Visual detection method for electromagnetic wire

Abstract

The invention relates to a visual detection method for an electromagnetic wire, which belongs to the technical field of electromagnetic wire detection and comprises three main parts, namely hardware, software and an algorithm. The detection method comprises the following steps: S1, transmitting a collected picture to the industrial switch through the visual detection box; transmitting the picture to a server through a visual server for processing by an algorithm program of the server; and then the collected defect information is synchronized to the operation terminal in real time through the visual server. According to the visual detection method for the electromagnetic wire, each surface of the electromagnetic wire can be respectively shot by adopting a machine visual imaging principle, and uniform and clear images can be acquired; the images are analyzed through an efficient image processing algorithm, so that various defects are effectively found and judged, the whole process is high in speed and automatic, and the situations that the speed is low, precision is low, fatigue is prone to occurring, statistics is not available, and a large amount of waste silk is prone to being generated due to lagging caused by traditional manual inspection and spot check are avoided.

Description

Technical Field

[0001] The invention relates to the technical field of electromagnetic wire detection, in particular to a visual detection method of electromagnetic wire. Background Art

[0002] As a conductor covered with an insulating layer, electromagnetic wire is widely used in motors, transformers, new energy vehicles, photovoltaics, high-speed rail and other fields. In the production process, improper selection of wires, impurities or low purity, particles attached to the surface of the wires, uneven painting, etc., may form defects such as paint nodules, bubbles, scratches, etc. on the surface of the wires, resulting in a decrease in the pore insulation parameters of the electromagnetic wires, uneven resistance of the components, and easy breakdown.

[0003] In order to ensure the reliability of the insulation layer, the surface of the electromagnetic wire must be inspected and fed back during the production process. When large defects appear on the surface of the electromagnetic wire, the operator must be reminded in time to confirm whether it needs to be taken offline in advance.

[0004] Traditional detection methods mainly include manual inspections and random inspections by quality inspectors. These methods have disadvantages such as slow speed, low accuracy, easy fatigue, no statistics, etc., and have serious lags, which easily lead to a large amount of finished product waste and cannot meet the needs of mass production.

[0005] With the development of non-contact online detection technology, researchers have applied high-voltage pinhole detection technology to the online detection of electromagnetic wires. Although it has improved the level of automation integration and detection efficiency, it is ineffective for some special defects on the surface of the wire, such as scratches and bright spots, and there is a risk of missed detection for wires with relatively low process speeds, which affects the detection and quality control of product quality. Therefore, a visual detection method for electromagnetic wires is proposed to solve the above problems. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a visual inspection method for electromagnetic wires, which has the advantages of fast inspection speed, high inspection accuracy and easy removal of defective products. It solves the problems of the existing inspection methods such as slow speed, low accuracy, easy to generate a large amount of finished product waste and cannot meet the needs of mass production.

[0007] To achieve the above object, the present invention provides the following technical solutions: a method for detecting electromagnetic wire vision, including three main parts: hardware, software and algorithm;

[0008] The following detection method steps are also included:

[0009] S1. The collected images are transmitted to the industrial switch through the visual inspection box; the images are then transmitted to the server through the visual server for processing by the server algorithm program; and the collected defect information is then synchronized to the operation terminal in real time through the visual server;

[0010] S2. The server software filters the real-time images obtained by each visual inspection box according to the algorithm, classifies the defects, marks the size and location of the defects, and records the relevant information, and sends the defect information through MQTT;

[0011] S3, receiving MQTT defect data through client software, and displaying the defect information of each production line and issuing alarm reminders;

[0012] S4. Model making: Collect 100,000 surface images of various electromagnetic wires as training data samples.

[0013] S5. Classification and labeling: the collected samples are then classified, and the specific defect areas on the image are labeled using a defect labeling tool;

[0014] S6. Recognition training: Use the SqueezeNet model to perform classification and recognition training on the classified and labeled images. By analyzing the detection and recognition results obtained from each training, the parameters of the SqueezeNet neural network model are continuously modified and improved until the model weight is optimized;

[0015] S7. Model usage: First use the trained algorithm model for modeling, and then use the detection software to analyze and process the real-time images collected on site.

[0016] Furthermore, the visual inspection box is connected to a gigabit industrial switch via a network cable, the gigabit industrial switch is connected to a visual server via an optical fiber, and the visual server is connected to an operation terminal via a network (network cable, optical fiber or wireless).

[0017] Furthermore, the visual inspection box in S1 is mainly composed of a PLC controller, a ring light source, an adjustable screw and a camera.

