A strip passing abnormality monitoring method and system based on deep learning technology

Abstract

The application discloses a strip steel threading abnormality monitoring method based on a deep learning technology, and comprises the following steps: acquiring images of a surface of a strip steel to be detected in real time by using an imaging device, reducing the images to a specified size according to size requirements of a ResNet model; adding labels to abnormal images of the surface of the strip steel; deploying the ResNet model to a server by using an acceleration inference framework; and determining abnormalities in the images of the surface of the strip steel acquired in real time by the imaging device by using a defect detection model. The application has the following beneficial effects: 1. The application combines deep learning with hardware, intelligently judges whether an abnormal condition occurs during threading by using deep learning technology, and realizes intelligent management; 2. The application can prevent the occurrence of safety hazards by real-time monitoring, thereby further guaranteeing the safety of workers; and 3. The application improves the effect of industrial production, liberates manual supervision, and reduces the labor cost of enterprises.

Description

Technical Field

[0001] This invention relates to the field of strip steel inspection technology, specifically to a method and system for monitoring strip steel threading anomalies based on deep learning technology. Background Technology

[0002] In the steel industry, hot-rolled strip steel is a metallic material produced through multiple processes including smelting, cold rolling, hot rolling, elastic deformation, and finish rolling. This is a very complex and intricate process requiring multi-party coordination and precise control of parameters such as spatial positioning, temperature control, processing depth, and surface flatness at each step to ensure the quality and specifications of the hot-rolled strip steel meet requirements. After production, the strip steel needs to be transported from the production line to the coiler for coiling.

[0003] After the strip head is rolled by the finishing mill stand, it completes the "threading" process in the finishing mill area (successfully passing through the finishing mill stand according to the set process parameters). The strip head then enters the laminar flow cooling zone, where it is rapidly cooled by water sprayed from the upper and lower cooling manifolds while moving at high speed towards the coiler, which is tens or even hundreds of meters away. From this moment until the strip head is successfully coiled on the coiler, the strip head is in the "threading" process in the laminar flow cooling zone. Threading anomalies include two parts: abnormalities on the strip surface and tracking and detection of the strip head during the threading process. During this process, because the strip head has no traction and needs to pass through the "water wall" formed by the water sprayed from the upper and lower cooling manifolds of the laminar flow cooling, the strip head will encounter significant resistance. For thinner and softer strips, the head sometimes has difficulty completing this "threading" action smoothly, resulting in abnormalities such as the head floating, folding, or insertion into the roller table. In severe cases, the inability of the strip head to enter the coiler can cause scrap metal shutdown accidents.

[0004] In traditional tape threading anomaly detection processes, real-time manual monitoring is required. When a tape threading anomaly causes a blockage and prevents the production line from operating normally, staff will intervene manually to restore normal operation. However, this traditional manual monitoring method has several problems:

[0005] 1. Insufficient intelligence: Traditional manual monitoring is mainly carried out by human inspection, which is inefficient and lacks intelligence, failing to meet the requirements of Industry 4.0 intelligence.

[0006] 2. There are safety hazards. In the traditional manual inspection method, people are prone to fatigue. When fatigued or drowsy, staff may fail to notice abnormalities, which may lead to production stoppages and pose safety risks to staff.

[0007] 3. Lack of real-time monitoring and feedback: During manual inspection, it is impossible for people to check and work at all times. Therefore, there are omissions and untimely feedback. When a belt threading abnormality occurs, the abnormality will not be resolved for a period of time, which will lead to production stoppage.

[0008] 4. High labor costs: Each workstation requires a staff member for testing, which results in high labor costs and significant cost expenditures. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for monitoring strip threading anomalies based on deep learning technology, so as to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring strip threading anomalies based on deep learning technology, the detection method comprising:

[0011] The imaging equipment is used to acquire images of the surface of the strip steel to be inspected in real time, resulting in an image set of the surface of the strip steel to be inspected. The images in the image set are classified, and the images are scaled down to a specified size according to the size requirements of the deep learning neural network target detection model.

[0012] Labels are added to the abnormal images of the strip steel surface. First, a convolution operation is performed on the original input, followed by a max pooling operation. In the middle are 7 consecutive modules, each containing two deep learning PT models, followed by another max pooling operation. Then, the input is flattened, and finally a softmax function is applied to obtain the ResNet model for classifying the surface of the strip steel to be detected. The corresponding training weights are saved, and the training accuracy and loss function graph are used to determine whether the model has been trained well.

[0013] Deploy the PT model to the server using the accelerated inference box;

[0014] Anomalies in the strip surface images acquired in real time by the imaging equipment are identified using a defect detection model.

[0015] Furthermore, during the training of anomaly images using a deep learning neural network target detection model, the residual structure is obtained and fitted. The residual structure fitting formula is: F(x) = H(x) - x, where H(x) is the input and F(x) is the output.

