An image analysis and recognition system based on aircraft landing gear shape defects

Abstract

This invention relates to the technical field of image analysis and recognition systems for aircraft landing gear shape defects, specifically disclosing an image analysis and recognition system for aircraft landing gear shape defects. The system includes a high-resolution image acquisition device, a digital prototype storage unit, a finite element stress field mapping module, a visual attention guidance network, and a defect recognition decision unit. It generates a high-stress region heat map through finite element simulation as prior guidance information, embeds a visual attention network to dynamically enhance the feature response of key areas, and combines a multi-scale fusion mechanism to achieve accurate classification and location of defects such as cracks and plastic deformation. This invention improves detection robustness and sensitivity through a mechanical-visual coupling mechanism, suppresses false alarms in complex backgrounds such as oil stains and reflections, reduces dependence on imaging hardware, and provides efficient and intelligent technical support for aircraft structural health monitoring.

Description

Technical Field

[0001] This invention belongs to the field of aircraft inspection and machine vision technology, specifically relating to an image analysis and recognition system based on aircraft landing gear shape defects. Background Technology

[0002] With the rapid development of the aviation industry, aircraft landing gear, as a core load-bearing component supporting the weight of the fuselage and absorbing landing impact, is crucial for ensuring flight safety due to its structural integrity. Landing gear endures high-frequency alternating loads over long periods in complex service environments, making its metal surfaces prone to fatigue cracks, plastic deformation, or geometric damage. Accurate detection of these defects has become an important part of the aviation maintenance and support system. Currently, high-precision, non-destructive structural health monitoring technology has become a key research focus in the field of aviation safety.

[0003] Utilizing image analysis technology to automatically identify landing gear morphological defects is a mainstream technological direction for improving inspection efficiency and reliability. This technology acquires visual representations of key landing gear components using high-resolution sensors, and combines digital image processing and pattern recognition methods to monitor and quantitatively evaluate the structural state of the landing gear. To improve inspection coverage, the system needs to perform multi-dimensional feature modeling of the complex geometry of the landing gear, thereby enabling morphological analysis of key stress points and connection parts.

[0004] Existing pixel-based visual algorithms often struggle to distinguish between genuine fatigue cracks and intrusive oil stains or surface scratches when dealing with the complex operating conditions of landing gear, resulting in poor robustness of the identification results. Traditional detection schemes lack deep integration of the landing gear's physical structure and mechanical properties, making the system insufficiently capable of detecting minute nascent defects hidden in high-stress concentration areas and unable to accurately focus on key hazardous areas under limited image resolution. Conventional scanning analysis modes, lacking guidance from prior mechanical data, suffer from limitations such as large detection blind spots, weak targeting, and difficulty in "predicting" potential fracture risk points. Therefore, an image analysis and recognition system based on the shape defects of aircraft landing gear is desired. Summary of the Invention

[0005] The purpose of this invention is to provide an image analysis and recognition system based on aircraft landing gear shape defects, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An image analysis and recognition system based on aircraft landing gear shape defects includes a high-resolution image acquisition device, a digital prototype storage unit, a finite element stress field mapping module, a visual attention guidance network, and a defect recognition decision unit, wherein: The high-resolution image acquisition device is configured to perform multi-angle, full-coverage optical imaging of the aircraft landing gear surface, acquire raw image data containing potential shape defects, and transmit the raw image data to the defect identification decision unit. The digital prototype storage unit is configured to store three-dimensional geometric model data of the aircraft landing gear. The three-dimensional geometric model data is constructed based on computer-aided design and fully represents the spatial distribution characteristics of the landing gear's structural outline, connection nodes, and key stress-bearing parts. The finite element stress field mapping module is configured to receive the three-dimensional geometric model data provided by the digital prototype storage unit, and combine it with the historical load spectrum information of the landing gear under typical service conditions to generate a high-stress area heat map through solid mechanics simulation calculation. The high-stress area heat map identifies areas with stress concentration higher than a preset threshold in a spatial distribution form. The visual attention guidance network is configured to receive the raw image data output by the high-resolution image acquisition device and the high-stress area heat map output by the finite element stress field mapping module. The high-stress area heat map is embedded as prior guidance information into the feature extraction layer of the convolutional neural network, and the feature response weights at different spatial locations are dynamically adjusted so that the network gives higher attention to the high-stress area during the feature learning process. The defect identification decision unit is configured to receive the enhanced feature map processed by the visual attention-guided network, perform classification and localization judgment of cracks, plastic deformation or geometric anomalies, and output the identification results of defect existence, type and spatial coordinates.

