FOD calculation method for precise identification of fallen objects based on fixed tower and application of FOD calculation method

By deploying a multimodal sensor array on a fixed tower, all-weather, high-precision automatic detection of objects thrown from airport runways and aprons has been achieved. This solves the problems of low efficiency and high false alarm rate in existing technologies, improves detection sensitivity and classification accuracy, and meets the rapid response requirements of airports.

CN120822023APending Publication Date: 2025-10-21SHANDONG EAGLE INFORMATION ENG CO LTD

Patent Information

Application Number
CN202511044298.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-28
Publication Date
2025-10-21

AI Technical Summary

Technical Problem

Existing technologies for detecting objects thrown from airport runways and aprons suffer from low efficiency, high cost, and high false alarm rates, making it difficult to achieve real-time monitoring and high-precision identification around the clock.

Method used

Based on a fixed tower, a multimodal sensor array is deployed, including a high-definition imaging device, a laser scanning device, and a distributed acoustic sensor. Through synchronous acquisition of multi-source heterogeneous data, multimodal data preprocessing and feature extraction, cross-modal feature fusion, spatiotemporal graph neural network modeling, and multi-dimensional threat quantification, accurate identification and location of falling objects are achieved.

Benefits of technology

It achieves 24/7 fully automated data acquisition, improves detection sensitivity by 60%, increases classification accuracy from 82% to 94%, reduces the false negative rate to below 5%, improves emergency response efficiency by 90%, improves system stability by 60%, and meets centimeter-level positioning accuracy requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120822023A_ABST
    Figure CN120822023A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of industrial safety monitoring, in particular to a fallen object accurate identification FOD calculation method based on a fixed tower and application of the fallen object accurate identification FOD calculation method. Comprising the following steps: synchronously acquiring multi-source heterogeneous sensing data; carrying out multi-modal data preprocessing and feature extraction; performing cross-modal feature fusion and representation construction; modeling and reasoning a time-space diagram neural network; and performing multi-dimensional threat quantification and decision output. According to the design of the invention, multi-modal sensor groups are deployed at the tower bottom, the tower middle and the tower top of a fixed tower in a layered manner, and a high-definition multispectral camera, a high-density laser radar and a distributed acoustic sensor are integrated, so that 7 * 24-hour full-automatic data acquisition is realized; according to the design, the time and environment limitation of manual inspection is eliminated, and the complex weather detection capability is improved by 80%; meanwhile, the sensor synchronously triggers the mechanism and the grading detection process, the single-frame data processing delay is less than or equal to 50ms, and the real-time monitoring requirement of an aircraft take-off and landing high-frequency scene is met.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of industrial safety monitoring, and in particular to a method for calculating FOD (Field of Object) (FOD) for accurately identifying falling objects based on a fixed tower and its application. Background Art

[0002] Foreign object debris (FOD) is a major hidden danger threatening the safety of airport runways and aprons. It may cause serious accidents such as aircraft engine damage and tire explosions, posing a direct threat to aviation safety. "Falling objects" specifically refer to the falling of high-altitude / moving objects caused by human or equipment failure (such as parts falling off during aircraft takeoff and landing, and tools lost during ground operations). It is one of the most direct and frequent types of FOD. Traditional FOD detection methods mainly rely on manual inspections and single-technique monitoring. Among them, manual visual inspections of runways and aprons can intuitively detect foreign objects, but there are problems such as low detection efficiency, high labor intensity, and significant influence of light or weather. It is difficult to achieve all-weather real-time monitoring, and the ability to identify FOD in small or complex backgrounds is limited.

[0003] Single-technology monitoring refers to the use of equipment such as microwave radar to scan the runway area. Although it can achieve a certain degree of automated detection, its detection accuracy is insufficient and the false alarm rate is high due to the limitations of the equipment accuracy. For example, Chinese patent CN202210654237.6 discloses an airport foreign object detection system and an airport foreign object inspection system. The airport foreign object detection system and the airport foreign object inspection system include a static image acquisition device for acquiring multiple single-frame static images of a preset area at a preset sampling speed; a local recognition device for receiving multiple single-frame static images and performing online target recognition of the target foreign object based on the single-frame static images; and a cloud recognition device for classifying, identifying, and risk assessing the target foreign object based on the multiple single-frame static images and / or online target recognition results. The static image acquisition device is used to acquire single-frame static images, and the single-frame static images are recognized online in the local recognition device. In conjunction with the cloud recognition device, the single-frame static images and / or online recognition results are received, and further and more accurate classification, identification, and risk assessment are performed based on the single-frame static images and / or online recognition results, thereby ensuring recognition efficiency while ensuring the accuracy of the final recognition. For example, Chinese patent CN201810053990.3 discloses a method for detecting foreign objects at airports by radar based on time-based statistics. The method includes obtaining N groups of echo data for a single position unit on an airport pavement without foreign objects, and performing statistics on the data to obtain a probability density curve as a baseline probability density curve. During detection, for a single position unit, the echo data scanned by the radar at the current moment plus the N-1 groups of echo data scanned at the previous moment are extracted in real time as the current N groups of echo data, and statistics are performed on the data to obtain the current probability density curve. The current probability density curve is compared with the baseline probability density curve. If the horizontal coordinate of the intersection is greater than or equal to a preset threshold value, it is considered that there is a foreign object; otherwise, it is considered that there is no foreign object. The beneficial effects of this invention are: increasing the reliability of the statistical results, and being suitable for airport foreign object radar detection.

[0004] While the aforementioned existing technologies each have their own advantages, they suffer from low efficiency, high cost, and high false alarm rates when detecting fallen objects in areas such as airport runways and aprons. To address this, we propose a FOD calculation method for accurate fallen object identification based on a fixed tower and its application. Summary of the Invention

[0005] The purpose of the present invention is to provide a FOD calculation method for accurately identifying falling objects based on a fixed tower and its application, so as to solve the problems raised in the above background technology.

[0006] To solve the above technical problems, one of the objectives of the present invention is to provide a method for accurately identifying FOD (Field of View) of falling objects based on a fixed tower, comprising the following steps:

[0007] 1. Synchronous collection of multi-source heterogeneous sensor data;

[0008] S1: Deploy a multimodal sensor group with visual, spatial, and acoustic sensing capabilities on the main structure of the fixed tower and the surrounding monitoring area. The sensor group includes a high-definition imaging device, a laser scanning device, and a distributed acoustic sensing device.

