Multi-scene-oriented low-altitude navigation multi-source heterogeneous data adaptive fusion method

Abstract

The invention belongs to the technical field of low-altitude navigation safety monitoring, and particularly relates to a low-altitude navigation multi-source heterogeneous data adaptive fusion method for multiple scenes, which is a low-altitude navigation multi-source heterogeneous data adaptive fusion method for multiple scenes such as urban air traffic, low-altitude logistics and emergency rescue. Efficient integration of multiple types of monitoring data and position output of the trusted aircraft can be achieved, and accurate data support is provided for low-altitude navigation anomaly recognition and risk deduction. The method comprises the following specific steps: constructing a low-altitude navigation scene classification system and fusion demand mapping, collecting and preprocessing multi-source heterogeneous data, quantifying data credibility and resolving conflicts, constructing a hierarchical adaptive fusion framework and outputting a fusion result. According to the method, efficient and accurate fusion of multi-source data in different scenes is realized by constructing a layered scene adaptation framework and a credibility fusion model, and finally, a high-credibility aircraft position is output, so that reliable data support is provided for low-altitude navigation anomaly recognition and risk management and control.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude flight safety monitoring technology. Specifically, it relates to an adaptive fusion method for multi-source heterogeneous data of low-altitude flight in multiple scenarios, such as urban air traffic, low-altitude logistics, and emergency rescue. It can achieve efficient integration and reliable output of multiple types of monitoring data, and provide accurate data support for low-altitude flight anomaly identification and risk prediction. Background Technology

[0002] In recent years, the operational complexity of low-altitude airspace has gradually increased, with the intermingling of aircraft types, overlapping flight paths, and diverse mission requirements becoming increasingly prominent. To ensure navigation safety in this complex environment, it is essential to conduct real-time monitoring of multi-dimensional key low-altitude navigation data and to perform in-depth fusion analysis of the obtained multi-source heterogeneous low-altitude data to achieve effective control over the low-altitude situation.

[0003] Among existing low-altitude navigation multi-source data fusion methods, CN116124134B discloses a multi-source data fusion positioning and navigation system, which determines the combined navigation information of the UAV during flight and plans the UAV's flight navigation information based on the combined navigation information and the UAV's current position information. However, it does not consider the reliability differences of data from different sources, and is prone to completely incorporating low-reliability data into the fusion process, making it difficult to meet the high requirements of low-altitude safety management for data fusion accuracy. CN120740607A discloses a UAV visual positioning flight optimization system based on multi-source data fusion, which uses multi-source data fusion and state prediction compensation mechanisms to address issues such as positioning interruption caused by visual signal loss. However, it fails to design differentiated fusion strategies for the diverse task requirements and highly dynamic environmental characteristics in low-altitude scenarios.

[0004] Therefore, there is an urgent need for an adaptive fusion method for multi-source heterogeneous data in low-altitude navigation that can adapt to the characteristics of multiple scenarios, ensure the credibility of data fusion, and balance efficiency and accuracy, so as to provide reliable data fusion support for low-altitude navigation safety. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive fusion method for multi-source heterogeneous data in low-altitude navigation for multiple scenarios. By constructing a hierarchical scenario adaptation framework and a credibility fusion model, it achieves efficient and accurate fusion of multi-source data in different scenarios, providing reliable data support for anomaly identification and risk management in low-altitude navigation.

[0006] The technical solution of the present invention is as follows:

[0007] An adaptive fusion method for multi-source heterogeneous data in low-altitude navigation for multiple scenarios, the specific steps of which are as follows:

[0008] Step 1: Construct a classification system and fusion requirement mapping for low-altitude flight scenarios. The specific process is as follows:

[0009] Based on the risk level, data integrity, and operational complexity of low-altitude navigation, the scenarios are divided into low-complexity, medium-complexity, and high-complexity scenarios, and the corresponding fusion requirements are clearly defined for the characteristics of different scenarios.

[0010] Low-complexity scenarios include routine urban air travel and interference-free low-altitude logistics, characterized by a single aircraft type, stable operating environment, and high data integrity. Medium-complexity scenarios include low-altitude logistics in complex urban areas and routine emergency patrols, characterized by diverse aircraft types and minor environmental interference. High-complexity scenarios include special situations such as emergency rescue in mountainous areas and low-altitude operations under severe weather conditions, involving mixed aircraft types, severe environmental interference, and low data integrity.

[0011] Step 2: Multi-source heterogeneous data acquisition and preprocessing, the specific process is as follows:

[0012] Step 2.1, Data Acquisition

[0013] Data is collected using various monitoring devices, covering multiple dimensions such as aircraft, ground infrastructure, and airspace management systems, forming a multi-source heterogeneous data input system; the data includes airborne data, ground data, environmental data, and control data.

[0014] Airborne data is directly collected by sensors on board the aircraft, including trajectory and attitude data. Trajectory data records latitude, longitude, and altitude information, while attitude data captures changes in roll, pitch, and yaw angles. Different sampling frequencies are set for each type of airborne data according to monitoring requirements. Ground-based data originates from ground infrastructure and is divided into positioning and environmental data: Positioning data includes ADS-B and radar data. ADS-B data provides aircraft position and speed information, while radar data enables higher-precision positioning monitoring. The sampling frequencies of both are set differently based on positioning accuracy requirements. Environmental data encompasses meteorological and electromagnetic spectrum data. Meteorological data records environmental parameters such as wind speed and direction, while electromagnetic spectrum data monitors interference signal strength, providing a basis for analyzing the impact of the environment on navigation. Control data is uniformly provided by the airspace management system, including text-based and rule-based data. Text-based data clearly defines spatial and altitude constraints such as temporary no-fly zone coordinates and flight altitude restrictions, while rule-based data defines operational rules such as aircraft priority and airspace access permissions, providing compliance references for data fusion.

[0015] Step 2.2, Data Preprocessing

[0016] This includes data cleaning and spatiotemporal alignment, which aim to eliminate data noise, fill data gaps, and achieve the unification of multi-source data in the temporal and spatial dimensions.

