A method and system for tracking cast billets
By processing image data from the casting and rolling operation area using edge computing and deep learning models, the problems of low tracking accuracy and limited information in billet tracking technology have been solved. This has enabled multi-dimensional information acquisition and precise tracking of billets, improving the level of precision in production management and the efficiency of handling anomalies.
Patent Information
- Application Number
- CN202611403237.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing billet tracking technologies suffer from low tracking accuracy, limited information dimensions, and a lack of objective data. They struggle to accurately identify the end face shape, surface quality, and movement trajectory of billets in high-temperature, dusty, and vibrating industrial environments, and lack real-time monitoring and historical data management capabilities.
A deep learning model on an edge computing unit is used to preprocess and analyze the image data of the casting and rolling operation area to obtain the end face recognition result, shape recognition result, motion detection result, displacement detection result and arrangement recognition result of the billet. Combined with the status identification information, identity matching and trajectory tracking are performed to form tracking result data.
It enables the acquisition and precise tracking of multi-dimensional information on billets in harsh environments, providing objective data to support refined management and anomaly handling in casting and rolling production.
Smart Images

Figure CN123118700A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of metallurgy and image recognition technology, and in particular to a billet tracking method and system. Background Technology
[0002] In steel production, real-time tracking and monitoring of billet trajectories is a core element in ensuring production stability, product quality, and production efficiency. Casting and rolling production lines typically involve multiple billets simultaneously undergoing different processes (such as heating, rolling, cooling, cutting, and unloading), with a fast production pace and harsh operating conditions (high temperature, dust, vibration), placing extremely high demands on the real-time performance, reliability, and continuity of the tracking system. As the steel industry transforms and upgrades towards intelligent and digital processes, the need for precise tracking of billets throughout the entire process from continuous casting to rolling is becoming increasingly urgent.
[0003] However, the billet tracking technologies widely used by steel companies are still mainly traditional, including mechanical sensor tracking and hot metal detector tracking. Both have significant shortcomings. Mechanical sensors (such as limit switches and proximity switches) rely on contact triggering and are prone to wear, jamming, and fatigue damage in the high-temperature, dusty, and vibrating industrial environment, resulting in high maintenance costs and limited lifespan. More importantly, these sensors can only output discrete switching signals and cannot perceive visual information such as the end face shape and surface quality of the billet. While hot metal detectors achieve non-contact detection, they are essentially fixed-point presence detection, only outputting logic level signals at fixed installation locations and unable to acquire continuous trajectory information between adjacent detection points. Especially when multiple billets are densely arranged or end-to-end, these devices struggle to distinguish the boundaries and identities of adjacent billets, easily leading to tracking loss or sequence number errors. Furthermore, existing tracking systems generally lack the ability to integrate with real-time monitoring screens, preventing operators from directly viewing images of the casting and rolling area for manual verification. Simultaneously, the systems lack effective storage and management mechanisms for historical results after tracking is completed. When quality disputes or production anomalies occur, it is impossible to trace the complete movement trajectory, offline status, and related images of a specific billet, resulting in a lack of objective and traceable data. This severely restricts the level of precision in production management and the efficiency of anomaly handling. Clearly, a new billet tracking method is urgently needed to address at least one of these problems.
[0004] It should be noted that the above content only provides background information related to this application and does not necessarily constitute prior art. Summary of the Invention
[0005] This application provides a billet tracking method and system to solve the technical problems of low tracking accuracy, single information dimension and lack of objective data basis in related billet tracking technologies.
[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.
[0007] This application provides a billet tracking method, comprising: acquiring initial image data and status identification information of a billet in a casting and rolling operation area; preprocessing the initial image data using a preprocessing algorithm deployed on an edge computing unit to obtain preprocessed image data; processing and analyzing the preprocessed image data using a deep learning model deployed on the edge computing unit to obtain a set of detection results for the billet, wherein the set of detection results includes at least one of end face recognition results, shape recognition results, motion detection results, displacement detection results, and arrangement recognition results; and performing identity matching and trajectory tracking on the billet according to a preset tracking rule based on the set of detection results and the status identification information to obtain tracking result data and complete the billet tracking.
[0008] In one embodiment of this application, based on the aforementioned scheme, if the detection result set includes the end face recognition result, and the deep learning model includes an end face detection model, the method for obtaining the end face recognition result includes: inputting the preprocessed image data into the end face detection model, locating the end face region, and performing continuous feature detection on the end face region; if discontinuous features are detected, it is determined that there is a tear on the end face of the billet corresponding to the end face region, and the end face recognition result is output; wherein, the end face detection model is obtained by training a preset end face detection network through a first training set, and the end face detection network learns the feature differences between the normal end face shape and the torn end face shape through image feature extraction, and the first training set is billet image data with end face annotation information, wherein the end face annotation information annotates the normal area and the torn area of the end face.
[0009] In one embodiment of this application, based on the aforementioned scheme, if the detection result set includes the morphology recognition result, and the deep learning model includes a first contour segmentation model, the method for obtaining the morphology recognition result includes: inputting the preprocessed image data into the first contour segmentation model and outputting a first segmentation mask for the billet region; extracting contour geometric parameters based on the first segmentation mask, wherein the contour geometric parameters include at least the contour area, contour perimeter, and aspect ratio of the billet; comparing the contour geometric parameters with a preset normal threshold range, and if any parameter in the contour geometric parameters exceeds the normal threshold range, then determining that the morphology recognition result is an abnormal billet morphology, otherwise determining that the morphology recognition result is a normal billet morphology; wherein, the first contour segmentation model is obtained by training a preset first contour segmentation network through a second training set, the first contour segmentation network learns the morphological features of the billet contour through training, and outputs the first segmentation mask based on the input preprocessed image data, the second training set being billet image data with billet contour annotation information, wherein the billet contour annotation information annotates the contour boundary of the billet in the image.
[0010] In one embodiment of this application, based on the foregoing scheme, if the detection result set includes the motion detection result and the deep learning model includes an object detection model, the method for obtaining the motion detection result includes: inputting the preprocessed image data into the object detection model, performing object detection on the slab in the image, and outputting a first object detection result, the first object detection result including the slab bounding box and position information; according to a preset image acquisition cycle, acquiring the original image data of the slab again at the next acquisition time after outputting the first object detection result, and performing the preprocessing on the original image data to obtain processed image data; inputting the processed image data into the object detection model, and performing object detection on the slab in the image... The system performs target detection and outputs a second target detection result, which includes the bounding box and position information of the billet. Based on the first and second target detection results, the system determines the movement of the billet within the image acquisition period to obtain the motion detection result, which includes position and velocity changes. The target detection model is trained on a preset target detection network using a third training set. The target detection network learns the mapping relationship between the target detection features of the billet and the image content through image feature extraction. The third training set consists of billet image data with target annotation information, which annotates the bounding box and position information of the billet in the image.
