Enameled wire defect detection system and method based on deep learning

Abstract

The invention discloses an enamelled wire defect detection system and method based on deep learning, relates to the technical field of enamelled wire defect detection, and solves the technical problems of production material and energy waste and the like caused by finished product detection lag when detection is carried out in an enamelled wire finished product stage in the prior art. The method comprises the following steps: acquiring multidirectional to-be-processed images of the enameled wire in each technological process; preprocessing the multidirectional images to be processed, and splicing the images after preprocessing to obtain a spliced image; separating a target image from the spliced image; identifying defects of the target image; if yes, defect early warning is carried out; according to the technical scheme, accurate monitoring of the quality of the enameled wire in each technological process is realized, early warning is carried out in time, measures are taken, and defective products are ensured not to continue to enter the subsequent production technological process, so that material and energy waste caused by production of unqualified products is effectively restrained.

Description

Technical Field

[0001] The present invention belongs to the field of enameled wire defect detection, and specifically relates to an enameled wire defect detection system and method based on deep learning. Background Art

[0002] In recent years, as customers have become increasingly demanding on the quality standards of enameled wire products, the demand for inspection items for enameled wire products has shown a significant growth trend, which has continuously pushed manufacturers to add more detailed inspection links to meet the market's expectations for high-quality products.

[0003] The invention patent application with publication number CN114324567A discloses an enameled wire detection system and detection method. The detection system includes a power supply and display module, a detection module, a wire collection module and a controller. The power supply and display module are used to provide AC and DC for detection and display the detection results; the detection module is connected to the power supply and display module to detect defects in the enameled wire; the wire collection module collects the enameled wire detected by the detection module. Under the control of the controller, the detection module and the wire collection module automatically detect and collect the enameled wire, without relying on manual operation by personnel, which is time-consuming, occupies less inspection manpower, does not have the risk of missed scanning, has a high degree of automation, and has accurate detection results. The obtained enameled wire is of good quality.

[0004] The above-mentioned prior art is to perform defect detection on the finished enameled wire products after production. When defective products exist, it means not only a waste of raw materials, but also the use of production equipment, manpower, time and other resources; and when defects are found in the finished product stage, it is often difficult to trace them back to the specific production link or cause; therefore, the present invention provides an enameled wire defect detection system and method based on deep learning. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an enameled wire defect detection system and method based on deep learning, which is used to solve the technical problems that the prior art performs detection at the finished product stage of the enameled wire, resulting in delayed detection of the finished product, leading to waste of production materials and energy.

[0006] To achieve the above-mentioned object, a first aspect of the present invention provides an enameled wire defect detection system based on deep learning, comprising an image processing module, and an image acquisition module and a defect warning module connected thereto; Image acquisition module: used to collect multi-directional images of enameled wire in each process flow, and mark them as images to be processed; wherein the process flow includes wire laying stage, painting stage and baking stage; and the image acquisition cycle of each process flow is different; Image processing module: used to pre-process the multi-directional images to be processed, and then stitch the images to obtain a stitched image; and separate the target image from the stitched image; Defect recognition and warning module: used to identify whether there are defects in the target image; if yes, a defect warning is issued and processed; if no, the subsequent process flow operation is continued; and, Analyze and evaluate the defect status of each production stage during the cycle and issue defect warnings.

[0007] Preferably, the collecting of multi-directional images of the enameled wire in each process flow includes: The vertical plane image and side image of the enameled wire in the production stage are collected by an image acquisition tool; wherein the image acquisition tool includes a camera and an LED light; the side image is taken from the side of the copper wire; and the vertical image is taken from the vertical direction.

[0008] The present invention obtains the vertical image and side image of the enameled wire through the camera, and can comprehensively observe the state of the enameled wire from different angles, and more accurately identify any abnormal conditions on the surface and inside of the enameled wire, thereby ensuring product quality; secondly, the LED lamp, as an auxiliary lighting device, can provide a stable light source during the image acquisition process to improve the image quality. The LED lamp can provide uniform and controllable lighting conditions, reduce image blur or reflection caused by insufficient or excessive light, and make the image clearer and the details more prominent. This is very important for subsequent image processing and analysis, because high-quality images can improve the accuracy of defect detection.

