Motor insert defect detection method based on multi-sensor fusion

Abstract

The application discloses a motor insert sheet defect detection method based on multi-sensor fusion, relates to the technical field of motor insert sheet defect detection, and comprises the following steps: in the presence of boundary penetration, calculating edge matching mismatch degree, edge jump rate, edge inversion proportion and edge closure deviation to form a geometric conflict feature set of an insert sheet edge, which is used for representing image geometric expression differences caused by boundary penetration; inputting the geometric conflict feature set into a fusion confidence evaluation function to obtain a fusion confidence evaluation factor, which is used for identifying whether the insert sheet edge has a fusion distortion risk. The application solves the structural expression conflict problem caused by boundary penetration in the multi-view imaging fusion process of the motor insert sheet, constructs a fusion confidence evaluation mechanism to identify the fusion distortion risk, dynamically limits the image participation fusion condition, and thus improves the recognition accuracy of the edge structure and the stability of the fusion result in the insert sheet defect detection.

Description

Technical Field

[0001] This invention relates to the field of motor splice defect detection technology, and specifically to a method for detecting motor splice defects based on multi-sensor fusion. Background Technology

[0002] Multi-sensor fusion-based motor lamination defect detection refers to a comprehensive technical approach that utilizes multiple image sensors of different types to work collaboratively, combining traditional machine vision algorithms with deep learning models to perform fully automated defect detection on the lamination areas of a motor stator. This technology employs a multi-station image acquisition system to acquire high-resolution images of the lamination area during the lamination detection stage (stations 6 and 7). Image preprocessing operations (including scaling to 512×612 pixels and rotating by 270 degrees) ensure the image format meets the requirements of subsequent algorithms. Subsequently, the system uses Halcon vision tools to perform traditional image processing procedures, such as grayscale conversion, threshold segmentation, morphological closing operations, connected component analysis, shape filtering, and circle fitting, to extract lamination edge, integrity, and positional features. Simultaneously, the system uploads the images to the LinghuAI deep learning backend service via HTTP API, calling the YOLOv3 + MobileNetV3 object detection model and the SOLOv2 instance segmentation model built on the PaddleDetection framework to extract lamination defect information from the images, including various types such as excessive lamination height, incomplete cutting, skewed, and missing laminations. After obtaining the detection results from both traditional and deep learning algorithms, the system employs a result fusion mechanism for comprehensive judgment. Only when both algorithms deem the workstation acceptable will the detection result be output as OK; otherwise, it is judged as NG. Furthermore, the system uses a multi-workstation detection result aggregation logic to integrate information from all detection workstations and generate a quality assessment report for the entire machine. The detection results are ultimately fed back to the production line control system via the PLC communication module, enabling real-time defect alarms and synchronized updates with production cycle time. This multi-sensor fusion solution not only improves the accuracy and robustness of insert defect detection but also significantly reduces sensitivity to ambient lighting and positional deviations, thereby enhancing overall production efficiency and the level of automation in inspection.

[0003] The existing technology has the following shortcomings:

[0004] In the process of detecting defects in motor splice plates based on multi-sensor fusion, when the edge of the splice plate is located in the imaging intersection area of ​​multiple sensing perspectives, due to the differences in imaging direction, projection relationship, and visible structure under different perspectives, the same splice plate edge will exhibit intersecting boundary morphology in images from different perspectives. That is, it appears as an outer edge in one image and an inner edge in another image, thus forming a structural boundary conflict after multi-view information fusion. Since the above boundary intersecting phenomenon stems from the objective existence of inconsistent geometric representation of splice plate edges under multi-view imaging, existing multi-sensor fusion-based motor splice plate defect detection technology cannot identify whether there is a risk of fusion distortion in the splice plate edge based on the structural conflict characteristics of the splice plate edge under boundary intersecting conditions in multi-view images. Consequently, it still directly uses conflicting edge information for fusion judgment, resulting in distortion or cross-over of the splice plate edge structure in the fusion result. This can easily cause the detection system to fail to accurately distinguish actual defects such as notches, breaks, or warping at the edges of the splice, or to misjudge structural conflicts generated by fusion as defects, resulting in false detections or missed detections. Ultimately, this reduces the reliability and stability of motor splice defect detection results based on multi-sensor fusion and has an adverse impact on the quality assessment of motor splices.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting defects in motor inserts based on multi-sensor fusion, so as to solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting defects in motor inserts based on multi-sensor fusion, specifically including the following steps:

[0008] S1. Extract the direction information of the motor insert edge from multiple imaging perspectives, establish an insert edge direction mapping map, and construct an insert edge direction offset map by calculating the direction consistency index, which is used to determine whether the insert edge has boundary interpenetration in multi-view images.

[0009] S2. In the case of boundary interpenetration, calculate the edge matching mismatch degree, edge jump rate, edge inversion ratio and edge closure deviation to form a set of geometric conflict features of the interpenetrated edges, which are used to characterize the differences in image geometric expression caused by boundary interpenetration.

[0010] S3. Input the geometric conflict feature set into the fusion confidence evaluation function to obtain the fusion confidence evaluation factor, which is used to identify whether there is a risk of fusion distortion at the edge of the insert;

[0011] S4. Divide the fusion confidence level of the insert edge according to the fusion confidence evaluation factor, and screen the image acquisition path based on the fusion confidence level to limit the conditions for edge images to participate in fusion.

[0012] S5. Track the temporal change trend of the fusion confidence assessment factor, and adjust the directional consistency index and confidence factor division threshold based on the spectrum volatility to achieve dynamic control of the fusion confidence level of the insert edge.

[0013] Preferably, S1 is as follows:

[0014] Edge extraction processing is performed on the edge of the motor insert from multiple imaging perspectives. Gradient calculation is used to obtain the directional information of each pixel on the edge of the motor insert. The directional difference between the directional vector of each pixel and the reference directional direction of the insert is calculated to obtain a set of directional information.

[0015] Based on the position index of each edge pixel in the unified interpolation space coordinates, the direction information extracted from multiple imaging perspectives is aligned according to the spatial position, and the direction vectors of each perspective are summarized at each edge position to establish an interpolation edge direction mapping map.

[0016] For each edge position in the insertion edge direction mapping map, the maximum angle between different viewpoint direction vectors is calculated as the direction consistency index, and the direction consistency index is mapped across the entire insertion edge domain to construct the insertion edge direction offset map; when the direction consistency index of consecutive edge positions exceeds the first set threshold, it is determined that the insertion edge has boundary interleaving in the multi-view image.

[0017] Preferably, S2 specifically includes the following steps:

[0018] S201. In the case of boundary overlap, the edge contour points extracted from the edge of the insert in multiple imaging views are spatially aligned, the Euclidean distance between contour points of different views at the same spatial position is calculated, and the ratio of the average distance to the maximum distance at each position is used as the edge matching mismatch degree.

[0019] S202. Based on the edge region calibrated by the edge matching mismatch degree, collect the direction vectors corresponding to each spatial position, calculate the angle difference sequence between adjacent direction vectors in sequence according to the continuous spatial index, count the proportion of points in the angle difference sequence that are greater than the preset angle threshold as the edge jump rate, and count the proportion of points where the angle difference between the direction vectors is close to the opposite relationship as the edge reversal ratio. Calculate the spatial closure error between the starting point and the ending point of the edge contour, and use the spatial closure error as the edge closure deviation.

[0020] S203. The edge matching mismatch degree, edge jump rate, edge reversal ratio and edge closure deviation are combined to form a set of geometric conflict features of the interpolation edge, which is used to characterize the differences in image geometric expression caused by boundary interpolation.

