A method for online early warning of necking defects of a spherical tank connecting shell drawn forming

Abstract

The present application belongs to the technical field of online early warning, and particularly relates to a kind of necking defect online early warning method of ball tank connecting shell drawing forming necking, comprising the following steps: S1, obtaining a plurality of continuous image sequences in the drawing forming process of ball tank connecting shell, obtaining the three-dimensional geometric model corresponding to ball tank connecting shell;For each frame of image in the continuous image sequence, a plurality of initial necking defect candidate regions represented by boundary contour and initial confidence are obtained using instance segmentation network;S2, the initial confidence of weight, the relevance with the previous frame defect region and the spatial matching degree with the pre-defined high stress area of three-dimensional geometric model are calculated, and the comprehensive risk weight of each initial necking defect candidate region of current frame is obtained, and all initial necking defect candidate regions are arranged in descending order according to comprehensive risk weight.The present application realizes the pre-judgment of defect development, and can issue an alarm when the defect has not yet obviously deteriorated.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of online early warning, and particularly relates to a necking defect online early warning method for a drawing forming neck of a spherical tank connecting shell. BACKGROUND

[0002] In the drawing forming process of the spherical tank shell, local plastic instability is prone to occur in a specific area due to uneven material flow, stress concentration and other complex factors, and necking defects are formed. The necking defects are characterized by sharp thinning of the sheet metal, and are the main hidden danger leading to early failure and structural damage of the shell. Offline nondestructive testing technologies such as ultrasonic waves and X-rays have low accuracy and cannot meet the needs of online monitoring and early warning of product quality on the automatic production line, often resulting in the formation of defective products and waste of materials and working hours.

[0003] Target detection or instance segmentation networks can directly identify necking defects in images. However, in actual applications, the defect morphology is small and the features are fuzzy in the early stage of necking defect formation, and the neural network may output multiple candidate regions with different confidence levels, overlapping or adjacent positions, increasing the difficulty of accurate positioning. Moreover, general post-processing algorithms, such as non-maximum suppression (NMS), only rely on the confidence score and intersection over union (IoU) of the candidate region for screening, and this single criterion often leads to false suppression or incomplete segmentation of irregular or continuous defects. Most existing methods ignore the time sequence continuity of defect evolution and information associated with the three-dimensional model of the part, such as high stress areas, lack of evaluation of the defect development process, and cannot achieve real early warning, often triggering an alarm when the defect has developed significantly, missing the best opportunity for process adjustment. SUMMARY

[0004] The application provides a necking defect online early warning method for a drawing forming neck of a spherical tank connecting shell to solve the technical problem that the prior art lacks evaluation of the defect development process and cannot achieve real early warning.

[0005] In a first aspect, the application provides a necking defect online early warning method for a drawing forming neck of a spherical tank connecting shell, comprising the following steps:

[0006] S1, acquiring a plurality of continuous image sequences in the drawing forming process of the spherical tank connecting shell and acquiring a three-dimensional geometric model corresponding to the spherical tank connecting shell; for each frame of image in the continuous image sequences, an instance segmentation network is used to acquire a plurality of initial necking defect candidate regions represented by a boundary contour and an initial confidence level;

[0007] S2, calculate the initial confidence of the weight, the relevance to the defect region of the previous frame, and the spatial matching degree with the predefined high stress area of the three-dimensional geometric model, and obtain the comprehensive risk weight of each initial necking defect candidate region of the current frame, arrange all the initial necking defect candidate regions in descending order according to the comprehensive risk weight, select the one with the highest weight as the reference region, and initialize the reference region contour as the defect contour of the current frame; traverse the remaining initial necking defect candidate regions of the current frame, if the morphological-spatial composite intersection ratio with the defect contour of the current frame is greater than a preset threshold, fuse the initial necking defect candidate region through weighted vertex affine transformation to update the defect contour, otherwise suppress it; the morphological-spatial composite intersection ratio is determined by the intersection ratio, the centroid distance and the shape similarity of the initial necking defect candidate region and the reference region contour;

[0008] S3, generate a warning index according to the area growth rate and shape irregularity of the defect contour, combined with the comprehensive risk weight of the reference region; when the warning index presents continuous growth in continuous multiple frames and the cumulative increment exceeds the risk threshold, output a warning signal.

