A method for detecting surface defects of a main rotor hub of a helicopter

Abstract

The application discloses a kind of helicopter main hub surface defect detection methods, first obtain the helicopter main hub surface image photographed under controlled illumination condition and carry out pretreatment;Then based on the priori of main hub structure, the key detection area in the surface image after pretreatment is determined, and the corresponding area boundary or area mask is obtained;Again, the pretreated main hub surface image and its key detection area are adaptively segmented, multi-layer visual feature is extracted, dimension is aligned, saliency analysis and defect candidate generation are carried out, and token level saliency map, suspicious token sequence, abnormal probability and defect candidate area set are obtained;Then call multimodal large language model to generate inference output, and classify and finely segment the defect candidate area set, output structured detection result;Finally, the severity score is calculated and compared with the preset score threshold, to determine the disposal result.The application can improve the detection rate and stability of the small defects of main hub.

Description

Technical Field

[0001] This invention relates to the fields of industrial visual inspection and aviation equipment quality inspection, and in particular to a method for detecting surface defects in helicopter main rotor hubs. Background Technology

[0002] The helicopter main rotor hub is a typical critical load-bearing component in flight. Its complex geometry, high precision machining, and stringent surface quality requirements make it susceptible to surface defects such as microcrack propagation, corrosion pitting, coating damage, and burrs and scratches at orifices under long-term service environments (vibration loads, temperature and humidity changes, salt spray corrosion, etc.). These defects are often small in size, intricate in shape, and have low contrast, and are easily distributed in irregular areas such as curved surfaces, rounded corners, and orifice edges. Missing or misjudging these defects can lead to significant safety risks and maintenance costs. Therefore, establishing reliable surface defect detection methods for the main rotor hub is of great importance during manufacturing, assembly, warehousing acceptance, and scheduled maintenance.

[0003] Existing methods for detecting surface defects on main propeller hubs typically include manual visual inspection, non-destructive testing (such as penetrant testing, magnetic particle testing, eddy current testing, and ultrasonic testing), and machine vision inspection based on industrial cameras. Manual visual inspection relies on experience and is easily affected by factors such as fatigue, viewing angle, lighting, and workstation posture. It is difficult to establish stable and consistent judgment criteria for minute defects, and the traceability of defect locations is insufficient. While traditional non-destructive testing is sensitive to specific defects, it typically suffers from complex inspection processes, long processing times, high requirements for inspection equipment, sensitivity to certain surface conditions (coatings, oil stains, reflections), and limited ability to provide semantic interpretation of defect types and repair suggestions. Machine vision inspection is widely used in industrial scenarios, but it still faces imaging challenges such as image saturation due to reflections, shadow occlusion, uneven texture, and large scale variations on complex curved surfaces like main propeller hubs, highly reflective metal or coated surfaces, and structural areas such as orifices.

[0004] With the development of deep learning in the field of visual inspection, existing technologies have attempted to classify, detect, or segment defects using models such as convolutional neural networks or visual Transformers. These methods typically require a large number of labeled samples for training. However, the main rotor hub belongs to the aerospace manufacturing scenario with small batches and high reliability requirements, making it difficult to obtain a large number of defect samples and resulting in high labeling costs. Different batches, processes, and surface treatments lead to significant variations in data distribution, and insufficient model generalization ability can easily lead to false positives and false negatives. Furthermore, most existing defect detection methods typically output defect scores, bounding boxes, or masks, but lack interpretable outputs regarding questions such as "why it was judged as a defect," "where the defect evidence is mainly located," "what the possible causes are," and "what processes and repair measures should be taken." This necessitates secondary manual review by quality inspectors, impacting the efficiency of the inspection loop.

[0005] Meanwhile, defects on the main rotor hub surface often exhibit small, sparse, and localized features, requiring high-resolution images to ensure defect discernibility. However, high-resolution input significantly increases the computational load of the visual model and generates a large number of visual tokens in token-based visual models, leading to the following problems: First, a large number of irrelevant background or normal texture tokens dilute defect-related information, making it difficult for the model to focus on the real defect area; second, in scenarios requiring defect explanations and recommendations, directly inputting all visual tokens into the inference model can easily introduce noise and reduce the consistency between the inference text and the defect evidence area, making the output difficult to use for engineering decision-making.