[0009] Preferably, the image acquisition cycles of the various process flows are different, including: Collect the moving speed and pretreatment length of the enameled wire in each process; The image acquisition period TS of each process flow is calculated by the formula TS=Li / Vi-TC; wherein i is the process flow number, i=0, 1, ..., N, N is a positive integer; Li is the pretreatment length of the enameled wire in process flow i; Vi is the moving speed of the enameled wire in process flow i; TC is the shooting delay time of the camera.

[0010] The present invention can accurately calculate the image acquisition cycle of each process stage by directly measuring the moving speed and pretreatment length of the enameled wire in each process flow and combining the shooting delay time of the camera, so that the set image acquisition cycle can capture a complete image of the process flow and avoid repeated image capture and omission of uncaptured images.

[0011] Preferably, the moving speed is the material conveying speed during the enameled wire production process; the pre-processing length is the maximum enameled wire length that can be processed at one time in each process flow.

[0012] Preferably, the step of stitching the multi-directional images to be processed to obtain a stitched image includes: Step P1: Set the number label of the image to be processed; extract the feature points of two adjacent numbered images to be processed respectively, and obtain the matching feature points between the two adjacent images to be processed through a feature matching algorithm; Step P2: With the matching feature point as the center point, rectangular areas of the same area are delineated and marked as rectangular area j; wherein j is the number of the rectangular area, j=0, 1, ..., M, and M is a positive integer; Step P3: Calculate the pixel difference between the rectangular areas j and determine whether the pixel value difference is greater than a preset threshold; if not, the rectangular areas j are similar and proceed to step P4; if yes, the rectangular areas j are not similar and proceed to step P4; Step P4: When the rectangular area j is similar, the four sides of the rectangular area j are expanded outward by one unit pixel to obtain a new rectangular area j, and j=j+1, and it is determined whether the new rectangular area j is similar; if yes, repeat step P4; if no, the overlapping area between the two adjacent numbered images to be processed is the rectangular area j; When the rectangular area j is not similar, the four sides of the rectangular area j are reduced inward by one unit pixel to obtain a new rectangular area j, and j=j+1, and it is judged whether the new rectangular area j is similar; if yes, the overlapping area between the two adjacent numbered images to be processed is the new rectangular area j; if no, repeat step P4; Step P5: Overlapping the overlapping areas j between the adjacent numbered images to be processed to obtain a spliced ​​image.

[0013] The present invention obtains matching feature points between adjacent images through a feature matching algorithm, and defines a rectangular area with the matching feature points as the center, analyzes the similarity of the area, determines the overlapping area between the images, and realizes seamless splicing, so that the spliced ​​image is visually coherent and natural, without obvious seams or ghosting, thereby improving the overall quality and viewing effect of the image; and splicing multi-directional images to be processed helps to reduce the amount of data processing for subsequent image recognition.

[0014] Preferably, the order in which the numbered labels of the images to be processed are set is set according to the orientation relationship of the collected images to be processed.

[0015] The present invention numbers images according to their orientation relationships, such as from left to right, from top to bottom, and in a circumferential order, so that not only each image has a clear and unique identifier, but also subsequent image stitching operations are automatically performed in a preset logical order, ensuring that adjacent images can be matched in the correct order and position when stitching, avoiding misalignment or inversion, and improving the accuracy and efficiency of subsequent stitching processing.

[0016] Preferably, separating the target image from the stitched images comprises: The standard image of the enameled wire at each process stage is extracted from the image data, and the pixel values ​​of the pixel points of the standard image are obtained. The pixel values ​​of the pixel points of the stitched image are differenced with the pixel values ​​of the standard image at the corresponding process stage, and the pixel points whose difference values ​​are less than the difference threshold are marked as target pixel points; the target pixel points are fitted to obtain the target image.