[0009] Further, the initial confidence of the weight, the relevance to the defect region of the previous frame, and the spatial matching degree with the predefined high stress area of the three-dimensional geometric model are calculated to obtain the comprehensive risk weight of each initial necking defect candidate region of the current frame, including:

[0010] According to the formula , , , is a weight coefficient;

[0011] wherein the comprehensive risk weight of the initial necking defect candidate region is , is the initial confidence of the initial necking defect candidate region, is the spatial matching intersection ratio of the projection area of the initial necking defect candidate region and the predefined high stress area of the three-dimensional geometric model, is the spatio-temporal association intersection ratio of the initial necking defect candidate region and the defect region contour of the previous frame, and when there is no defect region of the previous frame, is 0.

[0012] Further, the morphological-spatial composite intersection ratio is determined by the intersection ratio, the centroid distance and the shape similarity of the initial necking defect candidate region and the reference region contour, specifically:

[0013] Calculate the pixel-level intersection ratio IoUp, the normalized centroid distance D c and the shape similarity S h of the initial necking defect candidate region and the reference region contour through the formula E1×IoUp+E2×(1-Dc ) + E3 x S h The weighted combination is performed to obtain a morphology-space composite intersection-over-union ratio, wherein E1, E2, and E3 are preset weight coefficients.

[0014] wherein the normalized centroid distance D c is a ratio of the Euclidean distance between the centroid of the initial necking defect candidate region and the reference region to the diagonal length of the image, and the shape similarity S h is calculated by the Hu invariant moments.

[0015] Further, the initial necking defect candidate region is fused by a weighted vertex affine transformation to update the defect contour, including:

[0016] The current defect contour and the profile of the initial necking defect candidate region to be fused are sampled into equal-length vertex sequences and ;

[0017] For each vertex in , the corresponding vertex with the closest Euclidean distance in is found;

[0018] According to the comprehensive risk weight of the initial necking defect candidate region to be fused, the fusion weight is calculated by the formula ;

[0019] The fusion is completed by updating the positions of all vertices, is the th vertex in the new defect contour vertex sequence after fusion.

[0020] Further, according to the area growth rate and the shape irregularity of the defect contour, in combination with the comprehensive risk weight of the reference region, a warning index is generated, including:

[0021] The pixel area and the perimeter P k of the defect contour in the current frame k are calculated;

[0022] The area growth rate is calculated, wherein when k is the frame in which the defect first appears, it is defined that = 1, otherwise , is the area of the previous frame;

[0023] The shape irregularity is calculated;

[0024] ; ​

[0025] obtaining a comprehensive risk weight of the reference area ;

[0026] The early warning index is calculated by the formula .

[0027] Further, the calculation process of the spatial matching degree is as follows:

[0028] Performing finite element analysis on the three-dimensional geometric model, the physical weak areas with the top 5% equivalent stress values in the drawing forming process are identified as three-dimensional high stress areas;

[0029] Using the camera calibration parameters, the three-dimensional high stress areas are projected onto the current image plane to form two-dimensional high stress projection areas;

[0030] The intersection-union ratio of the initial necking defect candidate area and the two-dimensional high stress projection area is calculated as the spatial matching degree.

[0031] Further, the instance segmentation network is a YOLOv8-seg model trained on a dataset containing necking defect annotation samples;

[0032] The current frame image is input into the YOLOv8-seg model, and the segmentation area with a confidence greater than 0.4 in the output result is taken as the initial necking defect candidate area.

[0033] Further, Abaqus software is used to perform finite element simulation of the drawing forming process of the three-dimensional CAD model of the spherical tank connecting shell.

[0034] Further, a pre-trained Mask R-CNN model is used to perform forward inference on each frame of image to obtain the initial necking defect candidate area.

[0035] Further, a high-speed industrial camera is installed above the drawing forming station of the spherical tank connecting shell, and a uniform light source is provided to eliminate reflections and shadows. Through camera development kit programming, the acquisition frame rate is set to continuously capture images of the drawing forming area of the spherical tank connecting shell, forming a timestamp-aligned image sequence.