[0017] For airborne time-series data such as flight paths and attitude, single-point noise exceeding reasonable limits is first identified and eliminated using the 3σ criterion. Then, a sliding window trend detection is used to capture potential gradual or abrupt anomalies in the data, correcting latent noise and ensuring the stability of the time-series data. For missing data, it is processed in stages according to the proportion of missing data. Low-missing scenarios are filled using interpolation, while medium-to-high-missing scenarios utilize a prediction model to generate imputed values ​​that conform to the trend, ensuring data continuity. When meteorological data shows anomalies, a weighted average is calculated by retrieving data from neighboring stations during the same period for correction. Simultaneously, the consistency of the data's time-series trend is verified to avoid logical contradictions caused by isolated corrections. ADS-B and radar data are compared bidirectionally to eliminate positioning drift values. Anomaly identification thresholds are adjusted based on environmental interference to ensure the continuity and accuracy of positioning data. For electromagnetic spectrum data, obvious outliers are first eliminated using sliding window variance analysis. Then, cross-validation with data from neighboring electromagnetic monitoring nodes during the same period corrects deviations caused by signal fluctuations.

[0018] In terms of time alignment, all time-series data are unified to the same timestamp, and data with different sampling frequencies are processed by a sliding window resampling method to synchronize data with large differences in sampling frequencies. In terms of spatial alignment, a unified WGS-84 coordinate system is used to transform various spatial data to eliminate errors caused by differences in coordinate systems of different devices. Then, continuous space is discretized by dividing the data into grids of fixed size to achieve accurate matching of spatial elements such as flight paths and no-fly zones, laying a spatial consistency foundation for subsequent data fusion.

[0019] The preprocessed multi-source data has achieved consistent alignment and noise removal in the time and spatial dimensions, and can be directly connected to the hierarchical adaptive fusion framework to adapt to the fusion needs of different scenarios.

[0020] Step 3, Data Reliability Measurement and Conflict Resolution, the specific process is as follows:

[0021] Step 3.1: Reliability Assessment of Track Data

[0022] A reliability assessment model is constructed based on data sources, equipment accuracy, and real-time status. Specific indicators include equipment accuracy indicators determined according to the inherent errors of the monitoring equipment. Real-time signal quality indicators determined based on signal strength and packet loss rate during data transmission. And the historical error rate index determined based on the deviation rate between historical equipment data and the true value. .

[0023] Extract the inherent error parameters, real-time transmission status parameters, and historical operation records corresponding to the trajectory data of each single mode. Then, score the equipment accuracy, real-time signal quality, and historical error rate according to the corresponding scoring criteria and normalize them to obtain the equipment accuracy score. Real-time signal quality scoring Historical error rate score The final confidence value of the weighted single-modal data is calculated using the following formula: The credibility value of each single modality data is bound to the data itself, providing a basis for subsequent fusion weight allocation and conflict resolution.

[0024] Step 3.2: Multi-source track data conflict resolution

[0025] When conflicts arise in the fusion results of multi-source track data, the DS evidence synthesis rule is used to resolve the conflicts: First, the evidence body is defined, and the fusion result of each modality data is treated as an independent evidence body. Next, calculate the conflict coefficient. To quantify the degree of conflict between pieces of evidence. (Setting) The conflict threshold is , which is if If determined to be of low conflict, the DS synthesis rules are directly applied to fuse and calculate the evidence. If an argument is deemed highly conflicting, credibility weights for each modality are introduced to modify the evidence body, reducing the influence of low-credibility evidence during the fusion process before proceeding with the synthesis operation. Finally, a final decision is made, selecting the result with the highest probability after synthesis as the final fusion output to ensure the accuracy and reliability of the fusion result.

[0026] Step 4: Construct a hierarchical adaptive fusion framework

[0027] Based on the scenario types identified in Step 1 and the data from Step 3, the corresponding fusion strategy steps are selected. The specific processes for each fusion strategy are as follows:

[0028] Step 4.1: Perform lightweight edge-side single-modal fusion for low-complexity scenarios.

[0029] For low-complexity scenarios that require high real-time performance and lightweight data fusion, the data fusion process can be completed independently at the edge without cloud involvement, thereby avoiding data transmission latency and reducing the waste of computing resources.

[0030] In terms of data selection, the focus is on the core and critical data of aircraft operation—airborne trajectory data and attitude data. These two types of data are directly collected by high-precision sensors carried by the aircraft, without relying on ground equipment for transmission. The data has the highest reliability and can completely reflect the aircraft's position and attitude status, meeting the requirement of not needing multimodal redundancy verification in low-complexity scenarios.

[0031] This fusion strategy employs the Kalman filter algorithm. Its core logic is to use attitude data as "observations" to correct errors accumulated during the acquisition and transmission of track data in real time. First, a state equation is constructed.

[0032]

[0033] in, represent The core status of the flight track at any given time (latitude, longitude, and altitude). The state transition matrix (describing the trajectory from) Time's up (The natural laws governing the changes of time) For control parameters (such as the real-time speed of an aircraft). To address process noise (based on sensor historical operational error calibration), the dynamic changing trend of the flight path state is depicted at this moment; secondly, observation equations are constructed.

[0034]

[0035] in, yes Attitude observations at any given time (roll angle, pitch angle). For the observation matrix (establishing the mapping relationship between track state and attitude observation values). To eliminate observation noise (determined based on the historical error rate of the attitude sensor), the correlation between the track state and the actual observation is established. Finally, the predicted value of the state equation and the actual value of the observation equation are weighted and fused through Kalman gain calculation to obtain the fused track value.

[0036] Step 4.2: For medium-complexity scenarios, adopt edge-cloud dual-modal collaborative fusion.

[0037] To address the need for a balance between real-time performance and robustness in medium-complexity scenarios, the fusion framework employs initial processing at the edge and optimization and correction in the cloud, relies on cross-validation of dual-modal data to improve reliability, and controls latency through edge-cloud division of labor.

[0038] The core task of the edge processing stage is to rapidly output basic fusion results to meet low latency requirements. Two types of core data are selected for fusion: airborne flight path data and ground-based ADS-B data (aircraft flight path data provided by ground infrastructure). A weighted algorithm is used for initial fusion, with weight allocation adjusted based on data credibility. Differentiated weights ensure that data with higher credibility has a greater impact on the fusion result. All calculations in this stage are completed at the edge (UAV flight control system, ground area monitoring nodes, etc.), without requiring data uploads to the cloud.