[0011] In one embodiment of this application, based on the aforementioned scheme, if the detection result set includes the displacement detection result, and the deep learning model includes a second contour segmentation model, the method for obtaining the displacement detection result includes: inputting the preprocessed image data into the second contour segmentation model and outputting a second segmentation mask for the billet region; extracting billet displacement information based on the second segmentation mask, the billet displacement information including the billet movement distance and the billet movement trajectory; determining the displacement detection result between adjacent detection times based on the billet displacement information; wherein, the second contour segmentation model is obtained by training a preset second contour segmentation network through a fourth training set, the second contour segmentation network learns the segmentation features of the billet contour through image feature extraction, and outputs the second segmentation mask based on the input preprocessed image data, the fourth training set being billet image data with billet start and end position annotation information, the billet start and end position annotation information annotating the position information of the billet in the image at the detection start time and end time.
[0012] In one embodiment of this application, based on the foregoing scheme, if the detection result set includes the arrangement recognition result, and the deep learning model includes a third contour segmentation model, the method for obtaining the arrangement recognition result includes: inputting the preprocessed image data into the third contour segmentation model, and outputting a third segmentation mask for the casting area; based on the third segmentation mask, identifying the sequence number information of each casting in the image, the sequence number information including the arrangement order of each casting in the image and the corresponding sequence number identifier; according to the sequence number information, obtaining the first arrangement order of the casting when entering the detection port and the second arrangement order when leaving the detection port; if the first arrangement order is consistent with the... If the second arrangement order is consistent, the arrangement recognition result is determined to be a normal billet arrangement; if the first arrangement order is inconsistent with the second arrangement order, the arrangement recognition result is determined to be an abnormal billet arrangement; wherein, the third contour segmentation model is obtained by training a preset third contour segmentation network on a fifth training set, the third contour segmentation network learns the segmentation features of the billet serial number through image feature extraction, and outputs the third segmentation mask based on the input preprocessed image data, the fifth training set is billet image data with billet serial number annotation information, the billet serial number annotation information annotates the serial number identifier and arrangement order of each billet in the image.
[0013] In one embodiment of this application, based on the aforementioned scheme, based on the detection result set and the status identification information, the casting billet is matched for identity and tracked according to a preset tracking rule to obtain tracking result data, including: assigning a corresponding virtual casting billet number to each casting billet based on the detection result set and the status identification information; updating the real-time position and path of each casting billet according to preset position rules and preset queue rules based on the virtual casting billet number to form a continuous movement trajectory of each casting billet; assigning an abnormal marking state to each casting billet based on the comparison result of the first virtual casting billet number when entering the preset verification area and the second virtual casting billet number when leaving the preset verification area; generating tracking details for each casting billet based on the status identification information, combined with the continuous movement trajectory and abnormal marking state of each casting billet, and outputting a hot delivery record or offline record according to the hot delivery status or offline status in the status identification information, thereby obtaining the tracking result data.
[0014] In one embodiment of this application, based on the foregoing scheme, the method further includes: deploying sensors and image acquisition devices at key nodes on the casting and rolling production line to collect various types of data in real time; constructing multiple data identification queues according to time sequence and data type; adding a timestamp and device identifier to each data; using a data fusion algorithm to concatenate the data in each data identification queue; establishing data connections based on identification logic and time consistency to form a complete dataset; and extracting features from the data in the complete dataset to obtain feature vectors; inputting the feature vectors into an anomaly detection model to obtain an anomaly detection result, wherein the anomaly detection model is trained on a preset anomaly detection network using training samples, and the training samples are historical data containing normal and abnormal states; if the anomaly detection result is an anomaly, generating warning information containing the time and location of the anomaly occurrence, and triggering a preset offline processing.
[0015] In one embodiment of this application, based on the foregoing scheme, the method further includes: real-time monitoring of the image quality of the initial image data, the device status of the preset device, and the data synchronization of the preset data; determining the fault level based on the image quality, the device status, and the data synchronization; if the fault level is a level one fault, adjusting the visual processing parameters; if the fault level is a level two fault, activating the backup image acquisition device and correcting the positioning deviation by adjusting the roller speed; if the fault level is a level three fault, triggering the mill to slow down or partially stop, and maintaining tracking continuity in conjunction with the positioning device.
[0016] This application also provides a billet tracking system, comprising: a data acquisition module for acquiring initial image data and status identification information of the billet in the casting and rolling operation area; a preprocessing module for preprocessing the initial image data using a preprocessing algorithm deployed on an edge computing unit to obtain preprocessed image data; an analysis and processing module for processing and analyzing the preprocessed image data using a deep learning model deployed on an edge computing unit to obtain a set of detection results for the billet, wherein the set of detection results includes at least one of end face recognition results, shape recognition results, motion detection results, displacement detection results, and arrangement recognition results; and a billet tracking module for performing identity matching and trajectory tracking on the billet according to a preset tracking rule based on the set of detection results and the status identification information to obtain tracking result data and complete the billet tracking.
[0017] The beneficial effects of this application are as follows: Preprocessing the initial image data can reduce the interference of invalid data on model inference and improve the efficiency, accuracy and stability of subsequent deep learning model analysis and processing; by processing and analyzing the preprocessed image data through the deep learning model, at least one of the following can be output: end face recognition result, shape recognition result, motion detection result, displacement detection result and arrangement recognition result. This allows for the simultaneous acquisition of multi-dimensional information such as the geometric shape, motion state, spatial displacement and arrangement logic of the billet, providing a rich data foundation for subsequent accurate tracking; based on the detection result set and status identification information, identity matching and trajectory tracking are performed according to preset tracking rules to obtain tracking result data. This allows for continuous tracking of the position change and path direction of each billet, forming a complete motion trajectory, and also provides objective data basis for subsequent processing.
[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.
[0020] In the attached diagram: Figure 1 This is a schematic diagram of a system architecture illustrated in an exemplary embodiment of this application; Figure 2 This is a schematic flowchart illustrating a billet tracking method according to an exemplary embodiment of this application; Figure 3This is a block diagram illustrating a billet tracking system as shown in an exemplary embodiment of this application. Detailed Implementation
[0021] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.
[0022] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0023] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.
[0024] First, it's important to clarify that synchronous fusion refers to a technical process in multi-sensor or multi-source data processing where data collected at different times and locations is aligned and processed along the same spatiotemporal dimension using a unified time reference and spatial coordinate system. The core principle is "first, spatiotemporal synchronization, then information fusion."
[0025] Currently, in the casting and rolling process of steel enterprises, it is difficult to efficiently and accurately determine the movement of the billet during the casting and rolling process. This often results in unclear billet movement trajectories, difficulty in identifying billet morphology, difficulty in determining the billet's final status, and an inability to track the movement logic. Furthermore, there is no effective method to trace historical results after the billet movement is completed, resulting in a lack of objective data and hindering effective visualization and intelligent operation of the casting and rolling process. To address these issues, this application provides a billet tracking method and a billet tracking system, the embodiments of which will be described in detail below.
[0026] Figure 1 This is a schematic diagram of a system architecture illustrated in an exemplary embodiment of this application.