[0009] S2: Achieve collaborative multimodal data acquisition through a synchronous trigger mechanism, obtain RGB image data, laser point cloud data, and acoustic signal data, and establish a spatiotemporal mapping relationship;

[0010] 2. Multimodal data preprocessing and feature extraction;

[0011] S3: Perform noise reduction and contrast enhancement on RGB image data to suppress noise and highlight the edges and texture features of the falling objects; remove outliers and downsample the laser point cloud data to retain valid spatial geometric features; perform frame processing and frequency domain feature extraction on the acoustic signal data to filter out environmental background noise;

[0012] S4: Extracting single-modal deep features through deep learning models: extracting visual feature vectors containing texture and shape from RGB images; extracting geometric feature vectors containing centroid and curvature from laser point cloud data; extracting temporal feature vectors containing frequency and amplitude changes from acoustic wave signals;

[0013] 3. Cross-modal feature fusion and representation construction;

[0014] S5: Dimensional alignment and semantic space fusion of bimodal features are performed, and a two-tuple feature vector containing the association between modalities is generated through a weighted strategy or attention mechanism;

[0015] S6: Based on modal reliability assessment, adaptive weighted fusion of all modal features is performed to generate a joint representation vector containing multi-dimensional information of vision, geometry, and acoustics;

[0016] 4. Spatiotemporal graph neural network modeling and reasoning;

[0017] S7: Map the unimodal features and fusion features into graph nodes, and construct a dynamic graph structure based on spatiotemporal correlation. The nodes include unimodal nodes, fusion nodes, and global nodes. Edge connections are determined by feature similarity and spatiotemporal correlation.

[0018] S8: Aggregate neighborhood node features through spatiotemporal graph convolutional networks, combine with temporal modeling units to capture the motion patterns of falling objects, and generate feature expressions that include spatiotemporal correlations;

[0019] 5. Multi-dimensional threat quantification and decision-making output;

[0020] S9: Based on the joint spatiotemporal feature expression, the physical model and data-driven algorithm are used to calculate the damage risk level of falling objects, predict the three-dimensional motion trajectory, and output the positioning confidence area;

[0021] S10: Generate a structured decision report including threat level, positioning results and emergency response recommendations, and output a visual threat situation map with a data traceability chain.

[0022] As a further improvement to this technical solution, multimodal sensor clusters are installed on the main structure of the fixed tower in S1 at the bottom, middle, and top of the tower. These multimodal sensor clusters achieve vertical and horizontal stereoscopic monitoring, reducing the risk of missed detections through overlapping fields of view and improving positioning accuracy.

[0023] As a further improvement of this technical solution, the multimodal sensor group deployed in S1 is specifically configured as follows:

[0024] High-definition camera array: integrated with polarized light imaging and multispectral imaging capabilities, with a resolution of no less than 3840×2160 and a frame rate of no less than 60fps;

[0025] 3D LiDAR: Using SPAD array detectors, the point cloud density is not less than 500 points / m², and the minimum resolvable distance difference is not greater than 2cm;

[0026] Millimeter-wave radar: uses a MIMO architecture, with a transmission frequency between 76-81 GHz and a 3D point cloud refresh rate of no less than 100 Hz;

[0027] Infrared thermal imager: Thermal sensitivity is no higher than 0.03°C at 30°C, and it features an integrated non-uniformity correction algorithm. High-definition cameras, lidar, and other high-performance sensors support multispectral imaging and high-density point cloud acquisition, adapting to complex environments and enhancing sensitivity for detecting even the smallest foreign objects.

[0028] As a further improvement of this technical solution, the preprocessing method of different modal data in S3 is as follows:

[0029] RGB image denoising: uses wavelet domain dual-tree complex transform and dynamically sets the denoising threshold based on the noise statistical model;

[0030] Laser point cloud processing: Using the improved RANSAC algorithm, outliers are adaptively removed according to the point cloud distribution;

[0031] Acoustic Signal Extraction: Extracts improved CQT time-frequency features consistent with the critical bandwidth theory of hearing. Dedicated algorithms are used for preprocessing data from different modalities to effectively reduce noise, remove outliers, and extract features, improving data quality and subsequent processing efficiency.

[0032] As a further improvement of this technical solution, the bimodal feature semantic space alignment in S5 is achieved by calculating the mutual information entropy matrix between modalities; the mutual information entropy calculation formula is: ,in and are the single modal feature entropy, The joint entropy is calculated by maximizing the mutual information entropy, ensuring that the weight of complementary features accounts for no less than 60% to suppress redundant information between modalities. Feature alignment is achieved by calculating the inter-modal mutual information entropy matrix, maximizing the weight of complementary features, suppressing redundant information, and enhancing inter-modal association and fusion effects.

[0033] As a further improvement of this technical solution, the closed-loop optimization method adopted in S6 is as follows:

[0034] Sensor drift compensation: Using a dual Kalman filter architecture, the main filter is used to update the state estimate, and the auxiliary filter is used to identify the deviation parameters online;

[0035] Parameter Optimization: A meta-learning framework is used to dynamically adjust model weights. Dual Kalman filtering is used to compensate for sensor drift, and meta-learning dynamic parameter adjustment improves system robustness and model adaptability.

[0036] As a further improvement to this technical solution, the dynamic graph structure constructed in S7 employs a hybrid propagation mechanism, where a message passing function fuses node features with edge weights. Node updates utilize a gated GRU module, integrating spatiotemporal residual connections. This hybrid propagation mechanism, combined with the gated GRU module and spatiotemporal residual connections, enhances the modeling capabilities of spatiotemporal correlations and improves the accuracy of motion trajectory capture.

[0037] As a further improvement of this technical solution, the threat assessment in S9 is performed as follows:

[0038] Damage degree calculation: integrating material stress-strain constitutive model;

[0039] Trajectory Prediction: Utilizing an interactive multi-model algorithm, the system integrates the CTRA and CS motion models. This system integrates material physics models with the multi-model algorithm to accurately calculate damage risk levels and dynamically predict motion trajectories, meeting the airport's rapid response positioning accuracy requirements. This system integrates material physics models with the multi-model algorithm to accurately calculate damage risk levels and dynamically predict motion trajectories, meeting the airport's rapid response positioning accuracy requirements.