[0007] Preferably, the high-resolution image acquisition device adopts a multispectral imaging mode to simultaneously acquire images in the visible and near-infrared bands, thereby enhancing the ability to distinguish between oil stains, oxide layers, and real cracks.

[0008] Furthermore, the historical load spectrum information used by the finite element stress field mapping module includes dynamic load sequences of the landing gear under typical operating conditions such as takeoff, landing, taxiing, and emergency braking. These dynamic load sequences are normalized and then used to drive the finite element simulation, ensuring that the generated high-stress region thermal map covers the most severe mechanical response scenarios throughout the entire life cycle.

[0009] Furthermore, the attention mechanism in the visual attention guidance network adopts a channel-spatial dual weighting structure, where the spatial weights are directly modulated by the heatmap of the high-stress region, while the channel weights are adaptively adjusted according to the local texture complexity of the input image, thereby enhancing the expression of detailed features in high-risk regions while preserving global contextual information.

[0010] Preferably, the defect identification decision unit has a built-in multi-scale feature fusion module, which can perform hierarchical aggregation of crack features under different receptive fields, effectively improving the detection sensitivity of small nascent cracks, and is especially suitable for detection environments with limited image resolution or severe surface reflection interference.

[0011] Furthermore, the digital prototype storage unit supports dynamic loading of the corresponding landing gear 3D geometric model according to the aircraft model version, ensuring that the system can be adapted to the landing gear structure of various aircraft models and has good engineering scalability and platform compatibility.

[0012] Furthermore, the heat map of the high-stress region represents the fracture risk level of each spatial point in the form of continuous probability density, rather than a binary mask, which enables the visual attention guidance network to achieve gradient attention allocation and avoid missed detection of edge regions due to hard boundary division.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The image analysis and recognition system based on aircraft landing gear shape defects provided by this invention breaks through the limitations of traditional pure pixel-driven image recognition methods in the detection of complex aerospace components by deeply integrating solid mechanics simulation and deep visual perception.

[0014] The system uses the high-stress region thermal map generated by finite element analysis as prior knowledge to guide the visual network to focus on the key parts most likely to experience fatigue fracture, thereby improving the recognition rate and robustness of small but high-risk defects.

[0015] Even when there are interfering factors such as oil stains, scratches, or uneven lighting in the image background, this mechanical-visual coupling mechanism can still effectively suppress false alarms, realizing a paradigm shift from "blind scanning" to "targeted detection".

[0016] The system can achieve sub-millimeter level defect detection capability in critical areas without relying on ultra-high resolution imaging hardware, reducing the performance requirements of on-site inspection equipment, while improving inspection efficiency and safety, and providing highly reliable and intelligent technical support for aircraft structural health monitoring. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the deep integration of mechanical simulation and visual perception in this invention; Figure 3 This is a flowchart illustrating the logical process of finite element stress field mapping and high-stress region thermal map generation in this invention. Figure 4This is a flowchart illustrating the logical flow of the visual attention guidance network dynamically weighted based on prior guidance information in this invention. Figure 5 This is a schematic diagram of the data flow of the defect identification decision unit in this invention, which performs multi-scale feature fusion and outputs the identification results. Detailed Implementation

[0018] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0019] An image analysis and recognition system based on aircraft landing gear shape defects includes a high-resolution image acquisition device, a digital prototype storage unit, a finite element stress field mapping module, a visual attention guidance network, and a defect recognition decision unit.

[0020] The high-resolution image acquisition device is configured at preset spatial points in the aircraft maintenance bay or landing gear maintenance support. It is used to perform multi-angle, full-coverage optical imaging of the aircraft landing gear surface, acquiring raw image data containing potential shape defects, and transmitting the raw image data to the defect identification decision unit. The high-resolution image acquisition device includes multiple sets of industrial-grade high dynamic range cameras, each connected to the system's central processing unit via a 10 Gigabit Ethernet interface with high bandwidth. To achieve seamless coverage of the complex geometric surfaces of the landing gear, the high-resolution image acquisition device is also equipped with a six-degree-of-freedom robotic arm or a circular guide rail system. A preset path planning algorithm controls the spatial pose of the cameras, ensuring that the imaging resolution for key components such as the main landing gear struts, piston rods, actuators, torque arms, and axles is no less than 0.05 mm per pixel.

[0021] The high-resolution image acquisition device further includes an intelligent supplementary lighting control subunit. This subunit employs a multispectral imaging mode, simultaneously acquiring images in the visible and near-infrared bands. The intelligent supplementary lighting control subunit is configured to automatically adjust the output power of the high-power LED array based on ambient light intensity to eliminate specular reflection interference caused by excessively smooth landing gear metal surfaces. In the near-infrared band, the system can penetrate surface oil or light oxide layers to detect true geometric abrupt changes on the metal substrate surface, enhancing the ability to distinguish between oil artifacts and real fatigue cracks.