[0027] Reference Figure 1As shown, the system architecture may include a data acquisition device 101 and a computer device 102. The computer device 102 may be at least one of a desktop graphics processing unit (GPU) computer, a GPU computing cluster, a neural network computer, or an edge computing unit. The data acquisition device 101 is used to collect initial image data and status identification information of the billet in the casting and rolling operation area. In this embodiment, after acquiring the above data, the data acquisition device 101 provides it to the computer device 102 for processing. Those skilled in the art can use the computer device 102 to preprocess the initial image data using a preprocessing algorithm deployed on the edge computing unit, obtaining preprocessed image data. Then, using a deep learning model deployed on the edge computing unit, the preprocessed image data is processed and analyzed to obtain a set of billet detection results. The set of detection results includes at least one of end-face recognition results, shape recognition results, motion detection results, displacement detection results, and arrangement recognition results. Based on the set of detection results and status identification information, the billet is matched for identity and tracked according to preset tracking rules to obtain tracking result data, thus completing the billet tracking. It should be noted that the data acquisition device 101 and computer device 102 provided in this embodiment are merely examples and should not impose any limitations on the functions and scope of use of the embodiments of this application.
[0028] It should be noted that the billet tracking method provided in this application embodiment is generally executed by computer device 102, and correspondingly, the billet tracking system is generally set in computer device 102.
[0029] Figure 2 This is a schematic flowchart illustrating a billet tracking method according to an exemplary embodiment of this application. The billet tracking method can be executed by a computing processing device, which may be... Figure 1 The computer device 102 shown is illustrated. (Refer to...) Figure 2 As shown, the billet tracking method includes at least steps S210 to S240, which are described in detail below: In step S210, the initial image data and status identification information of the billet in the casting and rolling operation area are obtained.
[0030] In this embodiment, image acquisition devices are installed around the casting and rolling equipment. For example, the image acquisition devices can be installed above the billet. It is understood that the image acquisition devices can also be installed above or to one side of the roller conveyor, perpendicular or parallel to the billet's running direction, or in other positions, depending on requirements. This application does not limit this. The image acquisition devices can be industrial cameras or network high-definition cameras, etc. Initial image data of the billet in the casting and rolling operation area is acquired through the image acquisition devices. Production signals from the casting and rolling operation area are obtained through the production signal receiving module to obtain the billet's status identification information, thereby determining the billet's current production state, i.e., whether it is in the hot delivery state or the offline state, and thus obtaining the status identification information. Real-time positioning information of the billet during the production process is acquired using sensors, such as positioning sensors or position sensors, to obtain the billet's position information. The initial image data, position information, and status identification information of the billet are transmitted to the edge computing unit in real time through wired and / or wireless interfaces. A verification mechanism is used during transmission to avoid data packet loss. This embodiment provides a continuous, real-time, multi-dimensional data source for subsequent visual inspection, replacing the traditional hot metal detector that can only provide discrete position signals, and the verification mechanism ensures the integrity of data transmission.
[0031] In step S220, the initial image data is preprocessed using a preprocessing algorithm deployed on the edge computing unit to obtain preprocessed image data.
[0032] In this embodiment, the preprocessing algorithm includes at least an image denoising algorithm, a region of interest (ROI) cropping algorithm, and a white balance correction algorithm. The image denoising algorithm eliminates image noise introduced by environmental factors such as high temperature, dust, and water mist in the casting and rolling operation area. The ROI cropping algorithm crops out the areas where the billet might appear from the initial image data to eliminate interference from irrelevant backgrounds and reduce the amount of data that the subsequent deep learning model needs to process. The white balance correction algorithm eliminates the influence of different lighting conditions on image color and brightness, ensuring consistent imaging quality for images acquired at different times. Preprocessing includes at least image denoising, ROI cropping, and white balance correction. This embodiment can improve the recognition accuracy and stability of the subsequent deep learning model, reduce interference from invalid data, improve detection efficiency, reduce the computational load and power consumption of the edge computing unit, and ensure system real-time performance.
[0033] In step S230, a deep learning model deployed on the edge computing unit is used to process and analyze the preprocessed image data to obtain a set of detection results for the cast billet.
[0034] The detection result set includes at least one of the following: end face recognition result, shape recognition result, motion detection result, displacement detection result, and arrangement recognition result.
[0035] In one embodiment of this application, if the detection result set includes end face recognition results and the deep learning model includes an end face detection model, the method for obtaining the end face recognition results includes: inputting preprocessed image data into the end face detection model, locating the end face region, and performing continuous feature detection on the end face region; if discontinuous features are detected, it is determined that there is a tear on the end face of the billet corresponding to the end face region, and the end face recognition result is output; wherein, the end face detection model is obtained by training a preset end face detection network through a first training set, and the end face detection network learns the feature differences between the normal end face shape and the torn end face shape through image feature extraction, and the first training set is billet image data with end face annotation information, which annotates the normal area and the torn area of the end face.
[0036] In this embodiment, an image acquisition device is set on either side of the roller conveyor along the billet running direction. The image acquisition device is controlled to acquire images along the billet running direction to obtain the original image of the billet. The original image of the billet is preprocessed to obtain billet image data. The billet image data is labeled to obtain billet image data with end face annotation information, thus obtaining the first training set. The first training set is input into the end face detection network for training to obtain the end face detection model. The training steps include image feature extraction and end face recognition. The end face recognition steps include: when the image in the billet image data is discontinuous, it is determined that the billet image data has been torn. The preprocessed image data is input into the end face detection model for target detection to complete the end face recognition.
[0037] In one embodiment of this application, if the detection result set includes morphology recognition results and the deep learning model includes a first contour segmentation model, the method for obtaining the morphology recognition results includes: inputting preprocessed image data into the first contour segmentation model and outputting a first segmentation mask for the billet region; extracting contour geometric parameters based on the first segmentation mask, wherein the contour geometric parameters include at least the contour area, contour perimeter, and aspect ratio of the billet; comparing the contour geometric parameters with a preset normal threshold range, and if any parameter in the contour geometric parameters exceeds the normal threshold range, then the morphology recognition result is determined to be an abnormal billet morphology, otherwise the morphology recognition result is determined to be a normal billet morphology; wherein, the first contour segmentation model is obtained by training a preset first contour segmentation network through a second training set, the first contour segmentation network learns the morphological features of the billet contour through training, and outputs a first segmentation mask based on the input preprocessed image data, and the second training set is billet image data with billet contour annotation information, wherein the billet contour annotation information annotates the contour boundary of the billet in the image.
[0038] In this embodiment, the original image of the cast billet is acquired, preprocessed to obtain cast billet image data, and the cast billet image data is annotated with cast billet contour annotation information to obtain cast billet image data with cast billet contour annotation information, thereby obtaining a second training set; the second training set is input into a first contour segmentation network for training to obtain a first contour segmentation model; the preprocessed image data is input into the first contour segmentation model for contour segmentation to obtain a segmentation mask of the cast billet region in the preprocessed image data; based on the segmentation mask, contour geometric parameters used to characterize the geometric shape of the cast billet are extracted, the contour geometric parameters including at least the contour area, contour perimeter and aspect ratio of the cast billet; according to the contour geometric parameters, the cast billet shape recognition is performed to obtain the shape recognition result.