[0021] Preferably, S201 specifically refers to:

[0022] In the case of boundary overlap, based on the pixel coordinates of the interpolation edge in the multi-view image, the edge contour points extracted from multiple imaging views are projected to a unified three-dimensional spatial coordinate system through extrinsic parameter transformation, and the edge contour points from different imaging views are paired according to the spatial position index.

[0023] For each pairing position, calculate the Euclidean distance between any two sets of edge contour points at that position, and collect the Euclidean distance data of all pairing positions to obtain the average Euclidean distance and the maximum Euclidean distance for each position.

[0024] For each pairing position, the ratio between the average Euclidean distance and the maximum Euclidean distance at that position is used as the edge matching mismatch degree, which serves as the basis for evaluating the spatial consistency of the insert edges in the case of boundary interleaving, and is used for subsequent conflict feature extraction.

[0025] Preferably, S202 specifically refers to:

[0026] Based on the numerical distribution of edge matching mismatch degree on the edge of the insert, the spatial location where the edge matching mismatch degree is greater than the second set threshold is selected as the edge region. The direction vector corresponding to each spatial location is collected in the edge region, and the direction vector is continuously spatially indexed and sorted according to the spatial arrangement order of the insert edge to form a direction vector sequence.

[0027] For adjacent direction vectors in the direction vector sequence, calculate the angle difference between the direction vectors one by one to form an angle difference sequence. In the angle difference sequence, count the proportion of spatial positions where the angle difference is greater than a preset angle threshold to obtain the edge jump rate. Count the proportion of spatial positions where the angle difference is in the near-reverse relationship interval to obtain the edge reversal ratio.

[0028] After calculating the edge jump rate and edge reversal ratio, the minimum and maximum spatial index positions in the edge contour are selected as the starting and ending points of the edge contour. The spatial distance between the two points is calculated as the spatial closure error, and the spatial closure error is used as the edge closure deviation.

[0029] Preferably, S3 is as follows:

[0030] The edge matching mismatch degree, edge jump rate, edge reversal ratio and edge closure deviation corresponding to each spatial location in the geometric conflict feature set of the insert edge are feature-encoded, and a uniform input vector is constructed. This input vector is then input into the fusion confidence evaluation function, which is a nonlinear feature mapping function constructed based on multi-feature correlation analysis.

[0031] In the fusion confidence evaluation function, the feature importance of each feature field in the input vector is weighted and calculated, and the fusion confidence evaluation factor is obtained through a nonlinear fitting model. The fusion confidence evaluation factor is output in numerical form and is used to measure the stability of the interpolation edge fusion.

[0032] A first confidence risk threshold is set based on the numerical range of the fusion confidence assessment factor, and the fusion confidence assessment factor of each insert edge spatial location is compared and judged. When the fusion confidence assessment factor is lower than the first confidence risk threshold, the edge of the insert is identified as having a risk of fusion distortion.

[0033] Preferably, S4 is as follows:

[0034] Based on the numerical distribution of the fusion confidence assessment factor, a level division interval is constructed. The complete numerical range of the fusion confidence assessment factor is divided into multiple non-overlapping segments, and an independent fusion confidence level label is set for each numerical segment to divide the fusion confidence level of each spatial location of the insert edge.

[0035] By combining the fusion confidence evaluation factor value corresponding to each spatial location of the insert edge, the fusion confidence level label of the insert edge is determined according to the level division interval, and a fusion confidence level labeling map is generated to spatially label the fusion confidence level of the insert edge.

[0036] Based on the fusion confidence level annotation map, the image acquisition path is optimized and screened. Image acquisition paths with a fusion confidence level higher than the first screening reference value are included in the image range to participate in the fusion, while image acquisition paths with a fusion confidence level lower than the first screening reference value are removed from the image range to participate in the fusion. The conditions for edge images to participate in the fusion are limited to control the image fusion quality.

[0037] Preferably, S5 is as follows:

[0038] According to the detection time sequence, the fusion confidence evaluation factors corresponding to each spatial position of the insert edge are continuously sampled and recorded, and the fusion confidence evaluation factors are serialized and arranged with the timestamp as the index to form a time series data set of fusion confidence evaluation factors, which is used to track the time series change trend of fusion confidence evaluation factors.

[0039] Based on the time series data set of fusion confidence assessment factors, a corresponding fusion confidence assessment spectrum is constructed. The magnitude and density of numerical changes in adjacent time periods in the spectrum are statistically analyzed to obtain the spectrum volatility parameter reflecting the stability of the fusion state. The threshold for determining the directional consistency index and the threshold for dividing the confidence factors are adjusted in conjunction with the spectrum volatility parameter.

[0040] The adjusted directional consistency index judgment threshold and confidence factor division threshold are reapplied to the fusion confidence level division process of the insert edge, and the fusion confidence level of each spatial location is periodically updated. When the fusion confidence assessment factor changes over time, the dynamic control of the fusion confidence level of the insert edge is realized.

[0041] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0042] 1. This invention constructs multi-view edge direction mapping maps and direction offset maps for inserts, and integrates spatial consistency indicators such as edge matching mismatch, edge jump rate, edge inversion ratio, and edge closure deviation. This systematically identifies geometric conflict areas at insert edges caused by viewpoint differences, thus avoiding the direct use of inconsistent structural representation information for fusion judgment. Compared to traditional consistency detection methods that rely on image pixel differences, texture features, or depth values, this scheme can more accurately capture potential fusion distortion risks caused by boundary interleaving during multi-sensor fusion, reducing the probability of edge distortion, breakage, ghosting, or overlap, thereby enhancing the structural representation reliability of the image fusion result.

[0043] 2. This invention introduces a nonlinear fusion confidence evaluation function to establish a mapping relationship between multi-feature inputs and edge fusion stability. Through the division and dynamic adjustment mechanism of fusion confidence levels, combined with spectrum fluctuations and temporal trend changes, it achieves intelligent screening and temporal adaptive adjustment of the fusion conditions for insert edges, exhibiting higher flexibility and robustness. In long-term operation scenarios, this scheme can effectively prevent fusion judgment deviations caused by changes in sensor state, viewing angle, or insert posture adjustments, thereby improving the accuracy, stability, and scalability of motor insert defect detection results under multi-sensor fusion, providing a more engineering-adaptable solution for industrial quality inspection scenarios. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0045] Figure 1This is a flowchart illustrating the method for detecting defects in motor inserts based on multi-sensor fusion according to the present invention. Detailed Implementation

[0046] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0047] This invention provides, for example Figure 1 The method for detecting defects in motor inserts based on multi-sensor fusion, as shown, specifically includes the following steps:

[0048] S1. Extract the direction information of the motor insert edge from multiple imaging perspectives, establish an insert edge direction mapping map, and construct an insert edge direction offset map by calculating the direction consistency index, which is used to determine whether the insert edge has boundary interpenetration in multi-view images.

[0049] In this embodiment, S1 specifically refers to:

[0050] Edge extraction processing is performed on the edge of the motor insert from multiple imaging perspectives. Gradient calculation is used to obtain the directional information of each pixel on the edge of the motor insert. The directional difference between the directional vector of each pixel and the reference directional direction of the insert is calculated to obtain a set of directional information.

[0051] After acquiring insert images from multiple imaging perspectives, edge extraction is first performed. Common image edge operators such as Sobel, Canny, or Prewitt can be used to detect insert edge regions through grayscale gradient changes. The extracted edge regions contain continuous pixel sequences. Next, the orientation gradient of the local neighborhood of each edge pixel is estimated using gradient calculation methods to obtain the orientation vector of each pixel. For example, in the Sobel operator, the dominant direction of a point can be calculated as diagonally upward or diagonally downward based on the intensity of pixel changes in the horizontal and vertical directions. Subsequently, a reference orientation model of the motor insert is constructed, usually based on the overall orientation reference axis determined by the insert's structural design. The angle difference between the orientation vector of each edge pixel and the reference direction is calculated to obtain the orientation difference value of that pixel. Each orientation difference value is organized according to the edge position to form a set containing multiple pixel orientation differences. For example, in a camera image, if the left edge pixel orientation of the insert is diagonally upward at 45°, while the reference direction is horizontal, then the orientation difference value of that point is 45°. Multiple such data are summarized to form the orientation information set for that perspective.