[0018] Furthermore, the PLC controller is mainly used to collect the vehicle speed, meter counting and plate changing data of the production line, and the PLC controller is also used to adjust the adjustable screw of the camera.

[0019] Furthermore, the annular light source is a 45-degree annular light source, and the angle between the camera and the line direction is also 45 degrees.

[0020] Furthermore, the included angle of the adjustable screw is consistent with the angle of the camera, and the adjustable screw is mainly used for program-controlled adjustment of the focal length of the camera.

[0021] Furthermore, the camera uses a true color 4K line scan camera, which can restore the surface features of the electromagnetic wire to the greatest extent. Each visual inspection box is equipped with 4 cameras, which can monitor the surface features of the electromagnetic wire in real time from 4 directions.

[0022] Furthermore, the software includes server software and client software. The algorithm model used by the server software for real-time visual image processing is an improved lightweight convolutional neural network architecture model based on SqueezeNet, which can maintain high performance while reducing the number of model parameters and computational complexity.

[0023] Furthermore, the algorithm mainly includes model making, classification and labeling, recognition training and model use; the surface images in S4 include various defects, paint tumors, black spots, bubbles, scratches and bright spots, and the samples of each category of surface images cannot be less than 8,000.

[0024] Furthermore, the specific method of the recognition training in S6 is as follows:

[0025] S6-1. Use the improved SqueezeNet neural network model to classify and identify the defective images of electromagnetic wires, analyze the detection and identification results obtained from each training, and continuously modify and improve the parameters of the SqueezeNet neural network model to achieve model weight optimization;

[0026] S6-2, obtain an initial sample through the model in S4, and then adjust the pixels of the image to 800*5000;

[0027] S6-3. Mark the defect area and type on each image by manual classification and labeling;

[0028] S6-4. Then, the improved SqueezeNet neural network model was used for classification and recognition training, and continuous adjustment and optimization were performed, ultimately achieving an average detection accuracy of 88.5% and an average detection speed of 42 m / s per image.

[0029] Compared with the prior art, the present invention provides a method for detecting electromagnetic wire vision, which has the following features:

[0030] Beneficial effects:

[0031] 1. The electromagnetic wire visual inspection method adopts the principle of machine vision imaging to photograph each surface of the electromagnetic wire separately, and can collect uniform and clear images; the image is analyzed by an efficient image processing algorithm, so as to effectively discover and determine various defects, and calibrate the specific type, size, and location, and push the alarm to the client to remind the on-site staff. The whole process is automated at high speed, avoiding the slow speed, low accuracy, easy fatigue, no statistics, and lag caused by traditional manual inspections and random inspections, which easily generate a large amount of waste wire, and achieves the advantages of fast detection speed, high detection accuracy and easy removal of defective products.

[0032] 2. This electromagnetic wire visual inspection method can clearly restore the surface features of the electromagnetic wire by using a 4K high-definition line scan camera, record all defects on the surface of the electromagnetic wire, and quickly classify and label them through algorithms to eliminate the risk of missed inspections and improve product quality.

[0033] 3. The electromagnetic wire visual detection method uses a hardware architecture that is light on the client and heavy on the server. The on-site operation terminal can be a tablet or a workstation touch screen connected to the server, and one operation terminal can be connected to up to 5 production lines and 20 cameras, which saves a lot of space on site. In addition, since the visual processing is on the server, the virtualized server can be quickly expanded according to the actual hardware configuration requirements of the site to meet the actual production needs of the site. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A visual hardware architecture diagram of the hardware in a method for detecting electromagnetic wire vision of the present invention;

[0035] Figure 2 It is a structural schematic diagram of a visual inspection box in a method for visual inspection of electromagnetic wires of the present invention;

[0036] Figure 3 A schematic diagram of the light source and camera angles in a detection method of electromagnetic wire vision of the present invention;

[0037] Figure 4 This is a schematic diagram of sample training in an electromagnetic wire vision detection method of the present invention.

[0038] In the figure: 1. PLC controller; 2. Ring light source; 3. Adjustment screw; 4. Camera. DETAILED DESCRIPTION

[0039] The following will be combined with the embodiments of the present invention and the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present invention. 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.

[0040] Embodiment 1:

[0041] See also Figures 1 to 4 ,In this embodiment, a method for detecting electromagnetic wire vision includes three main parts of hardware, software and algorithm;

[0042] The following detection method steps are also included:

[0043] S1. The collected images are transmitted to the industrial switch through the visual inspection box; the images are then transmitted to the server through the visual server for processing by the server algorithm program; and the collected defect information is then synchronized to the operation terminal in real time through the visual server;

[0044] S2. The server software filters the real-time images obtained by each visual inspection box according to the algorithm, classifies the defects, marks the size and location of the defects, and records the relevant information, and sends the defect information through MQTT;

[0045] S3, receiving MQTT defect data through client software, and displaying the defect information of each production line and issuing alarm reminders;

[0046] S4. Model making: Collect 100,000 surface images of various electromagnetic wires as training data samples.

