Low-altitude multi-source device data base fusion method and system based on deep learning

Abstract

The application provides a low-altitude multi-source equipment data base fusion method and system based on deep learning, belongs to the technical field of low-altitude data processing, adopts a multi-modal dynamic strategy generation engine to analyze multi-dimensional context information such as equipment modal, space environment and task intention and generate optimal fusion strategy instructions, adaptively adjusts fusion according to the current situation, automatically reduces the dependence on GPS signals when the GPS signals are disturbed, enhances the weight of other reliable signal sources, dynamically reconstructs the data processing pipeline when the optical sensor fails in heavy fog weather, bypasses the image processing microservice, preferentially processes radar data, improves the dynamic adaptability and resilience anti-interference ability of the low-altitude multi-source data fusion process, can effectively resist various interferences, always maintains high-precision fusion output in a complex and changeable low-altitude environment, flexibly adjusts the data source selection and weight according to the real-time dynamic environment and task demand, and improves the accuracy and response speed of low-altitude multi-source data fusion.

Description

Technical Field

[0001] This invention relates to the field of low-altitude data processing technology, and more specifically, to a method and system for fusing low-altitude multi-source equipment data bases based on deep learning. Background Technology

[0002] In fields such as low-altitude monitoring, drone logistics, and urban air traffic, it is often necessary to fuse data from multiple devices to form a consistent and comprehensive perception of the airspace situation. Currently, common fusion schemes are mostly based on traditional filtering algorithms, such as Kalman filtering and its variants, or fixed-weight fusion rules. Before implementation, the fusion parameters of these schemes, such as the weights of each data source, filtering coefficients, correlation thresholds, and processing flows, are often pre-set and remain unchanged. For example, it might be assumed that the accuracy of optical sensors is always higher than that of infrared sensors during the day, and the former is assigned a higher fixed weight. However, the low-altitude environment is highly complex and dynamic; equipment status, environmental factors, and mission requirements are constantly changing. Static fusion strategies cannot perceive these changes, let alone provide real-time responses. When the environment or equipment condition deteriorates, the system will still use low-quality data, leading to a sharp drop in the accuracy of the fused output and even erroneous decisions. Furthermore, when performing data fusion processing, the core goal of building a data foundation is to provide high-quality and trustworthy data services for upper-layer application systems. The endpoint of most fusion systems is to generate a fused target trajectory or state estimate and directly push or store this result, lacking a basis for judging the reliability of the data. Moreover, when there are doubts or errors in the data, how can the system quickly and accurately locate the root cause of the problem and troubleshoot it?

[0003] The following technical problems exist in the existing technology:

[0004] 1. Existing low-altitude multi-source data fusion methods use static, preset fusion rules and processing procedures, which cannot be dynamically and adaptively adjusted according to the real-time status of equipment, changes in the airspace environment, and the requirements of upper-level tasks. This results in decreased fusion accuracy and low reliability in complex and ever-changing low-altitude application scenarios.

[0005] 2. Existing low-altitude data fusion systems typically only output the final fusion result, lacking the ability to assess the quality and trace the source of the fused data itself. This results in the unknowable reliability of the output data and difficulty in tracing problems, failing to provide a reliable, transparent, and high-quality data foundation for intelligent applications with high reliability requirements. Summary of the Invention

[0006] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for fusing low-altitude multi-source device data bases based on deep learning, the method comprising:

[0007] Real-time acquisition and standardized processing of heterogeneous data from low-altitude multi-source devices and their corresponding multi-dimensional context information to generate standardized context vectors;

[0008] The standardized context vector is input into the multimodal dynamic policy generation engine to perform policy reasoning and output dynamically fused policy instructions.

[0009] The workflow engine receives the dynamic fusion strategy instructions and dynamically calls and links the corresponding functional microservices according to the microservice call chain list therein, constructing an elastic data processing pipeline. The data processing pipeline performs fusion calculations on heterogeneous data and generates preliminary fusion results.

[0010] An endogeneity quality assessment is performed on the preliminary fusion results to generate a comprehensive credibility score. The preliminary fusion results, the comprehensive credibility score, and the data lineage information are then encapsulated into a trusted data unit and stored in a low-altitude data base.

[0011] Furthermore, embodiments of the present invention also provide a low-altitude multi-source equipment data base fusion system based on deep learning, the system comprising:

[0012] The context-aware module is used to collect and standardize heterogeneous data from low-altitude multi-source devices and their corresponding multi-dimensional context information in real time, and generate standardized context vectors.

[0013] A multimodal dynamic policy generation engine is used to receive the standardized context vector input to the multimodal dynamic policy generation engine, perform policy reasoning, and output dynamically fused policy instructions;

[0014] The elastic fusion execution module includes a workflow engine and a microservice resource pool. It is used to receive the dynamic fusion strategy instructions through the workflow engine, and dynamically call and link the corresponding functional microservices according to the microservice call chain list therein to build an elastic data processing pipeline. The workflow engine is used to parse and execute the microservice call chain list in the dynamic fusion strategy instructions; the microservice resource pool contains multiple functional microservices.

