Image recognition-based industrial endoscope defect detection method

Abstract

The present application relates to the technical field of intelligent detection and fault diagnosis of industrial equipment, in particular to an industrial endoscope defect detection method based on image recognition, comprising: acquiring a target object two-dimensional image video stream collected by an image collection probe and a three-dimensional digital twin model; identifying a target object surface defect area; aligning the collection pose with the three-dimensional digital twin model to extract absolute three-dimensional coordinates of the defect area; acquiring an ambient light source angle, generating three-dimensional geometric feature parameters of the defect area using a depth estimation neural network; fusing with a finite element analysis stress cloud atlas to query stress concentration coefficients and principal stress directions; generating a disaster-causing risk index according to the defect trend, and generating a dynamic decision instruction to adjust the running state of the target object, thereby realizing closed-loop linkage between the defect detection result and the equipment operation control.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection and fault diagnosis technology for industrial equipment, specifically to a method for detecting defects in industrial endoscopes based on image recognition. Background Technology

[0002] With the increasing demands for the operation and maintenance of high-reliability equipment such as aero engines, gas turbines, and large pressure components, industrial endoscopes are widely used for defect detection in narrow cavities and complex curved surfaces. This type of detection method can acquire image information of the target object's surface without or with minimal disassembly. How to not only discover defects during the detection process, but also further determine the true location, geometric extent, and impact on equipment operation safety has become an important issue facing the current equipment condition monitoring field. Traditional industrial endoscope defect detection currently relies mainly on the following methods: manually observing two-dimensional images to interpret defects, outputting defect categories and outlines based on image recognition models, and manually classifying defect risks based on historical experience.

[0003] However, methods such as manual observation, image recognition and interpretation, and experience-based grading all have certain shortcomings. For example, manual observation mainly relies on the experience of inspectors and is easily affected by differences in perspective, lighting, and subjective judgment, making it difficult to reliably determine the spatial location and depth of defects. Recognition methods based on two-dimensional images can usually only provide the pixel area of ​​the defect in the image, making it difficult to accurately map it to the absolute three-dimensional coordinates of the target object, which makes it difficult to track the same defect across batches. Experience-based grading methods lack an analysis process combined with the structural stress state, making it impossible to effectively distinguish between ordinary surface damage and dangerous cracks in high-stress areas, and it is also difficult to directly form operational decisions such as shutdown, load reduction, or maintenance scheduling. Summary of the Invention

[0004] The purpose of this invention is to provide an image recognition-based method for detecting defects in industrial endoscopes, addressing the following technical problems: Geometric information usable for mechanical analysis is recovered from two-dimensional images with limited perspectives, and the stress environment of the defect is introduced into the evaluation link to achieve a quantitative correlation between visual inspection results and structural mechanical consequences. This allows for an accurate assessment of the true danger of defects under structural stress conditions, and the risk assessment results are directly transformed into executable operation control and maintenance actions, ultimately achieving a closed-loop linkage between defect detection and equipment operation and maintenance management.

[0005] The objective of this invention can be achieved through the following technical solutions: Image recognition-based methods for detecting defects in industrial endoscopes include: Acquire two-dimensional image and video streams of the target object captured by the image acquisition probe of the endoscope device, as well as a three-dimensional digital twin model of the target object; Image recognition is performed on the two-dimensional image video stream to detect defect areas on the surface of the target object; The acquisition pose of the two-dimensional image video stream is aligned with the three-dimensional digital twin model using feature point matching technology, and the absolute three-dimensional coordinates of the defect region on the three-dimensional digital twin model are extracted. The ambient light angle of the endoscope device where the image acquisition probe is located is obtained, and a pre-trained depth estimation neural network is used to perform morphological deduction on the defect area. Combined with the ambient light angle, the three-dimensional geometric feature parameters of the defect area are generated, wherein the three-dimensional geometric feature parameters include the defect depth value and the three-dimensional edge contour. The absolute three-dimensional coordinates and the three-dimensional geometric feature parameters are fused with a preset finite element analysis stress cloud map library to query the stress concentration factor and principal stress direction corresponding to the absolute three-dimensional coordinates. The defect orientation is determined based on the three-dimensional edge contour. When the defect direction is perpendicular to the principal stress direction, a first disaster risk index is generated based on the stress concentration factor, the preset risk amplification factor, and the defect depth value. When the defect direction is not perpendicular to the principal stress direction, the angle feature between the defect direction and the principal stress direction is extracted, and a second disaster risk index is generated based on the angle feature, the stress concentration coefficient and the preset basic risk weight. Based on the first disaster risk index or the second disaster risk index, a dynamic decision instruction is generated and sent to the control system to adjust the operating state of the target object.

[0006] As a further aspect of the present invention, the acquisition pose of the two-dimensional image video stream is aligned with the three-dimensional digital twin model using feature point matching technology, and the absolute three-dimensional coordinates of the defect region on the three-dimensional digital twin model are extracted, including: Two-dimensional local feature points are extracted from the two-dimensional image video stream, and three-dimensional spatial feature points are extracted from the three-dimensional digital twin model; The matching relationship between the two-dimensional local feature points and the three-dimensional spatial feature points is calculated using a synchronous localization and mapping algorithm, and the real-time pose matrix of the image acquisition probe is solved. Based on the real-time pose matrix and the two-dimensional local feature points, the absolute three-dimensional coordinates of the defect region are calculated through coordinate space mapping.

[0007] As a further aspect of the present invention, a deep estimation neural network is used to perform morphological deduction on the defect region, and combined with the angle of the ambient light source, the three-dimensional geometric feature parameters of the defect region are generated, including: The two-dimensional image containing the defective region is input into a pre-trained depth estimation neural network to extract pixel-level two-dimensional contour features of the defective region. Combining the ambient light angle, the pixel-level two-dimensional contour features are processed by shadow recovery to reconstruct the depth distribution matrix of the defect region; Based on the depth distribution matrix and the pixel-level two-dimensional contour features, the three-dimensional geometric feature parameters of the defect region are generated.

[0008] As a further aspect of the present invention, the absolute three-dimensional coordinates and the three-dimensional geometric feature parameters are fused with a pre-set finite element analysis stress cloud map library, and the stress concentration factor and principal stress direction corresponding to the absolute three-dimensional coordinates are queried, including: The absolute three-dimensional coordinates are mapped to a preset finite element analysis stress cloud library to locate the stress mesh node where the absolute three-dimensional coordinates are located. Extract the stress tensor matrix of the stress mesh nodes under standard operating conditions; The stress tensor matrix is ​​decomposed into eigenvalues ​​to extract the maximum principal stress value. The ratio of the maximum principal stress value to the standard nominal stress is calculated as the stress concentration factor, and the direction of the eigenvector corresponding to the maximum principal stress value is taken as the principal stress direction.