[0039] In some embodiments, after obtaining the segmentation mask of the billet region through the first contour segmentation model, the pixel region of the billet in the original image can be located according to the segmentation mask. Then, texture analysis and grayscale distribution statistics are performed on the pixel values within that region to extract feature information characterizing the surface quality of the billet. The billet surface information includes, but is not limited to, at least one of surface texture features, grayscale distribution features, and surface defect features. The billet surface information can be used for surface defect detection, etc., and this application does not impose any limitations on this.
[0040] In one embodiment of this application, if the detection result set includes motion detection results and the deep learning model includes an object detection model, the method for obtaining the motion detection results includes: inputting preprocessed image data into the object detection model, performing object detection on the slab in the image, and outputting a first object detection result, the first object detection result including the slab bounding box and position information; according to a preset image acquisition cycle, acquiring the original image data of the slab again at the next acquisition time after outputting the first object detection result, and preprocessing the original image data to obtain processed image data; inputting the processed image data into the object detection model to perform object detection on the slab in the image. The system outputs a second target detection result, which includes the bounding box and position information of the billet. Based on the first and second target detection results, the system determines the motion of the billet within the image acquisition period to obtain a motion detection result, which includes position and velocity changes. The target detection model is trained on a preset target detection network using a third training set. The target detection network learns the mapping relationship between the target detection features of the billet and the image content through image feature extraction. The third training set consists of billet image data with target annotation information, which annotates the bounding box and position information of the billet in the image.
[0041] In this embodiment, a third training set is acquired, which includes one or more ingot image data with target annotation information; the third training set is input into a target detection network for training to obtain a target detection model; preprocessed image data is input into the target detection model for target detection to obtain a first target detection result; real-time image secondary acquisition is performed according to a preset image acquisition cycle to obtain original image data, and the original image data is preprocessed to obtain processed image data; the processed image data is input into the target detection network for target detection to obtain a second target detection result; based on the first target detection result and the second target detection result, the movement of the ingot within the image acquisition cycle is determined, so as to obtain a motion detection result based on the movement.
[0042] In this embodiment, the motion of the billet within the image acquisition cycle is determined based on the first target detection result and the second target detection result. The specific implementation process for obtaining the motion detection result based on the motion can be as follows: First, a multi-target tracking algorithm (such as association matching based on intersection-union ratio or position similarity) is used to associate each billet in the first frame image with each billet in the second frame image to ensure that the position change of the same billet between adjacent frames is calculated. Alternatively, images of the same billet can be acquired during image acquisition. Then, using the center point or centroid of each billet's bounding box as the position reference, the pixel displacement in the horizontal and vertical directions is calculated respectively, and the pixel displacement is converted into the actual physical displacement by combining the camera calibration parameters. Subsequently, the physical displacement is converted into the actual physical displacement. Bit removal is used to obtain the actual movement speed of the billet by the image acquisition cycle (i.e., the time interval between two adjacent frames). The speed change can be calculated based on the actual movement speed, and the movement direction of the billet can be determined based on the displacement direction. Thus, the movement of the billet within the image acquisition cycle can be obtained. The movement includes at least one of the billet's position change, movement speed, and movement direction. Finally, the movement is compared with preset normal movement parameter thresholds (such as upper / lower speed limits, trajectory offset tolerance, etc.). If any parameter in the movement exceeds the threshold range, the billet's movement state is determined to be abnormal; otherwise, the movement state is determined to be normal. The movement detection result is then output for subsequent tracking steps to determine whether the billet has deviated from the predetermined path or has experienced abnormal movement.
[0043] In one embodiment of this application, if the detection result set includes displacement detection results and the deep learning model includes a second contour segmentation model, the displacement detection results are obtained by: inputting preprocessed image data into the second contour segmentation model and outputting a second segmentation mask for the billet region; extracting billet displacement information based on the second segmentation mask, the billet displacement information including the billet movement distance and the billet movement trajectory; and determining the displacement detection results between adjacent detection times based on the billet displacement information. The second contour segmentation model is obtained by training a preset second contour segmentation network using a fourth training set. The second contour segmentation network learns the segmentation features of the billet contour through image feature extraction and outputs a second segmentation mask based on the input preprocessed image data. The fourth training set is billet image data with billet start and end position annotation information, which annotates the position information of the billet in the image at the start and end times of detection.
[0044] In this embodiment, the specific implementation process of determining the displacement detection result between adjacent detection times based on the billet displacement information is as follows: First, based on the second segmentation mask output by the second contour segmentation model, the starting position of the billet at the start time of detection and the ending position at the end time of detection are obtained (e.g., using the centroid or the center of the circumscribed rectangle of the segmentation mask at each time as the position reference point of the billet); then, based on the coordinate difference between the starting position and the ending position, the moving distance of the billet between adjacent detection times is calculated, and the moving trajectory of the billet is determined based on the coordinate change direction and path from the starting position to the ending position; finally, the moving distance and / or moving trajectory are compared with a preset normal displacement threshold (e.g., maximum allowable moving distance, predetermined path offset tolerance, etc.). If the moving distance exceeds the preset threshold or the moving trajectory deviates from the predetermined path, it is determined to be an abnormal displacement; otherwise, it is determined to be a normal displacement, thereby outputting the displacement detection result for subsequent tracking steps to determine whether the displacement state of the billet between adjacent detection times is normal. This embodiment can provide a data analysis basis for real-time statistical analysis of billet movement.
[0045] In one embodiment of this application, if the detection result set includes arrangement recognition results and the deep learning model includes a third contour segmentation model, the arrangement recognition results are obtained by: inputting preprocessed image data into the third contour segmentation model and outputting a third segmentation mask for the billet region; based on the third segmentation mask, identifying the sequence number information of each billet in the image, the sequence number information including the arrangement order of each billet in the image and the corresponding sequence number identifier; according to the sequence number information, obtaining the first arrangement order of the billet when entering the detection port and the second arrangement order when leaving the detection port; if the first arrangement order is consistent with the second arrangement order, the arrangement recognition result is determined to be that the billet arrangement is normal; if the first arrangement order is inconsistent with the second arrangement order, the arrangement recognition result is determined to be that the billet arrangement is abnormal; wherein, the third contour segmentation model is obtained by training a preset third contour segmentation network through a fifth training set, the third contour segmentation network learns the segmentation features of the billet sequence number through image feature extraction, and outputs a third segmentation mask based on the input preprocessed image data, the fifth training set being billet image data with billet sequence number annotation information, the billet sequence number annotation information annotating the sequence number identifier and arrangement order of each billet in the image.