[0040] As a further improvement to this technical solution, the following data security and storage measures are adopted in the entire data processing flow from data collection in S2 to result output in S10:

[0041] Data encryption: RGB image data, laser point cloud data, acoustic signal data and processing results are encrypted using the AES-256 symmetric encryption algorithm during transmission and storage, with a key length of 256 bits;

[0042] Blockchain evidence storage: The SHA-256 hash value of the test results and decision reports is stored on the blockchain node network. The timestamp of the evidence is generated by a trusted timestamp server with an accuracy of no less than 1ms.

[0043] Data backup: Adopting an off-site disaster recovery backup strategy, the processed data is regularly backed up to a secure data center ≥50 km away from the monitoring area through an encrypted data transmission channel, with a backup cycle of 24 hours per time.

[0044] Access Control: A role-based access control system is established, with fine-grained access rights assigned to different roles through the permissions management module. The authentication mechanism combines username / password, dynamic password, and biometrics for multi-factor authentication. Encryption, blockchain storage, off-site backup, and multi-factor access control ensure data security, traceability, and system access security.

[0045] The second object of the present invention is to provide an application of a FOD calculation method for accurately identifying falling objects based on a fixed tower, such as the application of the above-mentioned FOD calculation method for accurately identifying falling objects based on a fixed tower in the runways and aprons of airports.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This design deploys a multimodal sensor cluster at the base, center, and top of a fixed tower. It integrates high-definition multispectral cameras, high-density lidar, and distributed acoustic sensors to create a 360-degree monitoring network, enabling 24 / 7 fully automated data collection. This design eliminates the time and environmental constraints of manual inspections, improving detection capabilities in complex weather by 80% and detecting small dropped foreign objects by 60% compared to manual inspections. Furthermore, the sensors' synchronized triggering mechanism and hierarchical detection process ensure single-frame data processing latency of ≤50ms, meeting the real-time monitoring requirements of high-frequency aircraft takeoffs and landings.

[0048] 2. Through data preprocessing, feature enhancement, and cross-modal intelligent fusion, this design automatically extracts visual, geometric, and temporal features, overcoming the subjective limitations of traditional manual feature extraction. This improves feature recognition in complex backgrounds by 40%. Furthermore, it increases FOD classification accuracy from 82% to 94% and reduces the missed detection rate from 18% to below 5%.

[0049] 3. The multispectral camera designed and configured in this invention supports visible light, infrared, and ultraviolet imaging, improving edge feature extraction in low-light environments by 40% and contrast in highly reflective scenes by 50%. Simultaneously, the infrared thermal imager and millimeter-wave radar compensate for visual blind spots, enabling "visible light + infrared + radar" all-weather detection, with environmental adaptability improved by 70% compared to single-modality solutions. Furthermore, a dual Kalman filter architecture is used to compensate for sensor drift in real time, and a meta-learning framework dynamically adjusts model weights, improving the system's long-term operational stability by 60%, extending the equipment maintenance cycle from one week to one month.

[0050] 4. The design of this invention integrates the material stress-strain constitutive model and the interactive multi-model algorithm, which improves the accuracy of damage risk level judgment by 25%, and the trajectory prediction error is ≤15cm, providing centimeter-level navigation coordinates for the foreign object removal robot; at the same time, the structured decision report and visual situation map shorten the manual decision-making time from 10 minutes to 30 seconds, and the emergency response efficiency is improved by 90%. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Flow chart of the FOD calculation method of the present invention;

[0052] Figure 2 Schematic diagram of the steps for synchronously collecting multi-source heterogeneous sensor data in the present invention;

[0053] Figure 3 Schematic diagram of the steps of multimodal data preprocessing and feature extraction in the present invention;

[0054] Figure 4 Schematic diagram of the steps of cross-modal feature fusion and representation construction in the present invention;

[0055] Figure 5 Schematic diagram of the steps of spatiotemporal graph neural network modeling and reasoning in the present invention;

[0056] Figure 6 Schematic diagram of the steps for multi-dimensional threat quantification and decision output in the present invention. DETAILED DESCRIPTION

[0057] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0058] Example 1

[0059] like Figures 1-6As shown, this embodiment provides a method for calculating FOD for accurate identification of falling objects based on a fixed tower, comprising the following steps:

[0060] 1. Synchronous collection of multi-source heterogeneous sensor data, such as Figure 2 As shown;

[0061] S1: Deploy a multimodal sensor group with visual, spatial, and acoustic sensing capabilities on the main structure of the fixed tower and the surrounding monitoring area. The sensor group includes a high-definition imaging device, a laser scanning device, and a distributed acoustic sensing device.

[0062] In this embodiment, a multimodal sensor cluster is installed on the main structure of the fixed tower in S1 at the bottom, middle, and top of the tower. This layered installation design enables three-dimensional monitoring in both the vertical direction (ground to high altitude) and the horizontal direction (360° without blind spots).

[0063] The equipment on the top of the tower can take aerial photos of a large area and capture foreign objects at a distance;

[0064] The equipment in the tower can achieve fine scanning at mid-level heights to avoid obstructions;

[0065] The equipment at the bottom of the tower is close to the ground, which improves the detection sensitivity of small foreign objects in low altitude and on the ground (such as debris less than 5 cm);

[0066] Redundant detection is formed by the overlapping fields of view of sensors at different heights, reducing the risk of missed detection from a single perspective. For example, when the camera on the top of the tower detects a suspected foreign object, the laser radar in the tower can quickly scan the area to obtain three-dimensional coordinates, thereby improving positioning accuracy.

[0067] In this embodiment, the multimodal sensor group deployed in S1 is specifically configured as follows:

[0068] High-definition camera array: Integrates polarized light imaging and multispectral imaging capabilities, with a resolution of at least 3840×2160 and a frame rate of at least 60fps. This design supports multispectral fusion of visible light, infrared, and ultraviolet light, adapting to complex lighting environments (such as nighttime and highly reflective scenes). It improves the ability to extract edge features of foreign objects in low-light conditions by 40%, ensuring continuous and clear imaging of dynamic scenes (such as objects dropped during taxiing), with a motion blur error of ≤2 pixels.

[0069] 3D LiDAR: Using SPAD array detectors, the point cloud density is no less than 500 points / m², and the minimum resolvable distance difference is no more than 2cm. The SPAD array detector's single-photon avalanche diode technology enables millimeter-level distance measurement with a minimum resolvable distance difference of 2cm, supporting 3D contour reconstruction of tiny foreign objects (such as 2cm×2cm metal pieces). The point cloud density is 2 times higher than that of traditional LiDAR, and the accuracy rate of foreign object contour recognition in complex terrain (such as runway seams and apron markings) is ≥95%.

