Aircraft fuel consumption prediction method and system based on fast access recorder data

Abstract

The invention provides an aircraft fuel consumption prediction method and system based on fast access recorder data, and is applied to the field of data processing application. The method comprises the following steps: based on various parameter information recorded in historical fast access recorder data, performing feature extraction on the fast access recorder data by using a data processing algorithm to generate a multi-source heterogeneous data set; processing the multi-source heterogeneous data set to generate a target directed multi-view dynamic graph; performing modeling processing on the target directed multi-view dynamic graph to generate node embedding information; node embedding information is injected into each layer of the initial fuel consumption prediction model, the model is adjusted by adopting an adaptive learning rate, and fusion graph data and multi-source data semantic information are generated; and processing the fast access recorder data of the target aircraft based on the target machine learning model in combination with the fusion graph data and the multi-source data semantic information, and generating aircraft fuel consumption prediction result information.

Description

Technical Field

[0001] The present invention relates to the field of data processing applications, and in particular to a method and system for predicting aircraft fuel consumption based on fast access recorder data. Background Art

[0002] Accurately forecasting fuel consumption is crucial to airline operations and management. It's not only about cost control; in a highly competitive market, accurately forecasting fuel consumption can significantly reduce costs and improve economic efficiency. It also involves environmental protection: reducing fuel consumption can reduce carbon emissions and help mitigate climate change. However, there are currently many challenges in predicting aircraft fuel consumption: aviation fuel consumption is affected by a combination of factors. For example, the choice of flight path is directly related to the length of the flight, which in turn affects fuel consumption. Different flight speeds change air resistance, affecting engine power and fuel consumption. The frequency and method of maneuvering operations also have an impact on fuel consumption. Meteorological conditions, such as temperature, air pressure, wind speed and direction, have significantly different effects on fuel consumption during different flight phases. These intertwined factors make fuel consumption forecasting extremely complex.

[0003] Traditional fuel consumption prediction methods often rely on a single algorithm, failing to fully consider factors such as engine power, attitude changes, and environmental changes during flight. For example, engine power and attitude vary significantly during takeoff, cruising, and landing. A single model cannot accurately describe the fuel consumption characteristics of each phase, limiting prediction accuracy. Furthermore, existing prediction models face adaptability issues when applied to actual aviation operations, making them difficult to meet the diverse needs of airlines, such as fuel consumption forecasting requirements for different routes, aircraft types, and seasons.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore includes information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for predicting aircraft fuel consumption based on fast access recorder data, which at least to some extent overcomes the problems existing in the prior art. By obtaining the QAR data and training sample set of the target aircraft, a multi-source heterogeneous data set is generated through a data processing algorithm, and then processed into a target directed multi-view dynamic graph. Next, a multi-view dynamic graph neural network is used to generate node embedding information, which is injected into the initial model and adjusted to achieve semantic fusion of graph data and multi-source data. Finally, based on the target machine learning model, the fused data is processed, analyzed and predicted, and the final prediction result is obtained after verification and adjustment. The fusion of multi-source data and multiple algorithm models effectively improves the prediction accuracy, helping airlines optimize fuel management and reduce operating costs.

[0006] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of the present application, a method for predicting aircraft fuel consumption based on quick access recorder data is provided, comprising: obtaining quick access recorder data and a training sample set of a target aircraft; extracting features from the quick access recorder data using a data processing algorithm based on various parameter information recorded in historical quick access recorder data to generate a multi-source heterogeneous data set; processing the multi-source heterogeneous data set based on a data mapping relationship and a data conversion method to generate a target directed multi-view dynamic graph; modeling the target directed multi-view dynamic graph based on a multi-view dynamic graph neural network to generate node embedding information; injecting the node embedding information into each layer of an initial fuel consumption prediction model, adjusting the model using an adaptive learning rate to generate fused graph data and multi-source data semantic information; processing the quick access recorder data of the target aircraft based on a target machine learning model in combination with the fused graph data and multi-source data semantic information to generate aircraft fuel consumption prediction result information.

[0008] Another aspect of the present application is a device for predicting aircraft fuel consumption based on quick access recorder data, characterized in that it includes: an acquisition module for acquiring the quick access recorder data and a training sample set of a target aircraft; a processing module for extracting features from the quick access recorder data using a data processing algorithm based on various parameter information recorded in the historical quick access recorder data to generate a multi-source heterogeneous data set; processing the multi-source heterogeneous data set based on a data mapping relationship and a data conversion method to generate a target directed multi-view dynamic graph; modeling the target directed multi-view dynamic graph based on a multi-view dynamic graph neural network to generate node embedding information; injecting the node embedding information into each layer of the initial fuel consumption prediction model, adjusting the model using an adaptive learning rate to generate fused graph data and multi-source data semantic information; processing the quick access recorder data of the target aircraft based on a target machine learning model in combination with the fused graph data and multi-source data semantic information to generate aircraft fuel consumption prediction result information.

[0009] According to another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a second processor, the method for predicting aircraft fuel consumption based on fast access recorder data is implemented.

[0010] The present application provides a method and system for predicting aircraft fuel consumption based on fast access recorder data. The server obtains the QAR data and training sample set of the target aircraft, generates a multi-source heterogeneous data set through a data processing algorithm, and then processes it into a target directed multi-view dynamic graph. Next, a multi-view dynamic graph neural network is used to generate node embedding information, which is injected into the initial model and adjusted to achieve semantic fusion of graph data and multi-source data. Finally, based on the target machine learning model, the fused data is processed, analyzed and predicted, and the final prediction result is obtained after verification and adjustment. This technology integrates multi-source data with multiple algorithm models to effectively improve prediction accuracy, helping airlines optimize fuel management and reduce operating costs.

[0011] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A flowchart showing a method for predicting aircraft fuel consumption based on rapid access to recorder data provided by one embodiment of the present application is shown;

[0013] Figure 2 A schematic structural diagram of an aircraft fuel consumption prediction device based on fast access recorder data provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0014] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0015] The following combination Figure 1 The following describes an aircraft fuel consumption prediction method based on fast access recorder data according to an exemplary embodiment of the present application. In one embodiment, the present application also proposes an aircraft fuel consumption prediction method and system based on fast access recorder data. Figure 1 As shown, the method is applied to the server and includes:

[0016] S101, obtaining the target aircraft's quick access recorder data and training sample set.

