A method and system for end-to-end prediction of fuel consumption

Abstract

The present application belongs to the technical field of aviation fuel consumption prediction, and relates to an aviation fuel consumption end-to-end prediction method and system, aiming to solve the problem of physical constraint and data-driven fragmentation and weak model generalization ability in the prior art. The present application comprises: converting the fast access recorder data into standardized time series data, generating an enhanced data set through an atmospheric model and wind field compensation; dividing the flight phase and calculating the slope angle through time series state decoding; calculating the aerodynamic coefficient and calculating the thrust, combining the engine recession coefficient to generate a physical constraint feature set; obtaining the core feature set through nonlinear screening and causal test; the encoder composed of a local feature extraction network and a time series memory network determines the physical baseline fuel consumption based on the thrust, lift-drag ratio and recession coefficient, and learns the residual correction term for superposition, and outputs the aviation fuel consumption prediction result. The present application realizes the deep integration of physical mechanism and data-driven, and takes into account the prediction accuracy and physical consistency.

Description

Technical Field

[0001] This invention belongs to the field of aviation fuel consumption prediction, and specifically relates to an end-to-end prediction method and system for aviation fuel consumption. Background Technology

[0002] Aviation fuel cost is a core component of airline operating costs, and accurate fuel consumption forecasting is crucial for dispatch release, flight plan optimization, fuel-efficient operations, and carbon emission accounting. With the widespread deployment of fast access recorders (FARs) in civil aircraft, fuel consumption forecasting methods based on flight data have become a research hotspot.

[0003] Currently, mainstream methods for predicting aviation fuel consumption mainly fall into two categories. One category is a purely data-driven approach based on fast access recorder (FAR) data. This method establishes a mapping relationship between flight parameters and fuel consumption by performing statistical regression, machine learning, or time-series network modeling on flight data such as altitude, speed, and engine parameters. This type of method has good fitting accuracy under normal operating conditions, but because it does not incorporate physical mechanisms such as aerodynamic characteristics, thrust balance, and engine performance degradation, it suffers from severe prediction drift, poor physical consistency, and insufficient extrapolation stability under extreme conditions such as high temperatures and altitudes, steep climbs, and waiting for turns. The other category is based on simplified physical models, relying on performance databases provided by engine manufacturers and standard atmospheric tables to calculate fuel consumption. This type of method has good physical interpretability, but it lacks sufficient characterization of actual wind fields, altitude deviations, engine performance degradation, and the refined operation of flaps and landing gear, thus limiting its prediction accuracy.

[0004] In addition, existing methods generally suffer from the following shortcomings: they only perform simple statistical processing such as mean and extreme values ​​on high-frequency time-series data from fast access recorders, losing key time-series features of dynamic transition phases such as climb, acceleration, and turning; the flight phase division is coarse, usually only divided into three major segments: takeoff, cruise, and landing, failing to characterize the fuel consumption differences in subdivided phases such as initial climb, acceleration, approach, and go-around; feature selection relies on human experience, making it difficult to quantify the causal link between aerodynamic characteristics, thrust output, and fuel consumption, resulting in the omission of key features or the introduction of redundant features; the models lack an online incremental update mechanism, quickly becoming invalid after aircraft type replacement or engine performance degradation, and do not support stable extrapolation under multiple aircraft types, multiple airports, and variable weather conditions.

[0005] In summary, existing technologies cannot simultaneously meet the requirements of civil aviation operations for high accuracy, high robustness, and high interpretability in fuel forecasting. There is an urgent need for an end-to-end fuel consumption forecasting method that deeply integrates physical constraints and data-driven approaches and has online adaptive capabilities. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, namely the disconnect between physical constraints and data-driven approaches, and the weak generalization ability of models, this invention provides an end-to-end prediction method and system for aviation fuel consumption.