[0006] Furthermore, as a complex structural component, the main rotor hub exhibits distinct structural prior characteristics in its defect distribution: for example, cracks and corrosion are more likely to occur at stress concentration points such as the transition fillets of the support arms and the edges of orifices; spline or keyway areas require more attention to tooth surface scratches, chipping, and wear; and flange end faces and bolt hole circumference areas require more attention to indentations, scratches, and coating defects. Existing general defect detection schemes often fail to fully utilize these structural priors to determine key detection areas and imaging strategies, leading to false positives and false negatives, or low detection efficiency, in complex structural contexts. Therefore, there is an urgent need for a surface defect detection scheme that addresses the structural characteristics of the main rotor hub, effectively highlights defect evidence at high resolution, and outputs interpretable reasoning results and recommendations to improve the reliability, traceability, and closed-loop efficiency of main rotor hub quality inspection. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a method for detecting surface defects in helicopter main rotor hubs. The specific technical solution is as follows:

[0008] A method for detecting surface defects in helicopter main rotor hubs includes:

[0009] S1: Acquire and preprocess a surface image of the helicopter main rotor hub taken under controlled lighting conditions;

[0010] S2: Based on the prior knowledge of the main rotor hub structure, determine the key detection areas in the preprocessed surface image of the main rotor hub, and obtain the corresponding area boundaries or area masks.

[0011] S3: The preprocessed main rotor hub surface image and its key detection areas are sequentially subjected to resolution adaptive segmentation, multi-layer visual feature extraction, dimension alignment, saliency analysis and defect candidate generation to obtain token-level saliency map, suspicious token sequence, anomaly probability and defect candidate area set, and the multimodal large language model is called to generate inference output.

[0012] S4: Classify and finely segment the set of defect candidate regions, output structured detection results, calculate severity scores, and then compare them with preset scoring thresholds to determine the handling results.

[0013] A device for detecting surface defects in helicopter main rotor hubs includes:

[0014] The image preprocessing module is used to preprocess the surface image of the helicopter main rotor hub captured under controlled lighting conditions.

[0015] The structure prior and region determination module is used to determine the key detection region in the preprocessed surface image of the main rotor hub based on the main rotor hub structure prior.

[0016] The segmentation module is used to perform resolution-adaptive segmentation on the preprocessed main rotor hub surface image and its key detection areas to obtain an image set.

[0017] The visual encoding module is used to extract multi-layer visual features of an image at the output of multiple network layers and to dimensionally align the multi-layer visual features using at least one feature adapter.

[0018] The Anomaly Expert module is used to calculate a token-level saliency map based on the perception description vectors of abnormal objects and normal objects.

[0019] The visual token selection module is used to generate query tokens based on the token-level saliency graph, and use the original visual token set as keys and values ​​to obtain suspicious semantic tokens through a query-key-value attention aggregator, forming a suspicious token sequence.

[0020] The scoring and prompting module is used to perform normalized weighted aggregation on suspicious token sequences to obtain a global anomaly representation, and calculate the anomaly probability through a classifier; at the same time, it generates anomaly degree indication information based on the anomaly probability and inserts prompt fragments.

[0021] The reasoning generation module receives the original visual token sequence, the suspicious token sequence, and the hint fragment, calls the multimodal large language model, and generates reasoning output.

[0022] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements a method for detecting surface defects on the main rotor hub of a helicopter.

[0023] The beneficial effects of this invention are as follows:

[0024] 1. Improve the detection rate and stability of minor defects in the main propeller hub: This invention improves the distinguishability and robustness of defects such as cracks, corrosion points, scratches, burrs at the orifice, and coating peeling by using controlled illumination imaging, resolution adaptive segmentation and multi-layer feature extraction, taking into account both the overall structural semantics and the details of local defects.

[0025] 2. Reduce redundant information interference under high resolution conditions: By generating a token-level saliency map and generating a sequence of suspicious tokens through saliency-guided token selection and attention aggregation, defect evidence-related information is filtered from a large number of visual tokens, thereby reducing false detections and inference biases caused by irrelevant backgrounds and normal textures.

[0026] 3. Achieving consistent output of defect evidence, defect category, and reasoning suggestions: This invention simultaneously provides anomaly probability, defect category label, and defect evidence area identification information in the output, and constructs a prompt sequence input language model by combining the suspicious token sequence and anomaly degree indication information, so that the generated defect explanation and suggestions can correspond to the evidence area, improving the traceability and verifiability of quality inspection results.