[0017] In one implementation, to optimize fuel management for its Boeing 737-800 aircraft, an airline selected 50 Boeing 737-800 aircraft from its fleet that had flown both long-haul and short-haul routes within the past month as target aircraft. The diverse and representative flight data from these aircraft provided rich information for subsequent fuel consumption forecasts. After each flight, ground crew used specialized equipment that met aviation data collection standards to connect to the QAR through the aircraft's data interface. Using the accompanying data collection software, the flight data stored in the QAR was exported in CSV format. For example, for a flight from Beijing to Shanghai, the collected data included detailed parameters such as the engine's rpm throughout the flight, as well as the aircraft's altitude, speed, and heading, recorded every five minutes. This data fully documented the flight's real-time status.

[0018] Relevant performance parameters were obtained from the technical manuals and maintenance records of the Boeing 737-800 passenger aircraft. These included the specific fuel consumption rate of the engine under different operating conditions, as well as aerodynamic parameters such as the lift-to-drag ratio of the aircraft's wings. For example, for a specific engine model, in cruise mode, when the engine speed remains stable within a certain range, the corresponding fuel consumption rate is X kg / hour. We collaborated with professional meteorological data providers to obtain meteorological data for the flight period and route. For example, for the Beijing-Shanghai flight mentioned above, we obtained the ground temperature in Beijing at takeoff: 25°C, the air pressure: 1010 hPa, and wind speed and direction data at different altitudes during flight, such as 30 knots and a wind direction of 270° at 9,000 meters. Route-related data was extracted from the airline's flight operations system. For the Beijing-Shanghai route, we obtained the actual distance of 1,088 kilometers, the altitude range during the climb phase (specific values from takeoff altitude to cruising altitude), the altitude range during the descent phase, and the flight's scheduled departure time, actual departure time, estimated arrival time, and actual arrival time.

[0019] The QAR data, performance parameter data, historical meteorological data, and route characteristics data for the same flight are integrated. The flight data is grouped into 10-minute segments, and the actual fuel consumption within that segment is annotated. For example, during the 30th to 40th minute of flight, the actual fuel consumption is determined to be Y kilograms based on the changes in the fuel gauge readings in the QAR data. This data set is used as a training sample. A data validation algorithm is used to check the integrity and rationality of the QAR data. For example, it checks whether a parameter is missing at multiple consecutive recording points. If engine speed data is missing for a specific 5-minute period, that portion of the data is marked for processing. The data is also checked for rationality. If speed data significantly exceeds the normal flight speed range for the aircraft model (for example, a speed value of 1500 km / h, compared to the normal cruising speed range of 800-900 km / h for a Boeing 737-800), it is considered an outlier. Missing values were supplemented using linear interpolation based on previous and subsequent data. Outliers were corrected based on other relevant flight parameters and the normal data range for similar flights. If correction was not possible, they were eliminated. Missing values in meteorological data were supplemented using Kriging interpolation based on data from surrounding meteorological stations. For route characteristic data, the logical consistency of flight time and route distance was checked. Any discrepancies in time and distance calculations were corrected by comparing them with the actual flight trajectory and flight plan.

[0020] S102 , based on various parameter information recorded in the historical fast access recorder data, a data processing algorithm is used to extract features from the fast access recorder data to generate a multi-source heterogeneous data set.

[0021] In one embodiment, parameter screening is performed on the various parameter information recorded in the historical quick access recorder data to identify core parameters with a high correlation with fuel consumption. After normalization, the dimensionality effect is eliminated. The key feature vectors are extracted by combining the principal component analysis algorithm to generate preliminary quick access recorder feature data. For the Boeing 737-800 passenger aircraft, the historical quick access recorder (QAR) data contains many parameters, such as engine speed, flight altitude, speed, fuel flow, etc. By studying the physical mechanism of fuel consumption and statistically analyzing a large amount of historical flight data, the core parameters with a high correlation with fuel consumption are determined. For example, the fuel flow of the engine directly reflects the fuel consumption, the engine speed affects the fuel combustion efficiency and thus affects the fuel consumption, and the flight altitude and speed are related to air resistance and indirectly affect the fuel consumption. These are all determined as core parameters. These core parameters are normalized to eliminate the dimensionality effect. Taking engine speed (unit: revolutions per minute) and flight speed (unit: kilometers per hour) as examples, assuming the engine speed range is [4000, 12000] and the flight speed range is [300, 900], a minimum-maximum normalization method is used. After normalization, the engine speed X is mapped to the interval [0, 1]. The same applies to flight speed. Next, the principal component analysis (PCA) algorithm is used to extract key eigenvectors. The PCA algorithm works by transforming the original data onto a new orthogonal basis through linear transformation, maximizing the variance of the data on the new basis. The directions with the largest variance are the principal component directions. PCA analysis is performed on the data matrix composed of the normalized core parameters. For example, three principal components are extracted from the original five core parameters. These three principal components retain the majority of the information in the original data and constitute the initial fast access recorder feature data.

[0022] Aircraft performance parameter data is processed using a performance index classification system, categorized into power performance and aerodynamic performance categories. Missing values are then imputed for each category to generate standardized aircraft performance data. For the Boeing 737-800, the performance index classification system is used to categorize aircraft performance parameter data into power performance and aerodynamic performance categories. The power performance category includes parameters such as engine power and thrust; the aerodynamic performance category covers parameters such as the wing's lift coefficient, drag coefficient, and aircraft angle of attack. In real-world data, these parameters may contain missing values. For engine power data in the power performance category, if power values are missing for a specific period, mean interpolation is performed using power data from similar engines under similar operating conditions. Assume that 100 sets of power data from the same engine under similar flight conditions are collected. The mean of these values is calculated and used to impute missing values. For the wing lift coefficient data in the aerodynamic performance category, if there is any missing value, the lift coefficient is estimated using the aircraft's design parameters combined with the current flight status (such as flight altitude, speed, angle of attack and other known parameters) to fill the missing value. The missing values of data in different categories are completed to generate standardized aircraft performance data.

[0023] When processing route data for Boeing 737-800 passenger aircraft, the route network model constructed using a route topology analysis algorithm is a graph model. This route network model is based on a graph structure consisting of nodes and edges. In this scenario, each navigation point is considered a node, and the routes connecting these nodes are considered edges. This represents the route topology and is used to analyze route-related characteristics. The model consists of a node layer and an edge layer. Specifically, at the node layer, each navigation point is a node, which contains its geographic location information (such as longitude and latitude). This information is the basis for calculating distance and analyzing route direction. Different navigation points play different roles in a route, such as starting point, intermediate key turning point, and destination. At the edge layer, the edges connecting navigation points represent the aircraft's flight path between two navigation points. Edges not only represent the connection between nodes but also contain directional information, namely the aircraft's flight direction. Furthermore, edge attributes can include route-related data such as distance and flight time.