[0016] Furthermore, during the fitting of the residual structure, the two 3*3 convolutional layers are replaced with 1*1+3*3+1*1.

[0017] Furthermore, the process of deploying the defect detection model to the server using the accelerated inference box refers to deserializing and parsing the defect detection model inference file to accelerate the inference process of the defect detection model, achieving a FPS of 40 to 50, enabling real-time algorithm inference, and thus completing the deployment of the model inference file to the server.

[0018] Furthermore, the process of determining anomalies in the strip surface image involves adjusting the image acquired by the imaging device and verifying it against the PT model to obtain an anomaly image.

[0019] Furthermore, the image adjustment includes scaling the image to 128*85 pixels and sequentially employing data enhancement methods such as flip transformation, rotation transformation, noise perturbation, and contrast transformation to expand the generalizability of the preprocessed image data.

[0020] A strip threading anomaly monitoring system based on deep learning technology includes:

[0021] The image acquisition module uses imaging equipment to acquire images of the surface of the strip steel to be inspected in real time, obtains an image set of the surface of the strip steel to be inspected, classifies the images in the image set, and reduces the images to a specified size according to the size requirements of the deep learning neural network target detection model.

[0022] The surface anomaly detection module adds labels to the images of anomalies on the strip surface. It first performs a convolution operation on the original input, followed by a max pooling operation. In the middle are seven consecutive modules, each containing two deep learning PT models, followed by another max pooling operation. Then, the input is flattened, and finally a softmax function is applied to obtain a ResNet model for classifying the surface of the strip to be detected. The corresponding training weights are saved, and the training accuracy and loss function graph are used to determine whether the model has been trained well.

[0023] The model improvement module uses an accelerated inference box to deploy the PT model to the server.

[0024] The alarm module uses a defect detection model to identify anomalies in the real-time images of the strip surface acquired by the imaging equipment.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. This invention combines deep learning with hardware, using deep learning technology to intelligently determine whether there are any abnormalities in the strapping, thus achieving intelligent management;

[0027] 2. This invention, through real-time monitoring, can prevent the occurrence of safety hazards, thereby further protecting the personal safety of staff.

[0028] 3. It improved the efficiency of industrial production, freed up manual supervision, and reduced labor costs for enterprises. Attached Figure Description

[0029] Figure 1 This is a flowchart of the training process for a deep learning neural network target detection model.

[0030] Figure 2 This is the overall flowchart of the strip threading monitoring system. Detailed Implementation

[0031] 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.

[0032] Please see Figure 1-2 This invention provides a technical solution: a method for monitoring strip threading anomalies based on deep learning technology, the detection method comprising:

[0033] The imaging device is used to acquire images of the surface of the strip steel to be detected in real time, and an image set of the surface of the strip steel to be detected is obtained. The images in the image set are classified. According to the size requirements of the model, the images are scaled down to a specified size. In this instance, images of different scenarios and different strip steels with abnormalities are collected. In this embodiment, 8,000 to 9,544 images are collected. In order to obtain more accurate data for training the model, the more images the better in theory.

[0034] Labels are added to the abnormal images of the strip steel surface to obtain abnormal images. First, a convolution operation is performed on the original input, followed by a max pooling operation. In the middle are 7 consecutive modules, each containing two ResNet model modules, followed by a max pooling operation. Then, the input is flattened, and finally a softmax function is applied to obtain the PT model for classifying the surface of the strip steel to be detected.

[0035] Deploying the PT model to the server using the accelerated inference box specifically refers to parsing the model inference file to generate a .trt file, and then deploying the .trt file to the server using C++. The .trt file can speed up the inference of the model file, enabling the model to achieve real-time detection.

[0036] The defect detection model identifies anomalies in the strip surface images acquired in real time by the imaging equipment. The C++ program deployed in the .trt file is packaged into a dynamic library DLL file and interfaced with the front end. The images captured by the camera are sent to the dynamic library for inference, and the inference results are sent to the front end.

[0037] Specifically, during the training of anomaly images using a deep learning neural network target detection model, the residual structure is obtained and fitted. The residual structure fitting formula is: F(x) = H(x) - x, where H(x) is the input and F(x) is the output.

[0038] Specifically, during the fitting of the residual structure, the two 3*3 convolutional layers are replaced with 1*1+3*3+1*1.

[0039] Specifically, the process of deploying the defect detection model to the server using the accelerated inference box refers to deserializing and parsing the defect detection model inference file, accelerating the inference process of the defect detection model, achieving a FPS of 40 to 50, realizing real-time algorithm inference, and thus completing the deployment of the model inference file to the server.

[0040] Specifically, the process of identifying anomalies in the strip surface image involves adjusting the image acquired by the imaging device and verifying it against the PT model to obtain an anomaly image.

[0041] Specifically, the image adjustment includes scaling the image to 128*85 pixels and sequentially using data enhancement methods such as flip transformation, rotation transformation, noise perturbation, and contrast transformation to expand the generalizability of the preprocessed image data.