[0052] Edge extraction processing refers to using image processing algorithms to identify areas with significant grayscale or color changes in an image, thereby locating the boundary contours of the motor inserts. This process provides a pixel-based foundation for subsequent orientation analysis. Gradient calculation is a method that estimates local orientation trends based on pixel intensity changes, often used to determine the orientation of edges. Orientation information refers to the main directional features exhibited by edge pixels in their local regions, and is important data for measuring edge directionality. Orientation difference calculation involves analyzing the angle between the orientation vector of each edge pixel and the overall reference direction of the motor inserts to obtain the degree of local orientation shift, reflecting the difference between the local edge orientation and the global trend. The orientation information set is an ordered set composed of the orientation differences of all edge pixels according to their spatial positions, used to express the overall orientation characteristics of the insert edges from that viewpoint, providing basic data support for consistency analysis and offset map construction in subsequent image fusion.

[0053] Based on the position index of each edge pixel in the unified interpolation space coordinates, the direction information extracted from multiple imaging perspectives is aligned according to the spatial position, and the direction vectors of each perspective are summarized at each edge position to establish an interpolation edge direction mapping map.

[0054] To effectively integrate the edge orientation information of the insert obtained from different imaging perspectives, it is necessary to first transform all image content into a unified insert space coordinate system. Using known geometric calibration relationships between multi-view cameras (such as rotation and translation matrices), the image coordinates of each pixel can be converted into its true spatial position within the insert structure, thus assigning a unique position index to each edge pixel. Next, this position index is used as a spatial anchor point to pair and summarize direction vectors at the same location from images from different perspectives. For example, at a certain insert edge location, the direction vector recorded by perspective A is diagonally upward, perspective B is vertically downward, and perspective C is horizontal; thus, these three direction vectors will be aggregated at that spatial location. Finally, by establishing a one-to-one mapping relationship between all edge spatial locations and their corresponding multiple direction vectors, an insert edge orientation mapping map is generated to characterize the orientation information of each position of the insert edge in space under multiple perspectives. This process realizes the transformation from single-view orientation features to spatial-level orientation fusion features, providing core data support for subsequent offset judgment and consistency analysis.

[0055] The unified splice spatial coordinate system's position index refers to using the motor splice itself as a unified reference frame to spatially project and number the positions of corresponding edge points in different images, ensuring that edge pixels at the same physical location in multiple viewpoints can be correctly mapped and categorized. The orientation vector for each viewpoint refers to the edge pixel orientation information extracted through gradient analysis at each imaging viewpoint, typically represented as a unit vector. The splice edge orientation mapping map is a data map constructed with the splice edge's spatial position as the horizontal axis and the orientation vectors of that position across all viewpoints as the vertical data content. Essentially, it's a spatial fusion graph of multi-view orientation features, used to characterize the orientation consistency and offset distribution of the motor splice edges under multi-view imaging. This map provides an analytical basis for subsequent judgments regarding boundary overlap and the existence of fusion risks.

[0056] For each edge position in the insertion edge direction mapping map, the maximum angle between different viewpoint direction vectors is calculated as the direction consistency index, and the direction consistency index is mapped across the entire insertion edge domain to construct the insertion edge direction offset map; when the direction consistency index of consecutive edge positions exceeds the first set threshold, it is determined that the insertion edge has boundary interleaving in the multi-view image.

[0057] For each edge location in the edge orientation mapping map, edge orientation vectors from multiple imaging viewpoints can typically be obtained. For this location, by traversing any pairwise combinations of viewpoints, the angle difference between their orientation vectors is calculated, and the pair with the largest angle is selected as the most severe indication of orientation divergence. This maximum angle value is the orientation consistency index for that edge location; the larger the value, the more inconsistent the orientation of the edge under different viewpoints. After calculating the consistency index for all edge locations, these values ​​are spatially mapped across the entire edge region to obtain the edge orientation offset map. This map can visually reflect which edge regions have significant concentrated distributions of orientation offset. For example, if multiple adjacent locations on a certain edge exhibit a maximum angle exceeding a specified angle, it indicates a persistent multi-view orientation conflict in that region, often a direct signal of boundary interleaving. Therefore, when the orientation consistency index of multiple consecutive edge locations exceeds a preset first threshold, potential boundary interleaving regions can be identified, providing input for fusion error warning.

[0058] The orientation consistency index is an indicator used to quantify the degree of difference in the orientation expression of insert edges across multiple imaging perspectives, reflecting whether the subjective orientation of the same physical edge location is consistent in different images. This index is not based on an average value, but rather selects the maximum angle value to keenly capture the most extreme orientation conflicts, thereby enhancing the ability to detect potential fusion errors. The insert edge orientation offset map is a visual representation of the orientation consistency index distribution of each edge location on the insert edge space. High offset areas can be identified through color coding or numerical gradients, providing accurate input for subsequent processing. The first set threshold is a criterion for judging whether the orientation consistency index is abnormal. This threshold is set according to the acceptable range of edge orientation variation in the actual scene; generally, a small angular deviation is considered normal, while exceeding the threshold indicates abnormal divergence in edge orientation. This threshold not only serves as a critical basis for boundary interleaving judgment but can also be used as a preliminary screening condition for fusion confidence level evaluation.

[0059] S2. In the case of boundary interpenetration, calculate the edge matching mismatch degree, edge jump rate, edge inversion ratio and edge closure deviation to form a set of geometric conflict features of the interpenetrated edges, which are used to characterize the differences in image geometric expression caused by boundary interpenetration.

[0060] In this embodiment, S2 specifically includes the following steps:

[0061] S201. In the case of boundary overlap, the edge contour points extracted from the edge of the insert in multiple imaging views are spatially aligned, the Euclidean distance between contour points of different views at the same spatial position is calculated, and the ratio of the average distance to the maximum distance at each position is used as the edge matching mismatch degree.

[0062] S202. Based on the edge region calibrated by the edge matching mismatch degree, collect the direction vectors corresponding to each spatial position, calculate the angle difference sequence between adjacent direction vectors in sequence according to the continuous spatial index, count the proportion of points in the angle difference sequence that are greater than the preset angle threshold as the edge jump rate, and count the proportion of points where the angle difference between the direction vectors is close to the opposite relationship as the edge reversal ratio. Calculate the spatial closure error between the starting point and the ending point of the edge contour, and use the spatial closure error as the edge closure deviation.

[0063] S203. The edge matching mismatch degree, edge jump rate, edge reversal ratio and edge closure deviation are combined to form a set of geometric conflict features of the interpolation edge, which is used to characterize the differences in image geometric expression caused by boundary interpolation.

[0064] Combining edge mismatch, edge jump rate, edge reversal ratio, and edge closure deviation to form a geometric conflict feature set for intercalation edges aims to establish a multi-dimensional quantitative index system in boundary interleaving scenarios, comprehensively assessing whether there are anomalies in the geometric representation of intercalation edges in an image. Specifically, a four-dimensional vector can be constructed for each spatial location, with each dimension corresponding to the edge mismatch, edge jump rate, edge reversal ratio, and edge closure deviation for that location. Integrating the four-dimensional vectors from all locations constitutes the geometric conflict feature set for intercalation edges. Based on this, clustering algorithms can be used to identify anomalous clustered regions, or a joint threshold can be set to determine the significance of geometric conflict regions. For example, when a region simultaneously meets the conditions of large edge mismatch, high edge jump rate, significant direction reversal, and large closure error, it can be inferred that this region is highly likely to cause image fusion errors due to viewpoint interleaving, thus marking it as a potential high-risk region. This combination not only supplements single geometric indicators but also enhances the model's robustness under complex interference conditions.