[0024] By combining nodes and edges, a directed graph is constructed, fully representing the route network from Guangzhou to Chengdu. This graph intuitively illustrates the route layout, the connections between navigation points, and the flight direction, providing a foundational framework for subsequent calculations of route distances and analysis of flight segment characteristics. The latitude and longitude coordinates of each navigation point are key parameters of the model. Assume that the coordinates of a departure navigation point in Guangzhou are (latitude1, longitude1), the coordinates of an arrival navigation point in Chengdu are (latitude2, longitude2), and the coordinates of each intermediate navigation point are (latitude_n, longitude_n). These coordinates are used to calculate the distance between navigation points. For example, the distance between two navigation points is calculated using the spherical distance formula d = r × arccos(sin(latitude1) × sin(latitude2) + cos(latitude1) × cos(latitude2) × cos(longitude2 - longitude1)), where r is the radius of the Earth. After calculating the distances between each navigation point using this spherical distance formula, these distances are accumulated to obtain the total route distance, a key quantitative characteristic of the route. For example, the route from Guangzhou to Chengdu passes through 5 navigation points. The distances d1, d2, d3, and d4 between adjacent navigation points are calculated in sequence, and the total route distance D = d1 + d2 + d3 + d4.

[0025] The starting and ending altitudes of the climb and descent phases are parameters used to calculate the climb and descent altitude characteristics of the flight segment. For example, during the takeoff climb phase, the climb altitude is 10,000 meters, starting from ground level (assuming 0 meters) and ending at a cruising altitude of 10,000 meters. During the landing descent phase, the descent altitude is 10,000 meters, starting from a cruising altitude of 10,000 meters and ending at ground level. A day is divided into the following time periods: early morning (12:00-6:00), morning (6:00-12:00), afternoon (12:00-18:00), and evening (18:00-24:00). The flight's departure and arrival times are used to determine the flight period characteristics. If a flight departs at 10:00 am and arrives at 1:00 pm, the flight period characteristic is marked as AM-PM.

[0026] Based on preliminary QAR feature data, standardized aircraft performance data, dynamic weather data, and structured route feature data, association rule mining algorithms (such as the Apriori algorithm) are used to establish data associations. The Apriori algorithm discovers association rules by identifying frequently occurring item sets within a dataset. For example, within a large amount of flight data, it was found that when the engine speed is within a certain range, the flight altitude is within a certain range, and the meteorological conditions are a specific combination of temperature and wind speed, the fuel consumption rate exhibits a specific pattern. After establishing association rules, outlier detection and cleaning are performed. Taking fuel flow data as an example, if the fuel flow value at a certain moment deviates significantly from the expected value calculated based on the association rules and exceeds a set threshold (e.g., ±20% of the expected value), the value is considered an outlier. If the outlier can be corrected based on other relevant data, it is corrected; if not, the data is discarded. After outlier detection and cleaning, the various data types are integrated according to a specific format to generate a multi-source heterogeneous dataset for aircraft fuel consumption prediction, providing high-quality data support for subsequent fuel consumption prediction model training.

[0027] S103 , based on the data mapping relationship and data conversion method, the multi-source heterogeneous data sets are processed to generate a target directed multi-view dynamic graph.

[0028] In one embodiment, for a Boeing 737-800 passenger aircraft, the generated multi-source heterogeneous dataset includes preliminary QAR feature data, standardized aircraft performance data, dynamic meteorological environment data, and structured route feature data. The preliminary QAR feature data covers multiple fuel consumption-related features, such as engine speed, flight altitude, and speed. Key features are selected based on their direct correlation with fuel consumption and data variation trends to form a subset. Three features, namely engine fuel flow rate, engine speed, and flight altitude, which have a significant impact on fuel consumption and exhibit distinct variations, are selected to form the QAR feature data subset. This allows for a more focused focus on core data during subsequent processing, improving processing efficiency and prediction accuracy. Standardized aircraft performance data is divided into power performance and aerodynamic performance categories. From the power performance category, engine thrust and engine power are selected, two parameters that directly affect fuel consumption; from the aerodynamic performance category, wing lift coefficient and drag coefficient are selected, as they relate to air resistance during flight and thus affect fuel consumption. These parameters are combined to form the aircraft performance data subset.

[0029] Dynamic meteorological environment data includes information such as temperature, air pressure, wind speed and direction. Taking into account the meteorological factors that have a greater impact on fuel consumption, temperature and wind speed are selected as the two parameters to generate the meteorological environment data subset. For example, when flying at high altitudes, temperature and wind speed have a significant impact on the aircraft's aerodynamics and engine performance, thereby affecting fuel consumption. Structured route characteristic data includes route distance, flight segment climb and descent altitude characteristics, flight period characteristics, etc. The flight segment climb altitude and flight period are selected as the two characteristics to generate the route characteristic data subset. Changes in climb altitude are directly related to the aircraft's power demand and fuel consumption, and the different meteorological conditions and flight control requirements during different flight periods will also indirectly affect fuel consumption.

[0030] The various factors affecting the fuel consumption of a Boeing 737-800 passenger aircraft are considered nodes. Factors that directly drive fuel consumption are considered source nodes. For example, engine thrust is the direct driving force behind the aircraft. Higher engine thrust generally increases fuel consumption, so engine thrust is considered a source node. Furthermore, engine speed affects fuel combustion efficiency and is also an important source node. Fuel consumption-related indicators affected by source nodes are considered target nodes. For example, fuel flow directly reflects fuel consumption and is affected by factors such as engine thrust and speed, so fuel flow is considered a target node. Fuel consumption rate, which comprehensively reflects the aircraft's fuel consumption efficiency under different operating conditions, is also considered a target node. Node information is generated based on the previously generated data subset. For each source and target node, the corresponding data value and related metadata (such as the data collection time and corresponding flight number) are combined to form the node information. For example, for the source node engine thrust, its node information may include the engine thrust value at a certain time in a flight (such as 10,000 pounds), as well as the flight number corresponding to the data (such as CA1234), the collection time (such as October 10, 2024 10:00:00), and other information.