[0070] Millimeter-wave radar: Utilizing a MIMO architecture, the transmission frequency ranges from 76 to 81 GHz, and the 3D point cloud refresh rate is no less than 100 Hz. The MIMO architecture's multiple-input, multiple-output technology enables 128-channel signal reception with an angular resolution of ≤1°, enabling the distinction between co-directionally moving foreign objects and background objects (e.g., a taxiing aircraft and ground objects). The high refresh rate (100 Hz) enables real-time capture of high-speed moving foreign objects (e.g., debris ingested by an engine during takeoff), with trajectory prediction latency ≤10ms.

[0071] Infrared thermal imager: The thermal sensitivity is no higher than 0.03°C at 30°C, and it features an integrated non-uniformity correction algorithm. This high sensitivity facilitates the detection of non-metallic foreign objects with temperature differences of 0.05°C or greater (such as hot brake pad fragments that have just fallen to the ground), addressing the blind spots of visible light cameras at night or in hidden locations. The non-uniformity correction algorithm eliminates temperature deviations between sensor pixels, reducing thermal image noise by 60% and improving detection reliability in low-temperature environments (below -20°C).

[0072] As a further illustration of this embodiment, a damping shock absorber is installed between the tower and the sensor mounting base, so that the natural frequency avoids the tower vibration frequency (50 Hz), and the point cloud data error caused by vibration noise is ≤5 mm.

[0073] S2: Achieve collaborative multimodal data acquisition through a synchronous trigger mechanism, obtain RGB image data, laser point cloud data, and acoustic signal data, and establish a spatiotemporal mapping relationship;

[0074] As a further illustration of this embodiment, the sensor synchronization and calibration mechanism may adopt a spatiotemporal synchronization solution, which includes: hardware-level synchronization and data alignment algorithm: hardware-level synchronization adopts the IEEE1588 precision clock protocol, and the multi-sensor timestamp error is ≤0.5ms; the tower coordinate system is calibrated by the target, and the spatial coordinate deviation of each sensor is ≤3cm; the data alignment algorithm establishes a multimodal data spatiotemporal mapping table, based on timestamp synchronization triggering (such as synchronously collecting full-modal data once every 50ms), to ensure that the RGB image pixels correspond one-to-one with the coordinates of the laser point cloud and millimeter wave point cloud.

[0075] It should be added that the sensor collaborative workflow adopts a hierarchical detection mechanism, including:

[0076] Coarse detection: The millimeter-wave radar scans in real time and triggers the ROI (region of interest) focus of the camera and lidar after detecting a moving target (speed > 0.5m / s);

[0077] Fine identification: The camera's multispectral image is integrated with the laser point cloud to extract the texture (such as metal reflections) and geometric features (such as length, width, and height) of the foreign object. The SVM classifier is used to distinguish the FOD type (metal / non-metal / organic matter), with a classification accuracy of ≥92%;

[0078] Position verification: Acoustic sensors detect impact noise (frequency 200Hz-5kHz) and use time difference of approach (TDOA) to verify the location of the landing of foreign objects, eliminating false alarms (such as those triggered by birds flying over).

[0079] Furthermore, in the multimodal data collaborative acquisition in step S2, the synchronization trigger module uses a dedicated ASIC chip (such as the Altera ACEX1K series) to achieve high-precision collaboration of the sensor group through a hardware-level synchronization mechanism. The specific design is as follows:

[0080] Synchronization mechanism: The ASIC chip generates a hardware trigger signal every 50ms, which synchronizes the acquisition actions of all sensors (such as HD cameras, LiDAR, and millimeter-wave radar) through the GPIO interface to ensure the spatiotemporal alignment of RGB images, laser point clouds, and acoustic signals.

[0081] Time synchronization: Integrates the IEEE1588 precision clock protocol, with multi-sensor timestamp error ≤ 0.5ms; the tower coordinate system is calibrated by a target, and the spatial coordinate deviation of each sensor is ≤ 3cm, meeting aviation-grade synchronization accuracy requirements;

[0082] Anti-interference design: The chip integrates ClockLock™ and ClockBoost™ technologies to reduce clock offset caused by tower vibration (natural frequency 50Hz), ensuring stable data acquisition in dynamic scenarios.

[0083] 2. Multimodal data preprocessing and feature extraction, such as Figure 3 As shown;

[0084] S3: Perform noise reduction and contrast enhancement on RGB image data to suppress noise and highlight the edges and texture features of the falling objects; remove outliers and downsample the laser point cloud data to retain valid spatial geometric features; perform frame processing and frequency domain feature extraction on the acoustic signal data to filter out environmental background noise;

[0085] In this embodiment, the preprocessing method of different modal data in S3 is as follows:

[0086] RGB image denoising: This system uses a wavelet-domain dual-tree complex transform (DTCT) to dynamically set the denoising threshold based on a noise statistical model. Compared to traditional wavelet transforms, DTCT boasts shift invariance and multi-directional selectivity, more accurately preserving image edge texture details (such as metallic reflective edges and irregular contours of FOD). It improves noise suppression by 30% in low-light environments, maintaining ≥95% edge feature integrity. Furthermore, it uses a real-time noise intensity estimate based on noise statistical models (such as the Gaussian mixture model) to adaptively adjust the denoising threshold, avoiding detail loss or residual noise caused by a fixed threshold. This improves contrast by 40% for objects in complex backgrounds (such as runway markings and tarmac oil stains).

[0087] Laser point cloud processing: Utilizing an improved RANSAC algorithm, outlier removal is adaptively performed based on point cloud distribution. To address high-frequency vibration environments at airports (such as tower tremors caused by aircraft takeoffs and landings), the improved RANSAC algorithm dynamically adjusts the inlier distance threshold (calculated in real time based on the point cloud density distribution), increasing outlier removal accuracy from 85% with traditional RANSAC to 95%. This algorithm also retains ≥90% of valid point cloud data, ensuring that even the outline features of tiny foreign objects (such as 2cm³ metal blocks) are not lost. Furthermore, by introducing a random sampling consistency acceleration strategy (such as prioritizing point cloud sampling in high-density areas), processing time is reduced by 40% compared to traditional algorithms, meeting real-time requirements (single-frame point cloud processing ≤20ms).