[0017] The present invention compares the pixel values ​​of the spliced ​​image pixels with the pixel values ​​of the image pixels of the corresponding process flow stage one by one, and sets a difference threshold to screen the target pixel points, effectively reducing the interference of background noise and other non-target factors, and improving the accuracy of target image extraction.

[0018] Preferably, the step of identifying whether defects exist in the target image includes: Extracting a number of target images of each process flow from the historical data; wherein the target images include defective target images and non-defective target images; The target image is used as the standard input data of the convolutional neural network model, and the defect target image is used as the standard output data of the convolutional neural network model; The convolutional neural network model is trained using standard input data and standard output data to obtain a defect recognition model; wherein the defect recognition model includes a defect recognition model for the wire laying stage, a defect recognition model for the painting stage, and a defect recognition model for the baking stage.

[0019] The present invention extracts defective target images and non-defective target images of each process flow from historical data as standard input and output data of the convolutional neural network, ensuring that the model can learn accurate defect features. This enables the trained defect recognition model to accurately identify defects at each stage, improving the accuracy of quality control; and separately training defect recognition models for different process flows reflects an in-depth understanding and refined management of the production process. Phased modeling enables each model to be optimized for specific defects in its respective stage, improving recognition efficiency and accuracy.

[0020] Preferably, the number of defective target images in each process flow within the evaluation period is counted, and if the number of defective target images exceeds a preset number, a defect warning is issued.

[0021] The present invention realizes real-time monitoring of production quality by counting the number of defective target images in each process flow within the evaluation cycle. Once the number of defective target images exceeds the preset threshold, a defect warning is immediately triggered, so that management and operators can quickly know the problem and take corresponding corrective measures. This rapid response mechanism helps to reduce the number of defective products and reduce the defect rate.

[0022] Preferably, the detection method of the enameled wire defect detection system based on deep learning comprises the following steps: Step 1: Collect multi-directional images of the enameled wire in each process flow and mark them as images to be processed; the process flow includes a wire laying stage, a painting stage, and a baking stage; and the image acquisition cycle of each process flow is different; Step 2: Evaluate the quality of the images to be processed in multiple directions, stitch the images after processing to obtain a stitched image; separate the target image from the stitched image; Step 3: Identify whether there are defects in the target image; if yes, issue a defect warning and handle it; if not, continue with the subsequent process flow, analyze and evaluate the defect status of each production stage within the cycle, and issue a defect warning.

[0023] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention uses the moving speed and pre-processing length information of the enameled wire, and takes into account the inherent delay of the camera when shooting, so as to achieve accurate calculation of the image acquisition cycle of each process stage, ensuring that the set image acquisition cycle can not only completely capture the image of the enameled wire passing through each process, but also effectively avoid redundant and repeated image acquisition and any possible image omission, significantly improving the integrity and accuracy of the image data; 2. The present invention analyzes the matching feature points between adjacent images, takes the matching feature points as the core, demarcates rectangular areas for in-depth analysis of the similarity between the rectangular areas, accurately defines the overlapping parts between adjacent images, and realizes seamless image stitching technology. The stitched images effectively reduce the amount of data required to be processed for subsequent image recognition tasks, thereby improving the overall processing efficiency and performance. 3. The present invention collects multi-directional images to be processed of the enameled wire in each process flow; pre-processes the multi-directional images to be processed, and stitches the images after pre-processing to obtain a stitched image; separates a target image from the stitched image; identifies whether there are defects in the target image; if yes, issues a defect warning and processes it; if not, continues with the subsequent process flow operation; achieves precise control of the quality of the enameled wire, and once a defect is found, immediately triggers an early warning mechanism and takes corresponding treatment measures to prevent the defective product from continuing with the subsequent process flow, avoids unnecessary waste of materials and energy in the production of defective products, optimizes resource allocation and utilization efficiency; and is conducive to tracing back specific production links or causes. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 It is a schematic block diagram of the system structure of the present invention; Figure 2 It is a schematic diagram of the flow of the image stitching method of the present invention; Figure 3 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0026] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] See also Figure 1 , the first aspect of the present invention provides an enameled wire defect detection system based on deep learning, including an image processing module, and an image acquisition module and a defect warning module connected thereto; The image acquisition module uses an image acquisition tool to collect vertical images and side images of the enameled wire in the production stage, and both are marked as images to be processed; the process flow includes the wire laying stage, the painting stage, and the baking stage; It should be noted that the image acquisition tool includes a camera and an LED light; the side image is taken from the side of the copper wire, that is, the images of the left and right sides; the vertical image is taken from the vertical direction, that is, the top view image and the bottom view image.