[0039] The core of the cloud-based collaborative optimization phase is to introduce environmental factor corrections to improve the robustness of the fusion results. The initial trajectory fusion results output from the edge device are uploaded to the cloud platform along with ground meteorological data. A pre-trained LSTM model is deployed in the cloud, learning the correlation between trajectories and meteorological data through extensive historical data. Based on these correlation patterns, the network dynamically corrects the initial fusion results from the edge device, outputting trajectory data compensated for by meteorological data. The total latency of the entire collaborative process meets the scenario requirements, and indirect multimodal validation significantly improves the reliability of data in complex environments.

[0040] Step 4.3: For highly complex scenarios, adopt edge-cloud collaborative multimodal large model fusion.

[0041] To address the need for high robustness and multimodal cross-validation in highly complex scenarios, a deep collaborative model of edge data cleansing and cloud-based multimodal deep fusion is adopted. This model relies on large multimodal models to mine data correlations and ensure fusion accuracy in complex environments.

[0042] The core of the edge preprocessing stage is data purification, which reduces the pressure on cloud integration. Airborne data (track, attitude, velocity) is validated, and sliding window variance analysis is used to remove obvious anomalies. Then, the validated airborne data, along with meteorological data, ground radar data, and electromagnetic spectrum data, are uploaded to the cloud.

[0043] The cloud-based multimodal large model fusion stage is the core of achieving high robustness, and it is accomplished through three steps: The first step is model construction, which builds a multimodal fusion framework based on the CLIP model. During the model training process, low-altitude flight physics rules and domain knowledge are embedded. The model input covers six types of data: track, attitude, velocity, radar, meteorology, and electromagnetic spectrum, to achieve comprehensive integration of multi-dimensional information. The second step is semantic alignment, which uses a modality-specific encoder to extract features to address the heterogeneity of different modal data. An attention mechanism is used to map all modal features to a unified semantic space, while dynamically adjusting modal weights according to scenario requirements. The third step is fusion output, which outputs multimodal correlated track data.

[0044] Step 5: Output the fusion result. The specific process is as follows:

[0045] Based on the actual needs of different scenarios, the system outputs fusion results in adapted formats. For low-complexity scenarios, it outputs lightweight trajectory data with core information such as latitude, longitude, and altitude, providing simple and efficient data support for real-time obstacle avoidance by aircraft. For medium-complexity scenarios, it outputs trajectory data with weather compensation, meeting the dual requirements of accuracy and reliability for airspace conflict monitoring. For high-complexity scenarios, it outputs high-precision trajectory data reports, providing information reference for emergency rescue decision-making.

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

[0047] 1) For scenarios with different levels of complexity, a dynamic matching and fusion strategy is adopted. In low-complexity scenarios, real-time performance is prioritized to meet the obstacle avoidance requirements of the aircraft. In high-complexity scenarios, multi-modal cross-validation is used to enhance robustness to cope with the interference of harsh environments. This effectively solves the problem of insufficient adaptability caused by the adoption of a unified strategy in existing technologies and achieves accurate fusion in diverse scenarios.

[0048] 2) By eliminating the deviation of data in the time and space dimensions through spatiotemporal alignment, establishing the association logic of different types of data with the help of semantic mapping, and highlighting the role of high reliability data by combining credibility modeling, the feature matching degree of multimodal data is greatly improved, and the false alarms in the anomaly identification process are reduced. Especially in complex interference scenarios, the ability to identify key anomalies is significantly enhanced compared with existing technologies.

[0049] 3) By relying on edge preprocessing to reduce invalid data transmission and combining it with deep cloud integration to achieve global information integration, the accuracy of global integration is not affected while reducing data transmission volume and network bandwidth usage. This provides efficient and stable data processing support for large-scale mixed operation of low-altitude aircraft and helps the low-altitude navigation system develop on a large scale. Attached Figure Description

[0050] Figure 1 This is an overall flowchart of the present invention;

[0051] Figure 2 This is a flowchart illustrating the hardware environment deployment and data acquisition process of this invention.

[0052] Figure 3 This is a schematic diagram of the multi-source heterogeneous data acquisition and data preprocessing process of the present invention;

[0053] Figure 4 This is a schematic diagram of the hierarchical adaptive fusion framework of the present invention. Detailed Implementation

[0054] The objectives and technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0055] This invention relates to an adaptive fusion method for multi-source heterogeneous data from low-altitude flight across multiple scenarios, such as... Figure 1 As shown, the specific steps are as follows:

[0056] Step 1: Determine the scene classification criteria

[0057] Based on the risk level, data integrity, and operational complexity of low-altitude navigation, the scenarios are divided into low-complexity scenarios, medium-complexity scenarios, and high-complexity scenarios.

[0058] Low-complexity scenarios include routine urban air travel and interference-free low-altitude logistics. The core characteristics are a single aircraft type, a stable operating environment (wind speed measured as a 60-minute average from a ground weather station, not exceeding 5 m / s and free from electromagnetic interference), and data integrity maintained above 95%. In other words, scenarios with wind speed ≤ 5 m / s, data missing rate < 5%, and a single aircraft type are automatically classified as low-complexity scenarios.

[0059] Medium-complexity scenarios include low-altitude logistics in complex urban areas and routine emergency patrols. These scenarios are characterized by diverse aircraft types, slight environmental interference (wind speed is calculated using a 60-minute average from a ground weather station, maintained between 5-8 m / s), and data completeness between 85% and 95%. Specifically, scenarios where 5 m / s < wind speed ≤ 8 m / s, 5% ≤ data missing rate < 15%, and multiple aircraft types are present are automatically classified as medium-complexity scenarios.

[0060] High-complexity scenarios include special situations such as emergency rescue in mountainous areas and low-altitude operations under severe weather conditions. These scenarios involve a mix of aircraft types and severe environmental interference (wind speed is measured using a 60-minute average from a ground weather station; wind speeds exceeding 8 m / s are accompanied by strong electromagnetic interference or signal blockage), resulting in data integrity below 85%. Specifically, scenarios with wind speeds > 8 m / s, data loss rates ≥ 15%, or strong interference / signal blockage are automatically classified as high-complexity scenarios.