[0009] As a further aspect of the present invention, when the defect orientation is perpendicular to the principal stress direction, the calculation method for generating the first disaster risk index is as follows: based on the stress concentration factor, the preset risk amplification factor, and the defect depth value in the three-dimensional geometric feature parameters. The first disaster risk index is equal to the product of the risk amplification factor, the stress concentration factor, and the ratio of the defect depth value to the preset depth threshold value.

[0010] As a further aspect of the present invention, when the defect orientation is not perpendicular to the principal stress direction, a second disaster risk index is generated based on the stress concentration coefficient and a preset basic risk weight, including: Calculate the cosine of the angle between the defect orientation and the principal stress direction; The second disaster risk index is generated by weighting and multiplying the cosine value of the included angle, the stress concentration factor, and the basic risk weight. The dynamic decision-making instructions generated based on the first or second disaster risk index include: Compare the first disaster risk index or the second disaster risk index with a preset risk threshold; If the first disaster risk index or the second disaster risk index is greater than or equal to the preset risk threshold, an immediate shutdown decision instruction is generated. If either the first disaster risk index or the second disaster risk index is less than a preset risk threshold, a prediction instruction for the remaining safe cycle count is generated.

[0011] As a further aspect of the present invention, generating the remaining safe cycle count prediction instruction includes: Obtain the material fatigue characteristic parameters of the target object from a pre-set material mechanical property database; A crack propagation rate model is constructed based on the stress concentration factor, the defect depth value, and the material fatigue characteristic parameters. The crack propagation rate model is used for iterative calculation to predict the number of cycles required for the defect depth value to reach the preset fracture threshold. The number of cycle periods is used as the remaining safe cycle count. A remaining safe cycle count prediction instruction containing the remaining safe cycle count is generated and output to the monitoring terminal.

[0012] As a further aspect of the present invention, adjusting the operating state of the target object using the dynamic decision-making instruction includes: In response to the immediate shutdown decision command, a shutdown control signal is generated to trigger the emergency power-off relay of the equipment control system to stop the operation of the target object, and a mandatory maintenance work order is generated; In response to the remaining safety cycle prediction command, the remaining safety cycle count is written into the dynamic maintenance schedule of the target object, and a load adjustment control signal is generated to reduce the operating load parameters of the target object, thus forming a closed loop of defect detection and operation control.

[0013] The beneficial effects of this invention are: 1. This invention utilizes feature point matching technology to align two-dimensional image video streams with three-dimensional digital twin models, extracting the absolute three-dimensional coordinates of defect areas. This method overcomes the shortcomings of traditional two-dimensional image recognition in accurately locating defects, achieving absolute spatial positioning of defects on target objects, and providing a unified and stable benchmark for continuous defect tracking and status monitoring across time and batches.

[0014] 2. This invention combines the angle of the ambient light source of the endoscope and uses a depth estimation neural network to perform morphological deduction on the defect area to generate three-dimensional geometric feature parameters containing depth. This method overcomes the limitations of manual observation which is affected by subjective experience and lighting perspective, and realizes the accurate recovery of the three-dimensional depth and geometric contour of the defect from two-dimensional images, transforming visual inspection into objective quantitative evaluation.

[0015] 3. This invention integrates the three-dimensional coordinates and geometric features of defects with the stress cloud map library of finite element analysis, and classifies and calculates the disaster risk index according to the angle relationship between the defect direction and the principal stress direction. This mechanism overcomes the shortcomings of traditional experience-based classification that lacks mechanical support, and quantitatively correlates the visual inspection results with the real structural stress environment, which can accurately distinguish dangerous cracks and ordinary damage in high-stress areas.

[0016] 4. Based on the calculated disaster risk index, this invention generates dynamic decision-making instructions such as immediate shutdown or prediction of remaining safe cycle count, and directly uses them to adjust the operating status of the target object. This process directly transforms the static risk assessment results into executable operation control and maintenance scheduling actions, avoiding the delays and omissions of human decision-making, and realizing the closed-loop linkage between defect detection and equipment operation management. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the steps of the image recognition-based industrial endoscope defect detection method in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 An image recognition-based method for detecting defects in industrial endoscopes includes: acquiring a two-dimensional image and video stream of a target object and a three-dimensional digital twin model of the target object from the image acquisition probe of an endoscope device; performing image recognition on the two-dimensional image and video stream to detect defect areas on the surface of the target object; and using feature point matching technology to align the acquisition pose of the two-dimensional image and video stream with the three-dimensional digital twin model to extract the absolute three-dimensional coordinates of the defect area on the three-dimensional digital twin model. The ambient light angle of the endoscope device where the image acquisition probe is located is obtained. A pre-trained depth estimation neural network is used to perform morphological deduction of the defect area. Combined with the ambient light angle, three-dimensional geometric feature parameters of the defect area are generated, including the defect depth value and three-dimensional edge contour. The absolute three-dimensional coordinates and three-dimensional geometric feature parameters are fused with a pre-set finite element analysis stress cloud map library to query the stress concentration factor and principal stress direction corresponding to the absolute three-dimensional coordinates. The defect direction is determined based on the three-dimensional edge contour; when the defect direction is perpendicular to the principal stress direction, the first disaster risk index is generated based on the stress concentration factor, the preset risk amplification factor and the defect depth value. When the defect direction is not perpendicular to the principal stress direction, the angle feature between the defect direction and the principal stress direction is extracted, and a weighted calculation is performed based on the angle feature, stress concentration coefficient and preset basic risk weight to generate a second disaster risk index; a dynamic decision command is generated based on the first disaster risk index or the second disaster risk index, and the dynamic decision command is sent to the control system to adjust the operating state of the target object using the dynamic decision command.