[0027] 4. By making full use of the prior knowledge of the main propeller hub structure and performing constrained imaging and detection on key monitoring areas, the detection accuracy and efficiency have been improved.

[0028] 5. Facilitates application throughout the entire manufacturing-assembly-service process: The suggested text output by this invention can be used to guide re-inspection, rework, repair, or process parameter optimization, improve the closed-loop efficiency of main rotor hub quality inspection, reduce maintenance costs, and enhance reliability assurance levels. Attached Figure Description

[0029] Figure 1 This is a flowchart of a method for detecting surface defects on the main rotor hub of a helicopter, according to an embodiment of the present invention.

[0030] Figure 2 This is a flowchart illustrating the resolution-adaptive segmentation, saliency analysis, and defect candidate generation for key detection areas in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram illustrating the process of calculating the token-level saliency graph.

[0032] Figure 4 This is a schematic diagram illustrating the process of calculating the probability of anomalies.

[0033] Figure 5 This is a schematic diagram of a surface defect detection device for helicopter main rotor hub according to an embodiment of the present invention. Detailed Implementation

[0034] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0035] In response to the characteristics of helicopter main rotor hub surface such as high reflectivity, complex curved surface and orifice edge structure, small defect size and diverse shape, this invention receives high-resolution surface images acquired under controlled lighting conditions, and determines key detection areas in the high-resolution surface images based on the prior knowledge of the main rotor hub structure. Within the key detection areas, defect candidates are generated and defect categories and evidence areas are identified. Further, anomaly probability scores and defect severity scores (SEV) are calculated and graded. Finally, the detection results are bound and archived with the evidence areas to achieve a traceable quality inspection closed loop.

[0036] On the one hand, such as Figure 1 As shown in the figure, this embodiment provides a method for detecting surface defects in helicopter main rotor hubs, including the following steps one through four.

[0037] Step 1: Acquire and preprocess a surface image of the helicopter main rotor hub taken under controlled lighting conditions.

[0038] Controlled illumination includes at least one of the following: ring light, coaxial light, oblique incident stripe light, and polarized anti-reflective lighting, to suppress highlights on metal surfaces and highlight the edges of small defects.

[0039] To make defects distinguishable at the pixel level, it is preferable to set the acquisition resolution to meet the condition of "the minimum number of pixels corresponding to the width of the defect of interest" (for example, the crack width corresponds to at least 2 to 3 pixels in the image), thereby avoiding missed detections due to insufficient resolution.

[0040] For curved surfaces such as the edge of the orifice and the rounded corner of the support arm, images are acquired using two different incident directions or two different viewing angles to reduce shadow occlusion and reflective saturation caused by single-direction incident light. Multi-view images can be assigned the same index of "part number-area number-angle number" for easy subsequent fusion and traceability.

[0041] Image preprocessing includes distortion correction, illumination equalization, denoising, and contrast enhancement of the acquired images. Distortion correction can be performed using camera calibration parameters; illumination equalization can be achieved using local histogram equalization or homomorphic filtering; denoising can be achieved using bilateral filtering or nonlocal mean filtering; and contrast enhancement is used to improve the visibility of cracks, scratches, and corrosion points.

[0042] After preprocessing, image quality is checked, and the quality indicators include sharpness, exposure saturation ratio, and occlusion ratio of key detection areas. When the above indicators do not meet the preset thresholds, re-acquisition can be triggered or manual adjustment of the light source angle and exposure parameters can be prompted.

[0043] Step 2: Based on the prior knowledge of the main rotor hub structure, determine the key detection areas in the preprocessed surface image of the main rotor hub, and obtain the corresponding area boundaries or area masks.

[0044] The key inspection areas include at least one or more of the following: the main propeller hub and its flange end face, the bolt hole circumference area, the spline or keyway area, the bearing housing or bore edge, and the support arm transition fillet area.

[0045] When determining key inspection areas, at least one reference structure is used, such as the center of the positioning bolt hole group, the circular boundary of the flange end face, the boundary of the spline or keyway, or the edge of the bearing seat hole, to generate the area boundary or area mask of the key inspection area. The area boundary or area mask is then used as the area constraint for subsequent segmentation, saliency calculation, defect candidate generation, and defect identification.