[0061] Step 2, Hardware environment deployment and data acquisition, such as Figure 2 As shown. The methods for collecting all data are explained here; specific data collection should be performed according to the needs of each scenario:

[0062] Step 2.1 Hardware Environment Deployment

[0063] The hardware environment of this invention includes airborne equipment, ground equipment, a cloud platform, and edge processing nodes, with the following configurations: The airborne equipment uses a multi-rotor UAV and an electric vertical takeoff and landing aircraft equipped with a GNSS positioning module, an IMU inertial measurement unit, and a 4G / 5G dual-mode communication module to collect multi-dimensional data related to aircraft operation; the ground equipment consists of an ADS-B receiving station, an X-band radar, a ground weather station (including wind speed and direction), and an electromagnetic spectrum monitor to achieve comprehensive collection of ground positioning information and low-altitude environmental data; the cloud platform uses Alibaba Cloud ECS servers to deploy the inference and operation environment for LSTM and CLIP multimodal models; the edge processing nodes include two types: one is the edge computing unit built into the UAV flight control system, and the other is an industrial-grade ground area monitoring gateway, which undertakes the initial data processing tasks at the edge.

[0064] Step 2.2, Basic Data Preparation:

[0065] The UAV and electric vertical takeoff and landing (eVTOL) aircraft collect track data through their onboard GNSS positioning module. The track data includes longitude, latitude, and altitude, in °, °, and m, respectively. Attitude data is collected through the IMU inertial measurement unit, which includes roll angle, pitch angle, and yaw angle, in °. The eastward and northward velocities are calculated using the Doppler effect of the onboard GNSS module, and the vertical velocity is calculated by fusing the barometric altimeter and the IMU. These three data together constitute the onboard velocity data.

[0066] By acquiring the aircraft's flight path information (including longitude, latitude, and altitude, in °, °, and m respectively) through an ADS-B receiving station, and combining it with X-band radar data (including longitude, latitude, and altitude, in °, °, and m respectively), positioning accuracy can be further improved. Ground meteorological stations collect two characteristics of the low-altitude environment: wind speed (in m / s) and wind direction (in °, measured clockwise with true north as 0°). An electromagnetic spectrum monitor monitors the intensity of interference signals in the airspace (in dBm), providing a direct reflection of the degree of electromagnetic interference in the airspace and offering data support for environmental interference analysis.

[0067] The airspace management system uniformly distributes control-related data, including text data and rule data. The text data clearly defines spatial and altitude constraints such as the coordinates of temporary no-fly zones and flight altitude restrictions. The rule data defines operational specifications such as aircraft priority rules and airspace access permissions, providing a compliance reference for data fusion.

[0068] Step 3, Data Preprocessing:

[0069] Data preprocessing includes two core steps: data cleaning and spatiotemporal alignment, such as... Figure 3 As shown:

[0070] To address anomalous noise in airborne flight path and attitude data, the 3σ criterion is used to identify and remove outliers exceeding reasonable limits. For extreme anomalies in wind speed and direction data collected from ground meteorological stations, concurrent wind speed and direction data from three neighboring stations are retrieved. When the current data deviates from the mean by more than 30%, the weighted mean calculated based on station weights (inversely proportional to distance) is used to replace the outlier.

[0071] In the time dimension, a sliding window resampling method is used to uniformly resample time-series data such as track data, attitude data, and velocity data at different sampling frequencies to 10Hz, ensuring data time synchronization. In the spatial dimension, various spatial data such as radar data and ADS-B data are uniformly converted to the WGS-84 coordinate system to eliminate systematic errors caused by differences in coordinate systems of different devices; a fixed-size 50m×50m grid is used to discretize the continuous airspace to achieve accurate matching of spatial elements such as track, no-fly zone, and terrain obstacles, thus building a foundation for spatial consistency.

[0072] Step 4: Data credibility measurement and conflict resolution

[0073] This step involves quantifying data reliability and resolving multimodal conflicts to prevent low-reliability data from interfering with the fusion results. The specific implementation is as follows:

[0074] Step 4.1, Data Trustworthiness Measurement

[0075] For the trajectory data collected in step 2, a three-dimensional reliability quantification model is constructed, comprising "equipment accuracy - real-time signal quality - historical error rate". The reliability quantification process for trajectory data collected from airborne and ground sources is as follows:

[0076] Step 4.1.1: Quantitative indicator parameter calibration

[0077] Using the Analytic Hierarchy Process (AHP) combined with low-altitude flight data characteristics, this embodiment sets the weights of each indicator as follows: Equipment accuracy weight. Real-time signal quality weights Historical error rate weighting Each indicator is scored out of 10, and then normalized to the maximum value to map it to the [0,1] interval as the indicator score (denoted as ). , , Based on historical experience and industry standards, the specific standards set in this embodiment are as follows:

[0078] Equipment accuracy score Based on the inherent error specified by the equipment manufacturer, the inherent error is calculated as follows: less than 1m, 10 points; 1-3m, 8 points; 3-5m, 6 points; and greater than 5m, 4 points.

[0079] Real-time signal quality score The score is calculated by weighting signal strength and packet loss rate equally (0.5 each). A signal strength greater than -60dBm gets 10 points, between -80dBm and -60dBm gets 8 points, and less than -80dBm gets 5 points. A packet loss rate less than 2% gets 10 points, between 2% and 5% gets 7 points, and greater than 5% gets 3 points.

[0080] Historical error rate score The deviation rate between the historical data of the statistical equipment in the past 3 months and the true value (GNSS reference station calibration data, radar standard verification data, etc.) is calculated. A deviation rate of less than 1% is worth 10 points, 1% to 3% is worth 8 points, 3% to 5% is worth 6 points, and greater than 5% is worth 4 points.

[0081] Step 4.1.2, Calculation of Reliability Measurement

[0082] Based on the above scoring criteria and methods, the equipment accuracy, real-time signal quality, and historical error rate were scored and normalized respectively, resulting in the following scores: , , The final confidence value of each track data is calculated using a linear weighting formula, which is as follows:

[0083]

[0084] The credibility value of each track data is bound to the data itself and stored, providing a basis for subsequent fusion weight allocation and conflict resolution.

