A method for identifying coil shape of high-speed wire laying machine based on computer vision

Abstract

The present invention relates to a method for identifying the coil type of a high-speed wire laying machine based on computer vision, wherein the method comprises the following steps: collecting grayscale images of a thermal imager of a high-speed wire laying machine; manually annotating the grayscale images to prepare a high-speed wire rolling production line identification training set; constructing and training a high-speed wire rolling production line identification model; using the trained high-speed wire rolling production line identification model to identify the rolling production line, and using the identification result to cut and save the original image; manually annotating the cut images of the high-speed wire rolling production line to mark different coil types, and preparing a coil type identification training set for the laying machine; constructing and training a coil type identification model for the laying machine; and using the trained high-speed wire rolling production line identification model and coil type identification model to identify the coil type of the laying machine. Rapid and accurate identification of the coil type of the high-speed wire laying machine is achieved.

Description

Technical Field

[0001] The invention relates to the technical fields of artificial intelligence, computer vision and image recognition, and in particular to a method for identifying the coil type of a high-speed wire laying machine based on computer vision. Background Art

[0002] High-speed wire rolling mill is a comprehensive product of metallurgical technology, electronic control technology and mechanical manufacturing technology. On the high-speed wire production line, after the wire is rolled, it needs to be spun into coils by the wire laying machine to complete the transformation from straight wire to coil. With the increase of wire rolling speed and output, the requirements for the coil shape of the wire laying machine are getting higher and higher. If the problem of unstable coil shape occurs during production, it will affect the product quality. Summary of the invention

[0003] The object of the present invention is to provide a method for identifying the coil type of a high-speed wire laying machine based on computer vision to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for identifying the coil type of a high-speed wire laying machine based on computer vision, the method comprising the following steps in sequence:

[0005] 1) Collect grayscale images of the thermal imager of the high-speed wire laying machine;

[0006] 2) Manually annotate the grayscale images and create a high-speed wire rolling production line identification training set;

[0007] 3) Improve the SSD network structure and build a high-speed wire rolling production line identification model;

[0008] 4) Training the high-speed wire rolling production line identification model;

[0009] 5) Using the trained high-speed wire rolling production line recognition model, identify the rolling production line, and use the recognition result to crop and save the original image, which is the original image including the production line and the coil type;

[0010] 6) Manually label the images of the high-speed wire rolling production line after cutting, mark the different coil types, and make a coil type recognition training set for the wire laying machine;

[0011] 7) Use the YOLO_V3 network structure to build a spinning machine coil type recognition model;

[0012] 8) Training a spinning machine coil type recognition model;

[0013] 9) Use the trained high-speed wire rolling production line recognition model and coil type recognition model to identify the coil type of the wire laying machine.

[0014] Preferably, the method for improving the SSD network structure is: the backbone network adopts VGG-16, replaces its last two fully connected layers with convolutional layers Conv6 and Conv7, and adds 4 convolutional layers to the replaced network, namely Conv8_2, Conv9_2, Conv10_2, and Pool, to build an SSD detection network.

[0015] Preferably, the Conv6 is obtained by: Conv5_3 in VGG-16, through a convolution layer with a convolution kernel size of 3*3*256 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*512 and a step size of 1;

[0016] Conv7: obtained by Conv6 through a convolution layer with a convolution kernel size of 1*1*512 and a step size of 1;

[0017] Conv8_2: obtained by Conv7 through a convolution layer with a convolution kernel size of 1*1*128 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*256 and a step size of 2;

[0018] Conv9_2: obtained by Conv8_2 through a convolution layer with a convolution kernel size of 1*1*64 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*128 and a step size of 2;

[0019] Conv10_2: obtained by Conv9_2 through a convolution layer with a convolution kernel size of 1*1*64 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*128 and a step size of 2;

[0020] The Pool is obtained by Conv10_2 through a convolution layer with a convolution kernel size of 1*1*64 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*128 and a step size of 2.

[0021] Preferably, the method for training the high-speed wire rolling production line recognition model is: using Windows 10 system, Intel Core i9-10900X processor, NVIDIA GeForce RTX3090 graphics card, 128G memory. Based on the PyTorch framework, the training process sets the batch size to 16, the input image size to 660*440, the momentum to 0.9, the initial learning rate to 0.0005, and the learning rate will be reduced to 30% of the original after every 5 epochs of training, with a total of 60 iterations.