[0022] The digital prototype storage unit, configured in a high-reliability solid-state storage array, stores the three-dimensional geometric model data of the aircraft landing gear. This three-dimensional geometric model data is constructed based on computer-aided design, fully representing the spatial distribution characteristics of the landing gear's structural outline, connection nodes, and key stress-bearing components. The digital prototype storage unit supports dynamically loading the corresponding landing gear three-dimensional geometric model according to aircraft model version, ensuring the system is adaptable to landing gear structures of various aircraft models. At the data storage level, the three-dimensional geometric model data not only includes accurate boundary representation data but also encapsulates the material properties of each component, such as elastic modulus, Poisson's ratio, and yield strength, providing physical parameter support for subsequent stress analysis. The digital prototype storage unit has a built-in model indexing engine that can automatically retrieve and call the latest version of the digital prototype matching the input fuselage number, ensuring the timeliness and accuracy of the testing benchmark.

[0023] The finite element stress field mapping module, configured within a server equipped with a high-performance graphics processing unit computing cluster, receives the three-dimensional geometric model data provided by the digital prototype storage unit and, in conjunction with the historical load spectrum information of the landing gear under typical service conditions, generates a high-stress region heat map through solid mechanics simulation calculations. The high-stress region heat map spatially identifies areas where stress concentration exceeds a preset threshold. The historical load spectrum information is provided by the aircraft's onboard data recording system and specifically includes dynamic load sequences of the landing gear under typical conditions such as takeoff, landing, taxiing, and emergency braking, involving the time-varying curves of vertical load, lateral load, and longitudinal braking torque.

[0024] During simulation calculations, the finite element stress field mapping module first adaptively meshes the 3D geometric model, employing a finer mesh at geometric singularities such as pin connections and circular arc transition zones to ensure the accuracy of stress gradient solutions. The module then applies a normalized dynamic load sequence to the corresponding boundaries of the prototype model and obtains the equivalent stress distribution field of the landing gear under the most severe mechanical response scenario throughout its lifecycle by solving the elasticity control equations. The generated high-stress region heatmap is not a simple binary mask but represents the fracture risk level of each spatial point in the form of a continuous probability density. This gradient representation smoothly guides the subsequent visual perception process, avoiding missed detections of edge transition regions caused by hard boundary divisions.

[0025] The visual attention guidance network, configured in a deep learning accelerator card, receives raw image data output by the high-resolution image acquisition device and high-stress region heatmaps output by the finite element stress field mapping module. This network embeds the high-stress region heatmaps as prior guidance information into the feature extraction layer of a convolutional neural network, dynamically adjusting the feature response weights at different spatial locations. In terms of algorithm architecture, the visual attention guidance network employs a multi-level convolutional structure, acquiring feature representations from low-level texture to high-level semantics through layer-by-layer downsampling.

[0026] The attention mechanism in the described visual attention guidance network employs a channel-spatial dual-weighting structure. The spatial weights are directly modulated by the heatmap of the high-stress region, i.e., the spatial resolution of the heatmap is downsampled to match the convolutional features. Figure 1 The weighting is applied to the feature response values ​​using element-wise multiplication, thus giving higher attention to high-stress areas (such as the root and actuator connection) during feature learning. Channel weights are adaptively adjusted based on the local texture complexity of the input image, with importance coefficients for each feature channel calculated through global average pooling and fully connected layers. This dual-weighting mechanism significantly enhances the detailed feature representation of high-risk areas while preserving global contextual information, enabling the system to maintain extremely high feature discrimination even when dealing with background scratches, uneven lighting, and other interference factors.

[0027] The defect identification decision unit, configured in the instruction sequence of the central processing unit, receives the enhanced feature map processed by the visual attention-guided network, performs classification and localization judgment of cracks, plastic deformations, or geometric anomalies, and outputs the identification results of defect existence, type, and spatial coordinates. The defect identification decision unit incorporates a multi-scale feature fusion module, capable of hierarchically aggregating features from different receptive fields. This multi-scale feature fusion module fuses deep semantic features with shallow geometric features through lateral connections, thereby improving the detection sensitivity for small, initial cracks while maintaining accuracy in detecting large-size deformations.