[0020] This embodiment provides an image recognition-based industrial endoscope defect detection mechanism; specifically, the main scenario is an endoscope inspection of a first-stage high-pressure turbine blade of an aero-engine within a scheduled inspection window without disassembly, and the inspection objects are the blade basin surface, back surface, leading edge and blade root transition area; The system consists of an endoscope host, an image acquisition probe, an edge computing unit, a 3D digital twin model library, a stress cloud image library, and an engine control and maintenance terminal. The image acquisition probe is pre-calibrated with a camera, and its intrinsic parameters, such as focal length, optical center, and distortion coefficient, are stored in the edge computing unit. After the endoscope probe enters the engine flow channel, it continuously outputs a two-dimensional image and video stream, and the edge computing unit synchronously calls the corresponding 3D digital twin model of the blades of this engine model. Furthermore, the two-dimensional image video stream enters the defect recognition module; this module can use a combination of convolutional neural network and segmentation network to output defect candidate regions for each frame of image; for ease of explanation, assume that three frames of images F1, F2 and F3 are acquired consecutively, where there is a suspected crack region R2 in F2; The recognition module outputs the pixel bounding box of the region as the top left corner (120, 80) and the bottom right corner (168, 96), and outputs the category as crack with a confidence level of 0.93. The system matches the local texture features of the current image frame with the pre-annotated geometric features on the surface of the 3D digital twin model to obtain the acquisition pose of the probe relative to the blade model. Taking a simplified simulation as an example, if the probe pose matrix can be represented by looking forward 20mm, rotating 5 degrees relative to the blade axis, and shifting upward 12mm relative to the blade root coordinate system, then R2 can be mapped to a certain absolute three-dimensional position on the digital twin model. The absolute three-dimensional coordinates here represent a fixed position relative to the blade reference coordinate system, so the same defect can be continuously tracked in different inspection batches. After acquiring the spatial location, the system continues to read the ambient light angle of the endoscope device; industrial endoscopes are usually equipped with ring or side supplementary lighting, and the direction of the light source relative to the optical axis of the probe is known; assuming that the current incident angle of the light source relative to the image plane is 35 degrees, the depth estimation network no longer infers the depth of the defect based solely on grayscale differences, but combines the incident angle to recover the morphological relationship between the shadow and the edge. For ease of explanation, if the pixel-level grayscale of the defect region R2 is lower than the preset grayscale difference threshold in the central region, the network outputs a simplified 3×3 depth distribution matrix: the first row is 0.02, 0.08, 0.03, the second row is 0.05, 0.21, 0.07, and the third row is 0.01, 0.06, 0.02. The numerical unit can be set to mm. The matrix center value of 0.21 indicates that the depth value of the center of the region is greater than the depth value of the surrounding region. Based on this, the system extracts the defect depth value of 0.21 mm, the edge contour length of approximately 2.8 mm, the width of approximately 0.4 mm, and the defect main extension direction vector. After obtaining the spatial location and geometry, the system calls the corresponding finite element analysis stress cloud library for the blade. This library is pre-established under standard speed, standard temperature and rated aerodynamic conditions, and the blade surface is divided into multiple stress grid regions. For the grid where coordinate P is located, the system queries the stress concentration factor Kt and the principal stress direction. To continue with a simplified example, if the stress concentration factor in the region corresponding to P is 2.6, the principal stress direction is approximately the same as the radial direction of the blade, and the angle between the main extension direction of the defect and this direction is close to 90 degrees, it means that the crack direction cuts off the principal stress transmission path, which is a high-risk condition. The system then employs the first type of risk calculation logic. For example, if the risk amplification factor is set to 1.8 and the depth threshold is 0.30 mm, the first disaster risk index can be expressed as follows: The result is approximately 3.276; if the defect direction is not perpendicular to the principal stress direction, then the second type of risk calculation logic is applied, for example, the basic risk weight is taken as 0.7, and the risk value is calculated as lower than the first disaster risk index in combination with the angle relationship. The system compares the obtained risk index with a preset threshold and outputs dynamic decision instructions, such as immediately shutting down, limiting thrust operation, shortening the maintenance cycle, or providing a prediction of the remaining safe cycle count. As an anomaly handling mechanism, when the confidence level of a defect in a two-dimensional image is lower than the preset value due to oil stains, fogging, or strong reflection, the system does not directly enter the risk calculation, but triggers the supplementary acquisition process, controls the probe to fine-tune the angle and re-capture; When pose alignment cannot be completed in a certain frame, the pose results of adjacent frames can be used for short-term interpolation. If alignment cannot be completed in multiple consecutive frames, the current defect is marked as pending manual review. When non-physical artifacts such as local negative depth or abrupt changes exceeding the preset gradient appear in the depth estimation result, the system uses morphological smoothing constraints to re-estimate. If the geometric consistency requirement is still not met, only the location-level risk warning is output instead of the precise depth. When no completely corresponding mesh node is found in the stress cloud library, the approximate stress parameters are obtained by weighted interpolation of the three neighboring nodes to avoid interruption of the entire process due to local mesh sparsity. During this scheduled maintenance of the aero-engine, the endoscope probe entered the first-stage high-pressure turbine area and found a fine crack with a contrast lower than the visual recognition threshold in the transition area near the blade root at the leading edge of the blade. The system maps it onto the high-stress area of ​​the digital twin model, recovers the crack depth of 0.21 mm by combining the light source angle, and determines that the crack direction is almost perpendicular to the principal stress direction. Although the crack is less than 3mm long in the two-dimensional image, due to its unfavorable location and stress relationship, the system generates a high-risk index and outputs a dynamic decision instruction to the maintenance terminal to stop the test run and generate a maintenance work order. The purpose of this step is to elevate the observation of defects to an understanding of the true danger of defects under structural stress, thereby achieving a closed-loop linkage between defect detection results and equipment operation control.

[0021] In a preferred embodiment of the present invention, the acquisition pose of the two-dimensional image video stream is aligned with the three-dimensional digital twin model using feature point matching technology, and the absolute three-dimensional coordinates of the defect area on the three-dimensional digital twin model are extracted by: extracting two-dimensional local feature points from the two-dimensional image video stream and extracting three-dimensional spatial feature points from the three-dimensional digital twin model. The matching relationship between two-dimensional local feature points and three-dimensional spatial feature points is calculated using a synchronous positioning and mapping algorithm, and the real-time pose matrix of the image acquisition probe is solved. Based on the real-time pose matrix and two-dimensional local feature points, the absolute three-dimensional coordinates of the defect area are calculated through coordinate space mapping.