[0031] For each data subset, edge information is generated based on the relationships that influence fuel consumption. Edge information represents the association and degree of influence between nodes. For example, considering the relationship between engine thrust and fuel flow, based on the engine's operating principle, an increase in engine thrust leads to an increase in fuel flow. Therefore, a directed edge is established between these two nodes, from the engine thrust node to the fuel flow node, indicating the positive impact of engine thrust on fuel flow.

[0032] By statistically analyzing large amounts of flight data, edge weights are determined to quantify the degree of this influence. Suppose analysis reveals that for every 1,000 lbf increase in engine thrust, fuel flow increases by an average of 50 kg / h, all other conditions remaining unchanged. The weight of the edge from engine thrust to fuel flow can then be set to 0.05 (i.e., 50 ÷ 1,000). Similarly, for the edge between engine speed and fuel consumption rate, data analysis reveals that for every 100 rpm increase in engine speed, fuel consumption rate increases by 0.1 kg / (hour·kilometer). Therefore, the weight of this edge can be set to 0.001 (i.e., 0.1 ÷ 100). In this way, each edge contains information such as the starting node, the ending node, and the weight representing the degree of influence, forming complete edge information.

[0033] Based on the node and edge information generated previously, a directed graph is constructed to generate the preliminary structure of a directed multi-view dynamic graph. The directed graph is divided into different views based on the data source and nature. For example, nodes and edges related to QAR data are divided into the QAR view, which includes nodes such as engine speed and fuel flow, and their connecting edges. Nodes and edges related to aircraft performance data are divided into the aircraft performance view, such as engine thrust and wing lift coefficient, and their edges. Meteorological and environmental data are divided into the meteorological view, which includes nodes such as temperature and wind speed, and their edges. Route characteristic data are divided into the route view, which includes nodes such as flight altitude and flight duration, and their edges. In each view, nodes and edges are connected based on their information. For example, in the QAR view, nodes such as engine speed and fuel flow are connected based on their edges, forming a local directed graph structure. Similar structures are constructed for other views. The directed graph structures of these different views are then integrated to form the preliminary structure of the directed multi-view dynamic graph. This preliminary structure demonstrates the basic relationships between different factors, but may contain some unreasonable or redundant components.

[0034] A graph optimization algorithm is used to integrate and optimize the initially constructed directed multi-view dynamic graph. This can be done using a graph optimization algorithm based on the minimum spanning tree (MST) principle, such as Prim's algorithm. The initial structure may contain some redundant edges that have little impact on the overall structure and fuel consumption prediction. For example, the relationships represented by some edges may have very little impact in practice, or some edges can be indirectly represented by combinations of other edges. The graph optimization algorithm calculates the contribution of each edge to the overall graph structure and the objective (fuel consumption prediction). Suppose, after calculation, it is found that an edge connecting two meteorological nodes (such as temperature and pressure) has a negligible impact on the fuel consumption prediction, considering all other edges. In this case, this edge is removed to simplify the graph structure and improve subsequent processing efficiency.

[0035] To make the graph structure more clearly reflect the impact of various factors on fuel consumption and their interrelationships, node positions were adjusted based on edge weights and the closeness of the connections between nodes. For example, nodes closely related to fuel consumption (such as fuel flow and engine thrust) were placed at the center of the graph, while nodes with relatively less influence were placed at the edges. Further optimization of edge weights was performed. When considering the combined effects of multiple factors, it was discovered that the original edge weight for engine speed on fuel consumption rate was not accurate. Through reanalysis of the data and recalculation, the weight of this edge was adjusted to better reflect actual conditions.

[0036] During data collection and processing, nodes corresponding to abnormal data may appear. These abnormal nodes can interfere with accurate fuel consumption analysis and prediction. For example, the data for a particular engine thrust node clearly exceeded the normal range. Upon inspection, it was found to be due to data collection errors. By setting appropriate thresholds and data validation rules, these abnormal nodes can be identified and removed. After the above integration and optimization process, a target directed multi-view dynamic graph is generated. This graph more accurately and clearly demonstrates the impact of various factors on the fuel consumption of the Boeing 737-800 passenger aircraft, providing a more reliable foundation for subsequent fuel consumption predictions.

[0037] S104: Modeling the target directed multi-view dynamic graph based on a multi-view dynamic graph neural network to generate node embedding information.

[0038] In one embodiment, the target directed multi-view dynamic graph is layered and decomposed to generate the graph's hierarchical structure information, node connection relationship information, and time series feature information, which together constitute comprehensive graph structure information. The target directed multi-view dynamic graph covers multiple views, such as the QAR data view, the aircraft performance view, the meteorological environment view, and the route feature view. From the overall architecture, these views form a multi-level structure. The bottom layer is the basic elements within each view, such as the engine speed, fuel flow, and other nodes in the QAR data view; the middle layer is the connection relationship between the nodes in each view, such as the edge between the engine speed and fuel flow; and the top layer is the association relationship between different views. By decomposing and analyzing these levels, we can clearly understand the overall architecture of the graph and the organization of each part, which is the hierarchical structure information of the graph.

[0039] In each view, nodes are connected by edges. In the QAR data view, changes in engine speed affect fuel flow, so there is a directed edge from the engine speed node to the fuel flow node. In the aircraft performance view, the engine thrust node and the engine power node may have an edge related to each other due to the engine's operating principle. A detailed analysis of the direction, weight, and connection method of the edges connecting these nodes constitutes node connectivity information. This information reveals the direct and indirect influence paths between different factors. Because the data is recorded during flight, it has a time dimension. For example, in flight data from a Boeing 737-800 passenger aircraft, parameters such as engine speed and altitude change over time during different flight phases, such as takeoff, cruise, and landing. By analyzing the timeline trends of these parameters, such as the rapid increase in engine speed during takeoff and relative stability during cruise, we can obtain time series characteristics. This information is crucial for understanding the changing patterns of fuel consumption during different flight phases. The above hierarchical structure information, node connection relationship information and time series feature information are integrated to form comprehensive graph structure information, which fully presents all the key features of the target directed multi-view dynamic graph and provides detailed and accurate basic data for subsequent neural network processing.