[0028] and the image acquisition cycles of each process are different; Collect the moving speed and pretreatment length of the enameled wire in each process; It should be noted that in the production process of enameled wire, the material is continuously transported to each process flow through a traction wheel, and its transportation speed is the moving speed; the pretreatment length is the maximum enameled wire length that can be processed at one time in each process flow.

[0029] The image acquisition period TS of each process flow is calculated by the formula TS=Li / Vi-TC; wherein i is the process flow number, i=0, 1, ..., N, N is a positive integer; Li is the pretreatment length of the enameled wire in process flow i; Vi is the moving speed of the enameled wire in process flow i; TC is the shooting delay time of the camera.

[0030] It should be noted that the setting of the image acquisition cycle needs to ensure that the images passing through the process flow can be fully acquired and avoid omissions or duplications; among them, through the ratio of the preprocessing length to the moving speed, the time required for the preprocessing length to be delivered to a certain process flow can be calculated, which is marked as the moving time; and there is a delay problem in camera shooting, so the image acquisition cycle needs to be less than the moving time, that is, the delay time of the camera is reduced on the basis of the moving time. Therefore, this image acquisition cycle can acquire complete images of the process flow and avoid omissions or duplications.

[0031] The image processing module pre-processes the multi-directional images to be processed, and then stitches the images to obtain a stitched image; and separates the target image from the stitched image; Specifically, graying, noise removal, edge detection and other operations are performed on the multi-directional images to be processed, which is conducive to extracting image features and performing subsequent image stitching operations.

[0032] The order of setting the number labels of the images to be processed is set according to the orientation relationship of the acquired images to be processed.

[0033] Specifically, the left image, the right image, the top view image and the bottom view image in the image to be processed; the order of setting the number labels can be starting from any image to be processed, and then the number labels of the remaining images to be processed are set in a clockwise or counterclockwise direction; For example: if the number label of the left image is set to 0, in a clockwise direction, the number label of the overhead image is set to 1, the number label of the right image is set to 2, and the number label of the bottom image is set to 3; if in a counterclockwise direction, the number label of the bottom image is set to 1, the number label of the right image is set to 2, and the number label of the overhead image is set to 3.

[0034] See also Figure 2 , through the setting of numbered labels, images with adjacent numbered labels can be spliced, and the splicing process is as follows: Step P1: Set the number label of the image to be processed; extract the feature points of two adjacent numbered images to be processed respectively, and obtain the matching feature points between the two adjacent images to be processed through a feature matching algorithm; Step P2: With the matching feature point as the center point, rectangular areas of the same area are delineated and marked as rectangular area j; wherein j is the number of the rectangular area, j=0, 1, ..., M, and M is a positive integer; Step P3: Calculate the pixel difference between the rectangular areas j and determine whether the pixel value difference is greater than a preset threshold; if not, the rectangular areas j are similar and proceed to step P4; if yes, the rectangular areas j are not similar and proceed to step P4; Step P4: When the rectangular area j is similar, the four sides of the rectangular area j are expanded outward by one unit pixel to obtain a new rectangular area j, and j=j+1, and it is determined whether the new rectangular area j is similar; if yes, repeat step P4; if no, the overlapping area between the two adjacent numbered images to be processed is the rectangular area j; When the rectangular area j is not similar, the four sides of the rectangular area j are reduced inward by one unit pixel to obtain a new rectangular area j, and j=j+1, and it is judged whether the new rectangular area j is similar; if yes, the overlapping area between the two adjacent numbered images to be processed is the new rectangular area j; if no, repeat step P4; Step P5: Overlapping the overlapping areas j between the adjacent numbered images to be processed to obtain a spliced ​​image.

