Vacuum induction furnace inner crucible block identification method and system

Abstract

The application discloses a kind of vacuum induction furnace inner crucible broken block identification method and system, it is related to special smelting technical field.The method includes: obtaining the metal liquid surface image in vacuum induction furnace;Crucible broken block identification model based on improved YOLOv8 model is constructed, the model includes cascaded improved Backbone main network, neck network and detection head, wherein improved Backbone main network is obtained by replacing part C2f module in original Backbone with C3CrossConv module and replacing the last SPPF module with hollow spatial pyramid pooling module;C3CrossConv module extracts feature maps of different scales from metal liquid surface image, and hollow spatial pyramid pooling module extracts enhanced feature maps from the feature maps output by the last C3CrossConv module;Neck network splices and fuses enhanced feature maps and initial feature maps of different scales;Detection head uses fused feature maps to identify targets in metal liquid surface image Crucible broken block.The present application can effectively improve the detection accuracy of tiny broken block in high-temperature melt environment.

Description

Technical Field

[0001] This application relates to the field of automated detection technology in the metallurgical industry, and in particular to a method for identifying crucible fragments inside a vacuum induction furnace. Background Technology

[0002] In vacuum induction melting, the crucible, as a key component for holding molten metal, directly affects the smelting quality due to its high-temperature resistance. However, prolonged exposure to high-temperature molten environments can easily cause the crucible surface to peel off, forming crucible fragments of varying sizes. These fragments float on the surface of the molten metal, and if not detected in time, can lead to a series of quality problems: the fragments may be re-absorbed into the melt, forming non-metallic inclusions, affecting the cleanliness of the molten metal, and reducing the mechanical properties of the finished product.

[0003] In the current technology, machine vision technology is gradually being applied in the metallurgical industry. It uses industrial cameras to capture images of the surface of molten metal inside a furnace, and then utilizes methods such as threshold segmentation or edge detection to identify and locate crucible fragments on the surface of the molten metal. This type of method typically relies on image information in the visible light band, extracting the color, shape, or texture features of the crucible fragments to achieve automatic detection of floating objects on the surface of the molten metal.

[0004] However, the surface of molten metal also contains a large number of interfering targets such as bubbles and metal splashes with different shapes and motion characteristics. These targets have a high degree of similarity to the appearance of the crucible fragments in the imaging results. In addition, the crucible fragments are extremely small in size, occupying only a few pixels in the image. Furthermore, their irregular shape and simple texture features make them lacking in stable and distinguishable appearance. When extracting the color, shape, or texture features of crucible fragments, the existing technologies mentioned above have difficulty distinguishing between the actual crucible fragments and interfering targets on the surface of molten metal. This results in insufficient detection accuracy of crucible fragments and makes it difficult to meet the monitoring requirements of actual industrial sites. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for identifying crucible fragments inside a vacuum induction furnace to address the aforementioned technical problems.

[0006] The present invention adopts the following technical solution: This invention provides a method for identifying crucible fragments inside a vacuum induction furnace, comprising: Acquire continuous frames of images of the molten metal surface inside a vacuum induction furnace; Construct a crucible fragment recognition model; use the trained crucible fragment recognition model to identify crucible fragments in the molten metal surface image to obtain the initial recognition result of crucible fragments in each frame of the molten metal surface image; A mapping relationship between the image pixel coordinate system and the actual physical coordinate system is established to determine the actual projected area of ​​the initial identification result of the crucible fragment in the molten metal surface image; the optical flow method is used to track the displacement of the detection box center point of the initial identification result of the crucible fragment in the continuous frame molten metal surface image to determine the movement speed of each initial identification result of the crucible fragment; the crucible fragment contour of the initial identification result of the crucible fragment in each frame molten metal surface image is extracted, and the average morphological change rate of the initial identification result of the crucible fragment in all frames is determined based on the crucible fragment contour of each frame; The initial identification results of crucible fragments are filtered based on their actual projected area, movement speed, and average morphological change rate in the image of the molten metal surface, to obtain the final identification results of the crucible fragments.

[0007] Preferably, optical flow is used to track the displacement of the center point of the detection frame of the initial identification result of the crucible fragments in the continuous frame images of the molten metal surface, in order to determine the movement speed of each initial identification result of the crucible fragments, specifically including: The Lucas-Kanade optical flow method is used to calculate the pixel displacement between the crucible fragment identification result in the current frame and the crucible fragment identification result in the previous frame; Obtain the frame interval time between two consecutive images of the molten metal surface, as well as the pre-calibrated pixel-to-physical distance conversion coefficient; The movement speed of the crucible fragment recognition result in physical space is calculated based on the pixel displacement, frame interval time, and conversion coefficient.

[0008] Preferably, the crucible fragment outlines of the initial identification results in the continuous frame images of the molten metal surface are extracted, and the average morphological change rate of the initial identification results of the crucible fragments in the continuous frame images of the molten metal surface is determined based on the crucible fragment outlines, specifically including: The crucible fragment outlines in adjacent frames of molten metal surface images are extracted using a mask, and the area of ​​the symmetrical difference region between the crucible fragment outlines of two adjacent frames is calculated; wherein, the area of ​​the symmetrical difference region is the difference between the area of ​​the union region and the area of ​​the intersection region of the crucible fragment outlines of two frames. The ratio of the area of ​​the symmetrical difference region to the area of ​​the intersection region of the two frame contours is determined as the crucible fragment contour overlap rate between adjacent frames. The average morphological change rate of the crucible fragment identification results within a sliding time window of consecutive frames is statistically analyzed. The average morphological change rate is the average of the overlap rate of crucible fragment contours between adjacent frames.

[0009] Preferably, the initial identification results of the crucible fragments are filtered based on the actual projected area, movement speed, and average morphological change rate of the crucible fragments in the image of the molten metal surface, to obtain the final crucible fragment identification results, specifically including: When the actual projected area of ​​the crucible fragment identification result in the molten metal surface image is less than the preset area threshold, the crucible fragment identification result is judged as a bubble or noise and is removed. For crucible fragment identification results filtered by area, if the real-time movement speed of the center point of the crucible fragment is greater than the preset speed threshold, the crucible fragment identification result will be judged as a fast-moving bubble or splash and rejected. For crucible fragment identification results that pass the speed screening, if the average morphological change rate of crucible fragment identification results in consecutive frames is greater than the preset morphological change rate threshold, the crucible fragment identification results are judged as unstable interference objects and removed to obtain the final crucible fragment identification results.