[0046] Step 3: Perform resolution adaptive segmentation, multi-layer visual feature extraction, dimension alignment, saliency analysis, and defect candidate generation on the preprocessed main rotor hub surface image and its key detection areas in sequence to obtain a token-level saliency map, a suspicious token sequence, anomaly probability, and a set of defect candidate areas. Then, call a multimodal large language model to generate inference output.

[0047] Step three includes the following sub-steps:

[0048] S301: The preprocessed image of the main propeller hub surface and its key detection areas are adaptively segmented to obtain an image set; and for each image in the image set, the visual encoder extracts multi-layer visual features at the output of multiple network layers, and performs dimensional alignment of the multi-layer visual features through at least one feature adapter.

[0049] Here, the image set I is represented as:

[0050]

[0051] Where I0 is the scaled original image, I j This is a partial cropped image. The local cropped image includes at least image patches cropped from the key detection areas to preserve local details of defects such as fine cracks, burrs at the orifices, corrosion spots, and coating peeling.

[0052] Multi-layer visual features are represented as follows:

[0053]

[0054] Where i is the network layer index and j is the image index. Let represent the feature extraction function of the visual encoder in the i-th network layer.

[0055] S302: Based on the multi-layer visual features after dimension alignment, calculate the abnormal object perception description vector and the normal object perception description vector, and calculate the saliency weight of each visual token based on the abnormal object perception description vector and the normal object perception description vector; and perform cross-layer averaging or weighted averaging on the saliency weights of multiple network layers to obtain the token-level saliency map.

[0056] The formula for calculating the significance weight is as follows:

[0057]

[0058] In the formula, Represents cosine similarity or normalized inner product. Here, i represents the preset temperature coefficient, i is the network layer index, and j is the visual token index.

[0059] Among them, the j-th visual token of the i-th layer Significance weight Represented as:

[0060]

[0061] in, This represents the abnormal object perception description vector of the i-th layer. This represents the normal object perception description vector of the i-th layer. Represents cosine similarity or normalized inner product. This is the preset temperature coefficient.

[0062] S303: Based on the token-level saliency graph, the visual token selector performs saliency-guided pooling on the visual tokens to generate the query token q. j ; to query token q j As a query, the original visual token set V is used as the key and value. Suspicious semantic tokens are obtained through the query-key-value attention aggregator, forming a suspicious token sequence.

[0063]

[0064] in, Let V be the aggregation function based on the attention mechanism, and let V be the original set of visual tokens.

[0065] S304: Perform normalized weighted aggregation on the suspicious token sequence to obtain the global anomaly representation r(I), and then pass it through a classifier. Calculate the anomaly probability score(I):

[0066]

[0067] S305: Based on the token-level saliency map, the suspicious token sequence and its spatial mapping relationship in the original image, generate a set of defect candidate regions; and based on the anomaly probability score(I), generate anomaly degree indication information; and construct a prompt sequence by combining the original visual token sequence, the suspicious token sequence and the anomaly degree indication information, and input it into the multimodal large language model to generate inference output.

[0068] Specifically, tokens with a saliency weight greater than a preset threshold are mapped back to their corresponding spatial locations in the original image or a locally cropped image, forming initial defect candidate regions. These initial candidate regions are then filtered based on at least one constraint among connectivity, area, aspect ratio, elongation, and boundary regularity to obtain a set of defect candidate regions. For example, for crack-type defects, elongated candidates are prioritized; for corrosion point-type defects, multi-point candidates are prioritized; and for coating peeling, candidates with irregular boundaries and larger areas are prioritized.

[0069] The anomaly level indication information may include at least one of highly suspicious, moderately suspicious, or slightly suspicious.

[0070] The inference output includes at least the anomaly probability score(I), defect category label, defect evidence area identification information, and recommended text for main rotor hub machining, assembly, and service.

[0071] Step 4: Classify and finely segment the set of defect candidate regions, output structured detection results, calculate the severity score, and then compare it with the preset score threshold to determine the handling result.

[0072] Step four includes the following sub-steps:

[0073] S401: Receive the defect candidate region set, suspicious token sequence, token-level saliency map, anomaly probability score(I) generated in step three, and the inference output generated by the multimodal large language model.

[0074] Among them, the defect candidate region set is used to limit the spatial range of subsequent classification and segmentation; the suspicious token sequence and token-level saliency map are used to characterize visual evidence related to defects; the anomaly probability score(I) is used to characterize the probability that there are defects in the current image or the current key detection region; and the inference output is used to provide semantic constraints for defect category, cause inference and suggestion text.