[0028] In the final decision-making stage, the defect identification decision unit employs a region-proposal-based detection framework to perform bounding box regression prediction on potential outliers in the enhanced feature map. For each identified candidate defect, the system not only provides a confidence score indicating its belonging to a specific defect category but also outputs the defect's precise three-dimensional spatial coordinates based on the spatial coordinate system of the digital prototype. This result can be directly mapped back to the actual landing gear surface, guiding maintenance personnel to conduct targeted review and repair. The defect identification decision unit is also equipped with an anomaly alarm subunit. When the detected defect size or stress coupling risk value exceeds a preset safety alarm limit, the system will issue a real-time red warning through a graphical interface.

[0029] Example 2: Based on Example 1, this example further provides an image analysis and recognition system based on aircraft landing gear shape defects with distributed computing capabilities, aiming to solve the problems of real-time processing performance and data consistency in large-scale fleet maintenance scenarios.

[0030] In this embodiment, the system adopts a layered architecture combining edge-side acquisition and cloud-based deep analysis. The high-resolution image acquisition devices are distributed across multiple hangar operation terminals, each equipped with an independent edge computing node. The edge computing nodes are responsible for preliminary image quality assessment, defect removal, and data compression to reduce bandwidth consumption on the backbone network. The digital prototype storage unit is deployed using a distributed file system to achieve cross-regional data synchronization and version control, ensuring that inspection systems in different hangars can identify defects based on unified engineering specifications.

[0031] In this embodiment, the finite element stress field mapping module is configured as a high-performance parallel computing service. Since finite element simulation consumes a significant amount of computational resources, this system pre-establishes a stress response database for typical load combinations of different aircraft types and service periods. When a specific aircraft is being inspected, the finite element stress field mapping module does not execute the complete finite element solution process in real time. Instead, based on the real-time load statistical characteristics uploaded by the airborne system, it quickly retrieves and fits the high-stress region heat map from the stress response database using a multi-dimensional interpolation algorithm. This data-driven mapping method significantly shortens the system's response time, enabling it to support rapid transit inspections of the landing gear while in port.

[0032] In this embodiment, the visual attention guidance network incorporates an adaptive environment compensation submodule. Considering the significant differences in lighting and shadow environments inside the hangar under different seasons and climates, the adaptive environment compensation submodule improves image quality in low-contrast or backlit environments by performing histogram equalization and dark channel prior dehazing on the original image. The processed image, along with thermal analysis of high-stress areas... Figure 1 The same input attention guidance mechanism is used. The attention mechanism is further refined, and its spatial weight modulation logic introduces a temporal stability factor. By comparing the changes in thermodynamic mapping of the same part in multiple historical tests, the sensitive area with the most drastic stress evolution is dynamically locked, thereby realizing the dynamic tracking of the fatigue damage evolution process.

[0033] In this embodiment, the defect identification decision unit enhances its flexibility in adapting to multiple aircraft models. Its built-in classification model employs an incremental learning architecture, enabling it to automatically update weight parameters based on continuously accumulated actual detection samples. When the system detects a novel atypical defect on a particular aircraft model, the defect identification decision unit feeds back the sample data and corresponding manual verification results to the cloud training center. After fine-tuning the model, the training center distributes the updated network parameters to all terminals. The defect identification decision unit also integrates a structural health residual life assessment submodule. This submodule, by coupling detected defect features (such as crack length and depth) with the local stress level provided by the finite element stress field mapping module, estimates the safe service life of the landing gear before the next major overhaul based on the stress intensity factor criterion in fracture mechanics, providing data support for the airline's maintenance scheduling plan.

[0034] In terms of hardware physical connectivity, the various modules of the system in this embodiment are interconnected using a redundant fiber optic ring network. This ensures that even if some network nodes experience physical failures, the data stream can still be transmitted to the defect identification decision unit via a backup path. The photosensitive unit of the high-resolution image acquisition device is also equipped with an automatic cleaning component, which periodically cleans away metal dust and oil mist condensation using a high-pressure air curtain or mechanical wiper, ensuring image clarity during long-term operation.

[0035] Example 3: Based on the above examples, this example focuses on describing an image analysis and recognition system for aircraft landing gear shape defects that integrates real-time sensor feedback and dynamic load correction functions, in order to achieve a higher-dimensional mechanical-visual coupling depth.

[0036] In this embodiment, a new sensor feedback front-end is added to the system. This sensor feedback front-end interfaces with strain gauges, pressure sensors, and accelerometers on the aircraft landing gear. The sensor feedback front-end acquires transient impact load data of the landing gear during the most recent landing process and transmits this data to the finite element stress field mapping module in real time.