[0010] Preferably, the crucible fragment identification model is an improved YOLOv8 model, which includes: a cascaded improved backbone network, a neck network, and a detection head; the improved backbone network is obtained by replacing the C2f module in the original backbone network with the C3CrossConv module, and replacing the SPPF module at the end of the backbone network with a void space pyramid pooling module; Each C3CrossConv module extracts crucible fragment feature maps at multiple scales from the surface image of the molten metal; the void space pyramid pooling module extracts crucible fragment enhancement feature maps from the crucible fragment feature maps output by the last C3CrossConv module; The neck network stitches and fuses the enhanced feature map of the crucible fragments and the initial feature maps of the crucible fragments at different scales to obtain a fused feature map of the crucible fragments. The detection head uses the fused feature map of crucible fragments to identify the target and obtain the crucible fragment identification result.

[0011] Preferably, the crucible fragment identification model further embeds SE modules at the connection between the backbone network and the neck network, as well as in some nodes of the neck network; The SE module compresses the spatial dimension features of the backbone network output into channel descriptors through global average pooling. The correlation between channels is learned by using fully connected layers, channel weight coefficients between 0 and 1 are generated, and the learned weight coefficients are multiplied with the original feature map.

[0012] Preferably, each C3CrossConv module includes two parallel convolutional branches with convolutional kernels of different sizes; each C3CrossConv module extracts feature maps of crucible fragments of different scales from the molten metal surface image, specifically including: For each C3CrossConv module, parallel convolution branches with two different sized convolution kernels in the C3CrossConv module are used to extract two initial crucible fragment feature maps of different scales from the molten metal surface image. The two initial crucible fragment feature maps of different scales are stitched together, and the stitched initial crucible fragment feature maps are then subjected to channel shuffling to obtain crucible fragment feature maps.

[0013] Preferably, the hollow spatial pyramid pooling module includes: multiple parallel branches of hollow convolution with different dilation rates; the hollow spatial pyramid pooling module extracts the crucible fragment enhancement feature map from the crucible fragment feature map output by the last C3CrossConv module in the improved backbone network, specifically including: The hollow space pyramid pooling module uses its parallel branches of hollow convolution with different expansion rates to extract the initial enhanced feature map of crucible fragments corresponding to each expansion rate from the feature map of crucible fragments output by the C3CrossConv module. The initial enhanced feature maps of crucible fragments corresponding to each expansion rate are spliced ​​and fused to obtain the enhanced feature map of crucible fragments.

[0014] Preferably, the loss function used in the training process of the crucible fragment recognition model is determined by the IoU loss term, the center point distance loss term, the width and height loss term, and the similarity metric term based on the normalized Wasserstein distance. The IoU loss term characterizes the ratio of the intersection area to the union area of ​​the predicted box and the actual box of crucible fragment identification. The center point distance loss term characterizes the ratio of the Euclidean distance between the center point of the predicted box and the center point of the actual box to the diagonal distance of the minimum closure region. The width and height loss term characterizes the ratio of the difference between the width of the predicted box and the width of the actual box to the width of the minimum closure region, and the ratio of the difference between the height of the predicted box and the height of the actual box to the height of the minimum closure region. The similarity metric based on normalized Wasserstein distance provides the gradient direction when the predicted box and the actual box do not overlap.

[0015] This invention provides a crucible fragment identification system inside a vacuum induction furnace, comprising: The data acquisition module is used to acquire continuous frames of images of the molten metal surface inside the vacuum induction furnace; The fragment recognition module is used to construct a crucible fragment recognition model; the trained crucible fragment recognition model is used to identify crucible fragments in the molten metal surface image to obtain the initial recognition result of crucible fragments in each frame of the molten metal surface image; The fragment screening module is used to establish the mapping relationship between the image pixel coordinate system and the actual physical coordinate system to determine the actual projected area of ​​the initial crucible fragment identification result in the molten metal surface image; it uses optical flow to track the displacement of the detection box center point of the initial crucible fragment identification result in consecutive frames of molten metal surface images to determine the movement speed of each initial crucible fragment identification result; it extracts the crucible fragment contour of the initial crucible fragment identification result in each frame of molten metal surface image, and determines the average morphological change rate of the initial crucible fragment identification results in all frames based on the crucible fragment contour of each frame; it filters the initial crucible fragment identification results based on the actual projected area, movement speed, and average morphological change rate in the molten metal surface image to obtain the final crucible fragment identification result.

[0016] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: In the crucible fragment identification method provided by this invention, a mapping relationship from pixel space to physical space is constructed to transform the detection result into a projected area with actual physical meaning. Furthermore, the displacement field of the target between consecutive frames is tracked using an optical flow field to quantify its motion inertia. Simultaneously, the morphological evolution of the crucible within a short time window is characterized based on the fragment contour, forming three dimensions of apparent features of the crucible fragments in the image of the molten metal surface: projected area, motion speed, and morphological change rate. The crucible fragments themselves possess three inherent physical behavioral characteristics: small geometric size, low-speed flow, and stable contour. Based on these physical behavioral characteristics, the three dimensions of projected area, motion speed, and morphological change rate are used to filter the crucible fragment identification results output by the deep learning model. This effectively eliminates other interference targets in the crucible fragment identification results that are similar in appearance to the fragments but have different physical behaviors, thereby reducing the false recognition rate of crucible fragment identification and significantly improving the detection accuracy of small fragments in the complex molten metal surface environment inside the vacuum induction furnace. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0018] Figure 1 A flowchart illustrating a method for identifying crucible fragments inside a vacuum induction furnace, provided by the present invention. Figure 2A schematic diagram of a vacuum induction furnace system for a method of identifying crucible fragments inside a vacuum induction furnace provided by the present invention; Figure 3 A model architecture diagram of an improved model for identifying crucible fragments inside a vacuum induction furnace provided by the present invention; Figure 4 The flowchart of the algorithm for the improved model of the crucible fragment identification method in a vacuum induction furnace and the fusion of optical flow method provided by the present invention; Figure 5 This is a schematic diagram of the overall operation architecture of a method for identifying crucible fragments inside a vacuum induction furnace provided by the present invention; Figure 6 This is a schematic diagram of a crucible fragment identification system in a vacuum induction furnace provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in the specification without creative effort are within the scope of protection of this application.

[0020] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of a method for identifying crucible fragments inside a vacuum induction furnace according to the present invention, which specifically includes the following steps: S101: Acquire continuous frames of images of the molten metal surface inside the vacuum induction furnace.