[0022] Preferably, the spinning machine coil type recognition model includes three parts: feature extraction, feature fusion and detection head. The feature extraction adopts Darknet-53 as the feature extraction network, including a DBL module and a res module. The res module is repeated 1 time, 2 times, 8 times, 8 times and 4 times respectively. The DBL module includes a convolutional layer, a batch normalization layer and a Leaky ReLU activation function. The res module includes two DBL modules and a residual connection module.

[0023] The feature fusion part includes a DEL module and an upsampling module. The deep features are fused with the shallow features through the DBL module and the upsampling module, and finally the deep high-semantic features and the shallow high-resolution features are fused.

[0024] The detection head is composed of a fully convolutional structure. The features of the detection coil type and position are fused, and then passed through the DBL module and convolution layer to generate detection results at multiple scales. The scales of the detection heads P1, P2, and P3 are 14*14, 28*28, and 56*56, respectively.

[0025] Preferably, the method for training the spinning machine coil type recognition model is: using Windows 10 system, Intel Core i9-10900X processor, NVIDIA GeForce RTX3090 graphics card, 128G memory, based on PyTorch framework, setting the batch size to 16 during the training process, the input image size to 448*448, the momentum to 0.9, using cosine learning rate, the initial learning rate to 0.001, the end learning rate to 0.00001, and a total of 100 iterations.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. It can replace manual work and realize the coil type recognition of high-speed wire laying machine;

[0028] 2. The pixels of the coil shape of the high-speed wire laying machine account for a small proportion of the pixels of the entire image. By adopting a cascade method, the high-speed wire rolling production line and the coil shape of the laying machine are identified separately to improve the accuracy of model detection;

[0029] 3. High-speed wire rolling production line identification deletes the shallow feature detection layer, pays more attention to deep features and global receptive field, implements feature extraction in a targeted manner, and at the same time designs the model in a lightweight manner to improve the detection speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of the method of the present invention;

[0031] Figure 2The overall architecture diagram of the method of the present invention is implemented;

[0032] Figure 3 This is a diagram of the SSD network structure improved by the method of the present invention;

[0033] Figure 4 This is the network structure diagram of the YOLO_V3 method of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose and technical solution of the present invention clearly and completely described, and the advantages more clearly understood, the embodiments of the present invention are further described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, rather than all of the embodiments, and are only used to explain the embodiments of the present invention, and are not used to limit 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.

[0035] See also Figures 1 to 4 The present invention provides a method for identifying the coil type of a high-speed wire laying machine based on computer vision, and the method comprises the following steps in sequence:

[0036] 1) Collect grayscale images of the high-speed wire laying machine thermal imager.

[0037] 2) Manually label the grayscale images to create a high-speed wire rolling production line identification training set. Specifically, use labeling tools such as LabelMe to label the grayscale images of the high-speed wire laying machine thermal imager, and use rectangular boxes to label the positions of the coil production line in the entire grayscale image to create a training set; preprocess the training set, and the preprocessing methods include uniform resizing to 660*440, random cropping and normalization.

[0038] 3) According to Figure 3 The improved SSD network structure diagram shown is used to build a high-speed wire rolling production line identification model.

[0039] The high-speed wire rolling production line identification model adopts an improved SSD network structure. The backbone network uses VGG-16, and its last two fully connected layers are replaced with convolutional layers. Four convolutional layers are added to the replaced network to build an SSD detection network. Since the high-speed wire rolling production line identification pays more attention to deep features and global receptive fields, the detection layer of shallow features in the original SSD network is deleted, and the detection results of the four convolutional layers Conv7, Conv8_2, Conv9_2, and Conv10_2 are used as the detection output results. The feature map sizes of the four convolutional layers are 42*28, 21*14, 11*7, and 6*4, respectively. Since the high-speed wire rolling production line identification is relatively simple, the original SSD network is designed to be lightweight. The improved SSD network structure is as follows: Figure 3 shown.

[0040] Input image: The input image size is 640*440*3.