[0037] The finite element stress field mapping module corrects the initial boundary conditions in the digital prototype in real time based on the received transient impact loads. For example, if there is a significant off-center load or heavy landing during the landing process, the finite element stress field mapping module will automatically adjust the distribution coefficients of the force matrix and recalculate the stress field distribution. Since heavy landing may cause instantaneous stress peaks in some non-traditional parts, this dynamic correction mechanism enables the generated high-stress area heat map to accurately capture these potential damage points caused by accidental loads, thereby guiding the visual attention of the network to focus on scanning these "atypical danger zones".

[0038] In this embodiment, the digital prototype storage unit not only stores three-dimensional geometric data but also integrates a digital twin entity of the landing gear. This digital twin entity records all maintenance records, replacement part serial numbers, and minor defects discovered during previous inspections since the landing gear's delivery. When performing the current inspection task, the visual attention guidance network uses historical damage points in the digital twin entity as secondary attention focuses, fusing them with a high-stress heatmap generated based on the current load. The spatial weight calculation logic of the attention mechanism consists of two superimposed parts: the first part is the real-time stress weight based on the current mechanical state, and the second part is the memory weight based on accumulated historical damage. The linear combination of these two parts constitutes the final guidance heatmap, ensuring that the system can both detect newly generated sudden defects and closely monitor the evolution dynamics of existing minor damage.

[0039] The visual attention guidance network further includes a feature reconstruction layer. This layer utilizes the principles of generative adversarial networks to virtually reconstruct occluded or poorly lit areas in the image. When visual blind spots exist within certain deep-hole structures or complex linkage mechanisms of the landing gear, this feature reconstruction layer generates a predicted feature map of the region by referencing the geometric priors and symmetry features of the digital prototype. Although the reconstructed features are not directly used for the final judgment, they serve as auxiliary features and can be compared with measured features to effectively identify geometric defects or severe plastic displacements in components.

[0040] In this embodiment, the defect identification decision unit incorporates a knowledge graph-based fault reasoning engine. This engine correlates identified external defects (such as cracks at specific angles or indentations in specific directions) with the landing gear's material failure mechanisms and assembly process deviations. When the defect identification decision unit detects an anomaly, the knowledge graph automatically retrieves failure cases of that location under similar service environments, providing suggestions for the main causes of the defect, such as "stress corrosion cracking" or "surface reinforcement layer peeling."

[0041] In this embodiment, the high-resolution image acquisition device is designed as a mobile intelligent inspection robot. Equipped with Mecanum wheels at its bottom for omnidirectional movement, it is also equipped with a 3D laser scanner. Simultaneously with optical imaging, the 3D laser scanner acquires point cloud data of the landing gear to obtain its actual geometric contours. The defect identification decision unit performs Boolean operations between the measured point cloud data and the nominal model in the digital prototype storage unit, enabling it to detect bending or torsional deformations in the overall landing gear structure with micron-level precision. This large-scale defect identification at the geometric level complements the microscopic crack identification at the visual level, constructing a full-scale inspection dimension.

[0042] Example 4: This example details a portable landing gear image analysis and recognition system implementation scheme suitable for emergency maintenance environments in the field, which is a supplement to the above-mentioned fixed or integrated architecture.

[0043] In this embodiment, the high-resolution image acquisition device is integrated into a handheld smart terminal and employs array-type structured light projection imaging technology. When maintenance personnel walk around the landing gear while holding the terminal, the device calculates the spatial trajectory of the equipment through an inertial measurement unit, enabling automatic stitching and 3D reconstruction of multiple frames. The digital prototype storage unit is pre-stored in the local storage of the handheld terminal in the form of a highly compressed parametric model.

[0044] The finite element stress field mapping module employs a simplified mechanical model within a portable architecture. Instead of performing fully-degree-of-freedom finite element iterative calculations, it utilizes a pre-trained deep operator network. This deep operator network takes a simplified landing gear geometry and estimated load vectors as input, and directly outputs an approximate stress field distribution through operator mapping. Its computational overhead is only 5% of traditional finite element analysis, enabling real-time operation on a handheld terminal's embedded processor and instant generation of thermal maps of high-stress areas.

[0045] The visual attention-guided network is configured as a lightweight convolutional structure, compressing the weight precision to eight-bit fixed-point through pruning and quantization techniques. Despite the reduced model size, the network can concentrate limited computational resources on high-risk windows indicated by the heatmap of high-stress regions, achieving recognition accuracy comparable to server-side systems while maintaining low power consumption, thanks to the strong prior space constraints provided by the heatmap. The attention mechanism particularly strengthens anti-shake compensation logic in handheld scenarios, using affine transformations at the feature level to counteract image blur caused by hand movement.