[0022] This embodiment provides a spatial alignment step for three-dimensional defect localization; specifically, in the aforementioned aero-engine blade inspection scenario, it is difficult to determine whether the crack is in the blade tip, mid-chord, or blade root transition zone based solely on the pixel position in the two-dimensional image, and these locations have significant mechanical differences. Therefore, this embodiment further introduces a matching process between two-dimensional local feature points and three-dimensional spatial feature points to stably map the image seen by the probe into the digital twin coordinate system of the blade; Furthermore, the system first extracts two-dimensional local feature points from the current frame of the video stream; optionally, corner points, edge intersections, abrupt changes in surface coating texture, or cooling hole contour points are used as two-dimensional features; At the same time, corresponding three-dimensional spatial feature points are extracted from the three-dimensional digital twin model, such as the extreme point of the blade leading edge curvature, the center point of the cooling hole, the inflection point of the platform boundary, and the end point of the blade root dovetail groove. For ease of deduction, assume that the current image frame extracts four two-dimensional feature points p1, p2, p3, and p4, with pixel coordinates of (20,30), (50,32), (22,70), and (52,72), respectively; and that the digital twin model extracts four candidate three-dimensional points b1, b2, b3, and b4, with model coordinates of (10,5,2), (40,5,2), (10,35,2), and (40,35,2), respectively. The synchronous localization and mapping algorithm determines that p1 to p4 correspond to b1 to b4 respectively based on the relative topological relationship and historical frame constraints, thereby calculating the real-time pose matrix of the current probe; this matrix can be understood as a transformation relationship that transforms the image coordinate system to the blade coordinate system, containing rotation and translation information; When the crack area is located at a point in the image When the point is nearby, the system uses the pose matrix to back-project the intersection position of the point on the model surface; Specifically, for example, the system first determines an observation ray corresponding to the pixel based on the probe's intrinsic parameters, then transforms the ray to the blade coordinate system using the pose matrix, and calculates the intersection point of the ray and the digital twin surface; if the intersection point falls on the model's triangular facet T17, the absolute three-dimensional coordinates of that point can be determined as follows: ; If the crack is not a single point but a linear region, the above process can be repeated for multiple pixels on its boundary to form a set of spatial points, and the crack center line, maximum depth location and endpoint location can be further obtained. As an anomaly handling mechanism, the previous solution may have a bottleneck under extreme conditions: when the blade surface has too little local texture due to carbon buildup, ablation or specular reflection, the number of feature points in a single frame may be insufficient, affecting the stability of pose solution; therefore, this embodiment introduces time continuity constraints in pose solution. If the number of available feature points in the current frame is less than a preset threshold, such as less than 3 pairs, the system will not calculate the new pose separately, but will refer to the pose of the previous frame and combine it with the short-term motion model of the probe to make a prediction. If enough matching points cannot be obtained within 5 consecutive frames, the current detection area is marked as a low-confidence positioning area, and the operator is reminded to adjust the probe posture or increase supplementary lighting. If there are multiple solutions for two-dimensional and three-dimensional feature points, the solution with the smallest reprojection error and continuous with the motion direction of the previous moment is retained first to reduce coordinate jumps caused by mismatches. During this blade inspection, the probe moved from the blade tip to the blade root. Due to the presence of cooling holes and edge chamfers at the blade root platform, the system extracted a stable set of feature points in this area and matched them with the hole center and boundary inflection point in the model. The crack area was accurately located approximately 15.6 mm above the blade root transition zone; this location is the basis for subsequent queries of the high-stress grid and is also a key spatial anchor point for distinguishing between surface scratches and dangerous cracks. The purpose of this step is to provide a unified absolute spatial benchmark for subsequent morphology estimation and risk assessment, thereby enabling the traceable location of defects across time, batches, and operating conditions.

[0023] In a preferred embodiment of the present invention, the method of using a depth estimation neural network to perform morphological deduction of the defect region and generating three-dimensional geometric feature parameters of the defect region by combining the angle of the ambient light source includes: inputting a two-dimensional image containing the defect region into a pre-trained depth estimation neural network to extract pixel-level two-dimensional contour features of the defect region. By combining the ambient light angle, pixel-level two-dimensional contour features are processed to recover shape from shadows, and the depth distribution matrix of the defect area is reconstructed. Based on the depth distribution matrix and pixel-level two-dimensional contour features, three-dimensional geometric feature parameters of the defect region are generated.

[0024] This embodiment provides a step for deducing the morphology of a three-dimensional defect from a two-dimensional defect image; specifically, after the aforementioned spatial positioning is completed, knowing only that the crack is located in the high-stress area at the blade root is not enough to accurately assess the degree of danger, because both shallow scratches and deep cracks may appear as thin, dark lines. To address this technical problem, this embodiment introduces a depth estimation neural network and lighting constraints to perform pseudo-3D reconstruction of the defective region; Furthermore, the system first extracts a local image patch containing the crack, for example, with a size of 64×64 pixels, and inputs it into a pre-trained depth estimation network. The pre-trained depth estimation neural network is trained using a set of industrial defect sample images containing various known depth distributions and elevation ground truth labels synchronously acquired by a 3D laser scanning device through supervised learning. The network outputs pixel-level two-dimensional contour features, which include both the crack boundary and the direction of the central depression. For a more intuitive explanation, the network output can be simplified into a 5×5 binary profile map, where a value of 1 represents a candidate crack boundary or crack centerline, and a value of 0 represents the background. Assuming the middle column and adjacent pixels are marked as 1, the system can initially identify the defect as a vertically extended shape; the system reads the current light source angle of the endoscope, for example, the main supplementary light direction is incident from the upper left at 35 degrees; Since the lower right side of the crack is more likely to form a shadow when the light shines from the upper left, the grayscale distribution and the true depth are not a simple linear mapping relationship, but are constrained by the direction of the light. Based on this, the system performs shape recovery processing from the shadow, corrects the grayscale gradient around the pixel-level contour, and obtains the depth distribution matrix. Specifically, the shadow recovery shape processing establishes the correspondence between observed grayscale and surface geometric properties by constructing a simplified Lambertian reflection model, that is, the observed grayscale of each pixel is represented as the inner product of the light source direction vector and the surface normal vector corresponding to that pixel; Based on the known 35-degree ambient light angle, the system calculates the surface normal matrix that minimizes the error between the reconstructed grayscale and the actual observed grayscale. The surface normal matrix is ​​then path-integrated along the gradient direction of the two-dimensional contour features to smoothly convert the normal vector into a continuous absolute depth increment, thereby obtaining a physically corresponding depth distribution matrix. For ease of deduction, if the reconstructed depth distribution matrix shows a maximum indentation value of 0.21 mm in the central region, and 0.16 mm, 0.11 mm and 0.05 mm in the adjacent regions, it indicates that the defect cross-section is approximately V-shaped, rather than a shallow and wide U-shaped scratch. Based on the depth distribution matrix and two-dimensional contour features, the system further generates three-dimensional geometric feature parameters, including maximum depth, average depth, three-dimensional edge contour, crack length, opening width, and main extension direction. The three-dimensional edge contour can be represented as a set of spatial boundary points arranged in sequence, such as E1, E2, E3, and E4 forming a slender groove boundary, which can be used for defect path analysis and fatigue propagation simulation. Specifically, the system uses principal component analysis or least squares method to linearly fit the spatial boundary point set of the three-dimensional edge contour, and extracts the direction vector of the fitted center line as the defect direction. As an anomaly handling mechanism, although the previous-level spatial positioning scheme can indicate where the defect is, there is a probability that directly inferring the depth based on grayscale on a highly reflective metal surface will misjudge the reflective bright area as a protrusion or the dark area as a deep crack. Therefore, this embodiment adds an illumination consistency check before restoring the shape from the shadow. If the proportion of saturated pixels in the bright areas of the current image exceeds a preset threshold, such as exceeding 15% of the local area, then highlight removal processing will be performed first; if the ambient light angle sensor data is missing, the default optical axis angle recorded during the device calibration stage will be used for approximate recovery, and the uncertainty metric level of the depth result will be reduced simultaneously. If the difference between adjacent pixels in the depth matrix changes too abruptly, for example, if the difference between adjacent points exceeds 0.5 mm and does not meet the known blade surface curvature constraint, then the region will be smoothed and corrected, and a second acquisition will be triggered if necessary. If a stable depth value still cannot be output, retain the two-dimensional contour and position results, and mark the depth value as pending verification to prevent erroneous depths from entering the risk calculation chain. In this leaf inspection, the crack located in the leaf root transition zone appeared as only a dark line about 3mm long in the two-dimensional image; After processing in this embodiment, the system recovers that the cross-section is approximately a sharp V-shape with a maximum depth of about 0.21 mm, and the edge contour is concentrated and stable. This indicates that the defect is closer to an early fatigue crack than a common scratch, providing key geometric basis for subsequent stress coupling analysis. The purpose of this step is to recover geometric information that can be used for mechanical analysis from two-dimensional images with limited perspectives, thereby achieving the conversion of defect severity from the visual level to the structural level.