[0047] S5. Classification and labeling: the collected samples are then classified, and the specific defect areas on the image are labeled using a defect labeling tool;

[0048] S6. Recognition training: Use the SqueezeNet model to perform classification and recognition training on the classified and labeled images. By analyzing the detection and recognition results obtained from each training, the parameters of the SqueezeNet neural network model are continuously modified and improved until the model weight is optimized;

[0049] S7. Model usage: First use the trained algorithm model for modeling, and then use the detection software to analyze and process the real-time images collected on site.

[0050] Specifically, the visual inspection box is connected to a gigabit industrial switch via a network cable, the gigabit industrial switch is connected to a visual server via an optical fiber, and the visual server is connected to an operation terminal via a network (network cable, optical fiber or wireless).

[0051] In this embodiment, the visual inspection box in S1 is mainly composed of a PLC controller 1, a ring light source 2, an adjustable screw 3 and a camera 4.

[0052] Specifically, the PLC controller 1 is mainly used to collect the vehicle speed, metering and plate changing data of the production line. The PLC controller 1 is also used to adjust the adjustable screw 3 equipped with the camera 4.

[0053] Specifically, the annular light source 2 is a 45-degree annular light source, and the angle between the camera 4 and the line direction is also 45 degrees.

[0054] Specifically, the included angle of the adjustable screw rod 3 is consistent with the angle of the camera 4 , and the adjustable screw rod 3 is mainly used for adjusting the focal length of the camera 4 by program control.

[0055] Specifically, the camera 4 uses a true color 4K line scan camera, which can restore the surface features of the electromagnetic wire to the greatest extent. Each visual inspection box is equipped with 4 cameras 4, which can monitor the surface features of the electromagnetic wire in real time from 4 directions.

[0056] In this embodiment, the software includes server software and client software. The algorithm model used by the server software for real-time visual image processing is an improved lightweight convolutional neural network architecture model based on SqueezeNet, which can maintain high performance while reducing the number of model parameters and computational complexity.

[0057] In this embodiment, the algorithm mainly includes model making, classification and labeling, recognition training and model use; the surface images in S4 include various defects, paint tumors, black spots, bubbles, scratches and bright spots, and the samples of each category of surface images cannot be less than 8,000.

[0058] Specifically, the specific method of recognition training in S6 is as follows:

[0059] S6-1. Use the improved SqueezeNet neural network model to classify and identify the defective images of electromagnetic wires, analyze the detection and identification results obtained from each training, and continuously modify and improve the parameters of the SqueezeNet neural network model to achieve model weight optimization;

[0060] S6-2, obtain an initial sample through the model in S4, and then adjust the pixels of the image to 800*5000;

[0061] S6-3. Mark the defect area and type on each image by manual classification and labeling;

[0062] S6-4. Then, the improved SqueezeNet neural network model was used for classification and recognition training, and continuous adjustment and optimization were performed, ultimately achieving an average detection accuracy of 88.5% and an average detection speed of 42 m / s per image.

[0063] It should be noted that the server-side algorithm in this embodiment can also be adjusted to the Pytorch-based target detection tool library MMDetection. The target detection tool library contains a wealth of target detection, instance segmentation algorithms and related components. In addition, it can also support the deep learning training library MMEngine and the computer vision basic library MMCV. It not only achieves the best balance between parameter quantity and accuracy in target detection models from tiny to extra-large sizes, but also achieves the most advanced results in real-time instance segmentation and rotation target detection tasks.

[0064] The beneficial effects of the present invention are:

[0065] 1. The electromagnetic wire visual inspection method adopts the principle of machine vision imaging to photograph each surface of the electromagnetic wire separately, and can collect uniform and clear images; the image is analyzed by an efficient image processing algorithm, so as to effectively discover and determine various defects, and calibrate the specific type, size, and location, and push the alarm to the client to remind the on-site staff. The whole process is automated at high speed, avoiding the slow speed, low accuracy, easy fatigue, no statistics, and lag caused by traditional manual inspections and random inspections, which easily generate a large amount of waste wire, and achieves the advantages of fast detection speed, high detection accuracy and easy removal of defective products.

[0066] 2. This electromagnetic wire visual inspection method can clearly restore the surface features of the electromagnetic wire by using a 4K high-definition line scan camera, record all defects on the surface of the electromagnetic wire, and quickly classify and label them through algorithms to eliminate the risk of missed inspections and improve product quality.