[0088] Acoustic Signal Extraction: This system extracts improved CQT (Constant-Q Transform) time-frequency features that conform to the critical bandwidth theory of hearing. By integrating this theory with the Bark scaling transform (e.g., Bark scaling), the improved CQT (Constant-Q Transform) time-frequency features offer enhanced discrimination for typical airport noise (100-500Hz low-frequency noise from aircraft engines and 5-10kHz high-frequency noise from ground vehicles). This improves the detection sensitivity of FOD impact signals (e.g., broadband pulse signals from falling metal) in noisy environments by 60%. By adaptively adjusting the Q factor (Q = 20-40, dynamically varying), balancing time-domain resolution (high-speed impact signals) with frequency-domain resolution (low-frequency vibration signals), the energy concentration of foreign object signals in the time-frequency plot is 25% higher than that of traditional CQT.

[0089] As a further illustration of this embodiment, in laser point cloud processing, an improved dynamic adaptive RANSAC algorithm is used to solve the noise problem in the high-frequency vibration environment of the airport: the local density of the point cloud is estimated in real time by the K nearest neighbor method (K=15), and the inlier distance threshold is dynamically calculated using the "mean + 2 times the standard deviation" ( ), increasing the accuracy of outlier removal from 85% to 95% with traditional algorithms while retaining ≥90% of valid point cloud data (e.g., 92% completeness of the contour features of a 2cm³ metal block). Furthermore, a high-density region-priority sampling strategy based on a KD tree reduces sampling times by 40%, reducing single-frame processing time to ≤20ms, meeting the processing requirements of real-time LiDAR data streams (100Hz). This algorithm, through the synergy of density-adaptive thresholding and stratified sampling, achieves dual optimization for vibration noise suppression and small foreign object feature retention, providing a high-precision point cloud data foundation for subsequent geometric feature extraction (such as centroid and curvature calculation).

[0090] S4: Extracting single-modal deep features through deep learning models: extracting visual feature vectors containing texture and shape from RGB images; extracting geometric feature vectors containing centroid and curvature from laser point cloud data; extracting temporal feature vectors containing frequency and amplitude changes from acoustic wave signals;

[0091] To further illustrate this example, visual features (texture + shape) address the question of "what" (FOD type identification), geometric features (center of mass + curvature) address the question of "where" (3D positioning), and temporal features (frequency + amplitude) address the question of "when" (dynamic trajectory tracking). These three elements form a comprehensive feature representation encompassing "type, location, and timing." Even when a single modality fails (e.g., camera occlusion), the other modal features still support detection, reducing the missed detection rate by 50%. Furthermore, the deep learning model, trained on large-scale airport scene data (including over 100,000 FOD samples and 500 hours of noise data), automatically optimizes feature extraction parameters, improving its adaptability to extreme environments such as low illumination (<5 lux), high vibration (tower displacement ±10 cm), and strong electromagnetic interference (signal distortion ≥20%) by 30%, eliminating the need for manual adjustment of feature engineering parameters. In addition, the feature vector dimensions are unified (2048 dimensions for vision / 1024 dimensions for geometry / 512 dimensions for time series) and scale differences are eliminated through layer normalization to ensure efficient feature interaction in subsequent fusion modules (such as attention mechanism and graph neural network). The multimodal fusion efficiency is improved by 40% compared with traditional manual features, and feature redundancy is reduced by 35%.

[0092] 3. Cross-modal feature fusion and representation construction, such as Figure 4 As shown;

[0093] S5: Dimensional alignment and semantic space fusion of bimodal features are performed, and a two-tuple feature vector containing the association between modalities is generated through a weighted strategy or attention mechanism;

[0094] In this embodiment, the bimodal feature semantic space alignment in S5 is achieved by calculating the mutual information entropy matrix between modalities; the mutual information entropy calculation formula is: ,in and are the single modal feature entropy, is the joint entropy; in this design, by maximizing the mutual information entropy, the weight of complementary features is made to account for no less than 60% to suppress redundant information between modalities. The mutual information entropy matrix is ​​used to calculate the dependency between modalities (such as the correlation between visual texture features and the coordinates of the centroid of the laser point cloud), and the abstract modal association is converted into a computable quantitative indicator, which improves the feature complementarity by 40% compared with the traditional weighted average method (such as simple addition). At the same time, by maximizing the mutual information entropy strategy (complementary feature weight ≥ 60%), redundant information (such as the background color in the visual image and the invalid coordinates of the laser point cloud) is effectively suppressed, which reduces the information entropy of the fused features by 30% and improves the efficiency of subsequent classifiers (such as the Softmax calculation speed is increased by 25%). In addition, this design solves the semantic gap problem between visual features (RGB pixel values) and geometric features (three-dimensional coordinates). Establish feature space mapping relationships, such as aligning the highly reflective visual features of metal foreign objects with the high reflectivity values ​​of the laser point cloud, to improve the clustering compactness of similar features (reducing the intra-class distance by 20%).

[0095] It should be added that in the bimodal feature semantic space alignment in step S5, the feature types involved in the mutual information entropy calculation are defined as follows:

[0096] is the visual feature vector extracted by the S4 step (2048 dimensions, including texture and shape information), is a geometric feature vector (1024 dimensions, including geometric properties such as centroid coordinates and curvature);

[0097] Unimodal entropy The calculation method is as follows: First, the visual features are reduced to 50 dimensions by PCA (keeping 95% variance), and the bandwidth of the Gaussian kernel density estimation is determined by Silverman's rule. ( is the sample standard deviation, is the number of training samples), calculate the probability density and estimate the entropy value based on the features after dimensionality reduction to ensure the accuracy of entropy calculation of high-dimensional features.

[0098] Joint Entropy Through mutual information neural network (MINE) approximation estimation, the network adopts two fully connected layers (hidden layer 256 dimensions), the input is visual-geometric splicing features (100 dimensions), and the output mutual information lower bound is used to derive the joint entropy to achieve quantitative modeling of the dependency relationship between modalities.

[0099] Therefore, through the above method, the weight ratio of complementary features can be automatically adjusted by maximizing the mutual information entropy strategy, effectively suppressing redundant information between modalities and improving the discriminative ability of bimodal fusion features.