[0022] Specifically, acquiring images of the molten metal surface inside a vacuum induction furnace includes: initializing parameters of an industrial camera mounted on the furnace body, setting the exposure time to dynamic adjustment mode, and setting an initial range for the exposure time; installing a narrowband filter with a specified center wavelength and bandwidth in front of the optical lens; acquiring real-time images of the molten metal surface through a sapphire viewing window on the furnace body using the industrial camera to obtain an original molten metal surface image in RGB color space format; using a multi-exposure fusion algorithm to perform high dynamic range imaging processing on the acquired original molten metal surface images, and synthesizing multiple original molten metal surface images with combinations of long and short exposures in real time; and using a bilinear interpolation algorithm to perform downsampling preprocessing on the original molten metal surface images after high dynamic range processing, smoothly scaling the image size to a preset size to obtain the molten metal surface image.

[0023] S102: Construct a crucible fragment recognition model; use the trained crucible fragment recognition model to recognize crucible fragments in the molten metal surface image to obtain the initial recognition result of crucible fragments in each frame of the molten metal surface image.

[0024] Optionally, the crucible fragment recognition model is an improved YOLOv8 model, which includes: a cascaded improved backbone network, a neck network, and a detection head; the improved backbone network is obtained by replacing the C2f module in the original backbone network with the C3CrossConv module, and replacing the SPPF module at the end of the backbone network with a void space pyramid pooling module.

[0025] Each C3CrossConv module extracts crucible fragment feature maps at multiple scales from the molten metal surface image step by step; the void space pyramid pooling module extracts crucible fragment enhancement feature maps from the crucible fragment feature maps output by the last C3CrossConv module; the neck network splices and fuses the crucible fragment enhancement feature maps and the initial crucible fragment feature maps at different scales to obtain a fused crucible fragment feature map; the detection head uses the fused crucible fragment feature map to perform target recognition and obtain the crucible fragment recognition result.

[0026] The crucible fragment recognition model based on the improved YOLOv8 architecture further embeds SE modules at the connection points between the backbone and neck networks, as well as at key nodes of the neck network. The SE modules compress the spatial dimension features output by the backbone network into channel descriptors using global average pooling. Fully connected layers are used to learn the correlations between channels, generating channel weight coefficients between 0 and 1, and these learned weight coefficients are multiplied by the original feature map. See the improved YOLOv8 model architecture diagram. Figure 3 .

[0027] Optionally, each C3CrossConv module includes two parallel convolutional branches with convolutional kernels of different sizes; each C3CrossConv module extracts crucible fragment feature maps of different scales from the molten metal surface image, specifically including: for each C3CrossConv module, using the parallel convolutional branches with two convolutional kernels of different sizes in the C3CrossConv module to extract two initial crucible fragment feature maps of different scales from the molten metal surface image respectively; stitching the two initial crucible fragment feature maps of different scales together, and performing channel shuffling on the stitched initial crucible fragment feature maps to obtain crucible fragment feature maps.

[0028] Optionally, the hollow spatial pyramid pooling module includes: multiple parallel branches of dilated convolution with different expansion rates; the hollow spatial pyramid pooling module extracts crucible fragment enhancement feature maps from the crucible fragment feature map output by the last C3CrossConv module in the improved backbone network, specifically including: the hollow spatial pyramid pooling module uses its parallel branches of dilated convolution with different expansion rates to extract the initial enhancement feature map of the crucible fragment corresponding to each expansion rate from the crucible fragment feature map output by the C3CrossConv module; the initial enhancement feature maps of the crucible fragment corresponding to each expansion rate are spliced ​​and fused to obtain the crucible fragment enhancement feature map.

[0029] Optionally, the loss function used in the training process of the crucible fragment recognition model is determined by an IoU loss term, a center point distance loss term, a width-height loss term, and a similarity metric based on normalized Wasserstein distance. The IoU loss term represents the ratio of the intersection area to the union area of ​​the predicted and actual crucible fragment boxes. The center point distance loss term represents the ratio of the Euclidean distance between the center point of the predicted and actual crucible fragment boxes to the diagonal distance of the minimum closure region. The width-height loss term represents the ratio of the difference between the width of the predicted and actual crucible fragment boxes to the width of the minimum closure region, and the ratio of the difference between the height of the predicted and actual crucible fragment boxes to the height of the minimum closure region. The similarity metric based on normalized Wasserstein distance provides the gradient direction when the predicted and actual crucible fragment boxes do not overlap.

[0030] S103: Establish the mapping relationship between the image pixel coordinate system and the actual physical coordinate system to determine the actual projected area of ​​the initial identification result of the crucible fragment in the molten metal surface image; use optical flow method to track the displacement of the detection box center point of the initial identification result of the crucible fragment in the continuous frame molten metal surface image to determine the movement speed of each initial identification result of the crucible fragment; extract the crucible fragment contour of the initial identification result of the crucible fragment in each frame molten metal surface image, and determine the average morphological change rate of the initial identification result of the crucible fragment in all frames based on the crucible fragment contour of each frame.

[0031] Optionally, optical flow is used to track the displacement of the detection box center point of the initial crucible fragment identification result in consecutive frames of molten metal surface images to determine the movement speed of each initial crucible fragment identification result. Specifically, this includes: using the Lucas-Kanade optical flow method to calculate the pixel displacement between the crucible fragment identification result in the current frame and the crucible fragment identification result in the previous frame; obtaining the frame interval time between two consecutive frames of molten metal surface images, as well as the pre-calibrated pixel-to-physical distance conversion coefficient; and calculating the movement speed of the crucible fragment identification result in physical space based on the pixel displacement, frame interval time, and conversion coefficient.

[0032] Optionally, the crucible fragment outlines of the initial identification results in consecutive frames of molten metal surface images are extracted, and the average morphological change rate of the initial identification results in consecutive frames of molten metal surface images is determined based on the crucible fragment outlines. Specifically, this includes: extracting the crucible fragment outlines of the identification results in adjacent frames of molten metal surface images using a mask, and calculating the area of ​​the symmetrical difference region of the crucible fragment outlines in two adjacent frames; determining the ratio of the symmetrical difference region area to the area of ​​the intersection region of the two frame outlines as the crucible fragment outline overlap rate between adjacent frames; and statistically analyzing the average morphological change rate of the crucible fragment identification results within a sliding time window of consecutive frames, wherein the average morphological change rate is the average of the crucible fragment outline overlap rates between adjacent frames; wherein the area of ​​the symmetrical difference region is the difference between the area of ​​the union region and the area of ​​the intersection region of the crucible fragment outlines in two frames.