[0036] ​The beneficial effects are as follows: This invention improves accuracy in the early stages of necking defects by constructing a comprehensive risk weight that integrates initial confidence, spatiotemporal correlation of the previous frame, and prior information from the 3D model to rank candidate regions. Furthermore, by combining the morphological-spatial composite cross-union ratio (CUNR), determined by the cross-union ratio, centroid distance, and shape similarity, it reduces the problem of false suppression or incomplete segmentation of irregular or contiguous defects by non-maximum suppression. Based on the cumulative change trends of contour area growth rate and shape irregularity over continuous time series, it achieves prediction of defect development and can issue an alarm before the defect has significantly worsened. Attached Figure Description

[0037] Figure 1 A flowchart of an online early warning method for necking defects in the drawing process of spherical tank connecting shells. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] An embodiment of the online early warning method for necking defects in the drawing process of spherical tank connecting shells provided by the present invention:

[0040] like Figure 1 As shown, the online early warning method for necking defects in the drawing process of spherical tank connecting shells includes the following steps:

[0041] S1. Obtain a multi-frame continuous image sequence during the drawing process of the spherical tank connecting shell, and obtain the three-dimensional geometric model corresponding to the spherical tank connecting shell; for each frame of the continuous image sequence, use an instance segmentation network to obtain multiple initial necking defect candidate regions represented by boundary contours and initial confidence.

[0042] A high-speed industrial camera is fixedly installed above or aside the spherical tank connecting shell drawing forming station, and is equipped with a uniform light source to eliminate reflections and shadows. Through camera development kit programming, a fixed acquisition frame rate, for example, 30 frames per second, is set, and the image of the spherical tank connecting shell drawing forming area is continuously captured to form a timestamp-aligned image sequence and stored in the memory buffer. At the same time, the standard three-dimensional CAD model of the spherical tank connecting shell of this type is loaded from the product design database, for example, a STEP or IGES format file, and the file surface mesh data is extracted for subsequent use. A pre-trained instance segmentation network, for example, a Mask R-CNN model, is used to forward infer each frame of input image. The instance segmentation network is trained on a large image dataset containing necking defect samples, and can simultaneously complete defect classification, positioning and mask generation. The network outputs a series of potential defect targets, each target containing a bounding box, a classification confidence score and a pixel-level mask. Each mask is converted into a closed boundary contour point set, and the confidence score is the initial confidence, thereby obtaining multiple initial necking defect candidate regions.

[0043] In an optional embodiment, the instance segmentation network is a YOLOv8-seg model trained on a dataset containing necking defect annotation samples;

[0044] The current frame image is input to the YOLOv8-seg model, and the segmentation region with a confidence greater than 0.4 in the output result is taken as the initial necking defect candidate region.

[0045] Preferably, the YOLOv8-seg model is used to complete the initial defect recognition task. YOLOv8-seg can quickly process each frame of image in the video stream and simultaneously obtain the outline of potential defects. In order to enable the YOLOv8-seg model to specifically recognize necking defects in the sheet metal drawing forming process, the YOLOv8-seg model needs to be trained or fine-tuned on a specially constructed dataset. The dataset contains thousands of images collected in a real environment, and the outline of the necking defect in each image is manually annotated by technicians.

[0046] In actual operation, the cameras on the production line capture images at a rate of 30 frames per second, and each frame of image is sent to the trained YOLOv8-seg model, and all the necking defect regions it considers are output. Each output contains two parts of information: a segmentation mask that identifies all the pixels of the defect, and a confidence score between 0 and 1. For example, there may be three potential targets in a frame of image, with confidence scores of 0.85, 0.52 and 0.28 respectively. A preset confidence threshold such as 0.4 is used to filter out less certain detection results. In this example, only the two necking defect regions with confidence scores of 0.85 and 0.52 are retained as initial necking defect candidate regions, which enter the subsequent risk assessment and contour fusion process. The necking defect region with a confidence score of 0.28 is considered as noise or false alarm and is ignored.

[0047] S2, calculate the initial confidence of the weight, the relevance to the defect region of the previous frame, and the spatial matching degree with the pre-defined high stress area of the three-dimensional geometric model, and obtain the comprehensive risk weight of each initial necking defect candidate region of the current frame, arrange all the initial necking defect candidate regions in descending order according to the comprehensive risk weight, select the one with the highest weight as the reference region, and initialize the contour of the reference region as the defect contour of the current frame; traverse the remaining initial necking defect candidate regions of the current frame, if the morphological-spatial compound intersection ratio of the initial necking defect candidate region and the defect contour of the current frame is greater than a preset threshold, update the defect contour by weighted vertex affine transformation to fuse the initial necking defect candidate region, otherwise suppress it; the morphological-spatial compound intersection ratio is determined by the intersection ratio, centroid distance and shape similarity of the initial necking defect candidate region and the reference region contour;