[0046] In this embodiment, based on the sequence number information, the first arrangement order of the billets when entering the inspection port and the second arrangement order when leaving the inspection port are obtained, and it is determined whether the first arrangement order and the second arrangement order are consistent, so as to determine the specific implementation process of the arrangement recognition result: First, based on the third segmentation mask output by the third contour segmentation model, the sequence number information of each billet in the image is identified. Specifically, image processing algorithms are used to extract the serial number identifiers (such as numerical codes or QR codes sprayed on the end face of the billet) of each billet region. The spatial position of each billet in the image determines its arrangement order in the queue. The serial number information includes the arrangement order of each billet in the image and its corresponding serial number identifier. Then, based on the serial number information, the first arrangement order of the billets entering the inspection port (i.e., the order and corresponding serial number of each billet when entering a certain process or area) and the second arrangement order when leaving the inspection port (i.e., the order and corresponding serial number of each billet when leaving the process or area) are obtained respectively. Finally, the first and second arrangement orders are compared to see if they are consistent. If they are consistent, it indicates that the arrangement logic of each billet during the casting and rolling process has not been disordered, and the arrangement is judged to be normal. If they are inconsistent (e.g., the serial number of a billet is advanced or delayed, or there is a missing or repeated number), it indicates that the arrangement has been disordered or the tracking has been lost during the casting and rolling process, and the arrangement is judged to be abnormal. The arrangement recognition result is then output for subsequent tracking steps to determine whether there are errors in the casting and rolling process, avoiding unnecessary losses caused by human judgment errors.
[0047] In another embodiment, based on preprocessed image data, the position of the billet is first detected to obtain a first detection result. The first detection includes at least one of billet end face recognition and billet shape recognition. Correspondingly, the first detection result includes at least one of end face recognition result and shape recognition result. Based on preprocessed image data, the image of the billet passing through the inlet and outlet is second detected to obtain a second detection result. The second detection includes at least one of billet motion detection, billet displacement detection, and billet arrangement recognition. Correspondingly, the second detection result includes at least one of motion detection result, displacement detection result, and arrangement recognition result. Based on the first detection result and / or the second detection result, visual tracking and recognition of the billet are performed, facilitating intelligent monitoring of the casting and rolling process. This allows for better traceability of historical data on billet transportation, is highly feasible, highly automated, and low-cost. It is understood that the methods for obtaining the end face recognition result, shape recognition result, motion detection result, displacement detection result, and arrangement recognition result in this embodiment have been explained in the aforementioned corresponding embodiments and will not be repeated here.
[0048] In another embodiment, belt image data is acquired, and by identifying and analyzing the belt image data, equipment failure of the belt conveyor can be detected, achieving visualized and automated real-time monitoring of the belt conveyor status. Specifically, image acquisition devices are deployed along the belt conveyor (e.g., above or to the side of the belt) to collect belt image data in real time during belt operation. The belt image data is then preprocessed to obtain preprocessed belt image data.
[0049] The pre-processed belt image data is input into a pre-trained deep learning model. The deep learning model is trained using pre-collected belt images encompassing normal and various fault states (such as belt misalignment, belt tearing, foreign object intrusion, and idler roller failure) as training samples, learning the image feature patterns corresponding to each fault state. During the inference phase, the deep learning model performs real-time inference on the input belt image data, identifying the presence of fault features in the images and outputting fault detection results. The fault detection results include the fault type (such as misalignment, tearing, or foreign object intrusion) and the fault's location information within the image.
[0050] Based on the fault detection results output by the deep learning model, and combined with preset fault judgment rules (such as whether the fault area exceeds a threshold, whether the fault feature confidence level reaches a preset threshold, etc.), the system determines whether the belt conveyor is currently experiencing a fault and the severity of the fault. If a fault is determined, an early warning message containing the fault type, fault location, and occurrence time is generated and displayed in real time on the user interface via the monitoring and alarm module, along with audible and visual alarms, to achieve visualized and automated real-time monitoring of the belt conveyor status.
[0051] This embodiment utilizes machine vision technology and deep learning models to replace the traditional manual inspection method, realizing online intelligent monitoring of the belt conveyor's operating status and automatic fault early warning, effectively improving the efficiency and accuracy of fault detection.
[0052] In step S240, based on the detection result set and status identification information, the billet is matched for identity and tracked according to the preset tracking rules to obtain tracking result data and complete the billet tracking.
[0053] In one embodiment of this application, based on the detection result set and status identifier information, the billet is matched for identity and tracked according to preset tracking rules to obtain tracking result data. This includes: assigning a corresponding virtual billet number to each billet based on the detection result set and status identifier information; updating the real-time position and path of each billet according to preset position rules and preset queue rules based on the virtual billet number, forming a continuous movement trajectory for each billet; assigning an abnormal marker status to each billet based on the comparison result of the first virtual billet number when entering the preset verification area and the second virtual billet number when leaving the preset verification area; generating tracking details for each billet based on the status identifier information, combined with the continuous movement trajectory and abnormal marker status of each billet; and outputting a hot delivery record or offline record according to the hot delivery status or offline status in the status identifier information, thereby obtaining tracking result data. It is understood that the virtual billet number includes a first virtual billet number and a second virtual billet number.
[0054] In this embodiment, the specific implementation process of obtaining tracking result data based on the detection result set and status identification information is as follows: First, based on the end face recognition results, shape recognition results, motion detection results, displacement detection results, and arrangement recognition results of each casting in the detection result set, combined with the status identification information, a unique virtual casting ingot number is assigned to each casting ingot (for example, the casting ingot successfully detected for the first time is assigned number 1, and subsequent numbers are incremented sequentially). This virtual casting ingot number serves as the unique identifier of the casting ingot throughout the entire tracking process and is continuously associated with all detection information and status identification information of the casting ingot; then, based on the virtual... The billet number is updated according to preset position rules (such as updating the spatial coordinates of the billet on the production line based on the movement distance and trajectory in the displacement detection results, and judging whether it deviates from the predetermined path) and preset queue rules (such as judging whether the relative position and order between multiple billets are correct based on the arrangement recognition results). At each detection moment or each image acquisition cycle, the real-time position and path of each billet on the casting and rolling production line are updated, and the positions at each moment are connected in chronological order to form the continuous movement trajectory of each billet. At the same time, when any billet enters the preset verification area (such as the entrance and exit of the heating furnace, the front and rear of the cutting machine), the real-time position and path of each billet on the casting and rolling production line are updated. When a critical node (such as a slab) enters the verification area, its first virtual slab number is recorded. When it leaves the verification area, its second virtual slab number is recorded. The first and second virtual slab numbers are compared. If they match, it means that the slab's identity has remained stable within the verification area, and there has been no tracking loss or confusion. The current tracking status is maintained, and the slab is determined to be tracking normal. An abnormal flag status (no abnormal flag) is assigned to the slab. If they do not match, it means that the slab has experienced identity confusion, tracking loss, or confusion with other slabs within the verification area. This is determined to be tracking abnormal, and an abnormal flag status (abnormal flag) is assigned to the slab. Finally, based on the status identification information (hot delivery status or offline status), combined with the continuous movement trajectory and anomaly marker status of each billet, tracking details for each billet are generated (tracking details include the billet's position information at each moment, complete movement trajectory, anomaly marker records, etc.). Based on the hot delivery status or offline status in the status identification information, corresponding hot delivery records or offline records are output (e.g., billets that need to continue to the next process are included in the hot delivery record, and billets that need to be offline and stacked are included in the offline record), thus obtaining the tracking result data and completing the entire process of identity association and trajectory tracking for each billet from acquisition to tracking. In other words, the tracking result data includes tracking details and hot delivery records or offline records generated based on the status identification information.