[0042] This invention also provides a strip threading anomaly monitoring system based on deep learning technology, comprising:

[0043] The image acquisition module uses imaging equipment to acquire images of the surface of the strip steel to be inspected in real time, obtains an image set of the surface of the strip steel to be inspected, classifies the images in the image set, and reduces the images to a specified size according to the size requirements of the deep learning neural network target detection model.

[0044] The surface anomaly detection module adds labels to the images of anomalies on the strip surface. It first performs a convolution operation on the original input, followed by a max pooling operation. In the middle are seven consecutive modules, each containing two deep learning PT models, followed by another max pooling operation. Then, the input is flattened, and finally a softmax function is applied to obtain a ResNet model for classifying the surface of the strip to be detected. The corresponding training weights are saved, and the training accuracy and loss function graph are used to determine whether the model has been trained well.

[0045] The model improvement module uses an accelerated inference box to deploy the PT model to the server.

[0046] The alarm module uses a defect detection model to identify anomalies in the real-time images of the strip surface acquired by the imaging equipment.

[0047] In this specification, similar or identical parts in the embodiments can be referred to interchangeably. In particular, the terminal embodiments are basically similar to the method embodiments, so the description is relatively simple, and relevant parts can be referred to the description in the method embodiments.

[0048] In the embodiments provided by this invention, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0049] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0050] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0051] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should also be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A strip steel threading abnormality monitoring method based on a deep learning technique, characterized by: The detection method comprises: An image acquisition module acquires images of a surface of a strip steel to be detected in real time by using an imaging device, obtains a set of images of the surface of the strip steel to be detected, classifies pictures in the set of images, and reduces the pictures to a specified size according to a size requirement of a deep learning neural network target detection model for the pictures; A surface anomaly detection module adds labels to the surface anomaly images of the strip steel, performs a convolution operation on original input, and then adds a maximum pooling; seven continuous modules are arranged in the middle, each of the modules comprises two deep learning PT models, and then a maximum pooling is added, then a flattening operation is performed on the input, and finally a softmax function is connected to obtain a ResNet model for classifying the surface of the strip steel to be detected, and corresponding training weights are saved, and whether the model has been well trained is determined according to training accuracy and a loss function graph; A model improvement module deploys the PT model to a server by using an acceleration inference framework; An alarm module determines anomalies in the surface images of the strip steel by using a defect detection model to acquire the surface images of the strip steel in real time by the imaging device. 2.The strip running abnormality monitoring method based on deep learning technology according to claim 1, characterized in that: Residual structures are obtained in a process of training the anomaly images by using the ResNet, and the residual structures are fitted, and a fitting formula of the residual structures is F(x) = H(x) - x, wherein H(x) is input and F(x) is output. 3.The strip running abnormality monitoring method based on deep learning technology according to claim 1, characterized in that: In the process of fitting the residual structures, two 3*3 convolution layers are replaced by 1*1+3*3+1*1. 4.The strip running abnormality monitoring method based on deep learning technology according to claim 1, characterized in that: The process of deploying the defect detection model to the server by using the acceleration inference framework refers to deserializing and analyzing the defect detection model inference file, accelerating the inference process of the defect detection model, making the fps reach 40 to 50, realizing real-time inference of the algorithm, and thus completing the deployment of the model inference file to the server. 5.The strip running abnormality monitoring method based on deep learning technology according to claim 1, characterized in that: The process of determining the anomalies in the surface images of the strip steel is to adjust the pictures acquired by the imaging device, verify the PT model, and obtain anomaly pictures. 6.The strip running abnormality monitoring method based on deep learning technology according to claim 5, characterized in that: The process of adjusting the pictures comprises scaling the pictures to 128*85 pixels, and using a data enhancement mode of flip transformation, rotation transformation, noise disturbance and contrast transformation in sequence to expand the generalization of the pictures after the preprocessing.

7. A strip steel threading abnormality monitoring system based on a deep learning technique, characterized by, Comprise: An image acquisition module acquires images of a surface of a strip steel to be detected in real time by using an imaging device, obtains a set of images of the surface of the strip steel to be detected, classifies pictures in the set of images, and reduces the pictures to a specified size according to a size requirement of a deep learning neural network target detection model for the pictures; A surface anomaly detection module adds labels to the surface anomaly images of the strip steel, performs a convolution operation on original input, and then adds a maximum pooling; seven continuous modules are arranged in the middle, each of the modules comprises two deep learning PT models, and then a maximum pooling is added, then a flattening operation is performed on the input, and finally a softmax function is connected to obtain a ResNet model for classifying the surface of the strip steel to be detected, and corresponding training weights are saved, and whether the model has been well trained is determined according to training accuracy and a loss function graph; A model improvement module deploys the PT model to a server by using an acceleration inference framework; An alarm module determines anomalies in the surface images of the strip steel by using a defect detection model to acquire the surface images of the strip steel in real time by the imaging device.

Images