[0067] 3. The electromagnetic wire visual detection method uses a hardware architecture that is light on the client and heavy on the server. The on-site operation terminal can be a tablet or a workstation touch screen connected to the server, and one operation terminal can be connected to up to 5 production lines and 20 cameras, which saves a lot of space on site. In addition, since the visual processing is on the server, the virtualized server can be quickly expanded according to the actual hardware configuration requirements of the site to meet the actual production needs of the site.

[0068] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting electromagnetic wire vision, characterized in that: It includes three main parts: hardware, software and algorithm; The following detection method steps are also included: S1. The collected images are transmitted to the industrial switch through the visual inspection box; the images are then transmitted to the server through the visual server for processing by the server algorithm program; and the collected defect information is then synchronized to the operation terminal in real time through the visual server; S2. The server software filters the real-time images obtained by each visual inspection box according to the algorithm, classifies the defects, marks the size and location of the defects, and records the relevant information, and sends the defect information through MQTT; S3, receiving MQTT defect data through client software, and displaying the defect information of each production line and issuing alarm reminders; S4. Model making: Collect 100,000 surface images of various electromagnetic wires as training data samples. S5. Classification and labeling: the collected samples are then classified, and the specific defect areas on the image are labeled using a defect labeling tool; S6. Recognition training: Use the SqueezeNet model to perform classification and recognition training on the classified and labeled images. By analyzing the detection and recognition results obtained from each training, the parameters of the SqueezeNet neural network model are continuously modified and improved until the model weight is optimized; S7. Model usage: First use the trained algorithm model for modeling, and then use the detection software to analyze and process the real-time images collected on site.

2. The electromagnetic wire visual detection method according to claim 1, characterized in that: The visual inspection box is connected to a gigabit industrial switch via a network cable, the gigabit industrial switch is connected to a visual server via an optical fiber, and the visual server is connected to an operation terminal via a network (network cable, optical fiber or wireless).

3. The electromagnetic wire visual detection method according to claim 1, characterized in that: The visual inspection box in S1 is mainly composed of a PLC controller (1), a ring light source (2), an adjustable screw rod (3) and a camera (4).

4. The electromagnetic wire visual detection method according to claim 3, characterized in that: The PLC controller (1) is mainly used to collect vehicle speed, metering and plate changing data of the production line. The PLC controller (1) is also used to adjust the adjustable screw (3) carrying the camera (4).

5. The electromagnetic wire visual detection method according to claim 3, characterized in that: The annular light source (2) is a 45-degree annular light source, and the angle between the camera (4) and the line direction is also 45 degrees.

6. The electromagnetic wire visual detection method according to claim 3, characterized in that: The included angle of the adjustable screw rod (3) is consistent with the angle of the camera (4), and the adjustable screw rod (3) is mainly used for program-controlled adjustment of the focal length of the camera (4).

7. The electromagnetic wire visual detection method according to claim 3, characterized in that: The camera (4) adopts a true color 4K line scan camera, which can restore the surface features of the electromagnetic wire to the greatest extent. Each visual inspection box is equipped with 4 cameras (4), which can monitor the surface features of the electromagnetic wire in real time from 4 directions.

8. The electromagnetic wire visual detection method according to claim 1, characterized in that: The software includes server software and client software. The algorithm model used by the server software for real-time visual image processing is an improved lightweight convolutional neural network architecture model based on SqueezeNet, which can maintain high performance while reducing the number of model parameters and computational complexity.

9. The electromagnetic wire visual detection method according to claim 1, characterized in that: The algorithm mainly includes model making, classification and labeling, recognition training and model use; the surface images in S4 contain various defects, paint tumors, black spots, bubbles, scratches and bright spots, and the samples of each category of surface images cannot be less than 8,000.

10. The electromagnetic wire visual detection method according to claim 1, characterized in that: The specific method of the recognition training in S6 is as follows: S6-1. Use the improved SqueezeNet neural network model to classify and identify the defective images of electromagnetic wires, analyze the detection and identification results obtained from each training, and continuously modify and improve the parameters of the SqueezeNet neural network model to achieve model weight optimization; S6-2, obtain an initial sample through the model in S4, and then adjust the pixels of the image to 800*5000; S6-3. Mark the defect area and type on each image by manual classification and labeling; S6-4. Then, the improved SqueezeNet neural network model was used for classification and recognition training, and continuous adjustment and optimization were performed, ultimately achieving an average detection accuracy of 88.5% and an average detection speed of 42 m / s per image.

Images