[0075] S402: Classify each candidate region in the defect candidate region set to obtain the defect category label and confidence level.

[0076] The defect category label must include at least one of the following: cracks, scratches, corrosion spots, burrs at the orifice, and coating peeling. Different category judgment thresholds can be used for different key inspection areas; for example, the sensitivity to orifice burrs and cracks can be increased in the orifice edge area, the sensitivity to tooth surface scratches and chipping defects can be increased in the spline or keyway area, and the sensitivity to indentations, scratches, and coating defects can be increased in the flange end face area.

[0077] S403: Perform fine segmentation on the defect candidate region to obtain the defect evidence region.

[0078] The defect evidence region may include bounding box coordinates, a pixel-level mask, or a combination of bounding box coordinates and a pixel-level mask. Furthermore, at least one structured parameter among the following can be calculated based on the pixel-level mask: defect area A, defect perimeter P, crack length L, defect center point coordinates, and defect contrast C.

[0079] S404: Combining the suspicious token sequence, token-level saliency map, anomaly probability score(I), and the inference output of the multimodal large language model, perform consistency verification on the defect category and defect evidence region: If the defect candidate region is consistent with the highly saliency region in the token-level saliency map, and the anomaly probability score(I) is greater than the preset threshold, then the defect candidate region is retained as a valid defect evidence region; if the defect candidate region is only caused by a single perspective or local highlights, and is inconsistent with the saliency map or inference output, then its confidence is reduced or it is marked as a region to be reviewed.

[0080] S405: Output structured detection results, calculate severity score based on structured detection results, and then compare the severity score with a preset scoring threshold to determine the treatment result.

[0081] The structured inspection results should include at least the defect category label, confidence level, defect evidence area, anomaly probability score (I), defect area A, defect contrast C, and suggestion text.

[0082] The defect category label must include at least one of the following: cracks, scratches, corrosion spots, burrs at the orifice, and coating peeling.

[0083] The structured detection results serve as input for subsequent severity scoring, threshold determination, treatment output, and traceability archiving.

[0084] The severity score is related at least to the defect confidence level p and the defect size, and can be further combined with indicators such as contrast and texture anomaly. This embodiment uses the following scoring calculation formula:

[0085]

[0086] Where p is the defect identification confidence level, with a value range of [0,1]; w1, w2, and w3 are preset weights, satisfying w1+w2+w3=1, or set according to quality standards (e.g., increasing the proportion of w2 and w3 for crack defects to highlight the influence of scale and contrast).

[0087] A norm To normalize the defect area, A norm = A / A ROI A is the pixel area of ​​the defect mask. ROI A represents the pixel area of ​​the corresponding detection region; when a bounding box is used instead of a mask, A can be the area of ​​the bounding box.

[0088] C norm To normalize the contrast, C norm = (C - C min ) / (C max - C min C represents the difference in grayscale mean, gradient energy, or texture contrast between the defect region and its neighboring background. min C max This is a preset boundary used to map the contrast to [0,1].

[0089] Optionally, for crack-type defects, a length term L is added. norm And adjust the weights; for example, L norm =L / L ROI Where L is the crack length or the longest connected path length after skeletonization, L ROI This represents the characteristic scale of the detection area. For corrosion point defects, a point density D can be added. norm And adjust the weights; for example, D norm =N / A ROI , where N is the number of erosion points or the number of connected components.

[0090] In this embodiment, two scoring thresholds, T1 and T2, are set, and the judgment rule is as follows:

[0091] If sev < T1, it is considered qualified;

[0092] When T1 ≤ sev < T2, it is determined that a review is required;

[0093] When sev ≥ T2, it is judged as unqualified or needs to be reworked.

[0094] The T1 and T2 thresholds are set based on the main rotor hub quality level, defect type risk level, and historical statistical data. For example, lower T1 and T2 thresholds are set for crack-like defects to increase conservatism; higher thresholds are set for minor coating peeling to reduce unnecessary rework. The system outputs a handling conclusion and a suggestion text, which must be related to the defect category and the ROI. For example, at the edge of the orifice, it may suggest "Re-inspect the orifice burrs and perform deburring treatment," and at the transition fillet of the support arm, it may suggest "Re-inspect the suspected crack at the fillet transition and recommend non-destructive re-inspection."