[0055] In one embodiment of this application, sensors and image acquisition devices are deployed at key nodes on the casting and rolling production line to collect various types of data in real time. Multiple data recognition queues are constructed according to time sequence and data type, and a timestamp and device identifier are added to each data point. A data fusion algorithm is used to concatenate the data in each data recognition queue, establishing data connections based on recognition logic and time consistency to form a complete dataset. Feature vectors are extracted from the data in the complete dataset. The feature vectors are input into an anomaly detection model to obtain anomaly detection results. The anomaly detection model is trained on a preset anomaly detection network using training samples, which are historical data containing both normal and abnormal states, trained using an unsupervised learning method. If the anomaly detection result is abnormal, a warning message containing the time and location of the anomaly is generated, and a preset offline processing is triggered. The anomaly detection network includes at least one of support vector machines, neural networks, or isolated forests.
[0056] In this embodiment, firstly, sensors (such as position sensors, temperature sensors, and speed sensors) and image acquisition devices (such as industrial cameras) are deployed at key nodes on the casting and rolling production line (such as the furnace inlet / outlet, before and after the rolling mill, before and after the cutting machine, and the branch point at the end of the line). These sensors collect various types of data in real time (including but not limited to at least two of image data, positioning data, equipment status data, temperature data, and speed data). Based on the time sequence and data type of the collected data, multiple data identification queues are constructed (for example, image data is arranged in chronological order to form an image data queue, positioning data in chronological order to form a positioning data queue, and equipment status data in chronological order to form an equipment status queue). A timestamp (for subsequent time alignment) and an equipment identifier (for tracing the data source) are added to each data point to ensure temporal consistency and traceability between data streams. Alternatively, the collected data is divided into multiple categories according to data type (such as image data queue, positioning data queue, and equipment status data queue), and within each category's queue, the data is sorted in chronological order, with a timestamp and equipment identifier added to each data point, thereby constructing multiple data identification queues.
[0057] Then, a data fusion algorithm (such as a timestamp-based synchronous fusion method) is used to concatenate the data in each data recognition queue. Specifically, using timestamps as the alignment benchmark, various types of data collected at the same time or within the same time window are associated according to recognition logic relationships (such as spatial correspondence between image data and positioning data generated by the same device at the same time) and temporal consistency to form a complete dataset containing multi-source heterogeneous data. For different types of data in the complete dataset, corresponding feature extraction algorithms are used to extract features (for example, convolutional neural networks are used to extract visual features from image data, and statistical methods are used to extract time-domain or frequency-domain features from positioning data and device status data). All types of raw data are transformed into feature vectors of a unified format or dimension to facilitate unified processing by subsequent models.
[0058] Next, the feature vectors are input into a pre-trained anomaly detection model. The anomaly detection model is pre-trained as follows: historical data containing both normal and abnormal states are collected as training samples. Unsupervised learning methods (such as autoencoders based on reconstruction error, isolated forests based on distance metrics, or support vector descriptions based on density estimation) are used to train the pre-set anomaly detection network. During training, the model only learns normal data patterns and does not rely on the label information of abnormal data. The model parameters are adjusted to enable the model to accurately identify the distribution patterns and inherent laws of normal data. Optimization algorithms such as cross-entropy loss function and stochastic gradient descent are used during training to accelerate convergence. After training, the model internally establishes a representation of normal data patterns (such as the feature distribution range, cluster centers, or reconstruction error thresholds of normal data).
[0059] Finally, during the inference phase, the anomaly detection model receives feature vectors and calculates the degree of deviation between the feature vectors and the normal data patterns learned during the training phase (such as the distance of the feature vectors to the normal cluster centers, reconstruction error values, or deviations from the normal distribution). It then compares the degree of deviation with a preset judgment threshold. If the degree of deviation does not exceed the preset threshold, the current data pattern is determined to be normal, and a "normal" anomaly detection result is output. If the degree of deviation exceeds the preset threshold, the current data pattern is determined to be abnormal, and an "abnormal" anomaly detection result is output. Simultaneously, an early warning message containing the time and location information of the anomaly is generated, and a preset offline processing procedure is triggered (such as pushing the anomaly information to the monitoring interface, triggering an audible and visual alarm, or automatically isolating the abnormal billet). This completes real-time monitoring and anomaly early warning of the entire casting and rolling production line's operating status.
[0060] In one embodiment of this application, the image quality of the initial image data, the device status of the preset device, and the data synchronization of the preset data are monitored in real time. Based on the image quality, device status, and data synchronization, the fault level is determined. If the fault level is a level one fault, the visual processing parameters are adjusted. If the fault level is a level two fault, the backup image acquisition device is activated, and the positioning deviation is corrected by adjusting the roller speed. If the fault level is a level three fault, the mill speed is reduced or partially stopped, and the tracking continuity is maintained in conjunction with the positioning device.
[0061] In this embodiment, image acquisition devices (such as industrial cameras) are deployed at key nodes such as the furnace inlet and outlet of the casting and rolling production line to collect billet position information and equipment linkage data in real time (such as collaborative data between image acquisition devices, positioning sensors, and roller control systems). Based on the collected data, fault phenomena such as positioning deviation (such as the deviation between the actual position and the expected position of the billet), parameter jumps (such as sudden changes in image brightness, contrast, and frame rate), or sequence disorder (such as the disordered arrangement of multiple billets entering and exiting the inspection port) are identified.
[0062] The edge computing unit monitors in real time the image quality of acquired images (e.g., whether the initial image data is blurry, overexposed, underexposed, has dropped frames, or has screen tearing), the device status of preset devices (e.g., whether the image acquisition device is operating normally, whether the communication link is unobstructed, and whether the processor load and memory usage of the edge computing unit are normal), and the data synchronization of preset data (e.g., whether the timestamps between image data, positioning data, and device status data are aligned, and whether there are temporal logic contradictions between visual inspection results and device operating parameters). The fault level is comprehensively determined based on the monitoring results. Fault levels include Level 1 faults (minor faults, such as slight noise or blurring in the image, but not affecting the overall recognition function), Level 2 faults (moderate faults, such as main camera failure or positioning deviation exceeding the tolerance range), and Level 3 faults (serious faults, such as overall system failure or sequence disorder).