[0085] Step 4.2: Multi-source data conflict resolution

[0086] To address inconsistencies in fusion conclusions caused by factors such as equipment system errors, complex environmental interference, and data transmission delays in multimodal data, DS evidence synthesis rules are needed to accurately resolve conflicts. The specific implementation is as follows:

[0087] First, determine the number of pieces of evidence based on the complexity of the scenario: For low-complexity scenarios, define the pieces of evidence. (Conclusion of track data fusion); For medium-complexity scenarios, additional evidence is provided. (ADS-B data fusion conclusion); For highly complex scenarios, additional evidence should be added. (Conclusion of radar data fusion).

[0088] Unified definition of the proposition space for track data conflicts ,in This indicates that "the deviation between the fusion result and the true value is ≤2m". This indicates that "the fusion result deviates from the true value by 2~5m". This indicates that "the deviation between the fusion result and the true value is >5m".

[0089] Then, the support level of each piece of evidence for the proposition is determined through a combination of historical data statistics and real-time status verification: First, the historical operational data of the evidence (such as the airborne track data corresponding to m1) in similar scenarios is statistically analyzed, and the deviation between its measured value and the true value (with GNSS base station calibration data as the true value) is calculated. The percentage of deviations falling in each interval A1, A2, and A3 is statistically analyzed and used as the initial support level. For example, if 60% of the historical data deviation is ≤2m and the current signal is normal, the corresponding support level is... Then, fine-tuning is performed based on the real-time signal quality of the evidence: if the real-time signal quality is better than the historical average, the support of A1 is increased by 0.1; if the signal quality is worse than the historical average, the support of A1 is decreased by 0.1, and the support of A2 or A3 is increased by 0.05 each; finally, the normalized support of each piece of evidence for the three propositions is obtained.

[0090] Calculate the conflict coefficient according to the DS synthesis rule:

[0091]

[0092] A and B are proposition spaces. The propositions that are mutually exclusive, namely, A and B are taken from {A1, A2, A3} and A ≠ B, , For the evidence body number, this embodiment sets a conflict threshold. .

[0093] In low-conflict scenarios ( In the results, the consistency among the various results is high, so the DS synthesis rule is directly adopted for fusion, and the synthesis formula is as follows:

[0094]

[0095] High-conflict scenarios First, the credibility weight calculated in step 3.1 is used to correct the evidence body. The correction formula is as follows:

[0096]

[0097] in For the first The credibility of individual pieces of evidence This represents the total number of pieces of evidence in the current scenario. This is the serial number of the evidence.

[0098] After correction, the conflict coefficient is recalculated. Then execute the DS synthesis rule.

[0099] The proposition with the highest global support in the final output is taken as the conclusion for resolving track data conflicts (denoted as Result, with a value of...). This serves as the basis for subsequent layered integration, and its core application rules are as follows:

[0100] (Bias ≤ 2m, high reliability): Maintain the current data weight allocation, and the fusion result directly enters the subsequent output process without additional correction;

[0101] (deviation 2~5m, medium confidence): Triggers fine-tuning of data weights and increases constraint strength in the physical verification process;

[0102] (Bias > 5m, low confidence): It is determined that there is low-quality data interference, and weight reconstruction is forcibly started. Data with confidence < 0.6 is removed, and its weight is distributed proportionally to high confidence data.

[0103] Step 5: Perform layered adaptive fusion for different scenarios, such as... Figure 4 As shown:

[0104] Step 5.1: Perform lightweight edge-side single-modal fusion for low-complexity scenarios.

[0105] To address the core requirements of high real-time performance and lightweight computing in low-complexity scenarios, the fusion process is entirely completed independently at the edge, as detailed below:

[0106] Focusing on the core data of aircraft operation, only airborne trajectory data and attitude data are extracted, eliminating the need to introduce redundant ground data and simplifying the data processing process.

[0107] Data fusion is achieved using the Kalman filter algorithm, specifically by constructing the following equation:

[0108] Equations of state:

[0109]

[0110] in lon, lat, and alt respectively represent The system contains the longitude, latitude, and altitude information of the flight track at any given time; A is a 3×3 identity matrix, used to describe the natural variation of the track due to the short-term stability of the track in low-complexity scenarios; B is the control matrix, with values ​​set in this embodiment. ; The control variable is the real-time speed of the aircraft. This is process noise, and the value is based on statistical analysis of historical operating data of the GNSS positioning module over the past 6 months.

[0111] Observation equation:

[0112]

[0113] in roll and pitch respectively represent The observed values ​​of roll and pitch angles at each moment; For a 2×3 observation matrix, the value in this embodiment is... This is used to establish the mapping relationship between track state and attitude observations; The noise level was determined based on the factory calibration report of the IMU inertial measurement unit and on-site test data.

[0114] By calculating the Kalman gain of the predicted values ​​of the state equation and the actual values ​​of the observation equation, the fused track data is finally output.

[0115] Step 5.2: For medium-complexity scenarios, adopt edge-cloud dual-modal collaborative fusion.

[0116] To address the need for a balance between real-time performance and robustness in medium-complexity scenarios, a dual-modal collaborative fusion model is adopted, consisting of initial processing at the edge and optimization and correction in the cloud. The specific implementation is as follows:

[0117] Airborne track data and ground-based ADS-B data are selected as the data sources for fusion. All calculations in this stage are completed at the edge, and preliminary track data is output.

[0118] Adjust the fusion weights based on the conflict resolution conclusions in step 4.2. If the result is of high confidence, follow step 3.1 to allocate fusion weights for each data point based on data confidence assessment, and use a weighted method for initial fusion of track data to ensure that high-confidence data plays a dominant role in the fusion result. If the result is of medium confidence, increase the weight of the high-confidence modality by 0.03 while decreasing the weight of the low-confidence modality. If the result is of low confidence, remove the data with the lowest confidence.