[0048] For each initial necking defect candidate region, a comprehensive risk weight of the initial necking defect candidate region is calculated. Firstly, a correlation is calculated, i.e. a flow displacement vector between the initial necking defect candidate region and the last defect region determined in the previous frame is calculated, and the smaller the displacement is, the higher the correlation is. Secondly, a spatial matching degree is calculated. A high stress region pre-marked on a three-dimensional model through finite element analysis is projected to a current image plane, and an overlapping area ratio of the initial necking defect candidate region and the projected high stress region is calculated, and the larger the ratio is, the higher the matching degree is. The initial confidence, the spatio-temporal correlation score and the spatial matching degree score are weighted and summed to obtain the comprehensive risk weight. All the initial necking defect candidate regions are sorted from high to low according to the weight, and the initial necking defect candidate region with the highest weight is selected as a reference, and a contour of the initial necking defect candidate region is an initial defect contour. The remaining initial necking defect candidate regions are traversed from the initial necking defect candidate region with the second highest weight. For each initial necking defect candidate region to be processed, a morphological-spatial composite intersection-over-union ratio of the initial necking defect candidate region and the current defect contour is calculated. The intersection-over-union ratio is a weighted combination of three aspects: a pixel area intersection-over-union ratio of the two contour regions; a normalized inverse of a Euclidean distance between the two centroids; and a shape similarity of the two contours, for example, obtained by calculating a difference between the two Hu moments. If the morphological-spatial composite intersection-over-union ratio is greater than a preset threshold, for example, 0.6, it is considered that the two should be fused. In the fusion, a weighted vertex affine transformation is adopted, i.e. new vertex coordinates are generated by weighted averaging of coordinates of corresponding points on the two contours according to the comprehensive risk weights of the two regions, and the defect contour is updated. If the morphological-spatial composite intersection-over-union ratio is less than the preset threshold, the initial necking defect candidate region is directly discarded, and suppression is achieved.

[0049] In an optional embodiment, a comprehensive risk weight of each initial necking defect candidate region in the current frame is obtained according to a weighted initial confidence, a correlation with a defect region in a previous frame and a spatial matching degree with a high stress region pre-defined on a three-dimensional geometric model, and the comprehensive risk weight comprises:

[0050] The weighted sum is obtained according to a formula , wherein , , is a weight coefficient;

[0051] wherein is an initial confidence of the initial necking defect candidate region, is a spatial matching intersection-over-union ratio of the initial necking defect candidate region and a projected region of the high stress region pre-defined on the three-dimensional geometric model, is a spatio-temporal correlation intersection-over-union ratio of the initial necking defect candidate region and a contour of the defect region in the previous frame, and when there is no defect region in the previous frame, is 0.

[0052] The initial confidence C represents the judgment of the instance segmentation network that the defect region is a necking defect, and is preferably set to the highest weight of 0.5. The spatial matching IoU The continuity of the current initial necking defect candidate region and the confirmed defect of the last frame is measured, and the weight is preferably 0.3. The spatial matching IoU According to whether the defect appears in the high stress area predicted by the finite element analysis, the weight is preferably 0.2. An initial necking defect candidate region is detected, and the initial confidence C given by the segmentation model is 0.8. By comparing with the defect region of the previous frame, the spatial and temporal correlation IoU of the initial necking defect candidate region is calculated 0.9, indicating that the defect is continuous in time. At the same time, the spatial matching IoU of the initial necking defect candidate region and the pre-calculated two-dimensional high stress projection area 0.6, indicating that it also appears in a location prone to damage. According to the formula, the comprehensive risk weight W of the initial necking defect candidate region is 0.79.

[0053] In an optional embodiment, the calculation process of the spatial matching degree is as follows:

[0054] Perform finite element analysis on the three-dimensional geometric model to identify the top 5% of physical weak areas of equivalent stress value in the drawing forming process as three-dimensional high stress areas;

[0055] Project the three-dimensional high stress areas to the current image plane using the camera calibration parameters to form a two-dimensional high stress projection area;

[0056] Calculate the intersection over union of the initial necking defect candidate region and the two-dimensional high stress projection area as the spatial matching degree.