[0007] A first aspect of the present invention provides an end-to-end prediction method for aviation fuel consumption, comprising:

[0008] Acquire fast access recorder data and convert it into standardized time-series data, wherein the standardized time-series data includes at least vacuum speed, ground speed, total temperature, timestamp, latitude and longitude location data, altitude data, speed data, flap position data and landing gear status data; The vacuum velocity and ground velocity are corrected based on atmospheric models and wind field data. The calculated atmospheric environmental parameters, the corrected vacuum velocity and ground velocity, and all parameters in the standardized time series data are used together as an enhanced dataset. Altitude data, speed data, flap position data, and landing gear status data are used as observation sequences. Flight phases are divided by time-series state decoding and bank angle is calculated. Time-series data with phase labels are output. The drag is calculated from the time-series data with stage annotations, and the lift coefficient, drag coefficient, and side force coefficient are calculated. The thrust is calculated through force decomposition, and the engine decay coefficient is determined to generate a set of physical constraint features. A candidate feature pool is constructed from the augmented dataset and the physical constraint feature set, and the core feature set is obtained through nonlinear filtering, redundancy removal and causality test. The unit thrust fuel consumption rate is obtained, and the core feature set is input into an encoder composed of a local feature extraction network and a temporal memory network. The physical baseline fuel consumption is determined based on the thrust, the unit thrust fuel consumption rate, the ratio of the lift coefficient to the drag coefficient, and the engine decay coefficient. The encoder learns the residual correction term, and the physical baseline fuel consumption and the residual correction term are superimposed to output the fuel consumption prediction result.

[0009] Furthermore, the method for acquiring fast access recorder data and converting it into standardized time-series data is as follows: Identify the message format of the fast access logger data and extract parameter fields according to the field mapping rules of the corresponding format; The extracted data is processed by a combination of median deviation filtering based on sliding window, standard deviation criterion and first difference gradient test to detect and remove outliers and frozen values. By utilizing the correlation between parameter channels, the missing channels are reconstructed using the Kalman smooth interpolation algorithm, and the interpolation error is controlled within a preset range. The engine parameters at the first sampling frequency and the atmospheric data at the second sampling frequency are resampled to the same reference frequency. The resampled sequence is then denoised using a polynomial fitting smoothing filter to obtain the standardized time series data.

[0010] Furthermore, the augmented dataset is constructed as follows: By calling up the altitude data and total temperature from the standardized time-series data, the real-time atmospheric environment is retrieved according to the atmospheric model to determine the air density and temperature deviation relative to the standard atmosphere under the current flight conditions. Using the timestamp and latitude / longitude location data as indexes, the reanalysis wind field data is interpolated in the spatiotemporal dimensions to decompose the tailwind component acting on the aircraft's flight path and the crosswind component perpendicular to the flight path. The vacuum velocity and ground velocity recorded in the standardized time series data are calculated using the downwind component and crosswind component to obtain the corrected vacuum velocity and ground velocity. The air density, temperature deviation, downwind component, and crosswind component are added as new parameters, and together with the corrected vacuum velocity and ground velocity, as well as all parameters in the standardized time series data, constitute the enhanced dataset.

[0011] Furthermore, altitude data, speed data, flap position data, and landing gear status data are used as observation sequences. Flight phases are divided through time-series state decoding, and bank angles are calculated. Time-series data with phase labels are output. The method is as follows: The temporal state decoding adopts a hidden Markov model. The altitude data, speed data, flap position data, and landing gear state data are used as the observation sequence and input into the hidden Markov model for state decoding. The model identifies and labels each flight stage, including taxiing out, takeoff run, initial climb, climb, acceleration, cruise, descent, approach, go-around, and taxiing in. Latitude and longitude location data are extracted from the enhanced dataset, the turning radius is estimated using the differential curvature method, and the real-time slope angle is calculated by combining the corrected vacuum speed in the enhanced dataset. Flight phases with durations below a preset threshold are merged into adjacent flight phases. The phase labeling results and the slope angle are appended to the augmented dataset, and the time-series data with phase labels is output.