[0063] Based on the determined fault level, implement the corresponding fault tolerance strategy: If the fault level is Level 1, enable software noise reduction compensation (such as smoothing the image through a filter), adjust the visual processing parameters (such as adjusting the filter intensity, ROI (Region of Interest) range, or confidence threshold in the model inference stage), but do not adjust the casting and rolling process parameters (such as keeping the roller speed and rolling force unchanged) to restore system performance with minimal intervention. If the fault level is level 2, the backup image acquisition device is activated to achieve redundant switching of the image acquisition link; at the same time, the tracking error caused by visual positioning deviation is corrected and compensated by adjusting the roller speed (e.g., when the actual position of the billet is detected to be lagging behind the expected position, the roller speed is appropriately reduced so that the position of the billet returns to the expected trajectory), thereby maintaining the continuity of tracking. If the fault level is level three, the rolling mill will be slowed down or partially shut down to reduce the production pace and prevent the accident from escalating. At the same time, data from laser positioning equipment (such as laser rangefinders or lidar) will be used to assist in measuring and correcting the position of the billet to maintain the continuity and safety of tracking.
[0064] After completing the above fault-tolerance processing, the effectiveness of the intervention measures is verified through visual indicators such as positioning accuracy (e.g., the deviation between the actual position of the billet and the detection position) and defect detection rate (e.g., the accuracy of end-face tear detection and morphology detection), as well as process indicators such as speed stability (e.g., the range of roller speed fluctuation) and rolling force stability (e.g., the range of rolling force fluctuation). If all the above indicators meet the preset requirements, the current fault-tolerance strategy is confirmed to be effective, and the system returns to normal monitoring status; if any parameter of the above indicators does not meet the preset requirements, the current intervention measures are deemed insufficient, and the intervention level needs to be upgraded (e.g., from level one to level two, or from level two to level three, until the system returns to stable operation).
[0065] In some embodiments, this application also provides a casting and rolling tracking intelligent monitoring function. Specifically, the monitoring and alarm module is used to perform intelligent monitoring of casting and rolling tracking based on the data output by each module (including initial image data of each casting and rolling operation area, displacement detection results of each billet, tracking details of each billet, and offline / hot delivery records, etc.), obtain real-time identification results, and display them visually on the user interface; when an abnormal situation is detected (such as the tracking abnormality marker being triggered), the monitoring and alarm module prompts and outputs warning information on the user interface. In other words, if the monitoring and alarm module finds any abnormality in the end face recognition result, shape recognition result, motion detection result, displacement detection result, and arrangement recognition result, it prompts and outputs warning information on the user interface.
[0066] In some embodiments, this application also provides a historical record management function. Specifically, the historical record module is used to structurally store the initial image data, complete tracking time, offline status, and billet tracking details according to time sequence and virtual billet number, forming a traceable historical record, and supporting manual review and correction of historical detection results. The anomaly marker status includes states with and without anomaly markers; the tracking details status includes the end face recognition results, shape recognition results, motion detection results, displacement detection results, and arrangement recognition results of each billet at each detection node.
[0067] This application constructs a closed-loop system based on an edge intelligence architecture consisting of front-end perception and edge decision-making, comprising data acquisition devices and edge computing units. The system performs image preprocessing, deep learning model inference, and secondary processing according to preset tracking rules locally, outputting only structured results (i.e., a set of detection results) without uploading the original images. This avoids network transmission delays, ensures real-time tracking of the billet's motion trajectory, and balances data processing efficiency with the security of the original data. Simultaneously, multi-source data fusion perception is achieved by combining image information acquired by image acquisition devices with positioning information acquired by sensors. Verification mechanisms during transmission prevent data packet loss, and closed-loop feedback enables the system to adaptively adjust processing strategies, further ensuring the stability and accuracy of tracking. Based on this, the image acquisition module acquires real-time images of each casting and rolling operation area, and the production signal receiving module identifies the real-time status of the billet (hot delivery or offline). A virtual billet number is assigned to each billet for identity tracking, and the entry and exit sequence numbers of billets entering and leaving the casting and rolling area are verified to ensure that the identity of the same billet is consistent and the sequence number meets logical requirements. This enables real-time image monitoring and full-process tracking of the billets, outputting hot delivery and offline records. Furthermore, the historical record module records and statistically analyzes all tracking results, supporting subsequent viewing, traceability, and manual review and correction of historical inspection results. In summary, this application effectively solves the problems of poor real-time performance, limited information dimensions, high data security risks, and lack of objective historical traceability in traditional billet tracking schemes, providing an efficient, accurate, and traceable intelligent tracking solution for casting and rolling production.
[0068] Figure 3 This is a block diagram illustrating a billet tracking system as shown in an exemplary embodiment of this application. The system can be applied to... Figure 1 The implementation environment shown is specifically configured in computer device 102. This system can also be applied to other exemplary implementation environments and specifically configured in other devices; this embodiment does not limit the implementation environment to which the system is applicable.
[0069] like Figure 3 As shown, the exemplary billet tracking system includes: a data acquisition module 310, a preprocessing module 320, an analysis and processing module 330, and a billet tracking module 340.
[0070] The system includes a data acquisition module 310 for acquiring initial image data and status identification information of the billet in the casting and rolling operation area; a preprocessing module 320 for preprocessing the initial image data using a preprocessing algorithm deployed on the edge computing unit to obtain preprocessed image data; an analysis and processing module 330 for processing and analyzing the preprocessed image data using a deep learning model deployed on the edge computing unit to obtain a set of detection results for the billet, wherein the set of detection results includes at least one of end face recognition results, shape recognition results, motion detection results, displacement detection results, and arrangement recognition results; and a billet tracking module 340 for performing identity matching and trajectory tracking on the billet according to preset tracking rules based on the set of detection results and status identification information to obtain tracking result data and complete the billet tracking.
[0071] It should be noted that the billet tracking system and the billet tracking method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the billet tracking system provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.
[0072] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0073] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.
[0074] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.
[0075] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the method according to the embodiments of this application.
[0076] Other embodiments of this application will readily conceive of by those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.
[0077] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.
Claims
1. A method for tracking cast billets, characterized in that, include: Acquire initial image data and status identification information of the billet in the casting and rolling operation area; The initial image data is preprocessed using a preprocessing algorithm deployed on an edge computing unit to obtain preprocessed image data; Using a deep learning model deployed on an edge computing unit, the preprocessed image data is processed and analyzed to obtain a set of detection results for the billet, wherein the set of detection results includes at least one of end face recognition results, shape recognition results, motion detection results, displacement detection results, and arrangement recognition results; Based on the set of detection results and the status identification information, the billet is matched for identity and tracked according to the preset tracking rules to obtain tracking result data and complete the billet tracking.
2. The billet tracking method according to claim 1, characterized in that, If the detection result set includes the end-face recognition result, and the deep learning model includes an end-face detection model, the methods for obtaining the end-face recognition result include: The preprocessed image data is input into the end face detection model to locate the end face region and perform continuous feature detection on the end face region. If a discontinuity feature is detected, it is determined that there is a tear on the end face of the billet corresponding to the end face region, and the end face recognition result is output. The end face detection model is obtained by training a preset end face detection network using a first training set. The end face detection network learns the feature differences between normal end face morphology and torn end face morphology through image feature extraction. The first training set is cast billet image data with end face annotation information, which annotates the normal area and torn area of the end face.