[0025] In a preferred embodiment of the present invention, the absolute three-dimensional coordinates and three-dimensional geometric feature parameters are fused with a preset finite element analysis stress cloud map library. The query for the stress concentration factor and principal stress direction corresponding to the absolute three-dimensional coordinates includes: mapping the absolute three-dimensional coordinates to the preset finite element analysis stress cloud map library and locating the stress mesh node where the absolute three-dimensional coordinates are located. Extract the stress tensor matrix of the stress mesh nodes under standard operating conditions; perform eigenvalue decomposition on the stress tensor matrix to extract the maximum principal stress value, calculate the ratio of the maximum principal stress value to the standard nominal stress as the stress concentration factor, and take the eigenvector direction corresponding to the maximum principal stress value as the principal stress direction.

[0026] This embodiment provides a step for fusing visual results with mechanical field information; specifically, if the spatial location and three-dimensional geometric features of the crack are known, and the judgment is still made only based on the length and depth, the stress difference in the crack area will be ignored. For example, the impact of a dent with a depth dimension less than a preset safety threshold in the middle of the blade on the lifespan is different from that of a crack of the same depth in the blade root transition zone; therefore, in this embodiment, the defect information is mapped to a preset finite element analysis stress cloud library to extract the local stress concentration factor and principal stress direction. Furthermore, the stress cloud library can be pre-generated during the engine design or testing phase. Essentially, it divides the three-dimensional model of the blade into numerous stress mesh nodes and records the stress tensor information of each node under standard operating conditions; when the defect coordinates... After input, the system first searches for grid cells containing that point; For ease of explanation, assume that P is located near mesh node N7. The stress tensor matrix of N7 under standard conditions can be simplified as follows: the first row is 120, 15, 0; the second row is 15, 80, 5; and the third row is 0, 5, 30. The unit can be set to MPa. The system performs eigenvalue decomposition on this matrix to obtain three principal stress values. If the maximum principal stress is 125 MPa and the standard nominal stress is 50 MPa, then the stress concentration factor is... The direction of the eigenvector corresponding to 125MPa is the direction of the most dangerous principal stress at that location. If the angle between this direction and the blade radial direction is only 8 degrees, it can be approximated as the direction of the strongest tension along the blade radial direction. Based on this, the system fits the three-dimensional edge profile of the crack into a principal extension line and calculates its direction vector. If the angle between the crack direction vector and the principal stress direction is close to 90 degrees, it indicates that the crack surface is facing the direction of principal tensile stress propagation and the stress intensity factor at the crack tip is relatively large. If the angle is small, although there is still a risk, the crack propagation rate is lower than the set rate threshold. Through this fusion, cracks in different locations and with different shapes can be converted into comparable risk input parameters, rather than just black lines or bright spots in the image; As an anomaly handling mechanism, the previous-level solution has already provided the three-dimensional position and geometry. However, if the stress cloud library is inconsistent with the current digital twin model version, it may cause position mapping deviation. Therefore, this embodiment performs model version verification before fusion. Only when the geometry version number, material grade, and blade serial number are consistent will the query be performed directly. If the versions are not completely consistent but the geometric differences are within the allowable range, the coordinates are scaled up before querying; if they are completely inconsistent, the result is downgraded to a location hint and no precise stress parameters are output; furthermore, if the defect is located at the intersection of two stress grids, the weighted average of the nearest nodes is used; if a node is missing complete tensor information, the nearest complete node is called to replace it, and the confidence level is adjusted accordingly. In this blade inspection, the system mapped the crack location to the stress cloud map of the blade root transition zone and found that the location had a principal tensile stress greater than the preset stress threshold under standard takeoff conditions, with Kt being approximately 2.5 to 2.6, and the principal stress direction was close to the blade radial direction. Since the main extension direction of the crack was transverse to this direction, the system no longer regarded it as ordinary surface damage, but identified it as a fatigue crack source that could propagate rapidly. The purpose of this step is to incorporate the stress environment of the defect into the evaluation process, thereby achieving a quantitative correlation between visual inspection results and structural mechanical consequences.

[0027] In a preferred embodiment of the present invention, when the defect direction is perpendicular to the principal stress direction, the first disaster risk index is calculated based on the stress concentration factor, the preset risk amplification factor, and the defect depth value in the three-dimensional geometric feature parameters. The calculation method is as follows: the first disaster risk index is equal to the product of the risk amplification factor, the stress concentration factor, and the ratio of the defect depth value to the preset depth critical value.

[0028] This embodiment provides a risk amplification calculation step for high-risk orientation defects; specifically, after the aforementioned fusion analysis, the system has obtained the defect depth value, stress concentration factor, and principal stress direction; While the previous solution can identify that the crack is located in a high-stress zone, if the crack direction cuts off the principal stress transmission path, the crack propagation rate may be underestimated if the unified risk assessment model is still used. Therefore, this embodiment sets a first disaster risk index for the case where the defect direction is perpendicular to the principal stress direction. Furthermore, the system first calculates the angle between the defect orientation vector and the principal stress direction vector; for ease of explanation, if the crack orientation vector is approximated as... The principal stress direction is approximately as follows: If the angle between the two is 90 degrees, they can be determined to be perpendicular. Considering the possibility of errors in actual measurements, in engineering practice, 85 to 95 degrees can be regarded as the vertical range; the system reads the preset risk amplification factor, which is used to reflect the increase in danger when the crack surface faces the principal stress; The risk amplification factor, basic risk weight, and deep critical value are empirical values ​​that are pre-stored in the system configuration file after comprehensive calibration based on the target object's historical failure case database, material yield limit test data, and multinomial nonlinear regression model. Continuing with the aforementioned data, if the risk amplification factor is taken as 1.8, the stress concentration factor Kt as 2.6, the defect depth as 0.21 mm, and the depth threshold as 0.30 mm, then the first disaster risk index is: The higher the value, the greater the likelihood of rapid propagation or induced structural failure under this operating condition. As a comparative example of the effect of depth ratio: if another crack in the same location has a depth of only 0.06 mm, the exponent becomes Two cracks may be similar in length, but their risk values ​​differ significantly due to their depth. By making the depth a ratio of the critical value, the risk threshold settings for different blade specifications, materials, or warning levels can be made more uniform. As an anomaly handling mechanism, if the corner is near the critical boundary, such as 84 degrees or 96 degrees, the system can adopt a hysteresis judgment mechanism to avoid frequent switching of risk type between different models due to small measurement errors; if the defect depth value is close to 0 or cannot be obtained stably due to insufficient acquisition quality, the depth term is not directly set to zero, but conservatively estimated according to the minimum credible depth, or the crack is marked as high-risk in direction but the depth needs to be verified. If the depth threshold is configured abnormally to be 0, the system will immediately stop the calculation of the formula and call the default configuration to prevent division by zero from causing erroneous instructions. During this routine blade inspection, a crack located in the blade root transition zone was determined to be nearly perpendicular to the direction of the principal tensile stress. Based on this, the system activated the first type of risk calculation model and obtained a risk index that was significantly higher than the conventional warning threshold. Therefore, the maintenance terminal marked the crack as the highest priority structural hazard and no longer treated it as a regular scratch. The purpose of this step is to amplify the risk of the most easily propagating crack orientation, thereby enabling sensitive identification of high-stress vertical cracks.