[0040] The adaptive parameter initialization algorithm automatically adjusts the initial parameters of the neural network based on the characteristics of the input data and the graph structure. For multi-view dynamic graph neural networks, the algorithm performs targeted initialization of neuron parameters at different levels and for different types, taking into account the differences in data distribution and node connectivity between views. When processing data from a Boeing 737-800 passenger aircraft, node data in the QAR data view changes more frequently, while data such as temperature and wind speed in the meteorological environment view remains relatively stable. Based on this difference, the adaptive parameter initialization algorithm sets a higher initial learning rate for neurons processing the QAR data view to more quickly capture data changes, and a lower initial learning rate for neurons processing the meteorological environment view to ensure model stability. The resulting model initialization information enables the neural network to better adapt to the data characteristics of different views during the initial training phase, improving training efficiency and model performance.

[0041] The graph convolutional attention mechanism combines graph convolution and attention mechanisms. Graph convolution is used to extract local features from the graph structure, while the attention mechanism automatically learns the importance weights for different nodes and edges. When processing the graph structure information of a Boeing 737-800 aircraft, the graph convolution operation aggregates information from a node and its neighboring nodes in the QAR data view. For example, when calculating the features of a fuel flow node, information from connected nodes such as engine speed and flight altitude is considered. Simultaneously, the attention mechanism assigns weights to each node based on its influence on the fuel flow node. If, under certain flight conditions, engine speed is more critical to fuel flow, the engine speed node will receive a higher weight in the calculation. By processing graph structure information with this graph convolutional attention mechanism, each node generates an initial feature representation containing rich spatiotemporal features. For example, the initial features of a fuel flow node not only include the node's own data but also incorporate information from surrounding nodes, weighted according to their importance. These features provide more valuable information for subsequent analysis.

[0042] An attention mechanism that integrates spatiotemporal dependencies processes the initial node feature information to generate time-series fused feature information. This mechanism considers dependencies in both temporal and spatial dimensions. In the temporal dimension, it captures the temporal trends of node features; in the spatial dimension, it focuses on the correlations between different nodes at the same moment. For a Boeing 737-800 passenger aircraft, changes in engine speed at different points in a flight affect fuel consumption. Furthermore, varying engine speeds interact with other spatial factors, such as flight altitude and wing lift coefficient. Taking the engine speed node as an example, the mechanism analyzes its characteristic changes during different flight phases (takeoff, cruise, and landing), while also considering its correlations with other relevant nodes (such as fuel flow rate and flight altitude) at each time point. This spatiotemporal information is then integrated to generate time-series fused feature information that reflects the comprehensive impact of engine speed throughout the flight. This feature information more comprehensively reflects the impact of various factors on fuel consumption in both temporal and spatial dimensions.

[0043] A multi-view collaborative attention mechanism is used to process temporal fusion feature information and generate node embeddings that reflect the impact of various factors on fuel consumption. This mechanism aims to integrate temporal fusion feature information from different views and achieve synergy between multiple views by learning the importance of features from different views. When processing data from a Boeing 737-800 passenger aircraft, the QAR data view reflects real-time flight parameters, the aircraft performance view reflects the aircraft's performance, the weather environment view displays external weather conditions, and the route feature view contains route-related information. These views affect fuel consumption from different perspectives.

[0044] The multi-view collaborative attention mechanism automatically learns the importance weight of each view's time-series fusion features for the final fuel consumption prediction. During the high-altitude cruise phase, the wind speed in the meteorological environment view has a greater impact on fuel consumption, so the weight of this view is increased accordingly. During the takeoff and landing phases, the parameters in the aircraft performance view and the QAR data view are more critical, and their weights are increased. By weightedly fusing the time-series fusion features of different views, node embeddings are generated to reflect the impact of each factor on fuel consumption. These node embeddings integrate key information from multiple views and can represent the role of each factor in the fuel consumption process, providing high-quality input data for the subsequent fuel consumption prediction model.

[0045] S105, injecting the node embedding information into each layer of the initial fuel consumption prediction model, adjusting the model using an adaptive learning rate, and generating fusion graph data and multi-source data semantic information.

[0046] In one embodiment, the node embedding information is dimensionally adapted and format converted to generate adapted node embedding information. The node embedding information generated by the multi-view dynamic graph neural network has dimensions and formats that cannot be directly accepted by the initial fuel consumption prediction model. Different models have different requirements for the dimensions and formats of input data. Without adaptive conversion, the model cannot effectively utilize this information for training and prediction. Suppose the node embedding information generated by the multi-view dynamic graph neural network is a tensor with a shape of (128,), while the input layer of the initial fuel consumption prediction model (assuming a multi-layer perceptron (MLP)) expects a vector with a shape of (64,) and a NumPy array data type. In this case, the principal component analysis (PCA) algorithm is used to reduce the dimensionality of the node embedding information. The PCA algorithm projects high-dimensional data into a lower-dimensional space through linear transformation while preserving the data's key features as much as possible. In this example, the 128-dimensional node embedding information is projected into a 64-dimensional space. Then, the processed tensor is converted into a NumPy array using the corresponding function to generate the adapted node embedding information.

[0047] The initial fuel consumption prediction model is loaded and frozen with pretrained weights. Key parameters of the adaptive learning rate adjustment component are initialized to generate the model's initial state information. A pretrained multilayer perceptron (MLP) is selected as the initial fuel consumption prediction model. This MLP model has an input layer, multiple hidden layers, and an output layer. It has been trained on a large amount of aircraft fuel consumption data and has learned some common fuel consumption patterns and regularities. A specific loading function is used to load the trained model parameters into the current prediction system.

[0048] To prevent excessive modification of the model's learned general knowledge during subsequent training, the model's pre-trained weights are frozen. For example, in the Keras framework, you can freeze the weights of each layer in the MLP model by setting layer.trainable = False. This ensures that during the initial training phase, these layer weights will not change due to new data input. The model will only optimize some parameters based on the new data during subsequent fine-tuning. The model is trained using the adaptive learning rate algorithm, Adam. The Adam algorithm combines the advantages of the momentum method and the RMSProp algorithm, adaptively adjusting the learning rate. During initialization, the learning rate learning_rate = 0.001, the exponential decay rates of the moment estimates beta_1 = 0.9, beta_2 = 0.999, and a small constant epsilon = 1e-8 to prevent division by zero are set. These parameters determine how the Adam algorithm adjusts the learning rate based on gradient information during training. Appropriate initialization can help the model converge to the optimal solution more quickly. After completing the above operations, the model's initial state information is generated to prepare for subsequent training.