[0119] The preliminary fusion results from the edge processing unit are uploaded to the cloud platform along with ground meteorological data. A pre-trained LSTM model with three hidden layers and 128 neurons per layer is deployed in the cloud. This model learns the correlation between flight paths and meteorological factors through extensive historical data. Based on this correlation, the preliminary fusion results are dynamically corrected to compensate for flight path deviations caused by meteorological factors. The final output is flight path data with meteorological compensation.

[0120] Step 5.3, Implementation of High-Complexity Scene Integration:

[0121] The core fusion object in this scenario is the flight track data (latitude, longitude, and altitude). Attitude, velocity, radar, meteorological, and electromagnetic spectrum data are all used as auxiliary constraint information for flight track accuracy optimization and reliability verification, and do not participate in the fusion calculation of flight track coordinates.

[0122] Step 5.3.1: Edge Data Cleaning

[0123] The core task at the edge is to improve data quality and provide a high-quality, low-redundancy data source for deep integration with the cloud. The specific implementation is as follows:

[0124] For the three core airborne data types—track, attitude, and velocity—a 5-step sliding window is set, and the data variance within the window is calculated. If the variance of the track data exceeds 1 m² (corresponding to excessive latitude and longitude drift), the variance of the attitude data exceeds 0.5°² (abrupt roll / pitch angle changes), or the variance of the velocity data exceeds 2 m² / s² (eastward / northward velocity jumps), it is identified as an abnormal segment and removed. The time sequence continuity is ensured by linear interpolation of three adjacent valid data points.

[0125] Step 5.3.2: Cloud-based multimodal fusion model construction and training

[0126] A multimodal fusion framework adapted to the low-altitude navigation domain is built based on the CLIP pre-trained model. By embedding domain knowledge and physical constraints, it solves the technical problem of general multimodal models being "detached from industry reality and output results not conforming to physical logic". The specific implementation is as follows:

[0127] Using CLIP as the backbone network, its cross-modal attention mechanism is retained, and the input layer is expanded to be compatible with six types of heterogeneous data: flight track time series (60×3, including longitude, latitude, and altitude), attitude data (60×3, including roll angle, pitch angle, and heading angle), velocity (60×3, including eastward velocity, northward velocity, and vertical velocity), radar positioning data (60×3, including longitude, latitude, and altitude), surface meteorological data (60×2, including wind speed and wind direction), and electromagnetic spectrum data (60×1, including interference intensity), to achieve full-dimensional access to multi-source data and preliminary feature extraction.

[0128] A trajectory-wind dynamics constraint term is added to the loss function during model training to force the model to output a fusion result that conforms to physical laws. Loss function The overall expression is:

[0129]

[0130] in Cross-entropy loss (basic loss term) for multimodal data is used to optimize the feature matching degree of each modality. For the constraint term weight coefficient, this embodiment takes... ; The track-wind dynamics constraint term is expressed as follows:

[0131]

[0132] in , The latitude and longitude offsets of the predicted flight path by the model. , The theoretical offset is calculated based on an aerodynamic formula, as follows:

[0133]

[0134]

[0135] in, The real-time ground speed of the aircraft (unit: m / s) is the combined speed of the aircraft's own power and the wind speed. The wind direction angle (unit: °) is the angle between the wind direction and due north, taken from real-time monitoring data of ground meteorological stations. The flight heading angle of the aircraft (unit: °) is taken from data collected by the IMU inertial measurement unit; The latitude of the aircraft's current position (unit: °) is taken from the latitude coordinates in the track data; The data sampling time interval (unit: seconds); The average radius of the Earth (unit: m) is fixed at 6,371,000 m (industry standard).

[0136] when When the value is greater than 5m, the loss amplification mechanism is triggered: the current loss value is amplified. Multiply by a magnification factor of 1.5 and then include the total loss to force backpropagation optimization of the model, correcting trajectory prediction results that do not conform to aerodynamic laws; if all three consecutive sampling periods meet the requirements... If the value is greater than 5m, the weight of the training data in that batch will be reduced simultaneously (from 1.0 to 0.6) to avoid abnormal data from interfering with the model parameters.

[0137] The model is fine-tuned using preprocessed historical data of high-complexity scenarios. When the model's trajectory prediction error is consistently below 5m and the physical rule compliance (the percentage of samples with a deviation ≤5m from the theoretical value) is greater than 98%, the model training is considered converged, training is stopped, and the optimal model parameters are saved. Finally, the trained multimodal fusion model is obtained and deployed to a cloud server.

[0138] Step 5.3.3: Cloud-based multimodal feature fusion and dynamic weight adjustment

[0139] This scenario uses the time sequence of flight tracks (latitude, longitude, and altitude) as the core fusion object, attitude and velocity as "track association constraints", and ground meteorological and electromagnetic spectrum data as "environmental interference constraints", all of which together serve to optimize the accuracy of flight tracks.

[0140] Airborne track and radar data (latitude, longitude, and altitude) are used as independent positioning references for the track. Both are encoded using a shared GRU framework with modal-specific parameters, outputting 64-dimensional core positioning feature vectors: F1 and F2. Attitude data (60×3, roll, pitch, and heading angles) and velocity data (60×3, east / north / vertical velocity) are used as motion state constraints for the track. Encoded using the same GRU architecture, they output 64-dimensional associated constraint feature vectors: F3 and F4. Ground meteorological data (60×2, wind speed and direction) and electromagnetic spectrum data (60×1, interference intensity) are used as environmental interference quantification indicators. They are first compressed into static features (3-dimensional meteorological and 1-dimensional electromagnetic) through temporal aggregation (taking the average of 60 steps), and then generated into 64-dimensional environmental constraint feature vectors: F5 and F6 through a 3-layer fully connected encoder (128→64→64, including BatchNorm and ReLU, with the output mapped to [-1,1] via tanh).

[0141] The 64-dimensional feature vectors output by the encoder for the six modalities are standardized respectively. For each vector... Calculate its modulus and obtain For any two standardized eigenvectors and Calculate the cosine similarity value:

[0142]

[0143] For the first The th modal eigenvector Each element has a value ranging from [-1, 1]. Values ​​closer to 1 indicate greater similarity between the two modal features, while values ​​closer to -1 indicate greater difference. All pairwise similarity values ​​are arranged by modal index to generate a 6×6-dimensional correlation matrix M covering track, attitude, velocity, radar, meteorology, and electromagnetic aspects.