[0065] The geometric conflict feature set of insert edges is a multi-dimensional collection used to comprehensively reflect the geometric inconsistencies of insert edges in images from different viewpoints caused by boundary interleaving. Edge matching mismatch reflects the degree of spatial matching of contour points; edge jump rate reveals the degree of abrupt change in direction vectors locally; edge inversion ratio measures the concentration of edge direction inversion; and edge closure deviation reflects the continuity of the contour. These four indicators form a complementary evaluation perspective from three angles: spatial consistency, directional stability, and structural continuity. By constructing the geometric conflict feature set of insert edges, the geometric anomalies of images caused by boundary interleaving can be fully expressed numerically, providing high-dimensional data support for subsequent risk identification and fusion credibility assessment.

[0066] In this embodiment, S201 specifically refers to:

[0067] In the case of boundary overlap, based on the pixel coordinates of the interpolation edge in the multi-view image, the edge contour points extracted from multiple imaging views are projected to a unified three-dimensional spatial coordinate system through extrinsic parameter transformation, and the edge contour points from different imaging views are paired according to the spatial position index.

[0068] In cases of boundary overlap, to achieve spatial alignment of insert edge contour points across multiple imaging perspectives, the first step is to extract the two-dimensional pixel coordinates of the insert edge pixels in each imaging perspective image. These coordinates represent the positional distribution of the insert edge points in the image. By pre-calibrating the extrinsic parameters corresponding to each sensing perspective, including the camera's spatial rotation and translation vectors, a transformation relationship from individual image coordinate systems to a unified world coordinate system can be established. Using each edge pixel coordinate as input, the extrinsic parameter transformation projects it from the image plane to a unified three-dimensional spatial coordinate system, enabling edge points from different perspectives to be mapped to spatial locations with consistent semantics. The unified three-dimensional spatial coordinate system is a reference coordinate framework used to eliminate spatial offsets caused by shooting angles and positions between different cameras, ensuring that all edge points are processed at the same scale and spatial reference. The projected three-dimensional coordinates are divided according to spatial position indexing, and points falling in the same or similar spatial cells are considered as multi-view contour point pairs at the same location. This method achieves precise spatial pairing of edge contour points from different sensing perspectives, providing a foundation for subsequent calculation of spatial geometric differences and construction of conflict features. For example, if the edge of a certain insert appears in two images from different viewpoints, but exhibits different contour curvatures due to the difference in viewpoints, its physical structure can be determined in three-dimensional space after unified projection, avoiding judgment errors caused by differences in two-dimensional images.

[0069] For each pairing position, calculate the Euclidean distance between any two sets of edge contour points at that position, and collect the Euclidean distance data of all pairing positions to obtain the average Euclidean distance and the maximum Euclidean distance for each position.

[0070] For each paired location, the coordinates of edge contour points from different imaging perspectives at that location are first extracted in a unified three-dimensional coordinate system. Then, these three-dimensional coordinates are combined pairwise, and their Euclidean distances are calculated. Euclidean distance refers to the straight-line distance between two points in three-dimensional space, which can intuitively reflect the degree of difference in the physical position of edge contour points from different perspectives. When there are three sets of contour points at a certain paired location, it is necessary to calculate the distance values ​​of all pairwise combinations between these three sets of points, resulting in three sets of Euclidean distance data. Next, all Euclidean distance values ​​at that location are aggregated. On the one hand, the arithmetic mean of these distance values ​​is calculated to form the average Euclidean distance, which is used to evaluate the overall spatial consistency of the location. On the other hand, the maximum value among all distance values ​​is extracted to obtain the maximum Euclidean distance, which reflects the maximum spatial offset of the location across all perspectives. The purpose of this setup is to simultaneously evaluate the overall consistency and extreme deviation of the contour points, so as to more finely characterize the consistency and potential conflict of the edge geometry at that spatial location. For example, if the average distance between contour points at a location is small but the maximum distance is large across multiple viewpoints, it indicates that the location presents a consistent boundary in most viewpoints, but exhibits an abnormal deviation in one viewpoint. This may indicate a potential risk of boundary overlap, and therefore requires close attention in subsequent processing. This process provides a fundamental data source for calculating edge matching mismatch and significantly improves the accuracy of edge conflict identification.

[0071] For each pairing position, the ratio between the average Euclidean distance and the maximum Euclidean distance at that position is used as the edge matching mismatch degree, which serves as the basis for evaluating the spatial consistency of the insert edges in the case of boundary interleaving, and is used for subsequent conflict feature extraction.

[0072] For each paired location, the edge matching mismatch is calculated by the ratio of the average Euclidean distance to the maximum Euclidean distance at that location. This ratio reflects the degree of geometric consistency of the contour points across different imaging perspectives. A value closer to one indicates smaller differences between contour points across all perspectives at that location, higher spatial overlap, and a highly consistent geometric shape at that edge location in multi-view imaging. Conversely, a smaller edge matching mismatch value indicates significant deviations in contour points at certain perspectives, particularly severe spatial differences between individual perspectives and other perspectives, leading to a significant decrease in spatial consistency. This could be a direct manifestation of boundary interleaving. The edge matching mismatch serves as a key indicator for spatial consistency assessment, allowing for rapid identification of significant geometric expression conflicts at the edges of the inserts across multiple perspectives. For example, if the distance between two groups of contour points is only one unit, while the distance between the third group and the other two groups is as high as ten units, the average distance is small, but the maximum distance is large, resulting in a significant decrease in the matching mismatch ratio, indicating strong conflict characteristics at that location. When forming the geometric conflict feature set of the insert edge, only those edge positions with a matching mismatch degree below the set threshold will be included in the subsequent feature extraction process, thereby ensuring that the boundary interpenetration phenomenon can be accurately detected and analyzed.

[0073] In this embodiment, S202 specifically refers to:

[0074] Based on the numerical distribution of edge matching mismatch degree on the edge of the insert, the spatial location where the edge matching mismatch degree is greater than the second set threshold is selected as the edge region. The direction vector corresponding to each spatial location is collected in the edge region, and the direction vector is continuously spatially indexed and sorted according to the spatial arrangement order of the insert edge to form a direction vector sequence.

[0075] After spatially mapping the edge matching mismatch degree of the insert edges, analyzing the numerical distribution of the edge matching mismatch degree can identify which spatial locations exhibit strong geometric consistency deviations. A second threshold is set as the control benchmark for edge consistency screening, used to filter spatial locations with edge matching mismatch degrees greater than this threshold from the overall edge region, as edge regions where boundary intersecting conflicts are more likely. Within this edge region, the corresponding direction vector information is extracted for each spatial location; the direction vector reflects the local orientation of the edge at that location. Based on this, to better capture the continuous orientation changes of the edge in space, these direction vectors need to be reorganized according to the spatial arrangement order of the insert edges. By establishing a continuous spatial index, the direction vectors of adjacent spatial locations have a strict sequential relationship on the index, thus forming a direction vector sequence. This direction vector sequence not only retains the direction information itself but also embeds the contextual relationship of spatial changes, providing a continuous foundation for subsequent angle difference calculation and edge geometric behavior recognition. For example, if a segment of edge direction vector abruptly changes to its opposite direction in space, a large angle value will be formed in the continuous sequence, thereby further identifying edge reversal regions or abrupt change regions. Therefore, this process can provide a strong foundation for constructing fine edge geometric conflict features.