[0049] Based on the model's initial state information, the adapted node embedding information is injected layer by layer based on the layer structure of the initial fuel consumption prediction model to generate model input information with node embeddings. Based on the layer structure of the initial fuel consumption prediction model (MLP), the adapted node embedding information is injected layer by layer, starting from the input layer. The MLP model has three hidden layers. The input layer receives the preprocessed QAR data feature vector with a dimension of (50,). The adapted node embedding information (with a dimension of (64,)) is concatenated with the QAR data feature vector of the input layer to generate a new input vector with a dimension of (50+64,)=(114,). This new vector serves as the actual input to the input layer and then passes through the three hidden layers and the output layer. At each layer, the node embedding information is calculated with the layer's weights and participates in the model calculation process. This allows the model to utilize the information on fuel consumption influencing factors carried by the node embedding information at each layer, thereby generating model input information with node embeddings.

[0050] An adaptive learning rate strategy is used to train and adjust model input information with node embeddings. The parameters of the adaptive learning rate adjustment component are dynamically updated during training to generate adjusted model parameter information. The model is trained using a training sample set containing flight data from a Boeing 737-800 aircraft, using the model input information with node embeddings. During training, the Adam algorithm is used for optimization. In the first few iterations of training, due to the large difference between the model parameters and the optimal solution and the large gradient, the Adam algorithm updates the model parameters at a relatively large learning rate (e.g., 0.001) based on the initialized parameters to accelerate model learning. As training progresses, the gradient gradually decreases. The Adam algorithm automatically reduces the learning rate based on the first and second moment estimates of the gradient to prevent the model from skipping the optimal solution when approaching the optimal solution.

[0051] In each training iteration, the Adam algorithm calculates the exponentially weighted moving average of the gradient (the first-order moment estimate m) according to the formula t and the second-order moment estimate v t ):m t =β1m t-1 +(1-β1)g t , Among them, g t is the gradient of the current iteration, m t-1 and v t-1 are the first-order and second-order moment estimates from the previous iteration. These estimates are then used to adjust the learning rate and update the model parameters. Through continuous training and parameter updates, the adjusted model parameters are generated, enabling the model to better fit the data and improve the accuracy of fuel consumption predictions for the Boeing 737-800.

[0052] Using the adjusted model parameters, joint semantic modeling is performed on the input graph data (i.e., the data information contained in the target directed multi-view dynamic graph) and multi-source data (such as QAR data, aircraft performance data, weather and environmental data, and route characteristics data). The model learns the semantic associations between different data and unearths the complex relationship between various factors and fuel consumption. In this process, the various layers of the model (input layer, hidden layer, and output layer) work together to extract and combine features from the input data. For example, the input layer receives multi-source data such as QAR data and weather data. The hidden layer transforms the data using nonlinear activation functions (such as ReLU) to discover the potential features in the data. The output layer then conducts a comprehensive analysis based on these features to obtain a semantic representation related to fuel consumption.

[0053] After joint semantic modeling, the model generates fused graph data and multi-source data semantic information. This information integrates the characteristics and relationships of various data sources, providing a more comprehensive and in-depth understanding of the various factors and interrelationships that influence the fuel consumption of a Boeing 737-800 passenger aircraft. For example, the model may discover that under low temperature and strong headwind conditions, when the aircraft is in cruise mode and the engine speed is maintained within a certain range, fuel consumption exhibits a specific pattern. This fused semantic information provides strong support for subsequent accurate fuel consumption predictions.

[0054] S106 , based on the target machine learning model combined with the fusion graph data and multi-source data semantic information, the rapid access recorder data of the target aircraft is processed to generate aircraft fuel consumption prediction result information.

[0055] In one embodiment, based on the target machine learning model, feature extraction processing is performed on the fused graph data and the semantic information of the multi-source data to generate fused data feature information. A convolutional neural network (CNN) is selected as the target machine learning model. CNN has the characteristics of local perception and weight sharing, and can effectively extract features from the data. It includes an input layer, a convolution layer, a pooling layer, and a fully connected layer. In this scenario, the input layer receives the fused graph data and the semantic information of the multi-source data, which integrates the intrinsic correlation between QAR data, aircraft performance data, meteorological environment data, and route feature data.

[0056] The fused data is input to the input layer of the CNN in the form of a two-dimensional matrix. In the convolution layer, convolution operations are performed on the data by designing convolution kernels of different sizes and weights. For example, a 3×3 convolution kernel is used to slide across different locations of the fused data to extract local features. The weights of the convolution kernel are continuously optimized during the training process to capture key features related to fuel consumption. For the portion of the fused graph data that represents the relationship between engine speed and fuel flow, the convolution kernel can learn the change pattern between the two. After alternating processing of multiple convolution layers and pooling layers, the features of the data are gradually extracted and compressed. Pooling layers (such as max pooling) are used to reduce the data dimension and retain the main features. Finally, the extracted features are integrated through the fully connected layer to generate fused data feature information. This feature information contains key information after the fusion of multiple data, such as a comprehensive feature representation of the impact of factors such as flight status, weather conditions, and aircraft performance on fuel consumption.

[0057] The flight parameter sequences and related multi-source data sequences in the target aircraft's Quick Access Recorder (QAR) data are parsed and processed to generate target event feature information. For a Boeing 737-800 passenger aircraft, the flight parameter sequences in the Quick Access Recorder (QAR) data include time-varying data such as engine speed, flight altitude, and speed. Related multi-source data sequences include aircraft performance parameters (such as engine power and wing lift coefficient), meteorological and environmental data (temperature, wind speed), and route characteristic data (route distance, flight duration). Taking a flight from Beijing to Shanghai as an example, the engine speed sequence recorded every 5 minutes in the QAR data is parsed to calculate statistical features such as its rate of change and average value. Simultaneously, the flight altitude and speed data from the same time period are combined to analyze their interrelationships. For meteorological and environmental data, the temperature and wind speed sequences at different altitudes during the flight are parsed to determine whether there are areas of sudden temperature changes or strong winds.

[0058] Based on the analysis results, target event signature information is generated. For example, if during a certain flight period, the engine speed continues to increase while the altitude increases and the speed remains stable, and weather data indicates a headwind, this information is combined to generate a target event signature representing the "headwind climb phase." This signature not only incorporates the changes in flight parameters but also incorporates the influence of multiple sources such as weather and route data, enabling a more accurate description of the potential correlation between specific flight events and fuel consumption.