[0033] S104: Based on the actual projected area, movement speed, and average morphological change rate of the crucible fragments in the image of the molten metal surface, the initial identification results of the crucible fragments are filtered to obtain the final identification results of the crucible fragments.

[0034] Optionally, optical flow is used to track the displacement change of the center point of the detection frame in a continuous frame sequence to determine the real-time movement speed of the center point of the crucible fragment identification result. Specifically, this includes: using the Lucas-Kanade optical flow method to calculate the pixel displacement between the crucible fragment identification result in the current frame and the crucible fragment identification result in the previous frame; obtaining the frame interval time between two consecutive frames of molten metal surface images, as well as the pre-calibrated pixel-to-physical distance conversion coefficient; and calculating the real-time movement speed of the crucible fragment identification result in physical space based on the pixel displacement, frame interval time, and conversion coefficient.

[0035] Optionally, the contours of the crucible fragment recognition results are extracted, and the overlap rate of the crucible fragment contours between adjacent frames is calculated. Specifically, this includes: extracting the contours of the crucible fragment recognition results in two consecutive frames of molten metal surface images using a mask; calculating the area of ​​the symmetrical difference region of the contours in two adjacent frames, wherein the symmetrical difference region is the union region of the contours in the two frames minus the intersection region; determining the ratio of the area of ​​the symmetrical difference region to the area of ​​the intersection region of the contours in the two frames as the overlap rate of the crucible fragment contours between adjacent frames; and calculating the average morphological change rate of the crucible fragment recognition results within a sliding time window of multiple consecutive frames, wherein the average morphological change rate is the average value of the overlap rate of the crucible fragment contours between each adjacent frame.

[0036] Furthermore, the complete execution process for implementing the method provided by the present invention is provided below.

[0037] The implementation of the method of the present invention is based on, as follows: Figure 5 The overall architecture is shown. The device layer includes a sapphire window, an industrial camera, an imaging unit, and a light source control module. The sapphire window is 2mm thick and 50mm in diameter, mounted at the observation window of the vacuum induction furnace, with a transmission wavelength of 330-1100nm, and is replaced periodically in the factory. The industrial camera has a resolution of 1920×1080, a frame rate of 70fps, and supports water cooling and dynamic exposure control. The imaging unit provides a 120dB dynamic range, ensuring no loss of image detail in complex environments with both strong light and shadow. The light source control module adjusts ambient light to ensure optimal imaging conditions.

[0038] The control layer includes an image processing unit, a model inference unit, and a post-processing unit. The image processing unit is responsible for image acquisition, preprocessing, and feature extraction; the model inference unit runs an improved YOLO model to detect fragments in the image; and the post-processing unit combines the YOLO detection results with optical flow motion trajectory analysis to perform multi-feature fusion judgment.

[0039] The workshop level includes an alarm module and a communication interface. The alarm module triggers an alarm signal when it detects fragments; the communication interface supports the OPC UA protocol to upload the detection results to the MES / SCADA system.

[0040] To address the complex imaging environment of a vacuum induction furnace, characterized by high temperature, high dynamic range, and strong electromagnetic interference, this embodiment constructs a standardized image acquisition and preprocessing pipeline. This process aims to acquire high-quality, low-noise image data that is compatible with deep learning model inputs through hardware parameter optimization and algorithm enhancement. The specific steps are as follows:

[0041] First, based on the actual operating characteristics of the vacuum induction furnace, see... Figure 2This is a schematic diagram of a vacuum induction furnace system, showing the initial parameter settings for the industrial camera. The exposure time is set to a dynamic adjustment mode, with an initial range limited to 0.1ms to 10ms to accommodate the drastic fluctuations in light intensity during the melting process. Furthermore, to effectively filter out background noise generated by the thermal radiation of the molten metal and enhance the contrast of fragments, a narrowband filter with a center wavelength of 650nm ± 10nm is installed in front of the optical lens.

[0042] After parameter configuration, the system acquires real-time images of the molten metal surface through a sapphire viewing window on the vacuum furnace. The acquisition process uses RGB color space format, and the original output image maintains a high resolution of 1920×1080 to ensure that the texture details of the molten pool surface, liquid surface ripples, and potential floating object features are fully recorded.

[0043] This example introduces a high dynamic range (HDR) imaging processing module to handle the high dynamic range scenario during smelting, where the center of the molten pool is hot and bright, while the edges and shadow areas are dark. This module employs a multi-exposure fusion algorithm to synthesize 3 to 5 original images with combinations of long and short exposures in real time. Through fusion processing, the overall dynamic range of the image is extended to 120dB. Next, this example uses a bilinear interpolation algorithm to perform downsampling preprocessing on the processed image, smoothly scaling the 1920×1080 image size to 416×416. During downsampling, the algorithm focuses on optimizing the preservation of spatial information, ensuring that the edge sharpness and contour features of the fragments are not severely blurred due to the size reduction. Simultaneously, the image data format is converted to BGR format to adapt to the input tensor requirements of subsequent models.

[0044] This embodiment establishes a standardized "physical-visual" joint data acquisition and annotation system to address the problems of low signal-to-noise ratio and ambiguous annotation in high-temperature environments. The specific construction process is as follows:

[0045] Data acquisition strictly followed the imaging parameter configuration described in the claims, utilizing an industrial camera deployed on the side wall of the furnace with a narrow-band filter for continuous recording. The acquisition covered the entire lifecycle of the smelting process, including the melting period with intense light flashes and violent fluctuations in the liquid surface, the refining period with high temperature and brightness and a calm surface, and the casting period with the liquid level dropping and the edges solidifying.

[0046] The LabelImg tool was used to annotate the acquired 1080P high-definition images with rectangular bounding boxes, generating JSON label files in PASCAL VOC format. During the annotation process, this example used camera-calibrated coordinate mapping combined with motion vectors calculated by optical flow for auxiliary correction. For the annotated object, if its average motion velocity v > 0.1 m / s over 10 consecutive frames and its morphological change rate... If the value is greater than 5%, it will be forcibly classified as a dynamic disturbance rather than a static target.

[0047] This example also establishes a dynamic exposure control mechanism based on image feedback to address the nonlinear changes in light intensity at different stages of melting. This mechanism calculates the global average brightness of the current frame image in real time and sets a standard brightness threshold range of 80-120 grayscale values. When the average image brightness exceeds this threshold range, the control module automatically fine-tunes the camera's exposure time, forming a closed-loop feedback.