[0015] The data governance and infrastructure construction module is used to perform endogeneity quality assessment on the preliminary fusion results to generate a credibility score, and encapsulate the preliminary fusion results, the comprehensive credibility score and the data lineage information into a trusted data unit and store it in the low-altitude data infrastructure.

[0016] The system includes a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0017] Based on the above, the beneficial effects of the deep learning-based low-altitude multi-source equipment data base fusion method and system of the present invention are as follows:

[0018] This invention employs a multimodal dynamic strategy generation engine to analyze multi-dimensional contextual information such as device modality, spatial environment, and mission intent, and generate optimal fusion strategy instructions. It adaptively adjusts the fusion process based on the current situation. When GPS signals are interfered with, it automatically reduces reliance on them and increases the weight of other reliable signal sources. When optical sensors fail in foggy weather, it dynamically reconstructs the data processing pipeline, bypassing image processing microservices and prioritizing radar data processing. This improves the dynamic adaptability and robust anti-interference capabilities of the low-altitude multi-source data fusion process, effectively resisting various types of interference and maintaining high-precision fusion output in complex and ever-changing low-altitude environments. By flexibly adjusting data source selection and weights according to real-time dynamic environment and mission requirements, it enhances the accuracy and response speed of low-altitude multi-source data fusion.

[0019] This invention utilizes a workflow engine to dynamically instantiate and link functional microservices from a microservice resource pool according to dynamic fusion strategy instructions, constructing an elastic data processing pipeline to generate preliminary fusion results. Subsequently, the results undergo endogenous quality assessment, and data lineage information is encapsulated as trusted data units and stored in a low-altitude data base. This constructs a scalable and modular fusion execution framework, forming an endogenous quality governance mechanism. When upper-layer application systems, such as UAV autonomous navigation and air traffic management, call data, they can simultaneously obtain the comprehensive credibility score and traceability information, thereby enabling them to make smarter and safer decisions, supporting efficient resource utilization and closed-loop feedback optimization. Through endogenous quality assessment and data traceability, the reliability and traceability of the fusion output are ensured, reducing the risk of data error propagation and improving the overall credibility and application value of the low-altitude data base. Attached Figure Description

[0020] Figure 1 This is a flowchart of a low-altitude multi-source device data base fusion method based on deep learning provided in an embodiment of the present invention.

[0021] Figure 2 This is a data transmission relationship diagram of the low-altitude multi-source device data base fusion method based on deep learning provided in this embodiment of the invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1:

[0024] like Figures 1 to 2 As shown, a deep learning-based method for fusing data from multiple sources in low-altitude devices is presented. The method includes:

[0025] Step S100: Real-time acquisition and standardization of heterogeneous data from low-altitude multi-source devices and their corresponding multi-dimensional context information to generate a standardized context vector; wherein:

[0026] Low-altitude multi-source equipment includes a variety of sensors and devices deployed in the low-altitude domain below 1000 meters above sea level. These sensors and devices are used to collect various types of data to support applications such as low-altitude monitoring, drone operation, and air traffic management. The various types of data include cameras for capturing image data, lidar for generating point cloud data, radar for measuring distance and speed, GPS or BeiDou modules for positioning information, and meteorological sensors for environmental data. These various types of data are generated in real time by the equipment itself and reported to the system via wireless links such as 5G or LoRa to provide raw observation data from multiple angles. Based on the inherent static properties of the equipment, the real-time dynamic status of the data is calibrated to obtain heterogeneous data.

[0027] Heterogeneous data refers to raw observation data from low-altitude multi-source equipment, which has diverse formats, structures, and semantics. Heterogeneous data includes structured data (such as numerical positioning coordinates), semi-structured data (such as sensor readings in JSON format), and unstructured data (such as images or point clouds). Specifically, this includes video frames captured by UAV cameras, echo signals from radar stations (distance, speed, angle), GPS and BeiDou positioning coordinates (latitude, longitude, and altitude), and sensor readings (such as point cloud data from lidar or infrared thermal imaging). Heterogeneous data is directly collected from low-altitude multi-source devices such as UAVs, radar stations, and weather sensors. During collection, the acquisition frequency range of 10-100Hz is adaptively selected according to the device type, such as 30fps for UAV video frames and 1Hz for radar scans. The data is transmitted in real time through the device's sensor interface, data stream header, or dedicated acquisition module and input to the data standardization processor. After standardization processing, it is assembled with multi-dimensional context information into a standardized context vector and passed to the multimodal dynamic strategy generation engine in step S200. As the core data for fusion computing, heterogeneous data provides observation information on low-altitude targets such as aircraft and obstacles, supports subsequent multi-source fusion decisions, captures multi-angle views of the real-time scene, and improves the comprehensiveness and accuracy of the fusion results.