[0076] For adjacent direction vectors in the direction vector sequence, calculate the angle difference between the direction vectors one by one to form an angle difference sequence. In the angle difference sequence, count the proportion of spatial positions where the angle difference is greater than a preset angle threshold to obtain the edge jump rate. Count the proportion of spatial positions where the angle difference is in the near-reverse relationship interval to obtain the edge reversal ratio.

[0077] To accurately reveal the potential geometric conflict features of insert edges during multi-view image fusion, it is necessary to extract the degree of continuous directional change from the direction vector sequence. Specifically, for each adjacent direction vector in the direction vector sequence, the angle between them is calculated sequentially, forming an angle difference sequence reflecting the magnitude of edge directional change. Each value in the angle difference sequence represents the turning angle between adjacent edge segments; the larger the value, the more obvious the abrupt change in edge directional change. By setting a preset angle threshold, the ratio of the number of spatial locations with values ​​greater than the threshold in the angle difference sequence to the entire sequence length is calculated to obtain the edge jump rate, which reflects the density of sudden changes in the edge directional change of the insert. Simultaneously, vector pairs with directional differences close to 180 degrees are further identified in the angle difference sequence, i.e., pairs of direction vectors whose angles are in the near-reverse relationship range. This type of change often occurs in edge intersecting areas, representing a significant reversal of edge directional change. The ratio of the number of near-reverse points to the total number of points is then calculated to obtain the edge reversal ratio. The edge jump rate and edge reversal ratio, as two types of perturbation indicators for edge continuity and directional consistency, can effectively characterize geometric anomalies caused by boundary intersecting. For example, in the boundary interleaving region, high-frequency abrupt changes and direction mirror reversals often occur in the direction sequence, the angle difference jumps frequently, and the values ​​are densely distributed in the high-angle segment and near-reverse interval, thus forming a typical geometric feature of increased jump rate and prominent reversal ratio, which is used to support the subsequent fusion confidence assessment.

[0078] After calculating the edge jump rate and edge reversal ratio, the minimum and maximum spatial index positions in the edge contour are selected as the starting and ending points of the edge contour. The spatial distance between the two points is calculated as the spatial closure error, and the spatial closure error is used as the edge closure deviation.

[0079] When assessing whether the edge of an insert exhibits geometric anomalies due to boundary interleaving, in addition to directional changes, it is also necessary to detect the geometric closure of the edge contour. To this end, the starting and ending points of the edge contour in a continuous spatial index can be used as the basis for analysis. Specifically, the minimum position in the spatial index is selected as the starting point of the edge contour, and the maximum position is selected as the ending point. The coordinates of these two positions in a unified three-dimensional spatial coordinate system are then extracted. The spatial distance between these two points is then calculated. If this distance is large, it indicates that the edge contour has not closed into a continuous edge and there is significant deformation. This spatial distance is the spatial closure error, which is further defined as the edge closure deviation. The edge closure deviation reflects whether the edge maintains its morphological continuity after multi-view fusion. If boundary interleaving causes the edge to appear broken, misaligned, or deformed in the image, the starting and ending points will usually shift, increasing the closure deviation. For example, if the edge of the insert is actually a closed arc, and interleaving causes a break in the image, the closure deviation will be much greater than zero. By using this closure analysis method based on spatial index boundaries, a one-dimensional structural continuity criterion can be added in addition to directional consistency, thereby more comprehensively identifying differences in image geometric representation caused by boundary interlacing.

[0080] S3. Input the geometric conflict feature set into the fusion confidence evaluation function to obtain the fusion confidence evaluation factor, which is used to identify whether there is a risk of fusion distortion at the edge of the insert;

[0081] In this embodiment, S3 specifically refers to:

[0082] The edge matching mismatch degree, edge jump rate, edge reversal ratio and edge closure deviation corresponding to each spatial location in the geometric conflict feature set of the insert edge are feature-encoded, and a uniform input vector is constructed. This input vector is then input into the fusion confidence evaluation function, which is a nonlinear feature mapping function constructed based on multi-feature correlation analysis.

[0083] In the geometric conflict feature set of the insert edge, edge matching mismatch, edge jump rate, edge reversal ratio, and edge closure deviation usually exist in different forms and numerical dimensions. To facilitate joint analysis within the same model, these features need to be standardized. Specifically, each feature can be normalized according to its statistical distribution range, ensuring it falls within a uniform interval, such as 0 to 1. Then, the four types of feature values ​​corresponding to each edge spatial location are concatenated in a fixed order into a one-dimensional vector of length four, forming a uniform input vector. By traversing the entire insert edge region, an input set consisting of a large number of such input vectors is obtained. This set is then input into the fusion confidence evaluation function, which analyzes the combination relationships between the features. This design ensures that different types of conflict features have a uniform expression dimension and numerical scale when entering the confidence evaluation function, thereby improving the consistency and accuracy of multi-feature information fusion.

[0084] The uniformly formatted input vector is a fixed-dimensional data structure composed of multiple normalized geometric conflict features, whose order strictly corresponds to the feature names. For example, the first position is the edge matching mismatch degree, the second is the edge jump rate, the third is the edge inversion ratio, and the fourth is the edge closure deviation, ensuring that the data has a consistent semantic interpretation before entering the model. The fusion confidence assessment function is a computational function that models the interaction relationships between input features, with its core being a nonlinear feature mapping mechanism. This nonlinear feature mapping function is constructed based on multi-feature correlation analysis, typically employing models such as neural networks or kernel function mapping to uncover higher-order coupling patterns between features, rather than simple weighted averaging. This nonlinear feature mapping function can identify the potential fusion instability trends exhibited by various complex feature combinations in different edge regions, thus providing more accurate support for subsequent numerical inference of confidence factors.

[0085] In the fusion confidence evaluation function, the feature importance of each feature field in the input vector is weighted and calculated, and the fusion confidence evaluation factor is obtained through a nonlinear fitting model. The fusion confidence evaluation factor is output in numerical form and is used to measure the stability of the interpolation edge fusion.

[0086] In the fusion confidence assessment function, each input vector in a uniform format consists of four dimensions, representing edge matching mismatch, edge jump rate, edge inversion ratio, and edge closure deviation. To differentiate the influence of different features on the final judgment result, feature importance weighting is first performed on each feature field. The weighting coefficients can be learned from historical data of fusion results in the training samples, reflecting the correlation between each feature and the actual fusion stability. After weighting, the processed vector is input into a nonlinear fitting model, which can be constructed using neural networks, multilayer perceptrons, or gradient boosting tree-based regression models, possessing the ability to extract high-order patterns from complex feature combinations. Through model calculation, a fusion confidence assessment factor is output. This factor is a continuous value used to measure the stability or distortion risk exhibited by the current edge region during the fusion process. This value varies within a certain range, with different value intervals representing different levels of fusion reliability, thus providing a decision-making basis for subsequent image processing strategies.

[0087] Feature importance weighting calculation refers to assigning a weighting coefficient proportional to the influence of each type of feature in multidimensional input data. This enhances the decision-making influence of key features and reduces the interference of low-relevance features on the final result. These weighting coefficients are typically derived from feature contribution analysis during model training. Nonlinear fitting models refer to function construction methods with nonlinear expressive capabilities. They can capture implicit nonlinear relationships in input data through hidden layer structures or kernel functions, enabling the model to distinguish complex fusion patterns. The fusion confidence assessment factor is a real-valued index output by the model, representing the degree of structural consistency that may occur when fusing multi-view information at the current edge spatial location. The closer the factor is to the stable range, the more reliable the edge fusion; conversely, a lower value indicates a risk of fusion distortion. Through the above process, complex geometric conflict features can be mapped into quantifiable confidence evaluation results, which helps guide subsequent processing paths.