[0046] In this embodiment, the defect identification decision unit integrates an augmented reality display module. When the system identifies a defect, it doesn't just output a text report; instead, it overlays the identification results (defect box, defect type, and hazard level) onto the live view screen via a handheld terminal. Maintenance personnel can clearly see the stress distribution colors and the specific location of the defect on the landing gear. The system supports voice interaction, allowing maintenance personnel to retrieve digital prototype drawings of specific parts for comparison using voice commands.

[0047] The portable system is connected to a remote technical support center via Wi-Fi or 5G mobile communication technology. When encountering difficult-to-determine defective problems, the high-resolution original image data collected on-site and the associated mechanical and thermal maps will be uploaded together. Through the collaborative analysis interface provided by the defect identification decision unit, remote experts can call more complex simulation models in the cloud for secondary verification. Since the uploaded data contains mechanical guidance information, the remote model can quickly locate the disputed area, greatly improving the efficiency of remote assistance.

[0048] Embodiment 5: This embodiment describes an image analysis and recognition system for quality control in the manufacturing stage of an aircraft landing gear, focusing on how to improve the reliability of first-piece inspection and batch sampling inspection by coupling finite element analysis and visual perception.

[0049] In a manufacturing environment, the digital prototype storage unit stores a theoretical three-dimensional model containing design tolerance requirements. The high-resolution image acquisition device is installed on a fully automated inspection production line in a controlled environment with constant lighting conditions. The finite element stress field mapping module not only simulates service loads but also the prestress distribution during the assembly process. For example, the internal stress fields generated during processes such as bolt tightening and interference fit are converted into high-stress thermal maps to guide the vision system to detect early cracks or material micro-damage that may be caused by assembly stress.

[0050] In this scenario, the visual attention guidance network introduces a binocular structured light vision sub-module to obtain high-precision depth features. The spatial weight of the attention mechanism is set as a three-dimensional spatial weight, that is, not only weighted in the planar dimension of the image but also attention is distributed in the depth direction. For parts such as the inner wall of holes and grooves on the landing gear with depth features, the system will automatically increase the sampling frequency and exposure time to ensure that these complex structures, which are shown as stress concentration points in finite element analysis, are visually inspected sufficiently.

[0051] The defect identification decision unit has specifically optimized the model for manufacturing defects and can accurately identify burrs, sand holes, machining tool marks, and surface color differences caused by uneven heat treatment. The system compares the geometric parameters of the identified defects with the design tolerances in the digital prototype in real time to automatically determine whether the part is qualified. For unqualified parts, the system will automatically generate a defective cause report containing stress correlation analysis to assist process engineers in optimizing the manufacturing process and reducing the tendency of stress concentration.

[0052] This embodiment of the system also features a self-calibration function. The high-resolution image acquisition device periodically scans the standard calibration block, and the defect identification decision unit automatically corrects the distortion parameters and color response curve of the optical system by calculating the residual between the measured image and the standard image. This closed-loop calibration mechanism ensures the constancy and consistency of the system's testing standards over a production cycle that can last for several years.

[0053] Example 6: This example explores a landing gear image analysis and recognition system with self-evolution capabilities for future intelligent maintenance.

[0054] The finite element stress field mapping module in this system has online learning capabilities. Its integrated solid mechanics operator library can reverse-correct load spectrum parameters based on the actual defect morphology found in previous tests. For example, if unexpected fatigue cracks frequently appear in a certain area, the system will determine that the actual stress condition of that area deviates from the preset historical load spectrum, and then automatically start the parameter identification program to refit a more realistic mechanical boundary condition, thereby generating a more predictive thermal map of the high-stress area.

[0055] The visual attention-guided network incorporates multimodal semantic understanding capabilities enhanced by a large language model. The defect identification decision unit not only outputs physical features but also receives natural language descriptions input by maintenance technicians (e.g., "Suspected oil leakage found at the left actuator"). The system transforms this semantic information into temporary attention bias weights, guiding the network to perform higher-resolution feature resampling in specific regions.

[0056] In this embodiment, the digital prototype storage unit evolves into a cross-platform knowledge base, storing not only geometric models but also encapsulating a statistical probability map of landing gear failures for similar aircraft models worldwide. When generating the guiding heatmap, the attention mechanism fuses the physical stress weights derived from finite element analysis with risk weights based on statistical probability. This dual guidance of physical mechanisms and statistical laws enables the system to perform highly reliable preventative detection even in the event of sensor failure or missing load data, relying on prior big data.

[0057] To further enhance the system's robustness, the defect identification decision unit incorporates a simulator that can generate virtual defect samples (such as synthetic crack features superimposed on high-stress areas of an image) in real time during the detection process and test whether the network can accurately identify them. This online adversarial testing mechanism can evaluate the system's perception sensitivity in real time. When it detects a decrease in the system's detection capability due to extreme changes in ambient light, it will automatically trigger an alarm and suggest manual intervention.