[0028] The standardization process for heterogeneous data from low-altitude multi-source devices and their corresponding multidimensional context information involves the following steps: Due to the diverse sources, the data formats, dimensions, and update frequencies are completely different (e.g., precision factors are dimensionless values, wind speed units are m / s, and task instructions are JSON strings). A data standardization processor is set up to perform normalization (mapping values ​​to the [0, 1] interval), encoding (encoding text instructions such as "refresh rate priority" into numerical labels such as "1"), and time alignment (applying a unified timestamp to all data) on various types of data. The processed data is then assembled into a structured, fixed-dimensional feature vector, i.e., a standardized context vector, for example: [Device A] Weights, Device B weights, visibility values, wind speed values, task mode encoding, ...], a fixed template is used when assembling vectors, for example, the vector structure is [Device modality sub-vector (10-dimensional), Spatial environment sub-vector (15-dimensional), Task intent sub-vector (5-dimensional), Heterogeneous data embedding (34-dimensional)]. The standardized context vector serves as the sole input to the subsequent policy engine, encapsulating all state information of the current system, ensuring that all data are compared and fused under a unified framework, avoiding fusion errors caused by data inconsistency. Multi-source information is integrated into a single vector through vector assembly, which is convenient for neural network processing. Metadata tags such as vector version are added through vector encapsulation, which is convenient for debugging and backtracking.

[0029] Multidimensional context information includes device modal context, spatial environment context, and mission intent context. This multidimensional context information is auxiliary metadata describing the data generation environment. It is a multi-layered nested metadata structure with dozens of dimensions. Multidimensional context information can capture multi-source information in real time, ensuring data freshness with a latency of <100ms. This helps the multimodal dynamic strategy generation engine dynamically adjust weights, avoiding the blind fusion of low-quality data in complex low-altitude scenarios such as foggy weather or no-fly zones. Specifically:

[0030] The device modal context includes the device's inherent static attributes and real-time dynamic status. The inherent static attributes include at least the device type, such as drone camera, unique identifier, and factory calibration accuracy. The real-time dynamic status includes at least the positioning accuracy factor of the GPS and Beidou modules, the signal strength and transmission delay of the communication link, and the load rate and operating temperature of the sensor itself. The device modal context data is read directly from the device management unit or data stream header information of each low-altitude device, such as drones and radar stations, in real time.

[0031] The spatial environment context includes at least static no-fly zone and dynamic isolated airspace information obtained from the airspace management system, digital elevation models and obstacle layers obtained from geographic information databases, and real-time visibility, wind speed, wind direction and precipitation data obtained from meteorological sensors or meteorological services; the spatial environment context is obtained from the airspace management system, geographic information system (GIS database), meteorological data services such as meteorological bureau API or local meteorological stations through API interfaces.

[0032] The task intent context receives task request instructions from the upper-layer application system to obtain guidance information. The guidance information is quantified into at least one operable fusion preference parameter. The fusion preference parameter includes refresh rate priority mode, positioning accuracy priority mode, and target recognition confidence priority mode, such as refresh rate priority mode (coded as 1, emphasizing high-frequency updates), positioning accuracy priority mode (coded as 2, emphasizing low error), and target recognition confidence priority mode (coded as 3, emphasizing high accuracy). These parameters are operable numerical labels. The upper-layer application system includes UAV management and control platforms, logistics scheduling systems, etc.

[0033] Step S200: Input the standardized context vector into the multimodal dynamic policy generation engine for policy reasoning, and output dynamically fused policy instructions, including:

[0034] The multimodal dynamic policy generation engine is based on a deep learning-based inference module. Internally, it has a policy knowledge base storing N context-policy mapping rules. These rules are learned through training on historical data and predefined by domain experts. The historical data primarily originates from past low-altitude data fusion task records. This historical data includes standardized context vectors as input features, corresponding dynamic fusion policy instructions as output labels, execution metadata of the fusion process such as microservice call chain lists and source selection and weight configuration tables, quality indicators of the initial fusion results such as real-time quality scores and overall uncertainty metrics, and performance evaluation indicators from upper-layer application systems such as accuracy and latency. This historical data is processed using deep learning algorithms, such as neural network training models, to learn the mapping relationship between context and policy, thereby generating more accurate rules and improving the engine's inference capabilities.

[0035] The context policy mapping rules are a set of rules for parameters and operations set for low-altitude scenarios. Rules predefined by domain experts are set by domain experts based on conventional experience and requirements, such as the experience-based rule "prioritize enabling radar data sources and assigning a weight of 0.8 in high-wind-speed spatial environment contexts to improve anti-interference capabilities." Rules learned through training on historical data, such as using standardized context vectors from past fusion tasks and corresponding quality indicators of preliminary fusion results to train a neural network model to generate rules like "in a task intent context prioritizing positioning accuracy, call the spatiotemporal calibration microservice followed by the decision-level fusion microservice, expecting an output quality standard of accuracy > 95%." For example, expert predefined rules include "if the device modal context shows a GPS accuracy factor > 2, reduce its dynamic weight coefficient to 0.3 and enable the backup BeiDou module," while training rules include "based on historical low-visibility meteorological data, map to the configuration of image source weights of 0.2 and point cloud source weights of 0.7 in the source selection and weight configuration table." The total number of these rules is N, ensuring coverage of multiple low-altitude scenarios and supporting rapid matching and inference by the multimodal dynamic policy generation engine.