[0035] For example, two adjacent images to be processed are randomly selected, and matching feature points in the images to be processed are obtained by using SIFT feature detection algorithm or other feature detection algorithms; based on the obtained feature points, feature matching algorithms such as BFMatcher, FLANN, etc. are used to match the feature points to obtain a number of matching feature points between the adjacent images to be processed; If there are 4 pairs of matching feature points between adjacent images to be processed, rectangular areas of the same area are delineated with the matching feature points as the center points, and are marked as rectangular area 0, rectangular area 1, rectangular area 2, and rectangular area 3 respectively; wherein the numbers of the rectangular areas demarcated by the matching feature points are set to be the same; taking one pair of matching feature points as an example, if the pixel difference between the rectangular areas 0 is greater than a preset threshold, the rectangular areas 0 are similar, otherwise the rectangular areas 0 are not similar; When rectangular area 0 is similar, all four sides of rectangular area 0 are expanded outward by one unit pixel to obtain a new rectangular area, which is marked as new rectangular area 1. If new rectangular area 1 is not similar, the overlapping area between the two adjacent numbered images to be processed is rectangular area 0; If the rectangular area 1 is similar, it continues to expand outward to obtain a new rectangular area 2. If the new rectangular area 2 is not similar, the overlapping area between the two adjacent numbered images to be processed is the new rectangular area 1; If the new rectangular area 2 is similar, it will continue to expand outward to obtain the new rectangular area 3, and continue to judge until the rectangular area is not similar and stop the analysis.

[0036] When rectangular area 0 is not similar, the four sides of rectangular area 0 are reduced by one unit pixel inward to obtain a new rectangular area, which is marked as new rectangular area 1. If new rectangular area 1 is similar, the overlapping area between the two adjacent numbered images to be processed is new rectangular area 1; If the new rectangular area 1 is not similar, continue to shrink inward to obtain the new rectangular area 2. If the new rectangular area 2 is similar, the overlapping area between the two adjacent numbered images to be processed is the new rectangular area 2. If rectangular area 2 is not similar, continue to shrink inward to obtain a new rectangular area 3, and continue to judge until similar rectangular areas appear and stop the analysis.

[0037] After the images to be processed are stitched together, the target image is separated from the stitched images; Specifically, standard images of the enameled wire at each process stage are extracted from the image data, pixel values ​​of the pixel points of the standard image are obtained, pixel values ​​of the pixel points of the stitched image are differenced with pixel values ​​of the standard image at the corresponding process stage, and pixels whose difference values ​​are less than a difference threshold are marked as target pixels; and the target pixels are fitted to obtain a target image.

[0038] It should be noted that the pixel values ​​of the standard images at each process stage are images of qualified products at that process stage.

[0039] The defect recognition and warning module identifies whether there are defects in the target image; if yes, a defect warning is issued and processed; if no, the subsequent process flow operation is continued; Specifically, a plurality of target images of each process flow are extracted from the historical data; wherein the target images include defective target images and non-defective target images; The target image is used as the standard input data of the convolutional neural network model, and the defect target image is used as the standard output data of the convolutional neural network model; The convolutional neural network model is trained using standard input data and standard output data to obtain a defect recognition model; wherein the defect recognition model includes a defect recognition model for the wire laying stage, a defect recognition model for the painting stage, and a defect recognition model for the baking stage.