[0048] To address the challenges of small crucible fragments in a vacuum induction furnace, including their small size, complex background, and strong radiation interference, this embodiment proposes a small target detection network architecture for high-temperature molten environments, building upon the YOLOv8 baseline architecture. This architecture significantly improves the model's feature extraction and localization capabilities for small fragments by introducing a multi-scale feature fusion module, expanding the receptive field, enhancing attention, and optimizing the loss function for small targets. (See [link to relevant documentation]). Figure 3 The specific improvements are detailed below:

[0049] In the YOLOv8 backbone network, while the original C2f module possesses certain feature extraction capabilities, it easily loses crucial spatial detail information when processing targets with extremely small scales and simple texture features, such as crucible fragments. This embodiment replaces some standard C2f modules in the backbone network with C3CrossConv modules. The C3CrossConv module introduces a cross-convolutional structure to process feature maps of different scales in parallel. Specifically, this module contains two parallel convolutional branches, using 3×3 and 5×5 convolutional kernels respectively, to capture local features within different receptive fields. Subsequently, cross-dimensional information interaction is achieved through feature concatenation and channel shuffling operations.

[0050] At the end of the backbone network, to expand the receptive field without reducing the feature map resolution and to accommodate the scale changes that may occur in the fragments in molten metal, this embodiment uses a SPPM (Spatial Pyramid Pooling) structure instead of the traditional SPPF module. The SPPM structure extracts features in parallel using dilated convolutions with different dilation rates. This embodiment designs three parallel dilated convolutional layers with dilation rates of 1, 3, and 5. A dilation rate of 1 is equivalent to standard convolution, used to preserve basic features; dilation rates of 3 and 5 can significantly expand the receptive field of each pixel on the feature map without increasing the number of parameters and computational cost. These feature maps of different scales are concatenated after output, and then channel fusion and dimensionality reduction are performed through a 1×1 convolutional layer, finally outputting a single feature map. The final image contains both the fine texture of the fragments from the small dilation rate branch and the global semantics from the large dilation rate branch in a complex background. This feature map is then passed as input to the next layer of the network.

[0051] Due to optical distortion caused by temperature gradients in the vacuum induction furnace viewing window and random splashes on the molten metal surface, the background is extremely complex. To enable the network to focus on key target areas, this embodiment introduces a channel attention mechanism. SE modules are embedded at the connection between the backbone and neck network, and at the end of the C2f module after concatenating top-down and bottom-up path features in the neck network. The SE module recalibrates the channel feature responses through a "compression-excitation" operation. It compresses spatial features into channel descriptors using global average pooling, capturing global contextual information. Then, it uses fully connected layers to learn the correlation between channels, generating weight coefficients between 0 and 1. Finally, the learned weight coefficients are multiplied by the original feature map to obtain an enhanced feature map with a higher signal-to-noise ratio and greater discriminative power. This mechanism can automatically learn and assign higher weights to channels containing crucible fragment features, while suppressing channel responses related to background noise such as thermal radiation spots and liquid surface reflections.

[0052] By constructing a crucible fragment recognition model based on an improved YOLOv8 architecture, this model replaces some C2f modules in the original Backbone network with C3CrossConv modules and uses their parallel multi-scale convolution branches to extract initial feature maps under different receptive fields in the molten metal surface image. At the same time, the SPPF module at the end is replaced with a hollow spatial pyramid pooling module to expand the receptive field and extract enhanced feature maps without reducing resolution. Then, the neck network splices and fuses the enhanced feature maps with the initial feature maps at different scales, enabling the detection head to accurately locate and identify tiny fragments using the fused multi-scale feature maps.

[0053] To address the challenges of imaging environments where high-temperature radiation from molten metal in vacuum induction furnaces leads to localized overexposure and loss of fragment details, as well as the complex conditions of extremely small fragment size, simple texture features, and background interference from bubbles and splashes, this improved model employs the synergistic effect of multi-scale parallel convolution and hollow pyramid pooling. It progressively enhances the spatial details and contextual semantic information of fragments from the bottom to the top layers of the backbone network. Furthermore, cross-layer fusion via the neck network ensures that shallow positional information and deep semantic information are fully complementary. This solves the detection problem of weak fragment features easily drowned out by background noise in high-temperature scenarios across the entire feature extraction and fusion chain. Based on this technical concept of multi-scale feature enhancement and cross-layer fusion, the detection accuracy of tiny crucible fragments in high-temperature molten environments can be effectively improved, while reducing the false alarm rate caused by dynamic interference such as bubbles and splashes. This provides reliable technical support for real-time online monitoring of the vacuum induction melting process.

[0054] During model training, IoU is defined as the ratio of the area of ​​intersection to the area of ​​union of the predicted bounding box Bp and the ground truth bounding box Bgt. Traditional loss function:

[0055] ; ; There is a drawback when handling small targets: when the predicted bounding box and the ground truth bounding box are only slightly offset, if they do not intersect, the IoU value is 0, leading to gradient vanishing and ineffective optimization. This positional deviation is extremely sensitive for crucible fragments with very small pixel areas. This example uses EIoU optimized by normalized Wasserstein distance as the bounding box regression loss function.

[0056] The loss function consists of three parts: the IoU loss term, the center point distance loss term, and the width and height loss term. Its formula is defined as follows:

[0057] ; The unfolded form is as follows: ; in: b and b gt These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. (·) represents the square of the Euclidean distance. c This represents the diagonal distance of the smallest closure region containing two boxes. w and h These represent the width and height of the prediction box, respectively; w gt and h gt This represents the width and height of the actual bounding box. C w and C h These represent the width and height of the smallest closure region, respectively.

[0058] Define prediction box B With real frame B gt The Wasserstein distance between them is W 2 By introducing a normalization factor, the distance metric is transformed into a similarity metric based on a probability distribution. The improved loss function LNWD−EIoU is defined as follows:

[0059] ; Among them, L EIoU The original loss is represented by λ, the balance coefficient is represented by C, and C is a constant. W 2The normalized distance measures the distance between two bounding boxes as a Gaussian distribution. Even when the predicted and ground truth boxes do not overlap at all, the normalized distance provides an effective gradient direction to guide the model's optimization. This formula preserves... Building upon geometric regression capabilities, a measure of probability distribution similarity is added, thereby significantly improving the model's sensitivity to the location of small fragments.