[0144] The core positioning weights directly affect track fusion. In this embodiment, they are set to 0.4 for track and 0.3 for radar, totaling 0.7, focusing on the fusion of track and reference positioning. The association constraint weights adjust track reliability, set to 0.1 for attitude and 0.1 for velocity, totaling 0.2, verifying track rationality through motion state. The environmental constraint weights correct for interference effects, set to 0.05 for meteorology and 0.05 for electromagnetic interference, totaling 0.1, quantifying the degree of environmental interference on the track. The fusion is adjusted according to the conflict resolution conclusions and historical experience in step 4.2, with the following rules:

[0145] High-confidence result (A1): If the similarity between the track and radar is ≥0.7 and the confidence is ≥0.8, the weights of both are increased by 0.03 and are deducted from the environmental constraints on average to enhance the core positioning fusion accuracy; if the similarity between the track and velocity is ≥0.7, the velocity weight is increased by 0.02 to enhance the support of mechanical constraints for the track.

[0146] Medium confidence result (A2): Increase attitude / velocity weights by 0.02 each, correct track drift through motion state constraints, and deduct part from environmental constraints by average deduction;

[0147] Low confidence result (A3): If the track confidence is <0.6, reduce its weight to 0.2, while increasing the radar weight to 0.4; if the electromagnetic interference is strong (>-60dBm), reduce the electromagnetic weight to 0.02, deducting part and allocating it to speed; if the radar packet loss rate is >10%, reduce the radar weight to 0.2, deducting part and allocating it to the track, restoring the airborne track to its dominant position.

[0148] The weighted sum of all 64-dimensional feature vectors is calculated using the adjusted weights to generate a 64-dimensional global feature vector.

[0149] Step 5.3.4: Physical rule verification and fusion result output

[0150] The global feature vector is input into the multimodal fusion model trained in step 5.3.2. The model decodes through the output layer to generate initial trajectory data (including longitude, latitude, and altitude).

[0151] Calculate the mechanical constraint loss of the preliminary trajectory data ,like ≤5m, judged as "conforming to physical laws", the result is directly retained; if If the distance is greater than 5m, the correction mechanism is triggered. The initial trajectory data is shifted 20% towards the theoretical mechanical value, and the trajectory is recalculated. If it is still >5m, reduce the weight of the low-quality modes (reduce it by another 0.05) and re-execute the feature fusion process in 5.3.3. If after 3 consecutive corrections... If the value is still greater than 5m, the radar and data weights will be increased by 0.05, and inference will be restarted.

[0152] The final output is the trajectory data optimized by multimodal deep fusion (example coordinates: N30°20′10″, E114°25′30″, altitude 450m), which provides accurate data support for rescue commanders to monitor the aircraft's operational status in real time.

Claims

1. A multi-scene-oriented low-altitude flight multi-source heterogeneous data adaptive fusion method, characterized in that, The specific steps are as follows: Step 1, construct a low-altitude flight scene classification system and fusion demand mapping, the specific process is: According to the risk level, data integrity, and operation complexity of low-altitude flight, the scene is divided into low complexity scene, medium complexity scene, and high complexity scene, and the corresponding fusion demand is determined according to the characteristics of different scenes; Step 2, multi-source heterogeneous data collection and preprocessing, the specific process is: Step 2.1, data collection Relying on multiple monitoring equipment for data collection, covering multiple dimensions of aircraft, ground infrastructure and airspace management system, forming a multi-source heterogeneous data input system; data includes airborne data, ground data, environmental data and control data; Step 2.2, data preprocessing Including data cleaning and time-space alignment, aiming to eliminate data noise, fill data gaps, and realize the unity of multi-source data in time and space dimensions; Step 3, data credibility quantification and conflict resolution, the specific process is as follows: Step 3.1, track data credibility evaluation A credibility evaluation model is constructed based on data sources, equipment accuracy, and real-time state, and specific indexes include an equipment accuracy index determined according to inherent errors of monitoring equipment , a real-time signal quality index determined according to signal strength and packet loss rate of data transmission , and a historical error rate index determined according to a deviation rate of historical data of equipment and a true value ; Extract the equipment inherent error parameters, real-time transmission state parameters and historical operation records corresponding to the track data of each single modality, and score and normalize the equipment accuracy, real-time signal quality and historical error rate according to the corresponding scoring standards to obtain the equipment accuracy score , real-time signal quality score and historical error rate score ; The weighted calculation of the single-modal data final credibility value is as follows: ; Step 3.2, multi-source track data conflict resolution When the multi-source track data fusion results conflict, the D-S evidence combination rule is used to resolve the conflict: first, define the evidence body, and take the fusion results of each modal data as independent evidence body , then calculate the conflict coefficient to quantify the conflict degree between evidence bodies; set as the conflict threshold, if , it is determined as low conflict, and the D-S combination rule is directly used to fuse and calculate each evidence body, if , it is determined as high conflict, the credibility weight of each modal data is introduced to modify the evidence body, and the influence weight of the evidence body with low credibility in the fusion process is reduced before the combination operation is performed; Finally, make a final decision and select the result with the maximum synthesized probability as the final fusion output; Step 4, build a hierarchical adaptive fusion framework According to the scene types divided in step 1, combined with the data in step 3, select the corresponding fusion strategy step, the specific process of each fusion strategy is as follows: Step 4.1, edge-end single-mode lightweight fusion for low complexity scenes According to the demand for high real-time and lightweight of data fusion for low complexity scenes, the data fusion process is completely completed relying on the edge, without the participation of the cloud; In data selection, select airborne track data and attitude data; The fusion strategy adopts Kalman filter algorithm, taking attitude data as observation value to correct the error accumulated in the collection and transmission process of track data in real time; first, construct the state equation, in, represent The core status of a flight track at any given time includes latitude, longitude, and altitude; The state transition matrix describes the trajectory from... Time's up The natural laws governing the changes of time; For control parameters, including the aircraft's real-time speed; To address process noise; secondly, to construct observation equations. wherein, is attitude observation value at the moment, including roll angle and pitch angle; is an observation matrix, establishing a mapping relationship between the track state and the attitude observation value; is observation noise; finally, the predicted value of the state equation and the actual value of the observation equation are weighted and fused through Kalman gain calculation to obtain the fused track value; Step 4.2, edge-cloud dual-mode collaborative fusion for medium complexity scenes According to the demand for real-time and robustness balance of medium complexity scenes, edge-end preliminary processing and cloud-end optimization correction are adopted to improve reliability through cross verification of dual-mode data, and time delay is controlled through edge-cloud division; The task of edge-end processing stage is to quickly output basic fusion results to meet the low time delay demand, select two types of data for fusion: airborne track data and ground ADS-B data; adopt weighted algorithm for preliminary fusion, weight distribution is based on data credibility adjustment, and through differentiated weight, ensure that data with higher credibility have greater impact on fusion results; all calculations are completed in the edge, without data uploading to the cloud; The cloud collaborative optimization stage introduces environmental factors for correction to improve the robustness of the fusion results; the preliminary track fusion results output by the edge are uploaded to the cloud platform together with the ground meteorological data; the cloud deploys a pre-trained LSTM model, which learns the correlation between track and weather from historical data; based on the correlation, the network dynamically corrects the preliminary fusion results of the edge, and outputs the track data compensated by the weather data; Step 4.3, edge-cloud collaborative multi-modal large model fusion for high complexity scenes To meet the needs of high robustness and multi-modal cross-validation in high complexity scenarios, a deep collaboration mode is adopted, which combines edge data purification with cloud multi-modal deep fusion. Relying on multi-modal large models to mine data correlations, the fusion accuracy in complex environments is ensured. The task of the edge preprocessing stage is data purification, which reduces the pressure on cloud fusion. The airborne data including track, attitude, and speed are checked, and the sliding window variance analysis is used to eliminate obvious abnormal data. Then the checked airborne data, weather data, ground radar data, and electromagnetic spectrum data are uploaded to the cloud. The cloud multi-modal large model fusion stage is completed through three steps: the first step is model construction, based on the CLIP model to build a multi-modal fusion framework, and embed low-altitude flight physical rules and domain knowledge in the model training process. The model input covers six types of data: track, attitude, speed, radar, weather, and electromagnetic spectrum, realizing comprehensive integration of multi-dimensional information. The second step is semantic alignment, aiming at the heterogeneity of different modal data, using modal-specific encoders for feature extraction, and mapping all modal features to a unified semantic space through attention mechanisms, while dynamically adjusting the modal weights according to the scene requirements. The third step is fusion output, outputting multi-modal correlated track data. Step 5, output the fusion results, the specific process is as follows: According to the actual needs of different scenarios, output the fusion results in adaptive format. For low complexity scenarios, output lightweight track data including latitude, longitude, and height information, providing simple and efficient data support for real-time obstacle avoidance of aircraft. For medium complexity scenarios, output track data with weather compensation, meeting the dual requirements of data accuracy and reliability for airspace conflict monitoring. For high complexity scenarios, output high-precision track data reports, providing information reference for emergency rescue decision-making.