[0012] Furthermore, drag is calculated from the time-series data with stage annotations, and the lift coefficient, drag coefficient, and side force coefficient are calculated. Thrust is calculated through force decomposition, and the engine decay coefficient is determined to generate a physical constraint feature set. The method is as follows: Acceleration and track angle are extracted from the time-series data with stage annotations. The forces acting on the aircraft are decomposed along the track direction, and the drag is calculated in reverse. The reference area of ​​the aircraft wing and the mass of the aircraft are obtained. Combined with the drag, the air density in the enhanced dataset and the corrected vacuum velocity, the lift coefficient, drag coefficient and side force coefficient are calculated respectively, and low-pass filtering is performed for smoothing. The acceleration composition term is determined based on the acceleration and the rate of change of wind speed. The thrust is calculated by adding the product of the aircraft mass and the acceleration composition term, the gravity component along the flight path, the drag, and the wind component along the thrust direction determined based on the tailwind component and the crosswind component. Identify the cruise phase from the time-series data with phase annotations, and extract the engine rotor speed corresponding to the cruise phase; The engine rotor speed is compared with the reference characteristic curve of the corresponding engine model to obtain the engine degradation coefficient; The lift coefficient, drag coefficient, lateral force coefficient, thrust, and engine decay coefficient are combined to form the physical constraint feature set.

[0013] Furthermore, a candidate feature pool is constructed from the augmented dataset and the physical constraint feature set. Through nonlinear filtering, redundancy removal, and causality testing, a core feature set is obtained. The method is as follows: The original flight parameters and atmospheric environment parameters are extracted from the augmented dataset, and the physical constraint parameters are extracted from the physical constraint feature set. The time-series statistics of the original flight parameters, atmospheric environment parameters and physical constraint parameters within the sliding window are calculated respectively. The original flight parameters, atmospheric environment parameters, physical constraint parameters and time-series statistics are merged to form a candidate feature pool. Calculate the distance correlation coefficient between each feature in the candidate feature pool and fuel consumption, and retain features whose distance correlation coefficient is higher than a preset correlation threshold; Calculate the conditional mutual information for the retained features, and remove features whose conditional mutual information is lower than a preset redundancy threshold; After removing redundant features, an additive noise model is used to test the causal direction. Features with a unidirectional causal relationship with fuel consumption are retained, forming the core feature set.

[0014] Furthermore, to obtain the fuel consumption rate per unit thrust, the core feature set is input into an encoder composed of a local feature extraction network and a temporal memory network. Based on the thrust, the fuel consumption rate per unit thrust, the ratio of the lift coefficient to the drag coefficient, and the engine degradation coefficient, the physical baseline fuel consumption is determined. The method is as follows: The local feature extraction network is composed of multiple convolutional layers cascaded together, with the number of convolutional kernels increasing layer by layer; the temporal memory network is a bidirectional long short-term memory layer. The core feature set is input into the encoder, and the local feature extraction network and the temporal memory network extract temporal features. The reference fuel consumption rate is obtained by multiplying the thrust by the fuel consumption rate per unit thrust. The correction factor is obtained by multiplying the ratio of the lift coefficient to the drag coefficient and the engine degradation coefficient. The fuel consumption correction coefficient is determined according to the correction factor and a preset ratio. The reference fuel consumption rate is divided by the fuel consumption correction coefficient to obtain the physical baseline fuel consumption.