[0088] A first confidence risk threshold is set based on the numerical range of the fusion confidence assessment factor, and the fusion confidence assessment factor of each insert edge spatial location is compared and judged. When the fusion confidence assessment factor is lower than the first confidence risk threshold, the edge of the insert is identified as having a risk of fusion distortion.

[0089] During the edge fusion evaluation process, a numerical benchmark is needed to determine whether the current edge carries a risk of fusion distortion. To this end, a first confidence risk threshold can be set based on the correspondence between different fusion confidence assessment factor values ​​and actual fusion stability in historical samples. This threshold can be determined by annotating and analyzing multiple patch samples, statistically analyzing the probability of fusion distortion or boundary errors occurring when the confidence factor is below a specific value, thereby identifying representative boundary points. After setting this threshold, the fusion confidence assessment factor is compared point-by-point for the spatial location of each patch edge. When the assessment factor value is below the first confidence risk threshold, it is determined that the edge location has a high probability of fusion distortion risk under the current fusion conditions. For example, when the fusion confidence assessment factor is less than the empirically set 0.35, it indicates significant geometric conflicts between orientation, contour, or closure features under multiple views, and this area should be avoided in the final image fusion process to prevent subsequent misjudgment or fusion anomalies.

[0090] The first confidence risk threshold is a numerical boundary point set for the fusion confidence assessment factor, serving to distinguish reliable fusion regions from potentially distorted regions at a numerical level. This threshold is not fixed but is derived from historical fusion success and failure data in the training samples, and can be adaptively fine-tuned based on the actual motor insert type, sensor parameters, and imaging quality. The fusion confidence assessment factor is a continuous numerical value representing the degree of fusion risk; lower values ​​correspond to higher uncertainty and conflict levels. By comparing it with the first confidence risk threshold, high-risk edge regions can be quickly identified. This judgment mechanism plays a crucial role in the edge fusion stage, effectively improving the robustness and accuracy of the overall judgment.

[0091] S4. Divide the fusion confidence level of the insert edge according to the fusion confidence evaluation factor, and screen the image acquisition path based on the fusion confidence level to limit the conditions for edge images to participate in fusion.

[0092] In this embodiment, S4 specifically refers to:

[0093] Based on the numerical distribution of the fusion confidence assessment factor, a level division interval is constructed. The complete numerical range of the fusion confidence assessment factor is divided into multiple non-overlapping segments, and an independent fusion confidence level label is set for each numerical segment to divide the fusion confidence level of each spatial location of the insert edge.

[0094] To ensure accurate quality control of edge images during the fusion process, a graded interval needs to be constructed based on the numerical distribution of the fusion confidence assessment factor. This can be achieved by statistically analyzing the numerical distribution of the fusion confidence assessment factor at all interpolated edge spatial locations and using quantile or clustering methods to set several dividing points, dividing the entire numerical range into multiple non-overlapping segments. Each segment represents a different fusion confidence level, such as "high confidence," "medium confidence," and "low confidence." In practical implementation, historical statistical data of the fusion confidence assessment factor can be analyzed, and typical values ​​can be selected as dividing nodes to ensure clear boundaries between each level. This graded division allows for hierarchical control of edge images based on confidence factor strength during subsequent processing, improving the stability and reliability of the fusion process.

[0095] The grading interval refers to dividing the continuous numerical space of the fusion confidence assessment factor into several non-overlapping numerical segments to express different levels of fusion confidence. The complete numerical range of the fusion confidence assessment factor is usually a set of continuous floating-point numbers generated by the algorithm model during training or prediction, used to quantify the ability of the patch edges to maintain structural consistency during image fusion. Multiple non-overlapping segments mean that these divided intervals have no intersection or repetition, ensuring that each value belongs to only one unique grade. An independent fusion confidence grade label refers to assigning a unique fusion evaluation identifier to each non-overlapping segment, which can be represented in the form of a grade code or text label, used for subsequent identification of the fusion stability grade of the patch edges in different spatial locations. The above division and labeling provide a quantitative basis for image path selection and fusion quality control.

[0096] By combining the fusion confidence evaluation factor value corresponding to each spatial location of the insert edge, the fusion confidence level label of the insert edge is determined according to the level division interval, and a fusion confidence level labeling map is generated to spatially label the fusion confidence level of the insert edge.

[0097] To clearly define the fusion stability of the interpolation edges spatially, it is necessary to match the fusion confidence evaluation factor value at each spatial location with a preset level division interval and determine its corresponding fusion confidence level label. Specifically, this can be achieved by traversing all spatial locations along the interpolation edges, using the fusion confidence evaluation factor at each location as an index input, and searching for the interval it falls into among multiple defined non-overlapping segments, thereby assigning the corresponding fusion confidence level label. This label information is then mapped back to the 3D interpolation edge space, ultimately generating a fusion confidence level annotation map covering the entire edge space. This map visually indicates the fusion stability at various points along the interpolation edges, helping to selectively decide whether to adopt image data from a particular viewpoint based on spatial location in subsequent image fusion tasks, thus effectively reducing the risk of fusion distortion caused by boundary conflicts.

[0098] A fusion confidence level annotation map is a type of map used in three-dimensional space to classify and label the fusion stability of patch edges. Using spatial coordinates as the basic unit, this map maps the fusion confidence assessment factor at each edge location to a pre-defined fusion confidence level label through numerical comparison. Common label formats include numerical labels, color-coded labels, or discrete symbolic labels. This annotation map not only possesses spatial continuity and visualization capabilities but also provides a direct basis for judging the usability of edge information in multi-view fusion, serving as core supporting data for subsequent path selection and fusion condition constraints. Through the fusion confidence level annotation map, dynamic partitioning control of edge regions can be achieved, ensuring structural consistency and image integrity during the fusion process.

[0099] Based on the fusion confidence level annotation map, the image acquisition path is optimized and screened. Image acquisition paths with a fusion confidence level higher than the first screening reference value are included in the image range to participate in the fusion, while image acquisition paths with a fusion confidence level lower than the first screening reference value are removed from the image range to participate in the fusion. The conditions for edge images to participate in the fusion are limited to control the image fusion quality.

[0100] To ensure that only structurally stable patch edge image information is used during image fusion, the image acquisition paths need to be filtered based on the fusion confidence level annotation map. Specifically, the image projection region corresponding to each image acquisition path is first traversed within the patch edge space, and the fusion confidence level labels for all spatial locations within that region are extracted, and the distribution characteristics of the label values ​​are statistically analyzed. When the overall level of the fusion confidence level labels in a path's coverage area is higher than a preset first filtering reference value, it is included in the set of image paths participating in fusion; otherwise, the image corresponding to that path is removed to avoid low-confidence-level regions affecting fusion accuracy. This filtering mechanism can dynamically eliminate potential distortion risk sources before fusion, thereby improving the continuity and geometric consistency of edge structures during image fusion and reducing structural distortion caused by boundary interleaving.

[0101] The first screening reference value is a numerical threshold used to select image acquisition paths for fusion eligibility, determining whether a particular image path possesses a sufficient fusion confidence level. It is set based on the numerical distribution of the fusion confidence assessment factor and the system's fault tolerance strategy, typically a fusion confidence level label or its corresponding lower bound. During the image acquisition path screening process, the fusion confidence level label at all edge locations must be compared with this reference value. When the proportion of paths within the path's coverage area that have a fusion confidence level higher than the first screening reference value reaches a set standard, the path is considered to have a stable edge structure and is suitable for participating in multi-view image fusion; otherwise, it will be excluded. This mechanism ensures that the image information participating in the fusion has high structural consistency in the boundary region, contributing to improved overall geometric stability of the fused image.

[0102] S5. Track the temporal change trend of the fusion confidence assessment factor, and adjust the directional consistency index and confidence factor division threshold based on the spectrum volatility to achieve dynamic control of the fusion confidence level of the insert edge.