[0040] For example: collect image data from the layout stage, painting stage, and baking stage from the historical database. These data should include defect-free and defective image samples; screen the collected images to remove blurry, duplicated, or non-compliant images.

[0041] The screened images are labeled to clearly distinguish defective target images from non-defective target images.

[0042] Convert the target image into a format suitable for convolutional neural network model input. Commonly used convolutional neural network models include AlexNet, VGG, ResNet, etc. The preprocessed target image is used as the standard input data, and the defective target image is used as the standard output data. The standard input data and standard output data are divided according to the process flow, and the convolutional neural network model is trained separately. Through iterative training, the weights and parameters of the model are continuously adjusted to minimize the loss function and improve the recognition accuracy of the model. The corresponding defect recognition model is trained.

[0043] Analyze the defect status of each production stage within the evaluation cycle and issue defect warnings; the evaluation cycle consists of several collection cycles; The number of defective target images of each process flow within the evaluation cycle is counted. If the number of defective target images exceeds the preset number, a defect warning is issued.

[0044] It should be noted that when the number of defect target images at a certain stage of a process exceeds a preset number, it indicates that there are defects in the process or there are quality problems with the product; for example, if defects occur in the wire-laying stage, there may be problems with the raw materials of the enameled wire or improper traction of the traction wheel, resulting in scratches, etc.; if a defect warning occurs in the painting stage, there may be problems with the painting operation process, resulting in uneven spraying, etc.; if a defect warning occurs in the baking stage, it is possible that the baking temperature, time or pressure settings are improper, resulting in problems such as an uneven surface and uneven color of the enameled wire.

[0045] See also Figure 3 The second aspect of the present invention proposes a detection method of an enameled wire defect detection system based on deep learning, comprising the following steps: Step 1: Collect multi-directional images of the enameled wire in each process flow and mark them as images to be processed; the process flow includes a wire laying stage, a painting stage, and a baking stage; and the image acquisition cycle of each process flow is different; Step 2: Evaluate the quality of the images to be processed in multiple directions, stitch the images after processing to obtain a stitched image; separate the target image from the stitched image; Step 3: Identify whether there are defects in the target image; if yes, issue a defect warning and handle it; if not, continue with the subsequent process flow; and analyze the defect status of each production stage within the evaluation cycle and issue a defect warning.

[0046] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0047] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. The enameled wire defect detection system based on deep learning is characterized by: It includes an image processing module, and an image acquisition module and a defect warning module connected thereto; Image acquisition module: used to collect multi-directional images of enameled wire in each process flow, and mark them as images to be processed; wherein the process flow includes wire laying stage, painting stage and baking stage; and the image acquisition cycle of each process flow is different; Image processing module: used to pre-process the multi-directional images to be processed, and then stitch the images to obtain a stitched image; and separate the target image from the stitched image; Defect recognition and warning module: used to identify whether there are defects in the target image; if yes, a defect warning is issued and processed; if no, the subsequent process flow operation is continued; and, Analyze and evaluate the defect status of each production stage during the cycle and issue defect warnings.

2. The deep learning-based enameled wire defect detection system according to claim 1, characterized in that: The multi-directional images of the enameled wire in each process flow are collected, including: The vertical plane image and side image of the enameled wire in the production stage are collected by an image acquisition tool; wherein the image acquisition tool includes a camera and an LED light; the side image is taken from the side of the copper wire; and the vertical image is taken from the vertical direction.

3. The deep learning-based enameled wire defect detection system according to claim 2, characterized in that: The image acquisition cycles of the various process flows are different, including: Collect the moving speed and pretreatment length of the enameled wire in each process; The image acquisition period TS of each process flow is calculated by the formula TS=Li / Vi-TC; wherein i is the process flow number, i=0, 1, ..., N, N is a positive integer; Li is the pretreatment length of the enameled wire in process flow i; Vi is the moving speed of the enameled wire in process flow i; TC is the shooting delay time of the camera.