[0095] On the other hand, this embodiment also provides a device for detecting surface defects in helicopter main rotor hubs. This device includes an image preprocessing module, a structure prior and region determination module, a segmentation module, a visual encoding module, an anomaly expert module, a visual token selection module, a scoring and prompt construction module, an inference generation module, and a training control module, wherein:

[0096] The image preprocessing module is used to preprocess the surface image of the helicopter main rotor hub captured under controlled lighting conditions.

[0097] The structure prior and region determination module is used to determine the key detection region in the preprocessed surface image of the main rotor hub based on the main rotor hub structure prior.

[0098] The segmentation module is used to perform resolution-adaptive segmentation on the preprocessed main rotor hub surface image and its key detection areas to obtain an image set.

[0099] The visual encoding module is used to extract multi-layer visual features of an image at the output of multiple network layers and to dimensionally align the multi-layer visual features using at least one feature adapter.

[0100] The Anomaly Expert module is used to calculate a token-level saliency map based on the perception description vectors of abnormal objects and normal objects.

[0101] The visual token (lexical) selection module is used to generate query tokens based on the token-level saliency map, and uses the original visual token set as keys and values ​​to obtain suspicious semantic tokens through the query-key-value attention aggregator, forming a suspicious token sequence;

[0102] The scoring and prompting module is used to perform normalized weighted aggregation on suspicious token sequences to obtain a global anomaly representation, and calculate the anomaly probability through a classifier; at the same time, it generates anomaly degree indication information based on the anomaly probability and inserts prompt fragments.

[0103] The reasoning generation module receives the original visual token sequence, the suspicious token sequence, and the hint fragment, calls the multimodal large language model, and generates reasoning output.

[0104] The training control module is used to perform two-stage training:

[0105] In the first stage, only the parameters of the anomaly expert module, the visual token selection module, and the scoring and prompting construction module are updated to obtain the defect detection capability under the condition that the sample size is less than a preset threshold.

[0106] In the second stage, while retaining the parameters from the first stage, visual instructions are adjusted to enable the reasoning generation module to output an instruction response that includes defect evidence area identification information, defect category description, cause inference, and suggestion text.

[0107] The second stage of visual instruction adjustments only requires minor tweaks.

[0108] In another aspect, embodiments of the present invention also provide a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the method for detecting surface defects of helicopter main rotor hubs as described in the above embodiments.

[0109] The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0110] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for detecting surface defects in helicopter main rotor hubs, characterized in that, include: S1: Acquire and preprocess a surface image of the helicopter main rotor hub taken under controlled lighting conditions; S2: Based on the prior knowledge of the main rotor hub structure, determine the key detection areas in the preprocessed surface image of the main rotor hub, and obtain the corresponding area boundaries or area masks. S3: The preprocessed main rotor hub surface image and its key detection areas are sequentially subjected to resolution adaptive segmentation, multi-layer visual feature extraction, dimension alignment, saliency analysis, and defect candidate generation to obtain a token-level saliency map, a suspicious token sequence, anomaly probability, and a set of defect candidate regions. A multimodal large language model is then invoked to generate inference output. S3 includes the following sub-steps: S301: The preprocessed image of the main propeller hub surface and its key detection areas are adaptively segmented to obtain an image set; and for each image in the image set, the visual encoder extracts multi-layer visual features at the output of multiple network layers, and performs dimensional alignment of the multi-layer visual features through at least one feature adapter. S302: Based on the multi-layer visual features after dimension alignment, calculate the abnormal object perception description vector and the normal object perception description vector, and calculate the salience weight of each visual token based on the abnormal object perception description vector and the normal object perception description vector; and perform cross-layer averaging or weighted averaging on the salience weights of multiple network layers to obtain a token-level salience map. S303: Based on the token-level saliency map, the visual token selector performs saliency-guided pooling on the visual tokens to generate query tokens; using the original set of visual tokens as keys and values, the query-key-value attention aggregator obtains suspicious semantic tokens, forming a suspicious token sequence; S304: Perform normalized weighted aggregation on the suspicious token sequence to obtain a global anomaly representation, and calculate the anomaly probability through a classifier; S305: Based on the token-level saliency map, the suspicious token sequence, and their spatial mapping relationship in the original image, generate a set of defect candidate regions; Based on the anomaly probability, anomaly level indication information is generated; the original visual token sequence, the suspicious token sequence, and the anomaly level indication information are combined to construct a prompt sequence, which is then input into a multimodal large language model to generate inference output; S4: Classify and finely segment the candidate defect region set, output structured detection results, calculate severity scores, and then compare them with preset scoring thresholds to determine the handling result; S4 includes the following sub-steps: S401: Receives the defect candidate region set, suspicious token sequence, token-level saliency map, anomaly probability, and inference output generated by the multimodal large language model generated by S3; S402: Classify each candidate region in the defect candidate region set to obtain the defect category label and confidence level; S403: Perform fine segmentation on the defect candidate region to obtain the defect evidence region; S404: Combining the suspicious token sequence, token-level saliency map, anomaly probability, and inference output of the multimodal large language model, perform consistency verification on the defect category and defect evidence region: If the defect candidate region is consistent with the highly saliency region in the token-level saliency map and the anomaly probability is greater than the preset threshold, then retain the defect candidate region as a valid defect evidence region. If a defect candidate region is caused by a single viewpoint or local highlights and is inconsistent with the token-level saliency map or inference output, its confidence level is reduced or it is marked as a region to be reviewed. S405: Output the structured detection results, calculate the severity score based on the structured detection results, and then compare the severity score with a preset scoring threshold to determine the treatment result.