[0060] In a vacuum melting environment, numerous tiny bubbles are generated on the surface of the melt. These bubbles often appear as bright spots in images and are easily misdetected as fragments. However, destructive crucible fragments typically have a certain physical size. To further address the dynamic interference problem under the complex operating conditions of a vacuum induction furnace, this embodiment designs a post-processing algorithm module based on multi-feature fusion. This module acts as a secondary filter for the detection model, aiming to utilize prior physical knowledge to logically verify the preliminary detection results output by the model, thereby accurately distinguishing between real crucible fragments and interference items such as bubbles and light shadows on the melt surface in both time and physical dimensions. The specific filtering steps are as follows:

[0061] The system first establishes a mapping between the image pixel coordinate system and the actual physical coordinate system through camera calibration. Based on this mapping, the area of ​​the bounding box detection frame projected onto the molten surface is calculated. In this example, the area threshold Tarea is set to 10 mm². For each detected target, its actual physical area Area is calculated. If Area < 10 mm², the target is determined to be a harmless small bubble or noise and is discarded; if Area ≥ 10 mm², the target is retained and proceeds to the next level of feature analysis.

[0062] After crucible fragments detach, they typically float or semi-suspend on the surface of the high-viscosity molten metal. Their movement is controlled by the overall flow of the molten metal, and their speed is relatively slow and smooth. In contrast, bubbles, rising and bursting under buoyancy or influenced by electromagnetic stirring, often exhibit rapid and erratic motion. This example uses the Lucas-Kanade optical flow method to track the motion vector of the preserved target. Figure 4 The flowchart illustrates the algorithm for improving the fusion of the model and optical flow method. The real-time motion velocity v is estimated by calculating the displacement change of the target center point in a continuous frame sequence.

[0063] ; Where xt and yt are the coordinates of the current frame. t is the frame interval, and Scale is the conversion coefficient from pixel to physical distance.

[0064] In this example, the velocity threshold Tvel is set to 0.1 m / s. If the instantaneous velocity v of the target is greater than 0.1 m / s, the target is determined to be a fast-moving bubble or splash and is excluded; only targets with velocities v ≤ 0.1 m / s are retained.

[0065] Crucible fragments are typically solid materials, and their geometry remains relatively rigid within a short observation window, without undergoing drastic deformation. In contrast, bubbles expand, merge, or burst as they rise to the surface, exhibiting a high degree of instability in their profiles.

[0066] The system maintains a sliding time window of length N=10 frames. The contours of the target are extracted using a mask, and the contour overlap rate between adjacent frames is calculated.

[0067] ; in, This represents the symmetric difference operation, which is the union minus the intersection, used to measure differences in shape.

[0068] Set the shape change rate threshold Tshape = 5%. The system calculates the average change rate of the target within consecutive frames. .like A value >5% indicates that the target's morphology is extremely unstable and is considered interference; if

[0069] If the value is ≤ 5%, the target is considered to have solid-state rigidity characteristics and conforms to the fragmentation property.

[0070] An AND logic gate strategy is used for the final decision. An object is only identified as crucible fragment if it simultaneously meets all three of the following conditions: Areal ≥ 10 mm², v ≤ 0.1 m / s, and ≤ 5%;

[0071] Once the above comprehensive judgment is triggered, the system immediately marks the target location on the monitoring interface, draws the motion trajectory, and triggers an audible and visual alarm signal to notify the operator to take slag removal measures.

[0072] This embodiment employs a distributed edge computing architecture, deeply integrating intelligent detection functionality into the existing VIM (Vacuum Induction Melting) production system. The system hardware deployment follows the principle of "data remaining within the workshop and computing power residing at the edge," with each vacuum induction furnace station independently configured with an industrial-grade high-definition camera and an NVIDIA Jetson AGX Orin embedded edge computing unit. All image acquisition, preprocessing, YOLOv8 model inference, and multi-feature fusion post-processing algorithms are completed within the local edge computing unit. The edge computing unit is installed in an industrial cabinet with electromagnetic shielding and heat dissipation capabilities, connected to the camera via shielded twisted-pair cables or fiber optics to resist strong electromagnetic interference generated during furnace operation. Simultaneously, to balance real-time monitoring requirements with storage costs, the system adopts a "results-driven" data transmission strategy. When the system receives continuous signals and detects no anomalies, it only sends a "heartbeat packet" and a "normal" status code to the central server, with a very small amount of data. If the system does not receive data for 3 consecutive seconds, it is judged as a "system failure," triggering a low-confidence level 1 alarm. The industrial screen will continuously flash a red alarm, awaiting manual confirmation. Once the system is judged to have "crucible fragments," the edge device immediately triggers a data upload mechanism. The uploaded content includes metadata such as alarm timestamp, workstation number, target confidence level, fragment coordinates, estimated area, and movement speed. Furthermore, the system automatically captures high-resolution images and labeled inference result diagrams within 5 seconds before and after the alarm for subsequent manual review and model iteration optimization.

[0073] This embodiment proposes a deep integration scheme based on standardized industrial communication protocols, aiming to achieve seamless interconnection between intelligent vision inspection systems, MES (Manufacturing Execution System), and SCADA (Supervisory Control and Data Acquisition) systems. By constructing an end-to-cloud collaborative data interaction architecture, the system achieves closed-loop control from bottom-level perception to upper-level decision-making. Specific functional extensions are as follows:

[0074] The system integration adopts the OPC UA open platform communication unified architecture protocol, leveraging its cross-platform and high reliability characteristics to establish a bidirectional communication link between the edge computing end and the factory's central control system. The edge end, acting as an OPC UA server, encapsulates key feature parameters such as detected event timestamps, 2D / 3D spatial coordinates within the molten pool, fragment projection area, and morphological characteristic factors into standardized data nodes. The MES / SCADA system, acting as a client, acquires the aforementioned status information with millisecond-level latency through a subscription mechanism, achieving real-time synchronization between the physical melting process and the digital monitoring system, ensuring that production management personnel have immediate awareness of the furnace's operating conditions. To transform passive monitoring into proactive defense, the system is designed with an event-triggered automated response mechanism. When the visual inspection module determines the presence of "crucible fragments" or "abnormal floating objects" and triggers a high-priority alarm, the alarm signal is immediately pushed to the SCADA system via the OPC UA protocol. Upon receiving the signal, the MES / SCADA system automatically executes corresponding intervention measures according to preset process safety logic, including:

[0075] The system automatically reduces the power of the induction coil or adjusts the electromagnetic stirring parameters to slow down the melt flow rate and prevent fragments from further eroding or getting caught deep into the melt.

[0076] When a significant risk of structural detachment is detected, an emergency stop procedure is triggered to cut off the power supply and activate safety protection mechanisms, thereby minimizing equipment damage and costly material loss.