[0036] The multimodal dynamic policy generation engine matches the received standardized context vectors with context policy mapping rules in the policy knowledge base to find the M most suitable rules. The results of these M rules are then weighted and synthesized to generate a dynamic fusion policy instruction. Specifically, a convolutional neural network extracts features from the standard context vectors, which are then compared with N rules in the policy knowledge base. Cosine similarity is used to find the top M most suitable rules with the highest similarity, where M is less than N. Next, a confidence weight is assigned to each rule based on historical training data or expert ratings. Finally, a weighted average is calculated based on the source selection and weight configuration table for these rule suggestions, the microservice call chain list, and the expected output quality standard. For example, if rule 1 suggests a radar weight of 0.8 and rule 2 suggests 0.6, the weighted average is 0.7, ultimately synthesizing a unified dynamic fusion policy instruction.

[0037] Dynamic fusion strategy instructions specify the processing method and specific instructions for the current task. Dynamic fusion strategy instructions include at least a source selection and weight configuration table, a microservice call chain list, and expected output quality standards, among which:

[0038] The source selection and weight configuration table contains a data source table of standardized context vectors. Each data source in the data source table is assigned a dynamic weight coefficient, and the dynamic weight coefficient takes the value of a continuous value between 0 and 1. The data sources of the standardized context vectors are selected according to the order of execution to obtain the list of enabled data sources. The context policy mapping rule assigns weight coefficients to the list of enabled data sources to obtain dynamic weight coefficients.

[0039] The expected output quality standard is used to limit the parameters of the data output, including indicators such as accuracy and latency threshold. For example, the accuracy threshold is 0.95 and the latency is <200ms. The expected output quality standard is derived from the task intent context and is used to guide optimization and evaluation to ensure that the output meets the application requirements. Specifically, it compares and verifies the comprehensive credibility score based on the preliminary fusion results to obtain the effect evaluation index. Based on the comprehensive credibility score, the feedback signal is filtered by the quality threshold to obtain the optimized feedback signal, thereby improving data accuracy and reducing errors.

[0040] The microservice call chain list is an ordered list. The ordered list specifies the types, execution order, and initialization parameter configurations of the P functional microservices that need to be called to execute this fusion task, such as [Spatiotemporal calibration (parameter: coordinate system WGS84), filtering and noise reduction (parameter: Kalman filter)];

[0041] Functional microservices are pre-developed, independently deployable, and manageable lightweight computing units. Their types include, but are not limited to, spatiotemporal calibration microservices for timestamping and coordinate system unification of data from different sources; data association microservices for determining whether observation data from different data sources belong to the same low-altitude target; filtering and noise reduction microservices for smoothing data and suppressing observation noise; feature extraction microservices for extracting abstract features from unstructured data such as images and point clouds; and decision-level fusion microservices for fusing target recognition or tracking decision results from multiple sources.

[0042] Step S300: Receive dynamic fusion strategy instructions based on the workflow engine, and dynamically call and link the corresponding functional microservices according to the microservice call chain list to build a flexible data processing pipeline. The data processing pipeline performs fusion calculations on heterogeneous data and generates preliminary fusion results, including:

[0043] Step S301: The workflow engine dynamically instantiates and wakes up the corresponding microservice instances from the microservice resource pool according to the microservice call chain list, and links them in the order specified in the microservice call chain list. The output of the previous microservice is used as the input of the next, forming a directed acyclic computational data flow graph, which constitutes a flexible data processing pipeline. The original heterogeneous data is injected into the data processing pipeline and processed by each functional microservice in sequence, finally producing a preliminary fusion result. The preliminary fusion result is a unified data representation after fusion, including target trajectory, fusion identification label, etc., such as target trajectory [coordinates + velocity + confidence], fusion identification label [category + bounding box], representing the unified output of multi-source data;

[0044] Step S302: After the data processing pipeline finishes processing the current batch of data, based on the decision of the workflow engine to receive the next policy instruction, it chooses to dynamically destroy or retain the data processing pipeline. If it chooses to dynamically destroy the data processing pipeline, the pipeline is destroyed to achieve elastic allocation and release of computing resources. If it chooses to retain the pipeline for a period of time, the pipeline can be used for subsequent similar data batches, making the system resource utilization extremely high and able to cope with sudden peaks in data flow. The next policy specification includes the type specification of dynamic fusion policy instructions. The type specification is divided into retention type and destruction type. The retention type includes cases where the microservice call chain list is consistent with or highly similar to the current one, such as the same P functional microservice types, execution order, and initialization parameter configuration. The destruction type includes cases where the microservice call chain list has changed significantly, such as adding or deleting microservice types, adjusting the execution order, or modifying parameter configuration. These type specifications come from the dynamic fusion policy instructions generated by the multimodal dynamic policy generation engine based on the next standardized context vector. For example, when the next task intent context switches from positioning accuracy priority mode to refresh rate priority mode, the next policy instruction specifies the destruction type. The workflow engine first compares the differences between the current microservice call chain list and the new list. If it finds that the spatiotemporal calibration microservice needs to be replaced with the filtering and noise reduction microservice leading the link, it dynamically destroys the existing pipeline instance, releases the associated computing resources such as CPU and memory, and then waits for the new instruction to build a new pipeline, thereby achieving elastic allocation of resources and avoiding unnecessary occupation.