[0100] S6: Based on modal reliability assessment, adaptive weighted fusion of all modal features is performed to generate a joint representation vector containing multi-dimensional information of vision, geometry, and acoustics;

[0101] As a further explanation of this embodiment, this embodiment further supplements the details of mutual information entropy calculation: 1. The feature distribution estimation method includes: single modal entropy , uses kernel density estimation (KDE) to fit the feature vector distribution (such as the 2048-dimensional Gaussian mixture model of visual features), solving the dimensionality curse problem of high-dimensional feature entropy calculation, and reducing the computational complexity by 60% compared with the traditional histogram method. : Through the mutual information neural network (MINE) approximate estimation, the lower bound of mutual information is optimized using gradient ascent, and the convergence speed is increased by 3 times compared with the non-parametric method, which is suitable for real-time online computing (single frame processing ≤ 10ms); 2. Dimension alignment technology solution, the visual features (2048 dimensions) and geometric features (1024 dimensions) are reduced to 512 dimensions using a fully connected layer, and the acoustic time series features (128 dimensions) are expanded to 512 dimensions through zero padding to ensure the consistency of the mutual information entropy matrix dimension (512×512), avoiding calculation errors caused by dimensional mismatch.

[0102] 4. Spatiotemporal graph neural network modeling and reasoning, such as Figure 5 As shown;

[0103] S7: Mapping unimodal and fused features into graph nodes, constructing a dynamic graph structure based on spatiotemporal correlation. Nodes include unimodal nodes (unimodal), fused nodes (bimodal), and global nodes (triples). Edge connections are determined by feature similarity and spatiotemporal correlation. Hierarchical feature modeling from single modality to global fusion is achieved through unimodal (unimodal), bimodal (bimodal), and triad (all-modal joint representation) nodes. For example, unimodal nodes (such as visual features) capture local details, while triad nodes integrate multimodal complementary information. This addresses the problem of missing feature dimensions caused by the single node type in traditional graph models, improving feature representation capabilities by 40% in complex scenarios.

[0104] In this embodiment, the dynamic graph structure constructed in S7 adopts a hybrid propagation mechanism, in which the message passing function fuses node features and edge weights; and the node update adopts a gated GRU module, integrating spatiotemporal residual connections. The advantages of the hybrid propagation mechanism are: fusing node features (such as visual texture vectors) and edge weights (such as the inverse of spatiotemporal distances), and through the weighted aggregation formula (in is the node feature, is the edge weight), realizing the joint modeling of feature similarity and spatiotemporal proximity. Compared with the traditional GCN propagation method based only on node features, the efficiency of neighborhood information utilization is improved by 25%; at the same time, a temporal gating mechanism is introduced. , dynamically adjusts the fusion ratio of historical features and current features, and improves the ability to capture the temporal dependence of multi-frame falling object trajectories (such as 10 frames of continuous signals of rolling foreign objects) by 15% compared to LSTM, while reducing the computational complexity by 30%.

[0105] S8: A spatiotemporal graph convolutional network aggregates neighborhood node features, combined with a time series modeling unit to capture the motion patterns of falling objects, generating a feature representation that incorporates spatiotemporal correlations. Edge connections are constructed based on feature similarity (e.g., cosine similarity > 0.7) and spatiotemporal correlation (spatiotemporal distance between adjacent frames < 5 cm), dynamically adapting to changes in the trajectory of falling objects. Compared to static graph structures (with fixed edge connections), the accuracy of capturing the trajectory of high-speed moving foreign objects (such as objects falling near taxiing aircraft) is improved by 35%.

[0106] As a further illustration of this embodiment, by aggregating neighborhood node features through spatiotemporal graph convolution kernels (e.g., 3×3 spatiotemporal windows) and simultaneously capturing spatial proximity (nodes within 5 meters) and temporal continuity (a time window spanning 3 frames before and 2 frames after), the feature recovery capability for occluded scenarios (e.g., a foreign object briefly obscured by an aircraft landing gear) is improved by 50%, reducing the missed detection rate from 12% to below 6%. Furthermore, through the deep fusion of temporal modeling units and graph convolution, the classification and recognition of falling object motion patterns (parabolic, rolling, and sliding) is supported. For example, by combining temporal features extracted by LSTM with spatial features from graph convolution, the classification accuracy for complex motion trajectories reaches ≥92%.

[0107] For example, when a taxiing aircraft detects a high-speed moving target (20 m / s) from a millimeter-wave radar, triggering the creation of graph nodes: a vision node captures the reflective metallic texture near the tire, a laser point cloud node records 10 cm × 5 cm flake-like geometric features, and an acoustic node detects continuous impact signals at 50 ms intervals. The graph network aggregates these three types of nodes using spatiotemporal edges (with a time difference of 80 ms and a spatial distance of 3 m), identifying them as "brake pad debris." The trajectory prediction indicates that the target will enter the critical runway area in 3 seconds, triggering a red alert in advance.

[0108] 5. Multi-dimensional threat quantification and decision-making output, such as Figure 6 As shown;

[0109] S9: Based on the joint spatiotemporal feature expression, the physical model and data-driven algorithm are used to calculate the damage risk level of falling objects, predict the three-dimensional motion trajectory, and output the positioning confidence area;

[0110] In this embodiment, the threat assessment is performed in S9 as follows:

[0111] Damage level calculation: Integrating material stress-strain constitutive models; this design integrates material mechanics constitutive relations (such as the elastic-plastic model for metals and the viscoelastic model for plastics) with parameters such as falling object mass, velocity, and impact angle to calculate structural stress distribution (such as the maximum principal stress in runway concrete) and plastic deformation. This improves the accuracy of damage level assessment by 25% compared to purely data-driven methods (such as neural network regression), and particularly reduces the assessment error for new, untrained foreign objects (such as carbon fiber fragments) by 40%. A three-dimensional impact dynamics model is established by considering aerodynamic drag (calculated using the Reynolds number) and impact energy attenuation (based on the strain energy density formula). The kinetic energy conversion of falling objects (such as a component dropped from a 20-meter tower) is calculated with an accuracy of ±5%, addressing the risk underestimation caused by traditional methods that ignore physical mechanisms.

[0112] Trajectory Prediction: Utilizing an interactive multi-model algorithm, the system integrates the CTRA (Constant Roll Rate Acceleration) and CS (Constant Speed) motion models. This design dynamically switches between the CTRA (Constant Roll Rate Acceleration) and CS (Constant Speed) models through model probability weighting (e.g., Kalman filter residual covariance determination). This achieves trajectory prediction errors of ≤15cm (a 30% reduction compared to a single model) for complex motions such as straight-line taxiing (CS model, probability ≥0.7) and turning and rolling (CTRA model, probability ≥0.6), meeting the airport's positioning accuracy requirements for rapid response (error ≤20cm). Furthermore, the system utilizes a second-order Markov chain to model transition probabilities, with an update period of ≤20ms. This ensures trajectory continuity of ≥95% in scenarios such as drone interference and ground vehicle occlusion, avoiding the tracking loss associated with traditional single-model systems when motion patterns suddenly change.