[0041] Improved SSD network: The network structure is lightweight and designed to improve detection speed. The last two fully connected layers in the VGG-16 network are replaced with convolutional layers, that is, FC6 is replaced with Conv6, and FC7 is replaced with Conv7. The network structure after lightweight design and replacement is as follows:

[0042] Conv6: Conv5_3 in VGG-16, obtained by a convolutional layer with a kernel size of 3*3*256 and a stride of 1, and a convolutional layer with a kernel size of 3*3*512 and a stride of 1.

[0043] Conv7: obtained by Conv6 through a convolutional layer with a kernel size of 1*1*512 and a stride of 1.

[0044] Conv8_2: obtained by Conv7 through a convolutional layer with a kernel size of 1*1*128 and a stride of 1, and a convolutional layer with a kernel size of 3*3*256 and a stride of 2.

[0045] Conv9_2: obtained by Conv8_2 through a convolutional layer with a kernel size of 1*1*64 and a stride of 1, and a convolutional layer with a kernel size of 3*3*128 and a stride of 2.

[0046] Conv10_2: obtained by Conv9_2 through a convolutional layer with a kernel size of 1*1*64 and a stride of 1, and a convolutional layer with a kernel size of 3*3*128 and a stride of 2.

[0047] Pool: obtained by Conv10_2, a convolutional layer with a kernel size of 1*1*64 and a stride of 1, and a convolutional layer with a kernel size of 3*3*128 and a stride of 2.

[0048] 4) Train the high-speed wire rolling production line recognition model. Use Windows 10 system, Intel Core i9-10900X processor, NVIDIA GeForce RTX3090 graphics card, 128G memory. Based on the PyTorch framework, the training process sets the batch size to 16, the input image size to 660*440, the momentum to 0.9, the initial learning rate to 0.0005, and the learning rate will be reduced to 30% of the original after every 5 epochs of training, with a total of 60 iterations.

[0049] 5) Using the trained high-speed wire rolling production line recognition model, identify the rolling production line, and use the recognition result to crop and save the original image, which includes the original image of the production line and the coil type.

[0050] 6) Manually label the images cut from the high-speed wire rolling production line, mark different coil types, and prepare a training set for the coil type recognition of the spinning machine. Specifically, use a labeling tool such as LabelMe to label the image obtained in step 5 with a rectangular frame to prepare a training set for the coil type recognition of the spinning machine, and preprocess the training set. The preprocessing method includes uniformly resizing to 448*448, random cropping, and normalization.

[0051] 7) Construct a coil type recognition model for the spinning machine. Figure 4 The YOLO_V3 network structure diagram is shown to build a spinning machine coil type recognition model.

[0052] The input image size is 448*448*3.

[0053] The spinning machine coil type recognition model includes three parts: feature extraction, feature fusion and detection head. Feature extraction: Darknet-53 is used as the feature extraction network, including DBL module and res module repeated 1, 2, 8, 8 and 4 times respectively. The DBL module contains convolution layer, batch normalization layer and Leaky ReLU activation function. The res module contains two DBL modules and residual connection module.

[0054] The feature fusion part includes a DEL module and an upsampling module. The deep features are fused with the shallow features through the DBL module and the upsampling module, and finally the deep high-semantic features and the shallow high-resolution features are fused.

[0055] The detection head is composed of a fully convolutional structure. The features of the detection coil type and position are fused, and then passed through the DBL module and convolution layer to generate detection results at multiple scales. The scales of the detection heads P1, P2, and P3 are 14*14, 28*28, and 56*56, respectively.

[0056] 8) Train the spinning machine coil type recognition model. Use Windows 10 system, Intel Core i9-10900X processor, NVIDIA GeForce RTX3090 graphics card, 128G memory. Based on the PyTorch framework, the training process sets the batch size to 16, the input image size to 448*448, the momentum to 0.9, the cosine learning rate, the initial learning rate to 0.001, the final learning rate to 0.00001, and a total of 100 iterations.

[0057] 9) Use the trained high-speed wire rolling production line recognition model and coil type recognition model to identify the coil type of the wire laying machine.

[0058] Although the above describes the illustrative specific implementation methods of the present application so that technicians in this technical field can understand the present application, the present application is not limited to the scope of the specific implementation methods. For ordinary technicians in this technical field, as long as various changes are within the spirit and scope of the present application defined and determined by the attached claims, all application creations using the concept of the present application are protected.