[0058] The calculation and logic processing involved in the above embodiments are all described in text form. The parameter relationships within each module of the system, such as the mapping relationship between the weight values ​​and stress intensity in the heat map of the high-stress region, are configured as a nonlinear enhancement function. This function ensures that when the stress value exceeds a preset first threshold, the corresponding spatial weight increases exponentially, while when the stress value is below a second threshold, the weight converges to a very small preset constant, thereby suppressing irrelevant background noise.

[0059] In the visual attention-guided network, the feature transfer process between each convolutional layer follows the principle of energy conservation. That is, the sum of the weighted feature responses is normalized to maintain it within a preset numerical range, preventing numerical overflow due to excessively large local weights. When performing multi-scale fusion, the defect recognition decision unit uses layer-by-layer interpolation alignment to ensure precise spatial overlap of feature maps at different resolutions. This ensures that the spatial coordinate positioning accuracy deviation of micro-cracks is less than three times the diameter of a single pixel.

[0060] At the data interaction level, all signal transmissions within the system employ a checksum mechanism to ensure that every bit of data, from the high-resolution image acquisition device to the defect identification decision unit, is not tampered with or lost during transmission. Access to the digital prototype storage unit is controlled by a hardware-level encryption chip, allowing only legitimate, authenticated detection requests to access model data, thus ensuring the security of aerospace technical data.

[0061] This invention's system seamlessly integrates profound prior knowledge of solid mechanics into an advanced visual attention mechanism in the form of a heatmap, transforming the traditional "blind search" approach in landing gear inspection. This deep coupling of mechanics and vision improves the accuracy of identifying minute, hidden defects and endows the system with a degree of "engineering logic thinking," enabling it to prioritize critical areas most prone to problems, much like an experienced chief engineer. This paradigm shift provides a completely new solution for aircraft structural health monitoring, possessing significant engineering application value and social safety implications.

[0062] It should be noted that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Without departing from the essential meaning of the present invention, any combination of the various technical features in this embodiment to form a technical solution should fall within the scope of disclosure of the present invention. Furthermore, the qualifiers such as "first" and "second" used herein are only used to distinguish different logical units or parameters and do not represent any specific order or level of importance.

Claims

1. An image analysis and recognition system based on aircraft landing gear shape defects, characterized in that, include: A high-resolution image acquisition device is configured to perform multi-angle, full-coverage optical imaging of the aircraft landing gear surface to acquire raw image data containing potential shape defects. A digital prototype storage unit is configured to store three-dimensional geometric model data of the aircraft landing gear. The three-dimensional geometric model data is constructed based on computer-aided design and is used to characterize the spatial distribution features of the landing gear's structural outline, connection nodes, and key stress-bearing parts. The finite element stress field mapping module is connected to the digital prototype storage unit and is configured to receive the three-dimensional geometric model data and perform solid mechanics simulation calculations in combination with the historical load spectrum information of the landing gear to generate a high-stress area heat map. The high-stress area heat map identifies areas where the stress concentration is higher than a preset threshold in a spatial distribution form. A visual attention guidance network is connected to the high-resolution image acquisition device and the finite element stress field mapping module, respectively. It is configured to receive the original image data and the heat map of the high-stress region, embed the heat map of the high-stress region as prior guidance information into the feature extraction layer of the convolutional neural network, and output an enhanced feature map by dynamically adjusting the feature response weights of different spatial locations. The defect identification decision unit is connected to the visual attention guidance network and is configured to perform defect classification and localization judgment based on the enhanced feature map, and output the identification result.

2. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, The high-resolution image acquisition device includes multiple sets of industrial-grade high dynamic range cameras and a pose adjustment mechanism for supporting the cameras. The pose adjustment mechanism is configured as a six-degree-of-freedom robotic arm or a ring-shaped guide rail system, which controls the spatial pose of the cameras by executing a preset path planning algorithm to ensure that the imaging resolution of the main landing gear strut, piston rod, actuator, torque arm, and wheel axle is not lower than a preset single-pixel size accuracy. The multiple sets of industrial-grade high dynamic range cameras are connected to the system central processor through a 10 Gigabit Ethernet interface with preset bandwidth transmission capability to realize real-time reporting of raw image data.

3. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, The high-resolution image acquisition device further includes an intelligent supplementary lighting control subunit, which is configured to adopt a multispectral imaging mode to simultaneously acquire images in the visible light band and the near-infrared band; the intelligent supplementary lighting control subunit automatically adjusts the output power of the light-emitting diode array according to the ambient light intensity to eliminate specular reflection interference generated by the landing gear metal surface; In the near-infrared band, the system is configured to penetrate the oil or oxide layer on the landing gear surface to detect the real geometric abrupt changes on the metal substrate surface, thereby enhancing the ability to distinguish between oil artifacts and real fatigue cracks.

4. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, The digital prototype storage unit is configured in a high-reliability solid-state storage array, and the three-dimensional geometric model data encapsulates the material property parameters of each component, including elastic modulus, Poisson's ratio, and yield strength. The digital prototype storage unit has a built-in model indexing engine configured to automatically retrieve and call the latest version of the digital prototype that matches the input fuselage number. The digital prototype storage unit supports dynamically loading the corresponding landing gear model according to the aircraft model version, enabling the system to have engineering scalability to adapt to landing gear structures of various aircraft models.

5. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, Before performing simulation calculations, the finite element stress field mapping module is configured to adaptively mesh the three-dimensional geometric model. A fine mesh strategy is adopted at geometric singularities at pin joints and circular arc transition zones to improve the accuracy of stress gradient solutions. The historical load spectrum information includes dynamic load sequences of the landing gear under takeoff, landing, taxiing, and emergency braking conditions. The finite element stress field mapping module applies the normalized dynamic load sequences to the corresponding boundaries of the prototype model and obtains the equivalent stress distribution field of the landing gear under the full life cycle mechanical response scenario by solving the elasticity control equations.

6. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, The high-stress region heatmap represents the fracture risk level of each spatial point in the form of continuous probability density, rather than a binary mask; the visual attention guidance network implements gradient attention allocation based on the continuous probability density, ensuring that during feature learning, the feature response weights monotonically increase with the increase of stress concentration, avoiding missed detection of edge transition regions caused by hard boundary division; in the visual attention guidance network, the feature transfer process between each convolutional layer follows the principle of energy conservation, that is, the sum of the weighted feature responses is kept within a preset numerical range through normalization operations to prevent numerical overflow.

7. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, The attention mechanism in the visual attention guidance network adopts a channel-spatial dual-weighted structure. The spatial weights are directly modulated by the heatmap of the high-stress region, that is, the spatial resolution of the heatmap is downsampled to be consistent with the convolutional feature map, and applied to the feature response value in an element-wise multiplication manner. The channel weights are adaptively adjusted according to the local texture complexity of the input image, and the importance coefficient of each feature channel is calculated through global average pooling and fully connected layers. The channel-spatial dual-weighted structure is configured to enhance the expression of detailed features in high-risk regions while preserving global contextual information and suppressing the interference of scratches and uneven illumination in the background.

8. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, The defect identification decision unit incorporates a multi-scale feature fusion module, which aggregates deep semantic features and shallow geometric features hierarchically through lateral connections to improve the detection sensitivity of minute nascent cracks. The defect identification decision unit adopts a region-based detection framework to perform bounding box regression prediction on potential outliers in the enhanced feature map and outputs the precise three-dimensional spatial coordinates of the defect based on the spatial coordinate system of the digital prototype. The defect identification decision unit is also equipped with an abnormal alarm subunit. When the detected defect size or stress coupling risk value exceeds the preset safety alarm limit, an early warning prompt is issued through a graphical interface.

9. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, The system adopts a layered architecture that combines edge-side acquisition with cloud-based deep analysis. The high-resolution image acquisition device is connected to an edge computing node, which is responsible for preliminary image quality assessment, bad pixel removal, and data compression. The finite element stress field mapping module is configured as a high-performance parallel computing service. It pre-establishes a stress response database for load combinations of different models and service periods, and retrieves and fits a real-time high-stress area thermal map from the stress response database based on the real-time uploaded load statistical characteristics using a multi-dimensional interpolation algorithm. The defect identification decision unit has a built-in incremental learning architecture that automatically updates weight parameters based on accumulated detection samples.

10. The image analysis and recognition system based on aircraft landing gear shape defects according to claim 1, characterized in that, It also includes a sensor feedback front end, which is connected to strain gauges, pressure sensors and accelerometers on the landing gear to acquire transient impact load data of the landing gear during landing and transmit it to the finite element stress field mapping module for real-time correction of the initial boundary conditions of the three-dimensional geometric model. The digital prototype storage unit integrates a digital twin entity that records historical maintenance records and damage point information; The attention mechanism in the visual attention guidance network linearly combines real-time stress weights based on real-time mechanical state with memory weights based on historical damage accumulation to generate the final guidance heatmap, which is used to simultaneously monitor the evolution dynamics of new defects and existing damage.

Images