[0113] S10: Generate a structured decision report including threat level, positioning results and emergency response recommendations, and output a visual threat situation map with a data traceability chain.

[0114] As a further explanation of this embodiment, the structured decision report specifically includes the following content:

[0115] Multi-dimensional information integration: This includes threat level (1-5, combining damage energy and impact area), positioning results (3D coordinates ± error ellipse), and emergency response recommendations (such as the area of ​​partial apron blockade and the type of foreign object removal equipment). This improves decision-making efficiency by 40% compared to traditional single-alarm information, reducing manual analysis time.

[0116] Data traceability chain visualization: A data link is established through time and space stamps (accuracy of 1ms) and sensor ID codes (such as tower base camera C01 and tower radar L02). This allows for reverse retrieval of original sensor data (such as acoustic waveforms and laser point cloud clustering at the moment of impact) from the threat point, ensuring an event tracing error of ≤2 seconds, meeting the strict evidence collection requirements of aviation safety accident investigations.

[0117] It should be added that the impact energy calculation formula in this embodiment is: , where is the air resistance coefficient and is the critical speed, which solves the problem of calculating the kinetic energy attenuation of objects dropped from high altitude and ensures the energy calculation error of objects falling from a height of 200m.

[0118] Furthermore, this embodiment further optimizes the trajectory prediction algorithm, the specific contents are as follows:

[0119] The implementation details of the IMM algorithm are as follows: the model set is expanded to include a new CA (constant acceleration) model, forming a three-model combination (CS / CTRA / CA), which covers more than 95% of projectile motion scenarios (straight-line acceleration / deceleration, curved motion);

[0120] : ;

[0121] This design ensures smooth switching between models and reduces the misjudgment rate of motion modes (e.g., the misjudgment rate from CS to CA is ≤5%).

[0122] Initial state estimation: Use the first three frames of point cloud data to fit the initial velocity (least squares method, error ≤ 0.5m / s) as the initial input of the IMM algorithm;

[0123] Supplement on the calculation of the positioning confidence area: Based on the Kalman filter covariance matrix, a three-dimensional ellipsoid area with 95% confidence is generated (the major semi-axis = 2 times the position error standard deviation, the minor semi-axis = 1.5 times). For example, for a foreign object moving at 10m / s, the volume of the confidence area is 20% smaller than the traditional Gaussian distribution estimate, thereby improving the operating efficiency of the foreign object removal equipment.

[0124] Furthermore, in the construction of the data traceability chain in step S10, the Hyperledger Fabric consortium chain architecture is adopted to achieve the tamper-proof and full-process traceability of the test results. The specific design is as follows:

[0125] Consortium chain architecture: The consortium chain is composed of multiple nodes including airport management, air traffic control departments, equipment suppliers, etc. Each block contains the SHA-256 hash value of 100 test results, combined with the timestamp provided by the NTP server (accuracy ±0.5ms) to ensure the precise anchoring of event time.

[0126] Data traceability chain: A data link is established through sensor IDs (such as tower bottom camera C01 and tower radar L02) and time and space stamps (accuracy 1ms). This supports reverse retrieval of original sensor data (such as acoustic waveforms and laser point cloud clusters at the moment of impact) from the threat point. The event tracing error is ≤ 2 seconds, meeting the strict evidence collection requirements of aviation safety accident investigations.

[0127] Smart contracts: Chaincode automates the data on-chain process, hashing and storing key information such as threat levels and positioning results to ensure data cannot be tampered with. For example, when the system detects a foreign object entering a critical area of ​​a runway, the smart contract automatically packages the warning information and raw sensor data into a block and broadcasts it to consortium chain nodes, enabling real-time auditing and accountability.

[0128] Example 2

[0129] This embodiment also provides the application of a FOD calculation method for accurately identifying falling objects based on a fixed tower, based on the application of the FOD calculation method for accurately identifying falling objects based on a fixed tower in the above-mentioned embodiment 1 to runways and aprons in central airports.

[0130] Understandably, in this embodiment, the sensor deployment and monitoring focus differ significantly between airport runways and aprons. As the core area for aircraft takeoff and landing, the runway presents high security risks and time pressures. Therefore, a high-density deployment of towers (e.g., one every 50 meters) is required along both sides. Sensors at the top of the towers focus on high-altitude observation, capturing dynamic objects such as detached components and falling tires during takeoff and landing. Sensors at the bottom of the towers focus on the ground, detecting small foreign objects such as metal debris in runway cracks. Because targets on the runway move at high speeds (≥20 m / s), the sensor refresh rate must be ≥100 Hz to ensure continuous monitoring of the trajectory of high-speed dynamic targets. The apron, on the other hand, is an area for aircraft parking, maintenance, and cargo loading and unloading. Threats are primarily static or low-speed falling objects, such as lost maintenance tools and objects dropped during cargo loading and unloading. The towers can be evenly spaced 100-200 meters apart. Sensors at the top of the towers cover a large area where multiple aircraft are parked, while sensors in the towers focus on monitoring maintenance areas, such as those near the landing gear and cargo hold. To meet the needs of detecting non-metallic foreign objects (plastic parts, fabrics, etc.), it is necessary to optimize the sensitivity of the infrared and ultraviolet channels of the multispectral camera to improve the ability to identify such foreign objects. This is in stark contrast to the runway's focus on monitoring dynamic, high-speed targets, ensuring the accuracy and effectiveness of falling object monitoring in different scenarios.

[0131] Those skilled in the art will appreciate that the process of implementing all or part of the steps of the above embodiments may be accomplished by hardware, or by instructing related hardware through a program, which may be stored in a computer-readable storage medium.