[0077] The MES system has established a dedicated quality database that stores structured data and associated keyframe images from each inspection. Through in-depth analysis of historical data, a correlation model can be built between crucible wear curves and melting batches, temperature regimes, and raw material compositions. Based on accumulated big data analysis, the system can assist in optimizing crucible bridging processes, baking curves, and lifespan prediction models, driving a paradigm shift in production processes from "experience-driven" to "data-driven."

[0078] To ensure the long-term stability, robustness, and continuous algorithm evolution of the intelligent vision inspection system under the harsh industrial environment of vacuum induction melting, this embodiment constructs a comprehensive maintenance system covering hardware health monitoring, remote model iteration, data closed-loop optimization, and system security management. The specific implementation mechanism is as follows:

[0079] The system's embedded periodic self-diagnostic module scans the industrial camera's link connectivity, edge computing unit's CPU / GPU load, memory usage, and remaining storage space at set time intervals. Once the monitored indicators exceed preset thresholds, the system immediately triggers an alert, prompting maintenance personnel to intervene.

[0080] To adapt to changes in crucible material, furnace condition evolution, or the detection needs of new types of defects, the system also supports OTA (Over-The-Air) download technology for remote model upgrades. Maintenance personnel can train and verify new YOLO weight files on the server side, and then push them to the edge computing unit via an encrypted channel. The system employs a dual-partition update strategy to ensure automatic rollback to a stable version in case of upgrade failure, avoiding production line downtime due to upgrade failures. The edge computing unit supports dynamic hot loading of models, allowing switching to new models without restarting the operating system, minimizing the impact of maintenance on production cycle time.

[0081] The system establishes a closed-loop data mechanism for edge-side inference and cloud-side training. In addition to uploading alarm information, the system is configured with a flexible data sampling strategy, compressing and transmitting low-confidence samples, false positives, and randomly sampled normal samples back to the central training server. This transmitted real-world operating data, after manual review and annotation, is added to the training dataset for incremental training and fine-tuning of the model. This mechanism allows the system to continuously accumulate "experience" for specific production lines as its runtime increases, achieving a spiral increase in algorithm accuracy.

[0082] To ensure system configuration security and operational traceability, a role-based access control (RBAC) system was designed. The system defines different levels, including administrators, engineers, and operators. Administrators have the highest privileges and are responsible for system configuration and user management; engineers can adjust algorithm parameters and update models; operators only have real-time monitoring and alarm confirmation permissions. All operations, such as modifying thresholds, restarting services, and manually controlling critical operations, are recorded in the system log, including the operator, time, and specific instructions, ensuring complete audit trail capabilities in the event of anomalies.

[0083] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered to be within the scope of this invention. The above describes a method for identifying crucible fragments inside a vacuum induction furnace, provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding system for identifying crucible fragments inside a vacuum induction furnace, such as... Figure 6 As shown.

[0084] Figure 6 This is a schematic diagram of a crucible fragment identification system in a vacuum induction furnace provided by the present invention. The system includes: Data acquisition module 601 is used to acquire continuous frames of images of the surface of molten metal inside a vacuum induction furnace; The fragment recognition module 602 is used to construct a crucible fragment recognition model; the trained crucible fragment recognition model is used to recognize crucible fragments in the metal liquid surface image to obtain the initial recognition result of crucible fragments in each frame of the metal liquid surface image; The fragment screening module 603 is used to establish the mapping relationship between the image pixel coordinate system and the actual physical coordinate system to determine the actual projected area of ​​the initial identification result of the crucible fragment in the molten metal surface image; to use optical flow method to track the displacement of the detection box center point of the initial identification result of the crucible fragment in the continuous frame molten metal surface image to determine the movement speed of each initial identification result of the crucible fragment; to extract the crucible fragment contour of the initial identification result of the crucible fragment in each frame molten metal surface image, and to determine the average morphological change rate of the initial identification results of the crucible fragment in all frames based on the crucible fragment contour of each frame; and to screen the initial identification results of the crucible fragment based on the actual projected area, movement speed, and average morphological change rate of the initial identification results of the crucible fragment in the molten metal surface image to obtain the final crucible fragment identification result.

[0085] For specific limitations regarding the crucible fragment identification system in a vacuum induction furnace, please refer to the limitations of the crucible fragment identification method in a vacuum induction furnace described above, which will not be repeated here. Each module in the aforementioned crucible fragment identification system in a vacuum induction furnace can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0086] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A method for identifying crucible fragments inside a vacuum induction furnace is provided.

Claims

1. A method for identifying crucible fragments inside a vacuum induction furnace, characterized in that, include: Acquire continuous frames of images of the molten metal surface inside a vacuum induction furnace; Construct a crucible fragment recognition model; use the trained crucible fragment recognition model to identify crucible fragments in the surface image of molten metal, and obtain the initial recognition result of crucible fragments in each frame of the molten metal surface image; Establish a mapping relationship between the image pixel coordinate system and the actual physical coordinate system to determine the actual projected area of ​​the initial identification result of the crucible fragments in the image of the molten metal surface; Optical flow method is used to track the displacement of the detection box center point of the initial identification result of crucible fragments in consecutive frames of molten metal surface images to determine the movement speed of each initial identification result of crucible fragments; the crucible fragment contour of the initial identification result of crucible fragments in each frame of molten metal surface image is extracted, and the average morphological change rate of the initial identification results of crucible fragments in all frames is determined based on the crucible fragment contour of each frame. The initial identification results of crucible fragments are filtered based on their actual projected area, movement speed, and average morphological change rate in the image of the molten metal surface, to obtain the final identification results of the crucible fragments.

2. The method for identifying crucible fragments inside a vacuum induction furnace as described in claim 1, characterized in that, The method of using optical flow to track the displacement of the center point of the detection frame of the initial identification result of crucible fragments in consecutive frames of molten metal surface images to determine the movement speed of each initial identification result of crucible fragments specifically includes: The Lucas-Kanade optical flow method is used to calculate the pixel displacement between the crucible fragment identification result in the current frame and the crucible fragment identification result in the previous frame; Obtain the frame interval time between two consecutive images of the molten metal surface, as well as the pre-calibrated pixel-to-physical distance conversion coefficient; The movement speed of the crucible fragment recognition result in physical space is calculated based on the pixel displacement, frame interval time, and conversion coefficient.