[0045] The workflow engine includes a parser and a scheduler for building and managing system processes. Data is fed into microservice instances from dynamic fusion strategies. Functional microservices are instantiated based on the microservice call chain list to obtain microservice instances. These microservice instances are then linked according to their execution order to create a flexible data processing pipeline.

[0046] The data processing pipeline is a dynamically constructed, modular computational workflow framework, represented as a directed acyclic graph (DAG). It consists of multiple functional microservice instances linked in a specified order, used for end-to-end fusion processing of heterogeneous data. It includes node elements, edge elements, control elements, and auxiliary elements. Each node element has P (P=3-10) functional microservice instances, each an independent lightweight computational unit, such as a spatiotemporal calibration microservice, a data association microservice, a filtering and noise reduction microservice, a feature extraction microservice, and a decision-level fusion microservice. Each microservice has initialization parameters, such as the coordinate system "WGS84" and a filtering noise threshold of 0.05. Edge elements are connected via data flow, with the output of one microservice directly serving as the input of the next. Control elements contain scheduling metadata for the workflow engine, such as execution order, branch parallel points, and error handling hooks. Auxiliary elements include input buffers for injecting raw data and output buffers for temporarily storing preliminary fusion results. The overall structure supports elastic expansion, such as adding temporary microservice branches. The data processing pipeline is temporary unless S302 decides to retain it; the data processing pipeline enables adaptive fusion computation of heterogeneous data, transforming multi-source raw observations into unified preliminary fusion results, supporting real-time decision-making in complex low-altitude scenarios.

[0047] Step S400: Perform an endogeneity quality assessment on the preliminary fusion results to generate a comprehensive credibility score, including:

[0048] Step S401: Obtain the real-time dynamic state from the device modal context. Based on the data signal-to-noise ratio (SNR), data packet jitter rate, and positioning module accuracy factor, calculate the real-time quality score for each data source participating in the fusion. Data sources with low SNR, high jitter rate, and large accuracy factor have low real-time quality scores. The real-time quality score is a normalized value between 0 and 1, with parameters including SNR (SNR=dB, threshold>10 for high score), data packet jitter rate (jitter<50ms for low score), and positioning accuracy factor (PDOP<4 for low score). Extract the real-time dynamic state from the device modal context for calculation and input it into step S404 for synthesis function. The data source for fusion refers to the heterogeneous data sources enabled in the current fusion task. The data elements involved in fusion include various modal data generated by low-altitude multi-source devices, such as image data collected by cameras, point cloud data generated by lidar, distance and velocity data measured by radar, positioning information from GPS or Beidou modules, and environmental data from meteorological sensors. These data sources are selected from the list of enabled data sources in the source selection and weight configuration table according to the dynamic fusion strategy instructions. The list of enabled data sources originates from the inference output of the multimodal dynamic strategy generation engine on the standardized context vector, ensuring that only data sources adapted to the current device modal context, spatial environment context, and task intent context are used for fusion calculation.

[0049] Step S402: Receive and analyze the execution metadata of the data processing pipeline. The execution metadata includes the sequence of called functional microservices and their inputs and outputs. Perform uncertainty modeling on the output of each functional microservice in the data processing pipeline to generate an overall uncertainty metric. The overall uncertainty metric is a quantification of the uncertainty hidden in the data processing pipeline, such as numerical variance (σ²<0.01), perceived confidence (>0.9), or IOU (>0.7). Analyze the execution metadata as input and finally output it to the synthesis function in step S404. This step tracks the propagation and transformation process of the above uncertainty in the microservice call chain and calculates the overall uncertainty metric introduced by the entire processing flow for the preliminary fusion result.

[0050] Step S403: Obtain independent data sources from heterogeneous data, calculate the statistical consistency between the fusion results of different independent data sources. The statistical consistency is quantified by calculating the Euclidean distance variance between R result vectors, or the percentage of overlap of their respective confidence intervals. The vectors are fused coordinates [x1, y1; x2, y2]. The categories of independent data sources and the order of execution are derived from the data sources in the enabled data source list.

[0051] Step S404: Take the real-time quality score, overall uncertainty measure and statistical consistency as inputs, input a predefined synthesis function for calculation, and finally generate a comprehensive credibility score between 0 and 1 to represent the overall reliability of the fusion result, which serves as a benchmark for user trust.