[0057] In the product design stage, finite element simulation of the drawing process is performed on the three-dimensional CAD model of the part using software such as Abaqus or ANSYS. The simulation outputs the stress distribution cloud diagram of the entire part during the forming process. The 5% of the physical weak areas with the highest equivalent stress value are identified, and these physical weak areas are defined as three-dimensional high stress areas, which are the weakest links of the material. On the production line, the camera internal and external parameters are obtained through the camera calibration process. Using these parameters, a mapping relationship from the three-dimensional world coordinates to the image coordinates is established. The three-dimensional high stress areas determined previously are projected onto the real-time image captured by the camera through the mapping relationship, forming one or more two-dimensional high stress projection areas. When the vision system detects an initial necking defect candidate area, the pixel overlap, i.e. the intersection over union, of the initial necking defect candidate area and the two-dimensional high stress projection areas is calculated. For example, if an initial necking defect candidate area has 70% of its area falling within the two-dimensional high stress projection area, then the spatial matching degree of the initial necking defect candidate area is 0.7. As an input for calculating the comprehensive risk weight of the initial necking defect candidate area, a high matching degree means that the defect appears in a dangerous position as expected, greatly increasing the likelihood of a real defect.

[0058] In an optional embodiment, the morphology-space composite intersection over union is determined by the intersection over union, the centroid distance and the shape similarity of the initial necking defect candidate area and the reference area contour, specifically:

[0059] The pixel-level intersection over union IoUp of the initial necking defect candidate area and the reference area contour, the normalized centroid distance D c and the shape similarity S h are calculated, and the morphology-space composite intersection over union is obtained by weighted combination through formula E1 x IoUp + E2 x (1-D c ) + E3 x S h , wherein E1, E2 and E3 are preset weight coefficients.

[0060] The normalized centroid distance D c is the ratio of the Euclidean distance between the centroids of the initial necking defect candidate area and the reference area to the diagonal length of the image, and the shape similarity S h is calculated by Hu invariant moments.

[0061] The morphology-space composite intersection over union is used to track dynamically changing defects, and combines the overlap degree of the initial necking defect candidate area and the reference area, as well as their positional proximity and shape similarity. The pixel-level intersection over union IoUp serves as the basis, with the highest weight of 0.6, representing the spatial overlap. The normalized centroid distance D c then supplements the positional information, i.e. even if the initial necking defect candidate area and the reference area do not overlap but are very close, 1 minus D cThe value of shape similarity S will also be high, positively impacting the overall score. h By using Hu invariant moment calculation, the effects of rotation and scaling can be ignored, and it can be determined whether the initial necking defect candidate region and the reference region are similar in shape.

[0062] For example, suppose we want to compare an initial necking defect candidate region in the current frame with a defect region in the previous frame. Their pixel-level intersection-union (IoUp) is 0.7. The image diagonal length is 2200 pixels, and the distance between the centroids of the two regions is 110 pixels. Then the normalized centroid distance D c The value is 0.05. The shape similarity S obtained by calculating using Hu invariant moments is... h The value is 0.85. Substituting into the formula, the morphological-spatial composite intersection ratio is 0.78.

[0063] To smoothly update the defect contour and avoid abrupt changes due to the introduction of new detection results, in an optional embodiment, the defect contour is updated by fusing the initial necking defect candidate region through a weighted vertex affine transformation, including:

[0064] The current defect contour and the contour of the initial necking defect candidate region to be fused are sampled as equal-length vertex sequences. and ;

[0065] right Each vertex in ,exist Find the corresponding vertex with the closest Euclidean distance in the middle. ;

[0066] Based on the comprehensive risk weight of the initial necking defect candidate regions to be fused Through formula Calculate fusion weights ;

[0067] pass Update all vertex positions to complete the merging.

[0068] Both existing defect contours and newly discovered initial necking defect candidate region contours are represented as sequences containing, for example, 200 vertices, ensuring a one-to-one comparison. Fusion weights. It is an S-shaped function, the function value of which is determined by the comprehensive risk weight of the newly discovered initial necking defect candidate regions. The decision allows new information with high-risk weights to have a greater impact on the final contour, while information with low-risk weights is adopted more moderately. For example, the overall risk weight of an initial necking defect candidate region to be fused. The fusion weight λ is 0.88, indicating that the updated contour will be mainly determined by the shape of the newly found initial necking defect candidate region. For any vertex on the final contour , find the vertex on the new candidate contour closest to it , then the new position of the vertex =0.12× +0.88× . Repeat for all 200 vertices so that the whole contour is smoothly smoothed towards the new detection result with high confidence, thus reflecting the latest shape of the defect. Conversely, if is only 0.55, the value of λ will decrease to about 0.56, and the updating process will be more conservative.