Claims

1. A method for identifying the coil shape of a high-speed wire laying machine based on computer vision, characterized in that: The method comprises the following steps in sequence: 1) Collect grayscale images of the thermal imager of the high-speed wire laying machine; 2) Manually annotate the grayscale images and create a high-speed wire rolling production line identification training set; 3) Improve the SSD network structure and build a high-speed wire rolling production line identification model; 4) Training the high-speed wire rolling production line identification model; 5) Using the trained high-speed wire rolling production line recognition model, identify the rolling production line, and use the recognition result to crop and save the original image, which is the original image including the production line and the coil type; 6) Manually label the images of the high-speed wire rolling production line after cutting, mark the different coil types, and make a coil type recognition training set for the wire laying machine; 7) Use the YOLO_V3 network structure to build a spinning machine coil type recognition model; 8) Training a spinning machine coil type recognition model; 9) Use the trained high-speed wire rolling production line recognition model and coil type recognition model to identify the coil type of the wire laying machine.

2. The method for identifying coil type of a high-speed wire laying machine based on computer vision according to claim 1, characterized in that: The method to improve the SSD network structure is as follows: the backbone network adopts VGG-16, replaces its last two fully connected layers with convolutional layers Conv6 and Conv7, and adds 4 convolutional layers to the replaced network, namely Conv8_2, Conv9_2, Conv10_2, and Pool, to build the SSD detection network.

3. The method for identifying coil type of a high-speed wire laying machine based on computer vision according to claim 2, characterized in that: Conv6: obtained by Conv5_3 in VGG-16 through a convolution layer with a convolution kernel size of 3*3*256 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*512 and a step size of 1; Conv7: obtained by Conv6 through a convolution layer with a convolution kernel size of 1*1*512 and a step size of 1; Conv8_2: obtained by Conv7 through a convolution layer with a convolution kernel size of 1*1*128 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*256 and a step size of 2; Conv9_2: obtained by Conv8_2 through a convolution layer with a convolution kernel size of 1*1*64 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*128 and a step size of 2; Conv10_2: obtained by Conv9_2 through a convolution layer with a convolution kernel size of 1*1*64 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*128 and a step size of 2; The Pool is obtained by Conv10_2 through a convolution layer with a convolution kernel size of 1*1*64 and a step size of 1, and a convolution layer with a convolution kernel size of 3*3*128 and a step size of 2.

4. The method for identifying coil type of a high-speed wire laying machine based on computer vision according to claim 1, characterized in that: The method for training the high-speed wire rolling production line recognition model is as follows: using Windows 10 system, Intel Corei9-10900X processor, NVIDIA GeForce RTX3090 graphics card, 128G memory. Based on the PyTorch framework, the training process sets the batch size to 16, the input image size to 660*440, the momentum to 0.9, the initial learning rate to 0.0005, and the learning rate will be reduced to 30% of the original after every 5 epochs of training, with a total of 60 iterations.

5. The method for identifying coil type of a high-speed wire laying machine based on computer vision according to claim 1, characterized in that: The spinning machine coil type recognition model includes three parts: feature extraction, feature fusion and detection head. The feature extraction uses Darknet-53 as the feature extraction network, including a DBL module and a res module. The res module is repeated 1 time, 2 times, 8 times, 8 times and 4 times respectively. The DBL module includes a convolutional layer, a batch normalization layer and a Leaky ReLU activation function. The res module includes two DBL modules and a residual connection module. The feature fusion part includes a DEL module and an upsampling module. The deep features are fused with the shallow features through the DBL module and the upsampling module, and finally the deep high-semantic features and the shallow high-resolution features are fused. The detection head is composed of a fully convolutional structure. The features of the detection coil type and position are fused, and then passed through the DBL module and convolution layer to generate detection results at multiple scales. The scales of the detection heads P1, P2, and P3 are 14*14, 28*28, and 56*56, respectively.

6. The method for identifying coil type of a high-speed wire laying machine based on computer vision according to claim 1, characterized in that: The method for training the spinning machine coil type recognition model is as follows: using Windows 10 system, Intel Core i9-10900X processor, NVIDIA GeForce RTX3090 graphics card, 128G memory, based on the PyTorch framework, the training process sets the batch size to 16, the input image size to 448*448, the momentum to 0.9, the cosine learning rate, the initial learning rate to 0.001, the end learning rate to 0.00001, and a total of 100 iterations.

Images