[0052] The synthesis function is a predefined weighted linear combination function with configurable MLP extensions. It takes as input real-time quality score (Q), population uncertainty (U), and statistical consistency (C), and outputs a comprehensive confidence score [0, 1]. It takes as input validation (Q, C[0, 1], U[0, 1]), calculates positive / negative correlations, and then performs a weighted summation to interpret the output (contribution decomposition). A threshold > 0.8 is required for pass. The weights are fixed: α = 0.4(Q), β = 0.3(1-U), γ = 0.3(C). Based on expert tuning (minimum MSE in simulated data), the formula for calculating the comprehensive confidence score using the synthesis function is:

[0053] ;

[0054] Where T is the overall credibility score, 0.4 is the weight α, indicating that Q contributes 40%; Q is the vector mean or scalar of the quality score; 0.3 is the weight β, indicating that (1-U) contributes 30%; U is the uncertainty measure; 0.3 is the weight γ, indicating that C contributes 30%; cons is the consistency measure; linear synthesis is used, positive correlation Q / C increases T, and negative correlation U decreases T.

[0055] Step S405: The preliminary fusion result, the comprehensive credibility score, and their data lineage information are encapsulated into a trusted data unit and stored in a low-altitude data base. The data lineage information includes the original data and processing data involved in generating the preliminary fusion result. Specifically, this includes a list of source IDs for all original data, processing records of all functional microservices involved, their version numbers, and configuration parameters. This data is collected from the logs of the data processing pipeline and encapsulated in a data traceability body, chaining the entire process to provide transparency. The original data is processed by compiling an ID list of the functional microservice processing records to obtain a source ID list, and version number annotation is performed on the processing records according to the configuration parameters to obtain the data traceability body. The original data consists of heterogeneous data and multi-dimensional context information. Among them:

[0056] The trusted data unit adopts a standardized data encapsulation format, which includes a data header, a data body, and a data traceability body. The data header contains a globally unique identifier, timestamp, geospatial range, and comprehensive trust score for the trusted data unit. The data body stores the preliminary fusion results. The data traceability body records data lineage information in a structured form, including a list of source IDs for all original data, processing records of all functional microservices involved, their version numbers, and configuration parameters. The processing records refer to detailed log records of each functional microservice involved in the generation of the preliminary fusion results. The processing records include the microservice type (e.g., spatiotemporal calibration microservice or decision-level fusion microservice), the execution order in the microservice call chain list, input data digests (e.g., data source ID and input size), output data digests (e.g., processed feature vectors or fusion decisions), execution timestamp and duration, version number, configuration parameters (e.g., threshold settings or filtering algorithm type), and exception information. If an error occurs, these records, in a structured format, ensure complete traceability of the fusion path, thereby supporting debugging, auditing, and optimization.

[0057] The Low Altitude Data Base is a distributed data storage and management platform designed specifically for low-altitude multi-source fusion. It includes a storage layer, an index layer, a metadata layer, an access layer, and a backup layer. The Low Altitude Data Base provides a persistent and trusted data unit, offering a unified and traceable low-altitude data foundation. It supports real-time querying and analysis of upper-layer application systems, ensuring system compliance and scalability.

[0058] Step S500: Establish a closed-loop feedback optimization mechanism. The performance evaluation metrics of the trusted data unit after its use by the upper-layer application system, as well as the change in the trusted data unit's overall credibility score over time, are input as feedback signals to the multimodal dynamic policy generation engine. The multimodal dynamic policy generation engine uses these feedback signals to incrementally optimize and adjust the contextual policy mapping rules in its internal policy knowledge base through self-learning. The performance evaluation metrics are upper-layer quantitative feedback vectors (5-10 dimensions), such as accuracy, success rate, and resource efficiency. The change in the overall credibility score over time is represented by a time-series score trajectory, change rate, or trend, such as initial... The initial score of 0.95 gradually decreases to 0.80 after storage due to data timeliness, or briefly rises to 0.92 due to the introduction of new integrated data. These changes are represented by time-series data, such as curves with timestamps as the x-axis and score values ​​as the y-axis, or log records. These data serve as feedback signals input to the multimodal dynamic policy generation engine to detect the impact of outdated data or environmental changes. This allows the engine to adjust the context policy mapping rules through self-learning, thereby improving the accuracy of future policies and the overall adaptability of the system. The feedback signal is a composite signal packet of vectors and metadata, which is input to the self-learning module of the multimodal dynamic policy generation engine for comparison-driven policy updates.

[0059] In this embodiment, a multimodal dynamic strategy generation engine is used to analyze multi-dimensional contextual information such as device modes, spatial environment, and task intent to generate optimal fusion strategy instructions. The fusion is adaptively adjusted according to the current situation. When GPS signals are interfered with, the dependence on them is automatically reduced and the weight of other reliable signal sources is enhanced. When optical sensors fail in foggy weather, the data processing pipeline is dynamically reconstructed, bypassing the image processing microservice and prioritizing radar data processing. This improves the dynamic adaptability and robust anti-interference capability of the low-altitude multi-source data fusion process, effectively resisting various types of interference and maintaining high-precision fusion output in complex and ever-changing low-altitude environments. By flexibly adjusting the selection and weight of data sources according to the real-time dynamic environment and task requirements, the accuracy and response speed of low-altitude multi-source data fusion are improved.