[0029] In a preferred embodiment of the present invention, when the defect direction is not perpendicular to the principal stress direction, the second disaster risk index is generated by calculating based on the stress concentration factor and the preset basic risk weight. This includes: calculating the cosine value of the angle between the defect direction and the principal stress direction; and calculating the second disaster risk index by weighted multiplication of the cosine value of the angle, the stress concentration factor and the basic risk weight. The dynamic decision-making instruction generated based on the first or second disaster risk index includes: comparing the first or second disaster risk index with a preset risk threshold; if the first or second disaster risk index is greater than or equal to the preset risk threshold, generating an immediate shutdown decision instruction; if the first or second disaster risk index is less than the preset risk threshold, generating a remaining safe cycle prediction instruction.

[0030] This embodiment provides a risk calculation and dynamic decision-making process for non-vertical orientation defects; specifically, in addition to the aforementioned high-risk vertical cracks, scratches, cracks and corrosion grooves that are oblique to or nearly in the same direction as the principal stress direction often appear on site. If the amplified model is still applied to such defects, false alarms will be amplified; if the directional relationship is completely ignored, medium-risk defects may be missed. Therefore, this embodiment introduces a second disaster risk index for non-vertical cases and generates dynamic decision instructions based on it. Furthermore, the system calculates the cosine of the angle between the defect direction and the principal stress direction; for ease of explanation, if the defect direction vector is... The principal stress direction vector is If the angle is approximately 60 degrees, the cosine value is approximately 0.5. The system multiplies the cosine value by the stress concentration factor and the basic risk weight to obtain the second disaster risk index; assuming the basic risk weight is 0.7 and the stress concentration factor is 2.0, the index is approximately 0.5 × 2.0 × 0.7 = 0.7. If another defect forms an angle of only 20 degrees with the principal stress direction, the cosine value is approximately 0.94. Under the same Kt conditions, the risk value is higher, reflecting that after the defect deviates from the pure tensile opening mode, it has a stronger coupling relationship with the shear slip mode along the defect direction, i.e., type II crack damage. This weighted design based on cosine values ​​compensates for the inability of the vertical judgment model to capture the transverse shear component in terms of mechanical principles: that is, when the defect is not perpendicular to the principal tensile stress, the larger the absolute value of the cosine of the included angle, the larger the shear projection component of the principal stress on the defect direction. This serves as the mechanical basis for the second disaster risk index, avoiding the omission of high-risk slip-type abrasions in non-vertical states. After completing the calculation of the first or second type of risk, the system enters the dynamic decision-making stage; if the risk index is greater than or equal to the preset risk threshold, for example, the threshold is 1.5, then an immediate shutdown decision instruction is directly generated. If the risk index is lower than the threshold, the system will not be shut down immediately. Instead, a prediction instruction for the remaining safe cycle count will be generated, and the subsequent life model will further estimate the number of cycles that can be safely operated. The dynamism here is reflected in the fact that the same detection system does not output a fixed conclusion, but automatically switches the decision path based on the risk results coupled with position, geometry and orientation. As an anomaly handling mechanism, the aforementioned vertical crack model can capture the most dangerous defects. However, in non-vertical regions, noise in the angle measurement may cause the cosine value to fluctuate. Therefore, this embodiment can perform a moving average on the angle of multiple consecutive frames. For example, if the cosine values ​​of three consecutive frames are 0.48, 0.52, and 0.50, the mean value of 0.50 is used in the calculation to reduce the impact of single-frame fluctuations. If the risk index is close to the threshold, for example, between 1.45 and 1.55, the system can enter the critical value verification strategy, requiring an additional data collection and recalculation at a higher resolution. If both the first and second types of models generate candidate results due to boundary determination, the higher risk value will be used by default to ensure safety first. If the current equipment is in a ground test rather than a pre-flight mission state, the threshold can be dynamically adjusted according to the mission level, but the threshold adjustment process should be fully recorded in the audit log. During the same engine maintenance, in addition to the aforementioned high-risk crack, the system also detected an oblique scratch in the middle of another blade; the scratch was located in the medium stress zone, and its direction was at an angle of about 60 degrees to the principal stress direction. The calculated second disaster risk index was lower than the shutdown threshold. Therefore, the system did not issue an immediate shutdown command, but instead switched to the remaining safe cycle count prediction process to schedule subsequent re-inspection times; in contrast, the vertical crack located at the blade root directly triggered the shutdown decision; both conclusions came from the same set of dynamic decision-making links. The purpose of this step is to implement graded treatment based on the difference between the defect direction and the force relationship, so as to achieve adaptive decision-making of immediate blocking for high-risk defects and continued life assessment for defects below the threshold.

[0031] In a preferred embodiment of the present invention, generating the remaining safe cycle prediction instruction includes: obtaining the material fatigue characteristic parameters of the target object from a preset material mechanical property database; and constructing a crack propagation rate model based on the stress concentration factor, defect depth value, and material fatigue characteristic parameters. The crack propagation rate model is used for iterative calculation to predict the number of cycles required for the defect depth to reach the preset fracture threshold. The number of cycles is used as the remaining safe cycle count to generate a remaining safe cycle count prediction instruction, which is then output to the monitoring terminal.