[0069] S3, generating an early warning index according to the area growth rate and shape irregularity of the defect contour, in combination with the comprehensive risk weight of the reference region; when the early warning index presents a continuous growth and the cumulative increment exceeds the risk threshold in a plurality of consecutive frames, outputting an early warning signal;

[0070] For the defect contour obtained in the current frame t, the contour area and perimeter are calculated. The area growth rate is obtained by comparing the area of the current frame with the area of the previous frame. The shape irregularity is preferably obtained by calculating the ratio of the square of the perimeter to the area, the larger the ratio, the more irregular the shape. The normalized area growth rate, shape irregularity and the comprehensive risk weight of the selected reference region are weighted and summed to generate the dynamic early warning index value of the current frame. A fixed length time window is adopted, for example, the early warning index values of the last 10 frames are included. After the calculation of each new frame is completed, it is checked whether the index value sequence in the time window is monotonically increasing. If the number of consecutive increasing frames reaches a set value, for example, 5 frames, and the cumulative increment of the index values in the 5 frames, that is, the index value of the last frame minus the index value of the first frame, exceeds the preset risk threshold, an early warning signal is immediately output to the central control system or the operator interface.

[0071] In an optional embodiment, the early warning index is generated according to the area growth rate and shape irregularity of the defect contour, in combination with the comprehensive risk weight of the reference region, comprising:

[0072] calculating the pixel area and the perimeter P k of the defect contour in the current frame k;

[0073] calculating the area growth rate , wherein when k is the frame where the defect first appears, define =1, otherwise , is the area of the previous frame;

[0074] calculating the shape irregularity ;

[0075] Obtain the comprehensive risk weight of the benchmark area ;

[0076] Through formula Calculate early warning indicators .

[0077] The early warning indicator integrates information from three dimensions: the dynamic changes of defects, their geometric shape, and inherent risks. Area growth rate. This reflects the rate of defect propagation; rapidly growing defects are generally more dangerous. Shape irregularity. For a circle, the value is 1. The more irregular the shape, the larger the value, indicating a worsening of the defect morphology, such as the evolution from a stable necking defect region to an unstable crack. The comprehensive risk weight of the baseline region. This represents the severity of the defect itself, taking into account information such as the defect location and the confidence level of its occurrence. For example, in the k-th frame, the detected defect area... It is 300 pixels, while the area of ​​the previous frame is... If the pixel count is 200, then the area growth rate is... The perimeter is 1.5. The defect has a shape irregularity of 90 pixels. The value is approximately 2.15, indicating that the shape is highly irregular compared to a circle. This represents the overall risk weight for the defect baseline region. It is 0.8. Warning indicator. The value is approximately 2.58. This indicates that the defect is not only high-risk but also growing rapidly and is irregularly shaped. Based on a preset threshold, such as 2.0, a high-level alarm can be triggered, prompting the operator to intervene immediately.

[0078] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for online early warning of necking defects of a draw formed neck of a spherical tank connecting shell, characterized in that, The method comprises the following steps: S1, acquiring a plurality of continuous image sequences in a drawing forming process of a spherical tank connecting shell, and acquiring a three-dimensional geometric model corresponding to the spherical tank connecting shell; for each frame of image in the continuous image sequences, an instance segmentation network is used to acquire a plurality of initial necking defect candidate regions represented by a boundary contour and an initial confidence; S2, based on the initial confidence of the initial necking defect candidate region, the relevance of the initial necking defect candidate region and the defect region of the previous frame, and the spatial matching degree of the initial necking defect candidate region and the predefined high stress area of the three-dimensional geometric model, the comprehensive risk weight of each initial necking defect candidate region of the current frame is calculated, including: according to the formula Weighted sum is carried out to obtain wherein, 、 、 is the weight coefficient; wherein the comprehensive risk weight of the initial necking defect candidate region is , is the initial confidence of the initial necking defect candidate region, is the spatial matching intersection-union ratio of the initial necking defect candidate region and the projection area of the predefined high stress area of the three-dimensional geometric model, is the space-time correlation intersection-union ratio of the initial necking defect candidate region and the defect region profile of the previous frame, and when there is no defect region of the previous frame, is 0, arrange all initial necking defect candidate regions in descending order according to the comprehensive risk weight, select the one with the highest weight as the reference region, and initialize the reference region profile as the defect profile of the current frame; traverse the remaining initial necking defect candidate regions of the current frame, if the morphological-spatial composite intersection-union ratio of the defect profile of the current frame is greater than a preset threshold, the initial necking defect candidate region is fused to update the defect profile through weighted vertex affine transformation, otherwise it is suppressed; the morphological-spatial composite intersection-union ratio is determined by the intersection-union ratio, the centroid distance and the shape similarity of the initial necking defect candidate region and the reference region profile. S3, generating a warning index according to an area growth rate and a shape irregularity of the defect contour, in combination with a comprehensive risk weight of the reference region; when the warning index presents a continuous growth in a plurality of continuous frames and a cumulative increment exceeds a risk threshold, outputting a warning signal.