[0132] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The FOD calculation method for accurate identification of falling objects based on a fixed tower is characterized by: The steps include:

1. Synchronous collection of multi-source heterogeneous sensor data; S1: Deploy a multimodal sensor group with visual, spatial, and acoustic sensing capabilities on the main structure of the fixed tower and the surrounding monitoring area. The sensor group includes a high-definition imaging device, a laser scanning device, and a distributed acoustic sensing device. S2: Achieve collaborative multimodal data acquisition through a synchronous trigger mechanism, obtain RGB image data, laser point cloud data, and acoustic signal data, and establish a spatiotemporal mapping relationship; 2. Multimodal data preprocessing and feature extraction; S3: Perform noise reduction and contrast enhancement on the RGB image data to suppress noise and highlight the edges and texture features of the falling objects; Eliminate outliers and downsample laser point cloud data to retain valid spatial geometric features; Implement frame processing and frequency domain feature extraction on the sound wave signal data to filter out environmental background noise; S4: Extracting single-modal deep features through deep learning models: extracting visual feature vectors containing texture and shape from RGB images; extracting geometric feature vectors containing centroid and curvature from laser point cloud data; extracting temporal feature vectors containing frequency and amplitude changes from acoustic wave signals; 3. Cross-modal feature fusion and representation construction; S5: Dimensional alignment and semantic space fusion of bimodal features are performed, and a two-tuple feature vector containing the association between modalities is generated through a weighted strategy or attention mechanism; S6: Based on modal reliability assessment, adaptive weighted fusion of all modal features is performed to generate a joint representation vector containing multi-dimensional information of vision, geometry, and acoustics; 4. Spatiotemporal graph neural network modeling and reasoning; S7: Map the unimodal features and fusion features into graph nodes, and construct a dynamic graph structure based on spatiotemporal correlation. The nodes include unimodal nodes, fusion nodes, and global nodes. Edge connections are determined by feature similarity and spatiotemporal correlation. S8: Aggregate neighborhood node features through spatiotemporal graph convolutional networks, combine with temporal modeling units to capture the motion patterns of falling objects, and generate feature expressions that include spatiotemporal correlations; 5. Multi-dimensional threat quantification and decision-making output; S9: Based on the joint spatiotemporal feature expression, the physical model and data-driven algorithm are used to calculate the damage risk level of falling objects, predict the three-dimensional motion trajectory, and output the positioning confidence area; S10: Generate a structured decision report including threat level, positioning results and emergency response recommendations, and output a visual threat situation map with a data traceability chain.

2. The method for calculating FOD for accurate identification of falling objects based on a fixed tower according to claim 1 is characterized in that: On the main structure of the fixed tower in S1, a multimodal sensor group is installed at the bottom, middle and top of the tower respectively.

3. The method for calculating FOD for accurate identification of falling objects based on a fixed tower according to claim 1 is characterized in that: The specific configuration of the multimodal sensor group deployed in S1 is as follows: High-definition camera array: integrated with polarized light imaging and multispectral imaging capabilities, with a resolution of no less than 3840×2160 and a frame rate of no less than 60fps; 3D LiDAR: Using SPAD array detectors, the point cloud density is not less than 500 points / m², and the minimum resolvable distance difference is not greater than 2cm; Millimeter-wave radar: uses a MIMO architecture, with a transmission frequency between 76-81 GHz and a 3D point cloud refresh rate of no less than 100 Hz; Infrared thermal imager: Thermal sensitivity is no higher than 0.03°C at 30°C, and it has an integrated non-uniformity correction algorithm.

4. The method for calculating FOD for accurate identification of falling objects based on a fixed tower according to claim 1 is characterized in that: The preprocessing method of different modal data in S3 is as follows: RGB image denoising: uses wavelet domain dual-tree complex transform and dynamically sets the denoising threshold based on the noise statistical model; Laser point cloud processing: Using the improved RANSAC algorithm, outliers are adaptively removed according to the point cloud distribution; Acoustic signal extraction: Extract the improved CQT time-frequency features that conform to the critical bandwidth theory of hearing.

5. The method for calculating FOD for accurate identification of falling objects based on a fixed tower according to claim 1 is characterized in that: The bimodal feature semantic space alignment in S5 is achieved by calculating the mutual information entropy matrix between modalities; the mutual information entropy calculation formula is: ,in and are the single modal feature entropy, is the joint entropy; by maximizing the mutual information entropy, the complementary feature weight is made to account for no less than 60% to suppress redundant information between modalities.

6. The method for calculating FOD for accurate identification of falling objects based on a fixed tower according to claim 1 is characterized in that: The closed-loop optimization method adopted in S6 is as follows: Sensor drift compensation: Using a dual Kalman filter architecture, the main filter is used to update the state estimate, and the auxiliary filter is used to identify the deviation parameters online; Parameter optimization: Use a meta-learning framework to dynamically adjust model weights.

7. The method for calculating FOD for accurate identification of falling objects based on a fixed tower according to claim 1 is characterized by: The dynamic graph structure constructed in the S7 adopts a hybrid propagation mechanism, and the message passing function fuses node features and edge weights; and the node update adopts a gated GRU module and integrates spatiotemporal residual connections.

8. The method for calculating FOD for accurate identification of falling objects based on a fixed tower according to claim 1 is characterized in that: The method of performing threat assessment in S9 is as follows: Damage degree calculation: integrating material stress-strain constitutive model; Trajectory prediction: Adopts interactive multi-model algorithm, integrating CTRA and CS motion models.

9. The method for calculating FOD for accurate identification of falling objects based on a fixed tower according to claim 1 is characterized in that: In the entire data processing flow from data collection in S2 to result output in S10, the following data security and storage measures are adopted: Data encryption: RGB image data, laser point cloud data, acoustic signal data and processing results are encrypted using the AES-256 symmetric encryption algorithm during transmission and storage, with a key length of 256 bits; Blockchain evidence storage: The SHA-256 hash value of the test results and decision reports is stored on the blockchain node network. The timestamp of the evidence is generated by a trusted timestamp server with an accuracy of no less than 1ms. Data backup: Adopting an off-site disaster recovery backup strategy, the processed data is regularly backed up to a secure data center ≥50 km away from the monitoring area through an encrypted data transmission channel, with a backup cycle of 24 hours per time. Access control: Establish a role-based access control system and assign fine-grained access rights to different roles through the permission management module; The authentication mechanism combines username / password, dynamic password and biometrics for multi-factor authentication.

10. The application of the FOD calculation method for accurate identification of falling objects based on a fixed tower is characterized by: The application of the FOD calculation method for accurate identification of falling objects based on a fixed tower as described in any one of claims 1 to 9 to runways and aprons of medium-sized airports.

Citation Information

Patent Citations

  • Airport Foreign Object Radar Detection Method Based on Time-Directed Statistics

    CN108226890B

  • Airport foreign matter detection system and airport foreign matter inspection system

    CN117274652A

Cited By

  • Multi-modal detection information fusion method based on environment evaluation

    CN121679567A