[0032] This embodiment provides a step for predicting the remaining safe cycle count; specifically, in the previous embodiment, although some defects did not meet the conditions for immediate shutdown, they still could not be simply released. Without a life prediction component, maintenance personnel cannot know how many more takeoff and landing cycles the defect can withstand, making it difficult to develop accurate and reliable maintenance plans. To solve this technical problem, this embodiment introduces material fatigue characteristic parameters and crack propagation rate models to iteratively estimate the remaining life of the defect. Furthermore, the system first obtains the fatigue characteristic parameters of the current blade material, such as material constants C and m, threshold stress intensity factor range, fracture threshold, and environmental correction factor; the above fatigue characteristic parameters are obtained by parsing the material attribute tags bound to the three-dimensional digital twin model and retrieving and matching them from the system's pre-set material mechanical property database. Then, the current defect depth, stress concentration factor, and material fatigue parameters are input into the crack propagation rate model; specifically, the system uses material constants to construct a discretized propagation rate relationship: the depth increment for each cycle:

[0033] in, This represents the depth increment for each loop. and These are the material fatigue characteristic constants. The stress concentration factor is... Given the current defect depth, For standard operating nominal stress, Pi; Taking simplified iteration as an example: Assume the current defect depth mm, stress concentration factor The fracture threshold was set to 0.30 mm, and the depth increment was calculated in the first flight cycle based on the above formula. If the depth is approximately 0.002mm, then after one cycle the depth will be updated to 0.102mm; With crack depth As the value increases, the tip stress intensity factor also increases, and subsequent cycles... The depth increment may increase to 0.003mm; the system then increments the depth increment sequentially and counts in a step-by-step loop until the depth is reached. Continue until 0.30mm is reached; if a total of 80 iterations are required, then 80 is recorded as the remaining safe number of iterations. The number of cycles here corresponds to a standard operating cycle of the target object, such as a single start-high load operation-shutdown process of an aircraft engine, or a single complete load return process of a mechanical component. To better suit engineering applications, the system outputs not only a numerical value, but also a predicted confidence interval and recommended re-examination nodes; For example, if the model gives 80 remaining safety cycles, and the conservative lower bound after considering measurement errors and operating condition fluctuations is 60, the terminal can display a suggestion to complete the re-inspection within 50 cycles and the replacement before 80 cycles; this output is more convenient than simply low risk in providing a basis for accurate equipment operation and maintenance technical decisions. As an anomaly handling mechanism, if material fatigue parameters are missing, for example, if the blade has been overhauled and the batch record is incomplete, the system can call the standard parameters of the same type and process material as a substitute and mark the result as an estimate. If the crack depth is close to the fracture threshold, for example, more than 90% of the threshold, then there is no need to continue iterative prediction. The remaining safe number of cycles can be set to 0 and it is recommended to stop the system. If the crack growth is abnormally negative or remains unchanged for multiple consecutive cycles during the iteration process, it indicates that there may be an error in the parameter input or the model is not suitable. The system should terminate the prediction and prompt manual review. Furthermore, to prevent the risk from being underestimated in a few abnormal operating conditions, the system can provide a conservative correction value when outputting the number of cycles, such as multiplying the original value by a safety reduction factor. During this scheduled engine maintenance, an oblique scratch on the middle of another blade was assessed as not meeting the conditions for immediate shutdown. The system read the fatigue parameters of the nickel-based superalloy of the blade, combined with the current stress concentration factor and the defect depth of 0.10 mm, and iteratively determined that the remaining safe cycle count was approximately 80 cycles. It then output a predictive instruction to re-inspect within 50 cycles and replace the blade before 80 cycles. Based on this, maintenance personnel do not need to immediately replace the blade, but can reasonably schedule a re-inspection window in subsequent flight plans. The purpose of this step is to move beyond a static assessment of whether a low- or medium-risk defect is dangerous, and further advance to dynamic lifespan management, which considers how long the product can still operate safely.

[0034] In a preferred embodiment of the present invention, adjusting the operating state of the target object using dynamic decision instructions includes: in response to an immediate shutdown decision instruction, generating a shutdown control signal to trigger an emergency power-off relay of the equipment control system to stop the operation of the target object, and generating a mandatory maintenance work order; In response to the remaining safe cycle count prediction command, the remaining safe cycle count is written into the dynamic maintenance schedule of the target object, and a load adjustment control signal is generated to reduce the operating load parameters of the target object, thus forming a closed loop of defect detection and operation control.

[0035] This embodiment provides an execution step for applying the detection results in a closed loop to the equipment's operating status. Specifically, the aforementioned embodiments have completed defect identification, location, morphology reconstruction, stress fusion, risk assessment, and life prediction. However, if the results only remain at the screen display level, manual secondary judgment and transcription are still required, which can easily cause processing delays. Therefore, this embodiment directly links dynamic decision-making instructions to the equipment control system and maintenance scheduling system, forming a closed loop from defect detection to adjustment of operating status; Furthermore, when the system receives an immediate shutdown decision command, the edge computing unit outputs a shutdown control signal to the engine test control system; this signal can be sent to the emergency power-off relay or shutdown control channel via the safety isolation module to stop the current test or prevent the engine from entering the next operating stage; Meanwhile, the monitoring terminal automatically generates mandatory maintenance work orders, which record the defect location, image number, risk index, recommended measures and responsible team to avoid omissions from manual entry; if the received instruction is a prediction of the remaining safe cycle count, the system writes the remaining safe cycle count into the engine's dynamic maintenance schedule and generates a load adjustment control signal. For example, the maximum allowable thrust level can be lowered, the maximum duration of a single continuous run can be shortened, or the date of the next endoscopic re-examination can be brought forward. As a specific example of the closed-loop process, two different processing paths can be set: First, the risk index of the aforementioned high-risk blade root crack is calculated to be 3.276, which exceeds the threshold of 1.5. The system immediately sends a shutdown signal to the test control console and automatically generates a mandatory maintenance work order to replace the first-stage high-pressure turbine blade. Secondly, the remaining safe number of cycles for the diagonal scratch in the middle is 80. The system will write the re-inspection within 50 cycles into the maintenance schedule and reduce the load limit of the engine in the subsequent adjustment phase by 5%. In this way, the same system can handle high-risk situations such as sudden shutdowns and manage controllable progressive defects. As an abnormal handling mechanism, if the control system is temporarily offline or the communication link is interrupted, the system should not discard the decision results, but write the shutdown or load reduction command into the pending execution queue and issue multi-channel alarms through the monitoring terminal, duty SMS or audible and visual alarms. If no confirmation is received from the equipment side after the emergency stop command is sent, it will be sent again within the preset timeout period and upgraded to a manual intervention alarm; if the maintenance schedule is currently locked by other maintenance tasks, the system can first write to a temporary schedule cache, and automatically synchronize after the lock is released to ensure that the life prediction results are not missed due to system mutual exclusion. At the end of this scheduled maintenance process for the aero-engine, the system took different actions for the two defects: for the high-risk crack at the blade root, it triggered a shutdown and generated a mandatory maintenance work order; for the oblique scratch in the middle, it wrote the remaining safe cycle count and automatically adjusted the subsequent operating load. Thus, detection, diagnosis, decision-making and control were integrated into a complete link, rather than independent modules that were cut off from each other. The purpose of this step is to directly translate the risk assessment results into actionable operational controls and maintenance actions, thereby achieving a closed-loop linkage between defect detection and equipment operation and maintenance management.