2. The method of claim 1, wherein the method is characterized by: The shape-space compound intersection ratio is determined by the intersection ratio, the centroid distance and the shape similarity of the initial necking defect candidate region and the reference region contour, and specifically is: calculating a pixel-level intersection-over-union IoUp of the initial necking defect candidate region and the reference region profile, a normalized centroid distance D c and a shape similarity S h , and obtaining a morphology-space composite intersection-over-union by weighted combination through a formula E1xIoUp+E2x(1-D c )+E3xS h , wherein E1, E2 and E3 are preset weight coefficients. where the normalized centroid distance D c is the ratio of the Euclidean distance between the initial necking defect candidate region and the reference region centroid to the diagonal length of the image, and the shape similarity S h is calculated by the Hu invariant moments.

3. The method of claim 1, wherein the method is characterized by: The defect contour is updated by a weighted vertex affine transformation fusion of the initial necking defect candidate region, comprising: sampling the contour of the current defect and the contour of the initial necking defect candidate region to be fused into a sequence of equal length vertices and ; For each vertex in find the Euclidean distance closest corresponding vertex in , in find the Euclidean distance closest corresponding vertex in ; According to the comprehensive risk weight of the initial necking defect candidate region to be fused , the fusion weight is calculated by formula ;​ By updating all vertex positions to complete the fusion, the first vertex in the updated sequence of vertices of the new defect profile after fusion. the first vertex in the updated sequence of vertices of the new defect profile after fusion.

4. The method of claim 1, wherein the method is characterized by: The warning index is generated according to the area growth rate and the shape irregularity of the defect contour, in combination with the comprehensive risk weight of the reference region, comprising: calculating the pixel area of the current frame k defect profile and the perimeter P k ; Compute the area growth rate where k is the frame in which the defect first appears, is defined as = 1, otherwise , is the area of the previous frame. Computing shape irregularity ; ； Obtaining a comprehensive risk weight for a reference area ; The early warning index is calculated by the formula .​ 5. The method of claim 1, wherein the method is characterized by: The calculation process of the spatial matching degree is: The three-dimensional geometric model is subjected to finite element analysis, and the physical weak regions ranked in the top 5% of equivalent stress values in the drawing forming process are identified as three-dimensional high stress zones; The three-dimensional high stress zones are projected to the current image plane by using the camera calibration parameters to form two-dimensional high stress projection regions; The intersection ratio of the initial necking defect candidate region and the two-dimensional high stress projection region is calculated as the spatial matching degree.

6. The method of claim 1, wherein the method is characterized by: The instance segmentation network is a YOLOv8-seg model trained on a data set containing necking defect annotation samples; The current frame image is input into the YOLOv8-seg model, and the segmentation region with a confidence greater than 0.4 in the output result is taken as the initial necking defect candidate region.

7. The method of claim 1, wherein the method further comprises: Abaqus software is used to perform finite element simulation of the drawing forming process of the three-dimensional CAD model of the spherical tank connecting shell.

8. The method of claim 7, wherein the method further comprises: A pre-trained Mask R-CNN model is used to perform forward inference on each frame of image to acquire the initial necking defect candidate region.

9. The method of claim 1-8, wherein the method is characterized by, A high-speed industrial camera is installed above the drawing forming station of the spherical tank connecting shell, and is equipped with a uniform light source to eliminate reflection and shadow; a camera development kit is programmed to set the acquisition frame rate, and the images of the drawing forming area of the spherical tank connecting shell are continuously captured to form a timestamp-aligned image sequence.

Images