explain, Support Vector Machines (SVMs) are powerful machine learning algorithms suitable for classification tasks, effectively handling high-dimensional data and complex classification problems. By analyzing the parity changes of angles in the intersection, the type and severity of defects can be determined more accurately, avoiding misjudgments.

[0047] Further explanation: Defect intersection: If the rate of change of angles in all three included angle sequences exceeds the threshold (X), then the region is marked as a determined defect. Defect intersection to be determined: If the rate of change of angles in two of the three included angle sequences exceeds the threshold (X), then the region is marked as a defect to be determined. Feature defect intersection: If the rate of change of angles in one of the three included angle sequences exceeds the threshold (X), then the region is marked as a feature defect. Non-defect region: If the rate of change of angles in none of the three included angle sequences exceeds the threshold (X), then the region is marked as a non-defect region.

[0048] Further explanation involves obtaining the intersection of three included angle sequences, specifically defined as follows: **Definite Intersection:** Regions in the three included angle sequences simultaneously marked as definitive defects. **Defects to be Determined Intersection:** Regions in the three included angle sequences simultaneously marked as defects to be determined. **Featured Defect Intersection:** Regions in the three included angle sequences simultaneously marked as feature defects. **Non-Defect Regions:** Regions in the three included angle sequences simultaneously marked as non-defective. **Classification Rules:** **Definitive Defect:** If the rate of change of angles in all three included angle sequences exceeds a threshold (X), it is marked as a definitive defect. **Defects to be Determined:** If the rate of change of angles in two of the three included angle sequences exceeds a threshold (X), it is marked as a defect to be determined. **Featured Defect:** If the rate of change of angles in one of the three included angle sequences exceeds a threshold (X), it is marked as a feature defect. **Non-Defect Regions:** If the rate of change of angles in none of the three included angle sequences exceeds a threshold (X), it is marked as a non-defect region.

[0049] The entire system is constructed as a loop to achieve real-time online monitoring of flaw detection defects. This includes: detecting all known defects, undetermined defects, and characteristic defects according to classification criteria and intersection, and finally constructing the entire system as a loop. The specific classification criterion is: if the rate of change of the angle in the three included angle sequences exceeds the threshold X, it is marked as a known defect.

[0050] To further explain, by analyzing the odd-even change values ​​of the angles in the intersection, the type and severity of defects can be determined more accurately, avoiding misjudgment; the methods for determining the threshold X include, but are not limited to, historical data and statistical analysis, rules of thumb, machine learning methods, and dynamic adjustment.

[0051] Example 2: Based on Example 1, an online defect detection system combining image analysis and processing technology includes the following components.

[0052] Image acquisition module: used to acquire detection image data;

[0053] Image preprocessing module: used to preprocess the acquired image data;

[0054] Gridding and Feature Extraction Module: This module is used to uniformly divide the preprocessed image data into n*n grids and extract features.

[0055] Parity partitioning and angle calculation module: used to partition the gridded image into parity partitions and calculate the angles based on the partitioning results;

[0056] Predictive Model Building and Real-Time Prediction Module: Used to build predictive models and perform real-time predictive analysis;

[0057] Defect Feature Analysis and Result Output Module: This module is used to match the first defect feature with the second defect feature, determine the status of the defect feature, and output the final result.

[0058] The present application has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present application. Therefore, the content of this specification should not be construed as a limitation of the present application.

Claims

1. An on-line defect detection method for a flaw detection using a combined image analysis processing technique, characterized by, The method comprises the following steps: S1: acquiring detection image data and performing preprocessing, grid division on the preprocessed image and feature extraction; S2: performing odd-even division on the grid image, and calculating the angle according to the division result, comprising: marking the horizontal and vertical grids of n*n grids in odd-even order, and the marking result is that the horizontal grids are marked in the order of 1, 2, 1, 2, …, and the vertical grids are marked in the order of 1, 2, 1, 2, …; obtaining defect feature distribution, extracting the nearest defect features 1 and 2 in the horizontal direction, and extracting the nearest defect feature 3 in the vertical direction; connecting the defect feature 3 with the defect feature 1 by taking the defect features 1 and 2 as horizontal lines, forming an included angle between the connecting lines of the defect features 1, 2 and 3, and calculating the included angle; the classification of the included angle comprises: obtaining image data before the formation of the defect feature and defining it as A-class image data, and obtaining image data without forming a defect feature in the process of forming a defect feature and defining it as B-class image data; recording the image change data in the process from A-class image to B-class image, and classifying the included angle according to the image change data; if the horizontal direction changes in the process from A-class image to B-class image, the defect features 1 and 2 are modified, then a new included angle is formed and defined as a new horizontal included angle; if the vertical direction changes in the process from A-class image to B-class image, the defect feature 3 is modified, then a new included angle is formed and defined as a new vertical included angle; if neither the horizontal direction nor the vertical direction changes, the defect features 1, 2 and 3 are modified according to the angle change, then a new included angle is formed and defined as an oblique included angle; taking the three new included angles as the reference, three included angle sequences are formed; S3: constructing a prediction model and performing real-time prediction analysis, and determining the damage degree of the defect according to the prediction result; S4: designing the whole system as a closed loop to realize real-time online detection of the flaw detection defect.