[0036] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for detecting defects in industrial endoscopes based on image recognition, characterized in that, include: Acquire two-dimensional image and video streams of the target object captured by the image acquisition probe of the endoscope device, as well as a three-dimensional digital twin model of the target object; Image recognition is performed on the two-dimensional image video stream to detect defect areas on the surface of the target object; The acquisition pose of the two-dimensional image video stream is aligned with the three-dimensional digital twin model using feature point matching technology, and the absolute three-dimensional coordinates of the defect region on the three-dimensional digital twin model are extracted. The ambient light angle of the endoscope device where the image acquisition probe is located is obtained, and a pre-trained depth estimation neural network is used to perform morphological deduction on the defect area. Combined with the ambient light angle, the three-dimensional geometric feature parameters of the defect area are generated, wherein the three-dimensional geometric feature parameters include the defect depth value and the three-dimensional edge contour. The absolute three-dimensional coordinates and the three-dimensional geometric feature parameters are fused with a preset finite element analysis stress cloud map library to query the stress concentration factor and principal stress direction corresponding to the absolute three-dimensional coordinates. The defect orientation is determined based on the three-dimensional edge contour. When the defect direction is perpendicular to the principal stress direction, a first disaster risk index is generated based on the stress concentration factor, the preset risk amplification factor, and the defect depth value. When the defect direction is not perpendicular to the principal stress direction, the angle feature between the defect direction and the principal stress direction is extracted, and a second disaster risk index is generated based on the angle feature, the stress concentration coefficient and the preset basic risk weight. Based on the first disaster risk index or the second disaster risk index, a dynamic decision instruction is generated and sent to the control system to adjust the operating state of the target object.

2. The method for detecting defects in industrial endoscopes based on image recognition according to claim 1, characterized in that, Aligning the acquisition pose of the two-dimensional image / video stream with the three-dimensional digital twin model using feature point matching technology, and extracting the absolute three-dimensional coordinates of the defect region on the three-dimensional digital twin model includes: Two-dimensional local feature points are extracted from the two-dimensional image video stream, and three-dimensional spatial feature points are extracted from the three-dimensional digital twin model; The matching relationship between the two-dimensional local feature points and the three-dimensional spatial feature points is calculated using a synchronous localization and mapping algorithm, and the real-time pose matrix of the image acquisition probe is solved. Based on the real-time pose matrix and the two-dimensional local feature points, the absolute three-dimensional coordinates of the defect region are calculated through coordinate space mapping.

3. The method for detecting defects in industrial endoscopes based on image recognition according to claim 2, characterized in that, A depth estimation neural network is used to perform morphological deduction on the defect region, and combined with the angle of the ambient light source, the three-dimensional geometric feature parameters of the defect region are generated, including: The two-dimensional image containing the defective region is input into a pre-trained depth estimation neural network to extract pixel-level two-dimensional contour features of the defective region. Combining the ambient light angle, the pixel-level two-dimensional contour features are processed by shadow recovery to reconstruct the depth distribution matrix of the defect region; Based on the depth distribution matrix and the pixel-level two-dimensional contour features, the three-dimensional geometric feature parameters of the defect region are generated.

4. The method for detecting defects in industrial endoscopes based on image recognition according to claim 3, characterized in that, The absolute three-dimensional coordinates and the three-dimensional geometric feature parameters are fused with a pre-set finite element analysis stress cloud map library. The stress concentration factor and principal stress direction corresponding to the absolute three-dimensional coordinates are then queried, including: The absolute three-dimensional coordinates are mapped to a preset finite element analysis stress cloud library to locate the stress mesh node where the absolute three-dimensional coordinates are located. Extract the stress tensor matrix of the stress mesh nodes under standard operating conditions; The stress tensor matrix is ​​decomposed into eigenvalues ​​to extract the maximum principal stress value. The ratio of the maximum principal stress value to the standard nominal stress is calculated as the stress concentration factor, and the direction of the eigenvector corresponding to the maximum principal stress value is taken as the principal stress direction.

5. The method for detecting defects in industrial endoscopes based on image recognition according to claim 4, characterized in that, When the defect orientation is perpendicular to the principal stress direction, the first disaster risk index is calculated based on the stress concentration factor, the preset risk amplification factor, and the defect depth value in the three-dimensional geometric feature parameters. The calculation method is as follows: The first disaster risk index is equal to the product of the risk amplification factor, the stress concentration factor, and the ratio of the defect depth value to the preset depth threshold value.

6. The method for detecting defects in industrial endoscopes based on image recognition according to claim 5, characterized in that, When the defect orientation is not perpendicular to the principal stress direction, a second disaster risk index is generated based on the stress concentration coefficient and a preset basic risk weight, including: Calculate the cosine of the angle between the defect orientation and the principal stress direction; The second disaster risk index is generated by weighting and multiplying the cosine value of the included angle, the stress concentration factor, and the basic risk weight. The dynamic decision-making instructions generated based on the first or second disaster risk index include: Compare the first disaster risk index or the second disaster risk index with a preset risk threshold; If the first disaster risk index or the second disaster risk index is greater than or equal to the preset risk threshold, an immediate shutdown decision instruction is generated. If either the first disaster risk index or the second disaster risk index is less than a preset risk threshold, a prediction instruction for the remaining safe cycle count is generated.

7. The method for detecting defects in industrial endoscopes based on image recognition according to claim 6, characterized in that, The instruction for generating the remaining safe cycle count prediction includes: Obtain the material fatigue characteristic parameters of the target object from a pre-set material mechanical property database; A crack propagation rate model is constructed based on the stress concentration factor, the defect depth value, and the material fatigue characteristic parameters. The crack propagation rate model is used for iterative calculation to predict the number of cycles required for the defect depth value to reach the preset fracture threshold. The number of cycle periods is used as the remaining safe cycle count. A remaining safe cycle count prediction instruction containing the remaining safe cycle count is generated and output to the monitoring terminal.

8. The method for detecting defects in industrial endoscopes based on image recognition according to claim 7, characterized in that, Adjusting the operating state of the target object using the dynamic decision-making instructions includes: In response to the immediate shutdown decision command, a shutdown control signal is generated to trigger the emergency power-off relay of the equipment control system to stop the operation of the target object, and a mandatory maintenance work order is generated; In response to the remaining safety cycle prediction command, the remaining safety cycle count is written into the dynamic maintenance schedule of the target object, and a load adjustment control signal is generated to reduce the operating load parameters of the target object, thus forming a closed loop of defect detection and operation control.

Images