[0103] In this embodiment, S5 specifically refers to:

[0104] According to the detection time sequence, the fusion confidence evaluation factors corresponding to each spatial position of the insert edge are continuously sampled and recorded, and the fusion confidence evaluation factors are serialized and arranged with the timestamp as the index to form a time series data set of fusion confidence evaluation factors, which is used to track the time series change trend of fusion confidence evaluation factors.

[0105] In the edge fusion evaluation process of insert chips, to capture the pattern of fusion status changes during the detection process, it is necessary to continuously sample the fusion confidence evaluation factors and construct a time series. Specifically, after each insert chip detection, the fusion confidence evaluation factors corresponding to each spatial location of the insert chip edge are sampled once, and a unique timestamp is added to each sampling result. Then, according to the order of the timestamps, the fusion confidence evaluation factors obtained at the same spatial location at different detection times are arranged sequentially, forming a time-ordered data sequence. This continuous sampling and serialized arrangement method can intuitively reflect the changes in fusion confidence at a certain edge location during multiple detection processes. For example, during continuous production cycle operation, if the fusion confidence evaluation factor at a certain insert chip edge location continuously decreases at multiple adjacent time points, this trend can be captured in time to determine that the fusion status at that location is changing, thus providing a basis for subsequent threshold adjustment and confidence level updates.

[0106] The detection time sequence refers to arranging the sampling results of each fusion confidence assessment factor according to the actual occurrence order during the chip insertion detection process, ensuring a clear sequential relationship in the data sequence. Continuous sampling records refer to the uninterrupted acquisition of fusion confidence assessment factors at the same spatial location during multiple detection processes, enabling the data to reflect state changes rather than single-point results. Timestamps serve as indexes, assigning a unique time identifier to each fusion confidence assessment factor, thus supporting subsequent sorting, retrieval, and trend analysis. Serialization is the process of organizing fusion confidence assessment factors into an ordered data set according to timestamp order; this set is the time-series data set of fusion confidence assessment factors. The purpose of this time-series data set is to characterize the trend characteristics of fusion confidence assessment factors over time, transforming fusion assessment from static judgment to dynamic analysis, laying the data foundation for subsequent volatility-based threshold adjustments and fusion confidence level control.

[0107] Based on the time series data set of fusion confidence assessment factors, a corresponding fusion confidence assessment spectrum is constructed. The magnitude and density of numerical changes in adjacent time periods in the spectrum are statistically analyzed to obtain the spectrum volatility parameter reflecting the stability of the fusion state. The threshold for determining the directional consistency index and the threshold for dividing the confidence factors are adjusted in conjunction with the spectrum volatility parameter.

[0108] When constructing a dynamic evaluation mechanism for the fusion state, a fusion confidence evaluation map needs to be generated based on the time-series data set of fusion confidence evaluation factors to visualize the evolution of the fusion state over time. In implementation, the values ​​of the fusion confidence evaluation factors corresponding to each spatial location at different times can be mapped to the vertical axis data in a two-dimensional map, with the timestamp as the horizontal axis, presenting a continuous trend in the form of a line graph or heatmap. To quantify the stability of the map, the amplitude of the difference in fusion confidence evaluation factor values ​​between adjacent time periods needs to be statistically analyzed, and the number of significant changes in values ​​per unit time is counted using moving window technology to calculate the density of change. By jointly modeling the amplitude of change and the density of change, a map volatility parameter reflecting the stability of the fusion state can be obtained. For example, when a certain edge region frequently experiences large fluctuations in the fusion confidence value in several consecutive detections, the map volatility parameter will increase significantly. Based on this parameter, the judgment threshold of the directional consistency index can be adjusted in conjunction, making the boundary interleaving detection more adaptable to time-varying interference, and the confidence factor division threshold can be adjusted simultaneously to improve the sensitivity of the fusion confidence level judgment.

[0109] The fusion confidence assessment map is a graphical representation structure built upon time-series data of fusion confidence assessment factors, used to display the trajectory of edge fusion confidence status changes over time. The magnitude of numerical change between adjacent time periods refers to the absolute value of the difference in fusion confidence assessment factors at the same spatial location between two adjacent time points, reflecting the degree of local mutation. The density of change refers to the number of times changes exceed a preset threshold within a certain time window, used to measure the frequency of short-term fluctuations. The map volatility parameter is a quantitative indicator generated by combining the magnitude and density of numerical change through a statistical analysis model, used to describe the temporal stability of the fusion state. The threshold for the directional consistency index determines whether boundary overlap exists, while the confidence factor division threshold determines the boundary for fusion confidence level division. When the map volatility parameter indicates system instability, adjusting the above two thresholds in a coordinated manner can improve the overall assessment system's responsiveness to dynamic environments, thereby achieving dynamic control of the edge fusion confidence level.

[0110] The adjusted directional consistency index judgment threshold and confidence factor division threshold are reapplied to the fusion confidence level division process of the insert edge, and the fusion confidence level of each spatial location is periodically updated. When the fusion confidence assessment factor changes over time, the dynamic control of the fusion confidence level of the insert edge is realized.

[0111] To achieve dynamic adjustment of the confidence level of patch edge fusion, the previously adjusted directional consistency index threshold and confidence factor threshold, based on the map fluctuation parameter, need to be reapplied to the confidence level classification process of the entire edge region. Specifically, the new directional consistency index threshold is used to re-evaluate whether boundary overlap exists at each spatial location, thus affecting the corresponding geometric conflict feature composition. Then, the input vector is re-encoded using the updated geometric conflict feature set and input into the fusion confidence evaluation function to obtain a new fusion confidence evaluation factor. Based on the updated confidence factor values, the confidence level label is reassigned using the adjusted confidence factor classification threshold, generating a new fusion confidence level annotation map. This process can be set to be timed, for example, updated after several rounds of detection, to ensure timely correction of the confidence level evaluation results when the fusion state changes over time. For example, if a patch edge location experiences a short-term abnormal fusion confidence value due to vibration or illumination interference, periodic updates can prevent this area from being mislabeled as having a low confidence level for a long time, thereby improving system stability and fault tolerance.

[0112] The directional consistency index threshold is used to determine the degree of consistency in edge direction orientation across multiple perspectives; a smaller value indicates a more stringent requirement for edge direction consistency. The confidence factor classification threshold is used to divide the numerical range of the fusion confidence assessment factor into several level intervals, corresponding to different fusion confidence levels. Periodic updates refer to the periodic re-judgment process of edge fusion confidence levels under fixed time or event-driven conditions to respond to the impact of environmental changes on the fusion status. The fusion confidence level classification process is a mapping mechanism that transforms continuous values ​​into discrete level labels. It relies on the relationship between the assessment factor and the classification threshold; when the classification threshold is adjusted, the mapping process must be re-executed to reflect the current actual fusion status. Dynamic adjustment refers to adjusting relevant judgment criteria through continuous feedback based on the temporal changes of the assessment factor and periodically updating the results, thereby constructing an adaptive confidence level identification mechanism.