4. The deep learning-based enameled wire defect detection system according to claim 3, characterized in that: The moving speed is the material conveying speed during the enameled wire production process; the pre-processing length is the maximum enameled wire length that can be processed at one time in each process flow.

5. The deep learning-based enameled wire defect detection system according to claim 2, characterized in that: The step of stitching the multi-directional images to be processed to obtain a stitched image includes: Step P1: Set the number label of the image to be processed; k, and obtain the matching feature points between two adjacent images to be processed through the feature matching algorithm; Step P2: With the matching feature point as the center point, rectangular areas of the same area are delineated and marked as rectangular area j; wherein j is the number of the rectangular area, j=0, 1, ..., M, and M is a positive integer; Step P3: Calculate the pixel difference between the rectangular areas j and determine whether the pixel value difference is greater than a preset threshold; if not, the rectangular areas j are similar and proceed to step P4; if yes, the rectangular areas j are not similar and proceed to step P4; Step P4: When the rectangular area j is similar, the four sides of the rectangular area j are expanded outward by one unit pixel to obtain a new rectangular area j, and j=j+1, and it is determined whether the new rectangular area j is similar; if yes, repeat step P4; if no, the overlapping area between the two adjacent numbered images to be processed is the rectangular area j; When the rectangular area j is not similar, the four sides of the rectangular area j are reduced inward by one unit pixel to obtain a new rectangular area j, and j=j+1, and it is judged whether the new rectangular area j is similar; if yes, the overlapping area between the two adjacent numbered images to be processed is the new rectangular area j; if no, repeat step P4; Step P5: Overlapping the overlapping areas j between the adjacent numbered images to be processed to obtain a spliced ​​image.

6. The deep learning-based enameled wire defect detection system according to claim 5, characterized in that: The order in which the number labels of the images to be processed are set is based on the positional relationship of the collected images to be processed.

7. The deep learning-based enameled wire defect detection system according to claim 5, characterized in that: The step of separating the target image from the stitched images comprises: The standard image of the enameled wire at each process stage is extracted from the image data, and the pixel values ​​of the pixel points of the standard image are obtained. The pixel values ​​of the pixel points of the stitched image are differenced with the pixel values ​​of the standard image at the corresponding process stage, and the pixel points whose difference values ​​are less than the difference threshold are marked as target pixel points; the target pixel points are fitted to obtain the target image.

8. The deep learning-based enameled wire defect detection system according to claim 1, characterized in that: The step of identifying whether there are defects in the target image includes: Extracting a number of target images of each process flow from the historical data; wherein the target images include defective target images and non-defective target images; The target image is used as the standard input data of the convolutional neural network model, and the defect target image is used as the standard output data of the convolutional neural network model; The convolutional neural network model is trained using standard input data and standard output data to obtain a defect recognition model; wherein the defect recognition model includes a defect recognition model for the wire laying stage, a defect recognition model for the painting stage, and a defect recognition model for the baking stage.

9. The deep learning-based enameled wire defect detection system according to claim 1, characterized in that: The number of defective target images of each process flow within the evaluation cycle is counted. If the number of defective target images exceeds the preset number, a defect warning is issued.

10. A method for detecting enameled wire defects based on deep learning, used to execute the operation of the enameled wire defect detection system based on deep learning according to any one of claims 1 to 9, characterized in that: The following steps are involved: Step 1: Collect multi-directional images of the enameled wire in each process flow and mark them as images to be processed; the process flow includes a wire laying stage, a painting stage, and a baking stage; and the image acquisition cycle of each process flow is different; Step 2: Evaluate the quality of the images to be processed in multiple directions, stitch the images after processing to obtain a stitched image; separate the target image from the stitched image; Step 3: Identify whether there are defects in the target image; if yes, issue a defect warning and handle it; if not, continue with the subsequent process flow, analyze and evaluate the defect status of each production stage within the cycle, and issue a defect warning.

Images