3. The billet tracking method according to claim 1, characterized in that, If the detection result set includes the shape recognition result, and the deep learning model includes a first contour segmentation model, the method for obtaining the shape recognition result includes: The preprocessed image data is input into the first contour segmentation model, and the first segmentation mask of the billet region is output. Based on the first segmentation mask, the contour geometric parameters are extracted, and the contour geometric parameters include at least the contour area, contour perimeter and width-to-height ratio of the billet; The contour geometric parameters are compared with a preset normal threshold range. If any parameter in the contour geometric parameters exceeds the normal threshold range, the shape recognition result is determined to be an abnormal billet shape; otherwise, the shape recognition result is determined to be a normal billet shape. The first contour segmentation model is obtained by training a preset first contour segmentation network on a second training set. The first contour segmentation network learns the morphological features of the billet contour through training and outputs the first segmentation mask based on the input preprocessed image data. The second training set is billet image data with billet contour annotation information, which annotates the contour boundary of the billet in the image.
4. The billet tracking method according to claim 1, characterized in that, If the detection result set includes the motion detection results, and the deep learning model includes an object detection model, the methods for obtaining the motion detection results include: The preprocessed image data is input into the target detection model to perform target detection on the billet in the image, and a first target detection result is output, which includes the billet bounding box and position information. According to the preset image acquisition cycle, at the next acquisition time after the first target detection result is output, the original image data of the billet is acquired again, and the preprocessing is performed on the original image data to obtain the processed image data. The processed image data is input into the target detection model to perform target detection on the billet in the image, and a second target detection result is output, which includes the billet bounding box and position information. Based on the first target detection result and the second target detection result, the motion of the billet within the image acquisition cycle is determined, so as to obtain the motion detection result based on the motion, wherein the motion includes position change and velocity change; The target detection model is obtained by training a preset target detection network using a third training set. The target detection network learns the mapping relationship between the target detection features of the billet and the image content through image feature extraction. The third training set is billet image data with target annotation information, which annotates the bounding box and position information of the billet in the image.
5. The billet tracking method according to claim 1, characterized in that, If the detection result set includes the displacement detection result, and the deep learning model includes a second contour segmentation model, the displacement detection result can be obtained in the following ways: The preprocessed image data is input into the second contour segmentation model, and the second segmentation mask of the billet region is output. Based on the second segmentation mask, the billet displacement information is extracted, which includes the billet moving distance and the billet moving trajectory. Based on the billet displacement information, the displacement detection result between adjacent detection times is determined; The second contour segmentation model is obtained by training a preset second contour segmentation network on a fourth training set. The second contour segmentation network learns the segmentation features of the billet contour through image feature extraction and outputs the second segmentation mask based on the input preprocessed image data. The fourth training set is billet image data with billet start and end position annotation information. The billet start and end position annotation information annotates the position information of the billet in the image at the detection start time and end time.
6. The billet tracking method according to claim 1, characterized in that, If the detection result set includes the arrangement recognition result, and the deep learning model includes a third contour segmentation model, the method for obtaining the arrangement recognition result includes: The preprocessed image data is input into the third contour segmentation model, and the third segmentation mask of the billet region is output. Based on the third segmentation mask, the sequence number information of each casting billet in the image is identified. The sequence number information includes the arrangement order of each casting billet in the image and the corresponding sequence number identifier. Based on the serial number information, obtain the first arrangement order of the billets when they enter the inspection port and the second arrangement order when they leave the inspection port; If the first arrangement order is consistent with the second arrangement order, the arrangement identification result is determined to be that the billet arrangement is normal; if the first arrangement order is inconsistent with the second arrangement order, the arrangement identification result is determined to be that the billet arrangement is abnormal. The third contour segmentation model is obtained by training a preset third contour segmentation network on a fifth training set. The third contour segmentation network learns the segmentation features of the billet serial number by extracting image features and outputs the third segmentation mask based on the input preprocessed image data. The fifth training set is billet image data with billet serial number annotation information, which annotates the serial number identifier and arrangement order of each billet in the image.
7. The billet tracking method according to any one of claims 1 to 6, characterized in that, Based on the detection result set and the status identification information, the billet is matched for identity and tracked according to preset tracking rules to obtain tracking result data, including: Based on the set of detection results and the status identification information, a corresponding virtual billet number is assigned to each billet. Based on the virtual billet number, the real-time position and path of each billet are updated according to preset position rules and preset queue rules to form a continuous movement trajectory of each billet. Based on the comparison results of the first virtual billet number when entering the preset verification area and the second virtual billet number when leaving the preset verification area, an abnormal marker state is assigned to each billet. Based on the status identification information, combined with the continuous movement trajectory and abnormal marker status of each billet, tracking details for each billet are generated, and according to the hot delivery status or offline status in the status identification information, the corresponding hot delivery record or offline record is output, thereby obtaining the tracking result data.
8. The billet tracking method according to any one of claims 1 to 6, characterized in that, The method further includes: Sensors and image acquisition devices are deployed at key nodes on the casting and rolling production line to collect various types of data in real time. Multiple data recognition queues are constructed according to time sequence and data type, and a timestamp and device identifier are added to each data. A data fusion algorithm is used to connect the data in each data recognition queue, establish data connections based on recognition logic and time consistency, form a complete dataset, and extract features from the data in the complete dataset to obtain feature vectors; The feature vector is input into the anomaly detection model to obtain the anomaly detection result. The anomaly detection model is trained on a preset anomaly detection network using training samples, which are historical data containing normal and abnormal states. If the anomaly detection result is abnormal, an early warning message containing the time and location of the anomaly is generated, and a preset offline process is triggered.
9. The billet tracking method according to any one of claims 1 to 6, characterized in that, The method further includes: The image quality of the initial image data, the device status of the preset device, and the data synchronization of the preset data are monitored in real time. Based on the image quality, the device status, and the data synchronization, the fault level is determined. If the fault level is a Level 1 fault, adjust the vision processing parameters; If the fault level is a level 2 fault, activate the backup image acquisition device and correct the positioning deviation by adjusting the roller speed. If the fault level is level three, the mill will be slowed down or partially shut down, and the tracking continuity will be maintained in conjunction with the positioning equipment.
10. A billet tracking system, characterized in that, include: The data acquisition module is used to acquire initial image data and status identification information of the billet in the casting and rolling operation area; The preprocessing module is used to preprocess the initial image data using a preprocessing algorithm deployed on the edge computing unit to obtain preprocessed image data; The analysis and processing module is used to process and analyze the preprocessed image data using a deep learning model deployed on the edge computing unit to obtain a set of detection results for the billet, wherein the set of detection results includes at least one of end face recognition results, shape recognition results, motion detection results, displacement detection results, and arrangement recognition results; The billet tracking module is used to perform identity matching and trajectory tracking on the billet according to the detection result set and the status identification information, and obtain tracking result data to complete the billet tracking.