[0113] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0114] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0115] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0116] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0117] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0118] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0119] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A motor tab defect detection method based on multi-sensor fusion, characterized in that, Specifically comprising the following steps: S1, the direction of the motor insert edge is extracted in multiple imaging angles, the edge direction mapping graph is established, and the edge direction offset atlas is constructed by calculating the direction consistency index, which is used to judge whether the edge boundary insertion exists in the multiple view images, specifically: for each edge position in the edge direction mapping graph, the maximum included angle value between different view direction vectors is calculated as the direction consistency index, and the direction consistency index is mapped in the whole edge domain to construct the edge direction offset atlas; when the direction consistency index of the continuous edge position exceeds the first set threshold, it is judged that the edge boundary insertion exists in the multiple view images; S2, in the case of boundary insertion, the edge matching mismatch degree, edge jump rate, edge inversion ratio and edge closure deviation are calculated to form a geometric conflict feature set of the insert edge, which is used to represent the geometric expression difference caused by boundary insertion; S2 specifically comprises the following steps: S201, in the case of boundary insertion, the edge contour points extracted in multiple imaging angles are spatially aligned, the Euclidean distance between different view contour points at the same spatial position is calculated, and the ratio of the average distance to the maximum distance at each position is taken as the edge matching mismatch degree; S202, based on the edge region marked by the edge matching mismatch degree, the direction vector corresponding to each spatial position is collected, the angle difference sequence between adjacent direction vectors is calculated in sequence according to the continuous spatial index, the proportion of points greater than the preset angle threshold in the angle difference sequence is counted as the edge jump rate, and the proportion of points with nearly reverse relationship in the direction vector angle difference is counted as the edge inversion ratio. The spatial closure error between the starting point and the ending point of the edge contour is calculated, and the spatial closure error is taken as the edge closure deviation; S203, the edge matching mismatch degree, edge jump rate, edge inversion ratio and edge closure deviation are combined to form a geometric conflict feature set of the insert edge, which is used to represent the geometric expression difference caused by boundary insertion; S3, the geometric conflict feature set is input into the fusion confidence evaluation function to obtain the fusion confidence evaluation factor, which is used to identify whether the insert edge has fusion distortion risk; S4, the fusion confidence evaluation factor is used to divide the fusion confidence level of the insert edge, and the image acquisition path is selected based on the fusion confidence level to limit the participation condition of the edge image in fusion; S5, the time sequence change trend of the fusion confidence evaluation factor is tracked, and the direction consistency index and the confidence factor division threshold are adjusted based on the atlas fluctuation, so as to realize the dynamic regulation of the fusion confidence level of the insert edge.

2. The motor insert sheet defect detection method based on multi-sensor fusion according to claim 1, characterized in that, S1 specifically is: The edge of the motor insert is extracted in multiple imaging angles, the direction of each pixel point of the motor insert edge is obtained by gradient calculation, and the direction difference value of each pixel direction vector and the insert reference direction is calculated to obtain the direction information set. The direction mapping graph of the patch edge is established by aligning the direction information extracted from the multiple imaging perspectives according to the spatial positions based on the position index of each edge pixel in the unified patch space coordinates, and collecting the direction vectors of each edge position corresponding to each perspective. The maximum included angle value between the direction vectors of different perspectives is calculated as the direction consistency index for each edge position in the direction mapping graph of the patch edge, and the direction consistency index is mapped in the whole edge of the patch to construct the direction offset atlas of the patch edge.

3. The motor insert sheet defect detection method based on multi-sensor fusion according to claim 1, characterized in that, S201 specifically is: In the case of boundary insertion, the edge contour points extracted from multiple imaging perspectives are projected into a unified three-dimensional space coordinate system based on the patch edge pixel coordinates in the multi-perspective image, and the edge contour points from different imaging perspectives are paired according to the spatial position index. For each paired position, the Euclidean distance between any two groups of edge contour points at that position is calculated, and the Euclidean distance data of all paired positions is collected to obtain the average Euclidean distance and the maximum Euclidean distance corresponding to each position. For each paired position, the ratio between the average Euclidean distance and the maximum Euclidean distance is taken as the edge matching mismatch degree, which is used as the basis for evaluating the spatial consistency of the patch edge in the boundary insertion case for subsequent conflict feature extraction.

4. The motor insert sheet defect detection method based on multi-sensor fusion according to claim 1, characterized in that, S202 specifically is: According to the numerical distribution of the edge matching mismatch degree on the patch edge, the spatial positions with edge matching mismatch degree greater than the second set threshold are selected as the edge region, the direction vectors corresponding to each spatial position in the edge region are collected, and the direction vectors are sequentially indexed and sorted according to the spatial arrangement order of the patch edge to form a direction vector sequence. For adjacent direction vectors in the direction vector sequence, the included angle difference between the direction vectors is calculated one by one to form an included angle difference sequence, and the proportion of the number of spatial positions with an included angle difference greater than a preset angle threshold is calculated to obtain the edge jump rate, and the proportion of the number of spatial positions with a near reverse relationship is calculated to obtain the edge inversion ratio. After calculating the edge jump rate and the edge inversion ratio, the spatial index minimum position and the spatial index maximum position in the edge contour are selected as the starting point and the ending point of the edge contour, the spatial distance between the two points is calculated as the spatial closure error, and the spatial closure error is taken as the edge closure deviation.

5. The motor insert sheet defect detection method based on multi-sensor fusion according to claim 1, characterized in that, S3 specifically is: The edge matching mismatch degree, the edge jump rate, the edge inversion ratio and the edge closure deviation corresponding to each spatial position in the geometric conflict feature set of the patch edge are feature coded, and a unified format input vector is constructed, which is input into the fusion confidence evaluation function, which is a nonlinear feature mapping function based on multi-feature correlation analysis. In the fusion confidence evaluation function, the feature importance of each feature field in the input vector is calculated respectively, and the fusion confidence evaluation factor is obtained through a nonlinear fitting model. The fusion confidence evaluation factor is output in numerical form, which is used to measure the stability of the patch edge fusion. A first confidence risk threshold is set based on the numerical range of the fusion confidence evaluation factor, and the fusion confidence evaluation factor of each spatial position of the patch edge is compared and judged. When the fusion confidence evaluation factor is lower than the first confidence risk threshold, it is identified that the patch edge has a fusion distortion risk.

6. The motor insert sheet defect detection method based on multi-sensor fusion according to claim 1, characterized in that, S4 specifically is: According to the numerical distribution of the fusion confidence evaluation factor, the complete numerical range of the fusion confidence evaluation factor is divided into multiple non-overlapping sections, and an independent fusion confidence level label is set for each numerical section, which is used to divide the fusion confidence level of each spatial position of the patch edge. According to the numerical value of the fusion confidence evaluation factor corresponding to each spatial position of the patch edge, the fusion confidence level label of the corresponding spatial position is determined according to the grade division interval, and a fusion confidence level labeling map is generated to label the fusion confidence level of the patch edge in space. Based on the fusion confidence level labeling map, the image acquisition path is screened, and the image acquisition path with a fusion confidence level higher than the first screening reference value is included in the image range participating in fusion, and the image acquisition path with a fusion confidence level lower than the first screening reference value is excluded from the image range participating in fusion, thereby limiting the participation condition of the edge image in fusion to control the image fusion quality.

7. The motor insert sheet defect detection method based on multi-sensor fusion according to claim 1, characterized in that, S5 specifically is: According to the detection time sequence, the fusion confidence evaluation factors corresponding to each spatial position of the patch edge are continuously sampled and recorded, and the fusion confidence evaluation factors are sequentially arranged with time stamp as index to form a time sequence data set of the fusion confidence evaluation factors, which is used to track the time sequence change trend of the fusion confidence evaluation factors. Based on the time sequence data set of the fusion confidence evaluation factors, a corresponding fusion confidence evaluation graph is constructed, and the numerical change amplitude and change density of adjacent time periods in the graph are statistically analyzed to obtain a graph fluctuation parameter reflecting the stability of the fusion state, and the determination threshold of the direction consistency index and the confidence factor division threshold are adjusted according to the graph fluctuation parameter. The adjusted direction consistency index determination threshold and confidence factor division threshold are re-applied to the fusion confidence level division process of the patch edge, and the fusion confidence level of each spatial position is periodically updated, so that the dynamic regulation of the fusion confidence level of the patch edge is realized when the fusion confidence evaluation factor changes in time sequence.

Images