[0059] Based on the fused data feature information, a correlation analysis is performed on the target event feature information to generate correlation analysis results. The Pearson correlation coefficient method is used for correlation analysis. The Pearson correlation coefficient measures the degree of linear correlation between two variables and ranges from -1 to 1. In this scenario, the Pearson correlation coefficient is calculated pairwise between each feature in the fused data feature information and each feature in the target event feature information. For example, the correlation coefficient is calculated between the fused data feature representing the relationship between engine performance and fuel consumption and the "headwind climb phase" feature in the target event feature information. A series of correlation coefficients are calculated to form the correlation analysis results. If the fused feature representing the relationship between engine power and fuel consumption has a high correlation coefficient (e.g., 0.8) with the "headwind climb phase" target event feature, it indicates that engine power has a significant impact on fuel consumption during the headwind climb phase. These correlation analysis results reveal the degree and direction of correlation between different features, providing important information for subsequent fuel consumption prediction.

[0060] The association analysis results are fed into the prediction module of the target machine learning model, processed using the model's predictive capabilities, and a preliminary aircraft fuel consumption forecast is generated. In the selected CNN model, the prediction module consists of the final fully connected layer and the output layer. The fully connected layer further integrates and transforms the association analysis results. The output layer calculates a preliminary aircraft fuel consumption forecast based on the output of the fully connected layer using a specific activation function (such as a linear activation function). During training, the prediction module learns the mapping between the association analysis results and actual fuel consumption.

[0061] The correlation analysis results are input into the prediction module. For example, a vector containing the degree of correlation of various features is input into the fully connected layer. After weight calculation and nonlinear transformation, the output layer outputs preliminary aircraft fuel consumption forecast information. This forecast information may be a specific fuel consumption value or a trend in fuel consumption (such as increase, decrease, or stability). Suppose the model calculates that under current flight conditions, the fuel consumption of a Boeing 737-800 passenger aircraft is expected to be 500 kilograms in the next 10 minutes. This is the preliminary aircraft fuel consumption forecast information.

[0062] The preliminary aircraft fuel consumption forecast is verified and adjusted to generate the final aircraft fuel consumption forecast. This is done using a historical data comparison method. A large amount of actual fuel consumption data for Boeing 737-800 aircraft under similar flight conditions is collected as a historical reference. The preliminary forecast is compared with the historical data to calculate the forecast error. For example, the difference between the predicted fuel consumption of 500 kg and the actual fuel consumption for the same flight phase, similar weather conditions, and route is calculated, and the relative error is calculated. If the forecast error exceeds a preset threshold (e.g., a relative error exceeding 10%), the preliminary forecast is adjusted. This can be done by using a rule-based adjustment method or by fine-tuning the machine learning model based on the direction and magnitude of the error. For example, if the forecast value is consistently high and the error is relatively stable, the forecast value can be reduced by a certain percentage. After these adjustments, the final aircraft fuel consumption forecast is obtained. Assume that after verification and adjustment, the final fuel consumption for the next 10 minutes is determined to be 480 kg. This is the generated aircraft fuel consumption forecast, which can be used by airlines for fuel management and flight plan optimization.

[0063] This application proposes a method and system for predicting aircraft fuel consumption based on Quick Access Recorder (QAR) data. The system comprises three core components: data processing, model building, and prediction. During the data processing phase, QAR data and a training sample set are acquired from the target aircraft. The historical QAR data is then processed through parameter screening, normalization, and principal component analysis. This data is then combined with aircraft performance, meteorological environment, and route characteristics to generate a multi-source heterogeneous dataset. This dataset is then classified and filtered to construct a directed multi-view dynamic graph, providing the foundation for subsequent modeling.

[0064] In terms of model construction, a multi-view dynamic graph neural network is used to model the graph. Through operations such as layered decomposition, adaptive parameter initialization, and a graph convolutional attention mechanism, node embedding information is generated. This information is then injected into the initial fuel consumption prediction model, and adaptive learning rate adjustment is used to achieve semantic fusion of graph data and multi-source data. During the prediction phase, feature extraction is performed on the fused data based on the target machine learning model. Correlation analysis is performed based on the flight parameters in the QAR data and the results of multi-source data sequence analysis. The analysis results are input into the model prediction module to obtain preliminary prediction information. After verification and adjustment, the final aircraft fuel consumption prediction result is obtained. By integrating multi-source data and utilizing multiple algorithms and models, the accuracy of aircraft fuel consumption prediction is effectively improved, providing strong support for airlines to optimize fuel management and reduce costs.

[0065] In one embodiment, Figure 2 As shown, the present application also provides an aircraft fuel consumption prediction device based on fast access recorder data, comprising:

[0066] An acquisition module 201 is used to acquire the target aircraft's fast access recorder data and training sample set;

[0067] Processing module 202 is used to extract features from the historical quick access recorder data using a data processing algorithm based on the parameter information recorded in the quick access recorder data to generate a multi-source heterogeneous data set; process the multi-source heterogeneous data set based on data mapping relationships and data conversion methods to generate a target directed multi-view dynamic graph; model the target directed multi-view dynamic graph based on a multi-view dynamic graph neural network to generate node embedding information; inject the node embedding information into each layer of the initial fuel consumption prediction model, adjust the model using an adaptive learning rate, and generate fused graph data and multi-source data semantic information; process the quick access recorder data of the target aircraft based on the target machine learning model in combination with the fused graph data and multi-source data semantic information to generate aircraft fuel consumption prediction result information.

[0068] Each embodiment of this application is described in a related manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences between the other embodiments. In particular, the embodiments of the method, electronic device, electronic device, and readable storage medium for evaluating aircraft fuel consumption prediction based on rapid access recorder data are generally similar to the aforementioned embodiment of the method for predicting aircraft fuel consumption based on rapid access recorder data, so the description is relatively simple. For relevant portions, reference can be made to the partial description of the aforementioned embodiment of the method for predicting aircraft fuel consumption based on rapid access recorder data.

Claims

1. A method for predicting aircraft fuel consumption based on fast access recorder data, characterized in that: include: Obtaining fast access recorder data and training sample sets of target aircraft; Based on the parameter information recorded in the historical fast access recorder data, the data processing algorithm is used to extract the features of the fast access recorder data to generate a multi-source heterogeneous data set; Based on data mapping relationships and data conversion methods, multi-source heterogeneous data sets are processed to generate target directed multi-view dynamic graphs; Model the target directed multi-view dynamic graph based on the multi-view dynamic graph neural network and generate node embedding information; Inject node embedding information into each layer of the initial fuel consumption prediction model, adjust the model using an adaptive learning rate, and generate fused graph data and multi-source data semantic information; Based on the target machine learning model combined with fusion graph data and multi-source data semantic information, the rapid access recorder data of the target aircraft is processed to generate aircraft fuel consumption prediction result information.