2. The multi-scene-oriented low-altitude flight multi-source heterogeneous data adaptive fusion method according to claim 1, characterized in that, In step 1, low complexity scenarios include urban air travel and interference-free low-altitude logistics; medium complexity scenarios include urban complex area low-altitude logistics and emergency patrol; high complexity scenarios include mountain emergency rescue and low-altitude operation in adverse weather.

3. The multi-scene-oriented low-altitude flight multi-source heterogeneous data adaptive fusion method according to claim 1, characterized in that, In step 2.1, the onboard data is directly collected by the sensors carried by the aircraft, including flight path data and attitude data. The flight path data records latitude, longitude, and altitude information, and the attitude data captures changes in roll angle, pitch angle, and heading angle. The sampling frequency of each type of onboard data is set according to the monitoring requirements. Ground data is obtained from ground infrastructure and is divided into positioning data and environmental data. Positioning data includes ADS-B data and radar data. ADS-B data provides aircraft position and speed information, and radar data enables higher-precision positioning monitoring. The sampling frequency of these two types of data is set differently according to the positioning accuracy requirements. Environmental data includes meteorological data and electromagnetic spectrum data. Meteorological data records environmental parameters such as wind speed and direction, and electromagnetic spectrum data monitors interference signal strength, providing a basis for analyzing the impact of the environment on navigation. Control data is provided by the airspace management system and includes text data and rule data. Text data clearly specifies spatial and height constraints such as temporary flight restriction zone coordinates and flight height restrictions, and rule data defines operating rules such as aircraft priority and airspace access rights, providing compliance reference for data fusion.

4. The multi-scene-oriented low-altitude flight multi-source heterogeneous data adaptive fusion method according to claim 1, characterized in that, In step 2.2, For time-series data including onboard flight path and attitude, single-point noise outside the reasonable range is identified and removed using the 3σ criterion, and sliding window trend detection is used to capture possible gradual or sudden abnormalities in the data, correct implicit noise, and ensure the stability of time-series data. For missing data, low-missing scenarios are filled using interpolation methods, and high-missing scenarios use prediction models to generate filling values that conform to the trend, ensuring data continuity. When meteorological data is abnormal, the weighted mean of nearby station data is calculated to correct it, and the consistency of data time series trend is verified to avoid logical contradictions caused by isolated correction. ADS-B and radar data are compared bidirectionally to remove positioning drift values, and the abnormality recognition threshold is adjusted based on environmental interference to ensure the continuity and accuracy of positioning data. For electromagnetic spectrum data, obvious outliers are removed using sliding window variance analysis, and cross-validation with nearby electromagnetic monitoring nodes is performed to correct deviations caused by signal fluctuations. In terms of time alignment, all time-series data are aligned to the same timestamp, and sliding window resampling is used to handle data with different sampling frequencies, making data with large differences in sampling frequency synchronized. In terms of spatial alignment, a unified WGS-84 coordinate system is used to convert various types of spatial data, eliminating errors caused by differences in different device coordinate systems. Then, continuous space is discretized by dividing it into fixed-size grids, achieving precise matching of flight path and flight restricted area spatial elements.

Images