3. The method for identifying crucible fragments inside a vacuum induction furnace as described in claim 1, characterized in that, The extraction of the crucible fragment outline from the initial identification result of the crucible fragments in the continuous frame images of the molten metal surface, and the determination of the average morphological change rate of the initial identification result of the crucible fragments in the continuous frame images of the molten metal surface based on the crucible fragment outline, specifically includes: The crucible fragment outlines in adjacent frames of molten metal surface images are extracted using a mask, and the area of ​​the symmetrical difference region between the crucible fragment outlines of two adjacent frames is calculated; wherein, the area of ​​the symmetrical difference region is the difference between the area of ​​the union region and the area of ​​the intersection region of the crucible fragment outlines of two frames. The ratio of the area of ​​the symmetrical difference region to the area of ​​the intersection region of the two frame contours is determined as the crucible fragment contour overlap rate between adjacent frames. The average morphological change rate of the crucible fragment identification results within a sliding time window of consecutive frames is statistically analyzed. The average morphological change rate is the average of the overlap rate of crucible fragment contours between adjacent frames.

4. The method for identifying crucible fragments inside a vacuum induction furnace as described in claim 1, characterized in that, The initial identification results of the crucible fragments are filtered based on their actual projected area, movement speed, and average morphological change rate in the image of the molten metal surface, to obtain the final crucible fragment identification results. Specifically, this includes: When the actual projected area of ​​the crucible fragment identification result in the molten metal surface image is less than the preset area threshold, the crucible fragment identification result is judged as a bubble or noise and is removed. For crucible fragment identification results filtered by area, if the real-time movement speed of the center point of the crucible fragment is greater than the preset speed threshold, the crucible fragment identification result will be judged as a fast-moving bubble or splash and rejected. For crucible fragment identification results that pass the speed screening, if the average morphological change rate of crucible fragment identification results in consecutive frames is greater than the preset morphological change rate threshold, the crucible fragment identification results are judged as unstable interference objects and removed to obtain the final crucible fragment identification results.

5. The method for identifying crucible fragments inside a vacuum induction furnace as described in claim 1, characterized in that, The crucible fragment identification model is an improved YOLOv8 model, which includes a cascaded improved backbone network, a neck network, and a detection head. The improved backbone network is obtained by replacing the C2f module in the original backbone network with the C3CrossConv module, and replacing the SPPF module at the end of the backbone network with a void space pyramid pooling module. Each C3CrossConv module extracts crucible fragment feature maps at multiple scales from the surface image of the molten metal; the void space pyramid pooling module extracts crucible fragment enhancement feature maps from the crucible fragment feature maps output by the last C3CrossConv module. The neck network stitches and fuses the enhanced feature map of the crucible fragments and the initial feature maps of the crucible fragments at different scales to obtain a fused feature map of the crucible fragments. The detection head uses the fused feature map of crucible fragments to identify the target and obtain the crucible fragment identification result.

6. The method for identifying crucible fragments inside a vacuum induction furnace as described in claim 5, characterized in that, The crucible fragment identification model also embeds SE modules at the connection between the backbone network and the neck network, as well as at some nodes of the neck network. The SE module compresses the spatial dimension features of the backbone network output into channel descriptors through global average pooling. The correlation between channels is learned by using fully connected layers, channel weight coefficients between 0 and 1 are generated, and the learned weight coefficients are multiplied with the original feature map.

7. The method for identifying crucible fragments inside a vacuum induction furnace as described in claim 5, characterized in that, Each C3CrossConv module includes two parallel convolutional branches with kernels of different sizes; each C3CrossConv module extracts crucible fragment feature maps of different scales from the molten metal surface image, specifically including: For each C3CrossConv module, parallel convolution branches with two different sized convolution kernels in the C3CrossConv module are used to extract two initial crucible fragment feature maps of different scales from the molten metal surface image. The two initial crucible fragment feature maps of different scales are stitched together, and the stitched initial crucible fragment feature maps are then subjected to channel shuffling to obtain crucible fragment feature maps.

8. The method for identifying crucible fragments inside a vacuum induction furnace as described in claim 5, characterized in that, The hollow spatial pyramid pooling module includes: multiple parallel branches of hollow convolution with different dilation rates; the hollow spatial pyramid pooling module extracts crucible fragment enhancement feature maps from the crucible fragment feature maps output by the last C3CrossConv module in the improved backbone network, specifically including: The hollow space pyramid pooling module uses its parallel branches of hollow convolution with different expansion rates to extract the initial enhanced feature map of crucible fragments corresponding to each expansion rate from the feature map of crucible fragments output by the C3CrossConv module. The initial enhanced feature maps of crucible fragments corresponding to each expansion rate are spliced ​​and fused to obtain the enhanced feature map of crucible fragments.

9. The method for identifying crucible fragments inside a vacuum induction furnace as described in claim 1, characterized in that, The loss function used in the training process of the crucible fragment recognition model is determined by the IoU loss term, the center point distance loss term, the width and height loss term, and the similarity metric term based on the normalized Wasserstein distance. The IoU loss term characterizes the ratio of the intersection area to the union area of ​​the predicted box and the actual box of crucible fragment identification. The center point distance loss term characterizes the ratio of the Euclidean distance between the center point of the predicted box and the center point of the actual box to the diagonal distance of the minimum closure region. The width and height loss term characterizes the ratio of the difference between the width of the predicted box and the width of the actual box to the width of the minimum closure region, and the ratio of the difference between the height of the predicted box and the height of the actual box to the height of the minimum closure region. The similarity metric based on normalized Wasserstein distance provides the gradient direction when the predicted box and the actual box do not overlap.

10. A crucible fragment identification system for a vacuum induction furnace, characterized in that, include: The data acquisition module is used to acquire continuous frames of images of the molten metal surface inside the vacuum induction furnace; The fragment recognition module is used to construct a crucible fragment recognition model; the trained crucible fragment recognition model is used to identify crucible fragments in the surface image of molten metal, and the initial recognition result of crucible fragments in each frame of the molten metal surface image is obtained. The fragment screening module is used to establish the mapping relationship between the image pixel coordinate system and the actual physical coordinate system in order to determine the actual projected area of ​​the initial identification result of crucible fragments in the image of the molten metal surface; Optical flow was used to track the displacement of the detection box center point of the initial identification result of crucible fragments in consecutive frames of molten metal surface images to determine the movement speed of each initial identification result of crucible fragments. The outline of the crucible fragments in each frame of molten metal surface image was extracted, and the average morphological change rate of the initial identification results of all frames of crucible fragments was determined based on the outline of the crucible fragments in each frame. The initial identification results of crucible fragments were filtered based on the actual projected area, movement speed, and average morphological change rate of the initial identification results in the molten metal surface image to obtain the final identification result of crucible fragments.

Images