[0060] By leveraging a workflow engine to dynamically instantiate and link functional microservices from a microservice resource pool according to dynamic fusion strategy instructions, an elastic data processing pipeline is built to generate preliminary fusion results. Subsequently, the results undergo endogenous quality assessment, and data lineage information is encapsulated as trusted data units and stored in a low-altitude data base. This constructs a scalable and modular fusion execution framework, forming an endogenous quality governance mechanism. This allows upper-layer application systems, such as UAV autonomous navigation and air traffic management, to simultaneously obtain the comprehensive credibility score and traceability information when accessing data, thereby enabling smarter and safer decisions. It supports efficient resource utilization and closed-loop feedback optimization. Through endogenous quality assessment and data traceability, the reliability and traceability of the fusion output are ensured, reducing the risk of data error propagation and enhancing the overall credibility and application value of the low-altitude data base. Specific Implementation Example 2:

[0062] like Figures 1 to 2 As shown, based on the content of the above specific embodiments, the following content is further disclosed:

[0063] The deep learning-based low-altitude multi-source device data fusion system includes:

[0064] The context-aware module is used to collect and standardize heterogeneous data from low-altitude multi-source devices and their corresponding multi-dimensional context information in real time, and generate standardized context vectors.

[0065] The multimodal dynamic policy generation engine receives standardized context vector inputs to perform policy reasoning and outputs dynamically fused policy instructions.

[0066] The elastic fusion execution module includes a workflow engine and a microservice resource pool. It receives dynamic fusion strategy instructions through the workflow engine and dynamically calls and links the corresponding functional microservices according to the microservice call chain list therein to build an elastic data processing pipeline. The workflow engine is used to parse and execute the microservice call chain list in the dynamic fusion strategy instructions; multiple functional microservices reside in the microservice resource pool.

[0067] The data governance and infrastructure construction module is used to perform endogeneity quality assessment on the preliminary fusion results to generate a credibility score, and to encapsulate the preliminary fusion results, the comprehensive credibility score and their data lineage information into a trusted data unit and store it in the low-altitude data infrastructure.

[0068] This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code, which, when executed by the one or more processors, can perform the above-described deep learning-based low-altitude multi-source device data fusion method and system.

[0069] The methods or systems according to the embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as a ROM or hard disk, may store the deep learning-based low-altitude multi-source device data docking method and system provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components in the electronic device shown in this application may be omitted according to actual needs. Specific Implementation Example 3:

[0071] like Figures 1 to 2 As shown, based on the content of the above specific embodiments, the following content is further disclosed:

[0072] One embodiment of this application discloses a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the deep learning-based low-altitude multi-source device data docking method and system described in the above-described embodiments of this application can be performed. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0073] Furthermore, according to embodiments of this application, the processes described in the above-mentioned flowcharts can be implemented as computer software programs. For example, this application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided in this application, such as a deep learning-based method and system for fusing data from low-altitude multi-source devices. When this computer program is executed by a central processing unit (CPU), it performs the functions defined in the method of this application.

[0074] The deep learning-based low-altitude multi-source equipment data base fusion system includes a processor, a machine-readable storage medium, and the machine-readable storage medium and the processor are connected. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-mentioned method.

[0075] The above-mentioned hardware structure and its corresponding hierarchical structure are only one way to implement the deep learning-based low-altitude multi-source device data base fusion method and system in this application, including but not limited to the above-mentioned hardware structure and its corresponding parameters. In actual use, the hardware structure can be further set according to the specific scenario and usage requirements.

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0077] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0078] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A low-altitude multi-source device data base fusion method based on deep learning, characterized in that, The method comprises: Real-time acquisition and standardized processing of heterogeneous data from low-altitude multi-source equipment and its corresponding multi-dimensional context information, including device modal context, spatial environment context and task intention context, to generate a standardized context vector; Inputting the standardized context vector into a multi-modal dynamic strategy generation engine for strategy reasoning, outputting a dynamic fusion strategy instruction, the multi-modal dynamic strategy generation engine internally setting a deep learning algorithm and a strategy knowledge base, the strategy knowledge base storing N context strategy mapping rules; Based on the workflow engine, receiving the dynamic fusion strategy instruction, and dynamically calling and linking corresponding functional microservices according to the microservice call chain list in it to build a flexible data processing pipeline, which performs fusion calculation on heterogeneous data and generates a preliminary fusion result; Endogenous quality assessment of the preliminary fusion result to generate a comprehensive credibility score, encapsulating the preliminary fusion result, comprehensive credibility score and data bloodline information into a trusted data unit and storing it in a low-altitude data base.

2. The deep learning-based low-altitude multi-source device data base station fusion method according to claim 1, characterized in that, The device modal context includes inherent static attributes and real-time dynamic states of the device, the spatial environment context includes static no-fly zones and dynamic isolated airspace information, digital elevation models and obstacle layers, and real-time visibility, wind speed, wind direction and precipitation data, and the task intention context includes fusion preference parameters, including refresh rate priority mode, positioning accuracy priority mode and target recognition confidence priority mode.