2. The method according to claim 1, wherein Based on the various parameter information recorded in the historical fast access recorder data, the data processing algorithm is used to extract features from the fast access recorder data to generate a multi-source heterogeneous data set, including: Perform parameter screening on various parameters recorded in historical fast access recorder data to identify core parameters with a high correlation with fuel consumption. Normalization is performed to eliminate dimensionality effects, and key feature vectors are extracted using a principal component analysis algorithm to generate preliminary fast access recorder feature data. Process the aircraft performance parameter data through the performance index classification system and divide the data into power performance category and aerodynamic performance category; Perform missing value filling on different categories of data to generate standardized aircraft performance data; Use spatiotemporal interpolation algorithms to supplement missing data in historical meteorological and environmental data and generate dynamic meteorological and environmental data; Based on the route characteristic data, a route network model is constructed using a route topology analysis algorithm to calculate the route distance and the climb and descent altitude characteristics of each segment. The flight period characteristics are extracted by combining the flight schedule data to generate structured route characteristic data. Based on preliminary QAR feature data, standardized aircraft performance data, dynamic meteorological environment data, and structured route feature data, data associations are established through association rule mining algorithms. After outlier detection and cleaning, a multi-source heterogeneous dataset for aircraft fuel consumption prediction is generated.

3. The method according to claim 1, wherein Based on data mapping relationships and data conversion methods, multi-source heterogeneous data sets are processed to generate target directed multi-view dynamic graphs, including: Classify and filter the preliminary QAR feature data, standardized aircraft performance data, dynamic meteorological environment data, and structured route feature data in the multi-source heterogeneous data set to generate QAR feature data subsets, aircraft performance data subsets, meteorological environment data subsets, and route feature data subsets; Taking the various factors that affect fuel consumption as nodes, the factors that directly drive fuel consumption as source nodes, and the fuel consumption-related indicators affected by them as target nodes, node information is generated based on each data subset; Based on each data subset, the effects of the relationship affecting fuel consumption are processed to generate side information; Based on the generated node and edge information, a directed graph is constructed to generate the preliminary structure of a directed multi-view dynamic graph. The preliminarily constructed directed multi-view dynamic graph is integrated and optimized. By using a graph optimization algorithm, the node positions and edge weights are adjusted to make the graph structure reflect the influence of various factors on fuel consumption and their mutual relationships. At the same time, redundant edges and abnormal nodes are removed to generate the target directed multi-view dynamic graph.

4. The method according to claim 3, wherein The target directed multi-view dynamic graph is modeled based on a multi-view dynamic graph neural network to generate node embedding information, including: The target directed multi-view dynamic graph is decomposed into layers to generate the graph's hierarchical structure information, node connection relationship information, and time series feature information, which together constitute comprehensive graph structure information; For multi-view dynamic graph neural networks, an adaptive parameter initialization algorithm is used to initialize network parameters according to the graph structure characteristics and data features to generate model initialization information; Based on the model initialization information, an innovative graph convolution attention mechanism is used to process the graph structure information to generate initial node feature information containing rich spatiotemporal features; The initial node feature information is processed using the attention mechanism that integrates spatiotemporal dependencies to generate temporal fusion feature information; A multi-view collaborative attention mechanism is used to process the temporal fusion feature information and generate node embedding information that reflects the impact of various factors on fuel consumption.

5. The method according to claim 1, wherein The node embedding information is injected into each layer of the initial fuel consumption prediction model. The model is adjusted using an adaptive learning rate to generate fused graph data and multi-source data semantic information, including: Perform dimension adaptation and format conversion on the node embedding information to generate adapted node embedding information; Load and freeze pre-trained weights of the initial fuel consumption prediction model, initialize key parameters of the adaptive learning rate adjustment component, and generate the model's initial state information; Based on the model's initial state information, the adapted node embedding information is injected layer by layer into the layer structure of the initial fuel consumption prediction model to generate model input information with node embedding. Adopting an adaptive learning rate strategy, the model input information with node embedding is trained and adjusted. During the training process, the parameters of the adaptive learning rate adjustment component are dynamically updated to generate the adjusted model parameter information. Based on the adjusted model parameter information, the input graph data and multi-source data are jointly semantically modeled to generate fused graph data and multi-source data semantic information.

6. The method according to claim 1, wherein Based on the target machine learning model, combined with the fusion graph data and multi-source data semantic information, the rapid access recorder data of the target aircraft is processed to generate the aircraft fuel consumption prediction result information, including: Based on the target machine learning model, feature extraction and processing are performed on the fused graph data and multi-source data semantic information to generate fused data feature information: The flight parameter sequence and related multi-source data sequence in the target aircraft's fast access recorder data are parsed and processed to generate target event feature information: Based on the fusion data feature information, the target event feature information is subjected to correlation analysis and processing to generate correlation analysis result information: The association analysis results are fed into the prediction module of the target machine learning model, processed using the model’s predictive capabilities, and preliminary aircraft fuel consumption prediction information is generated: The preliminary aircraft fuel consumption prediction information is verified and adjusted to generate aircraft fuel consumption prediction result information.

7. An aircraft fuel consumption prediction device based on fast access recorder data, characterized in that: The device comprises: An acquisition module, used to obtain the target aircraft's fast access recorder data and training sample sets; The processing module is used to extract features from the historical quick access recorder data using a data processing algorithm based on the parameter information recorded in the quick access recorder data to generate a multi-source heterogeneous data set; process the multi-source heterogeneous data set based on data mapping relationships and data conversion methods to generate a target directed multi-view dynamic graph; model the target directed multi-view dynamic graph based on a multi-view dynamic graph neural network to generate node embedding information; inject the node embedding information into each layer of the initial fuel consumption prediction model, adjust the model using an adaptive learning rate, and generate fused graph data and multi-source data semantic information; process the quick access recorder data of the target aircraft based on the target machine learning model in combination with the fused graph data and multi-source data semantic information to generate aircraft fuel consumption prediction result information.

8. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; The first processor is configured to execute the aircraft fuel consumption prediction method based on fast access recorder data according to any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the method for predicting aircraft fuel consumption based on fast access recorder data as claimed in any one of claims 1 to 6 is implemented.

Images