2. The method of claim 1, wherein the method is characterized by: The acquisition of detection image data and preprocessing, grid division on the preprocessed image and feature extraction comprises: collecting image data at equal intervals t, performing data cleaning and standardization processing on the image data; dividing the image data into n*n grids, extracting image defect features using a CNN model, calculating the mean and variance of each grid defect feature, and recording the position of each defect feature in the grid.

3. The method of claim 1, wherein the method is characterized by: Taking the three new included angles as the reference, three included angle sequences are formed, comprising: in the horizontal direction, if the defect feature 3 is located on the left side of the defect feature 1 in the vertical direction or the defect feature 3 is located within the vertical direction range of the defect features 1 and 2, then the new angle data formed by the change of the included angle due to the change of the defect feature forms a horizontal sequence; in the vertical direction, the new angle data formed by the change of the vertical direction included angle forms a vertical sequence; in the oblique direction, the new angle data is sorted according to the size of the angle change to form an oblique sequence.

4. The method of claim 1, wherein the method is characterized by: The method comprises the following steps: acquiring defect characteristic data training set, constructing LSTM model, inputting training set data into LSTM model, obtaining trained LSTM model, predicting defect characteristics by using LSTM model, obtaining predicted defect characteristics, and marking the predicted defect characteristics as first defect characteristics; obtaining historical defect characteristic data set, obtaining predicted defect characteristic data set predicted by LSTM model, taking historical defect characteristic data set as denominator and predicted defect characteristic data set as numerator, constructing prediction factor, predicting defect characteristics by using prediction factor, obtaining predicted defect characteristics, and marking the predicted defect characteristics as second defect characteristics; analyzing three included angle angle sequences by using first defect characteristics and second defect characteristics; and prediction factor = predicted defect characteristic data set / historical defect characteristic data set.

5. The method of claim 4, wherein the method is characterized by: The step of analyzing three included angle angle sequences by using first defect characteristics and second defect characteristics comprises the following steps: matching first defect characteristics and second defect characteristics with three included angle angle sequences; if three included angle angle sequences do not conform to first defect characteristics and second defect characteristics, marking the defect characteristics as determined defects; if three included angle angle sequences conform to second defect characteristics but do not conform to first defect characteristics, marking the defect characteristics as to-be-determined defects; and if three included angle angles all conform to first defect characteristics and second defect characteristics, obtaining angle parity change values in three included angle angle sequences.

6. The method of claim 5, wherein the method is characterized by: The step of obtaining angle parity change values in three included angle angle sequences comprises the following steps: acquiring angle parity change value data training set, constructing SVM classification model, inputting training set data into SVM classification model, obtaining trained SVM classification model, and classifying angle parity changes by using SVM classification model; wherein, the classification standard is that, firstly, recording parity proportions in three included angle angle sequences, and secondly, taking the relationship between parity proportion quantity and angle change rate as the classification standard; obtaining the intersection of three included angle angle sequences; if determined defects, to-be-determined defects and feature defects are located in different intersections respectively, marking according to respective classification rules; and if the angle change rates in three included angle angle sequences all do not exceed threshold value (X), marking as a non-defect region.

7. The method of claim 1, wherein the method is characterized by: The step of constructing the entire system into a cycle to realize real-time online monitoring and defect detection comprises the following steps: according to the classification standard and the intersection, detecting determined defects, to-be-determined defects and feature defects, and finally constructing the entire system into a cycle system; and the specific classification standard is that: if the angle change rates in three included angle angle sequences exceed threshold value X, marking as determined defects.

8. An on-line defect detection system for the use of the combined image analysis processing technique according to any one of claims 1 to 7, characterized in that The method comprises the following steps: An image acquisition module is configured to acquire detection image data; An image preprocessing module is configured to preprocess the acquired image data; A grid division and feature extraction module is configured to uniformly divide the preprocessed image data into n*n grids and extract features; An odd-even division and angle calculation module is configured to perform odd-even division on the gridded image and calculate angles according to the division results; A prediction model construction and real-time prediction module is configured to construct a prediction model and perform real-time prediction analysis. The defect feature analysis and result output module is configured to match the first defect feature and the second defect feature, determine a state of the defect feature, and output a final result.

Images