2. The method for detecting surface defects in helicopter main rotor hubs according to claim 1, characterized in that, The preprocessing includes distortion correction, illumination equalization, noise reduction, and contrast enhancement of the image.

3. The method for detecting surface defects in helicopter main rotor hubs according to claim 1, characterized in that, The key inspection areas include at least one or more of the following: the main propeller hub and its flange end face, the bolt hole circumference area, the spline or keyway area, the bearing seat or orifice edge, and the support arm transition fillet area.

4. The method for detecting surface defects in helicopter main rotor hubs according to claim 1, characterized in that, The inference output includes at least the anomaly probability, defect category label, defect evidence area identification information, and suggested text for main rotor hub machining, assembly, and service.

5. The method for detecting surface defects in helicopter main rotor hubs according to claim 1, characterized in that, The defect evidence region includes bounding box coordinates, pixel-level masks, or a combination of bounding box coordinates and pixel-level masks.

6. A device for detecting surface defects on a helicopter main rotor hub, used to implement the method for detecting surface defects on a helicopter main rotor hub as described in any one of claims 1 to 5, characterized in that, include: The image preprocessing module is used to preprocess the surface image of the helicopter main rotor hub captured under controlled lighting conditions. The structure prior and region determination module is used to determine the key detection region in the preprocessed surface image of the main rotor hub based on the main rotor hub structure prior. The segmentation module is used to perform resolution-adaptive segmentation on the preprocessed main rotor hub surface image and its key detection areas to obtain an image set. The visual encoding module is used to extract multi-layer visual features of an image at the output of multiple network layers and to dimensionally align the multi-layer visual features using at least one feature adapter. The Anomaly Expert module is used to calculate a token-level saliency map based on the perception description vectors of abnormal objects and normal objects. The visual token selection module is used to generate query tokens based on the token-level saliency graph, and use the original visual token set as keys and values ​​to obtain suspicious semantic tokens through a query-key-value attention aggregator, forming a suspicious token sequence. The scoring and prompting module is used to perform normalized weighted aggregation on suspicious token sequences to obtain a global anomaly representation, and calculate the anomaly probability through a classifier; at the same time, it generates anomaly degree indication information based on the anomaly probability and inserts prompt fragments. The reasoning generation module receives the original visual token sequence, the suspicious token sequence, and the hint fragment, calls the multimodal large language model, and generates reasoning output.

7. The surface defect detection device for helicopter main rotor hub according to claim 6, characterized in that, It also includes a training control module for performing two-stage training: In the first stage, only the parameters of the anomaly expert module, the visual token selection module, and the scoring and prompting construction module are updated to obtain the defect detection capability under the condition that the sample size is less than a preset threshold. In the second stage, while retaining the parameters from the first stage, visual instructions are adjusted so that the reasoning generation module outputs an instruction response that includes defect evidence area identification information, defect category description, cause inference, and suggestion text.

8. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the method for detecting surface defects of helicopter main rotor hubs as described in any one of claims 1 to 5.

Images