3. The deep learning-based low-altitude multi-source device data base station fusion method according to claim 1, characterized in that, The dynamic fusion strategy instruction includes a source selection and weight configuration table, a microservice call chain list and an expected output quality standard; Based on the multi-modal dynamic strategy generation engine, the standardized context vector is matched with the context strategy mapping rules to find the most suitable M rules, and the results of the M rules are weighted and integrated to finally generate the dynamic fusion strategy instruction.

4. The deep learning-based low-altitude multi-source device data base station fusion method according to claim 3, characterized in that, The source selection and weight configuration table contains a data source table of the standardized context vector, and each data source in the data source table is assigned a dynamic weight coefficient; The microservice call chain list is an ordered list that specifies the types, execution order and initialization parameter configurations of P functional microservices required to be called for executing the current fusion task; The expected output quality standard includes accuracy and delay threshold indicators to limit the parameters of data output.

5. The deep learning-based low-altitude multi-source device data base station fusion method according to claim 4, characterized in that, The functional microservices include lightweight computing units, including space-time calibration microservices, data correlation microservices, filtering and noise reduction microservices, feature extraction microservices and decision-level fusion microservices.

6. The deep learning-based low-altitude multi-source device data base station fusion method according to claim 1, characterized in that, Using the workflow engine, corresponding microservice instances are dynamically instantiated from the microservice resource pool according to the microservice call chain list, and are linked into a directed acyclic computation data flow graph in the order specified by the microservice call chain list to form a flexible data processing pipeline; After processing the current batch of data, the data processing pipeline receives the next strategy instruction based on the workflow engine to decide whether to dynamically destroy or retain the data processing pipeline.

7. The deep learning-based low-altitude multi-source device data base station fusion method according to claim 1, characterized in that, The endogenous quality assessment of the preliminary fusion result generates a comprehensive credibility score, comprising: Real-time dynamic state in the device modality context is acquired to calculate real-time quality scores for each data source participating in fusion; Execution metadata of the data processing pipeline is received and analyzed, and the execution metadata comprises a sequence of called function microservices and their input and output, and the output of each function microservice in the data processing pipeline is modeled to generate an overall uncertainty metric; Independent data sources in the heterogeneous data are acquired to calculate statistical consistency between fusion results of different independent data sources; The real-time quality scores, the overall uncertainty metric and the statistical consistency are input into a predefined synthesis function to generate a comprehensive credibility score.

8. The deep learning based low-altitude multi-source device data base station fusion method according to claim 1, characterized in that, The trusted data unit adopts a standardized data packaging format, which comprises a data header, a data body and a data provenance body, the data header contains a globally unique identifier of the trusted data unit, a timestamp, a geospatial range and a comprehensive credibility score, the data body stores the preliminary fusion result, and the data provenance body records data lineage information in a structured form, including a list of source IDs of all original data, processing records of all function microservices experienced and their version numbers and configuration parameters, and the data lineage information includes original data and processing process data involved in generating the preliminary fusion result.

9. The deep learning-based low-altitude multi-source device data base station fusion method according to claim 1, characterized in that, The method further comprises: A closed-loop feedback optimization mechanism is established, and the effect evaluation index of the trusted data unit after being used by an upper application system and the change of the comprehensive credibility score of the trusted data unit over time are input as feedback signals into the multi-modal dynamic strategy generation engine; The multi-modal dynamic strategy generation engine uses the feedback signals to incrementally optimize and adjust the context strategy mapping rules in the strategy knowledge base inside the multi-modal dynamic strategy generation engine through self-learning.

10. A low-altitude multi-source device data base fusion system based on deep learning, characterized in that, The system comprises: A context perception module is configured to collect and standardize heterogeneous data from low-altitude multi-source devices and corresponding multi-dimensional context information in real time, generate a standardized context vector, and the multi-dimensional context information comprises device modality context, spatial environment context and task intention context; A multi-modal dynamic strategy generation engine is configured to receive the standardized context vector input into the multi-modal dynamic strategy generation engine, perform strategy reasoning, and output dynamic fusion strategy instructions, the multi-modal dynamic strategy generation engine is internally provided with a deep learning algorithm and a strategy knowledge base, and the strategy knowledge base stores N context strategy mapping rules; An elastic fusion execution module comprises a workflow engine and a microservice resource pool, configured to receive the dynamic fusion strategy instructions through the workflow engine, dynamically call and link corresponding function microservices according to a microservice call chain list in the dynamic fusion strategy instructions, and construct an elastic data processing pipeline, and the data processing pipeline performs fusion calculation on the heterogeneous data and generates a preliminary fusion result. The data governance and base construction module is used for endogenous quality assessment of the preliminary fusion result to generate a credibility score, encapsulates the preliminary fusion result, the comprehensive credibility score and data blood relationship information as a credible data unit, and stores the credible data unit to the low-altitude data base.

Images