[0015] Furthermore, the encoder learns a residual correction term, which is then superimposed on the physical baseline fuel consumption to output a predicted fuel consumption result. The method is as follows: The encoder outputs the residual correction term based on the temporal characteristics; The predicted fuel consumption is obtained by adding the physical baseline fuel consumption to the residual correction term. The encoder is obtained through training, and a joint loss function is used during training. The joint loss function includes a first loss term that represents the deviation between the predicted fuel consumption and the actual fuel consumption, a second loss term that constrains the residual correction term to approach zero, and a model parameter regularization term. The loss is weighted according to the flight phase, with the loss weight of the climb phase and the approach phase being greater than the loss weight of the taxiing, taxiing, and cruise phases. When new fast access recorder data arrives, the parameters of the local feature extraction network and the temporal memory network are fixed, and the parameters of the network layer in the encoder used to output the residual correction term are fine-tuned only to update the model online. The predicted fuel consumption results are output, including the fuel consumption of each flight phase and its proportion in the total fuel consumption, as well as the prediction uncertainty range generated based on deep integration.

[0016] Furthermore, a federated learning strategy is employed when updating the model online, the federated learning strategy including: Each airline node trains the encoder locally using its own fast access recorder data, and extracts the model gradient after training is complete. The extracted model gradients are uploaded to the central server, but the original fast access recorder data is not uploaded. The central server aggregates the received model gradients by weighted average based on the amount of data to obtain the global gradient, updates the encoder parameters using the global gradient, and distributes the updated parameters to each airline node.

[0017] In a second aspect, the present invention provides an end-to-end aviation fuel consumption prediction system for performing an end-to-end aviation fuel consumption prediction method, the system comprising: The data preprocessing module is used to acquire fast access recorder data and convert it into standardized time-series data. The standardized time-series data includes at least vacuum speed, ground speed, total temperature, timestamp, latitude and longitude location data, altitude data, speed data, flap position data, and landing gear status data. The atmospheric environment compensation module is used to correct the vacuum velocity and ground velocity based on the atmospheric model and wind field data. The calculated atmospheric environment parameters, the corrected vacuum velocity and ground velocity, and all parameters in the standardized time series data are used together as an enhanced dataset. The flight phase segmentation module is used to take altitude data, speed data, flap position data and landing gear status data as observation sequences, divide the flight phases through time-series state decoding and calculate the bank angle, and output time-series data with phase labels. The physical feature generation module is used to back-calculate drag and lift coefficient, drag coefficient and side force coefficient from the time series data with stage annotation, calculate thrust through force decomposition, determine engine decay coefficient, and generate physical constraint feature set; The feature selection module is used to construct a candidate feature pool from the augmented dataset and the physical constraint feature set, and obtain the core feature set through nonlinear filtering, redundancy removal and causality test; The end-to-end prediction module is used to obtain the fuel consumption rate per unit thrust. The core feature set is input into an encoder composed of a local feature extraction network and a temporal memory network. The physical baseline fuel consumption is determined based on the thrust, the fuel consumption rate per unit thrust, the ratio of the lift coefficient to the drag coefficient, and the engine decay coefficient. The encoder learns the residual correction term, and the physical baseline fuel consumption and the residual correction term are superimposed to output the fuel consumption prediction result.

[0018] The beneficial effects of this invention are: This invention embeds aerodynamic characteristic parameters, thrust calculation results, and engine decay coefficient as physical constraints into the prediction model, ensuring that fuel consumption predictions remain consistent with actual flight mechanics under extreme conditions such as high temperature, high altitude, and steep climbs. This avoids the problem of prediction value drift outside the operating condition boundary in purely data-driven methods and guarantees the physical consistency of the prediction results.

[0019] This invention constructs high-frequency standardized time-series data containing parameters such as vacuum speed, ground speed, altitude, velocity, flap position, and landing gear status. It then decodes the time-series status to divide the flight into multiple stages, including taxiing, takeoff run, initial climb, climb, acceleration, cruise, descent, approach, go-around, and taxiing. This fully preserves the time-series characteristics of dynamic transition stages such as climb, acceleration, and turning, avoiding the information loss caused by simply averaging or extreme value statistics of high-frequency data in existing technologies. This provides a more refined, staged data foundation for subsequent predictions.

[0020] This invention calculates drag, lift coefficient, drag coefficient, and lateral force coefficient by back-calculating from time-series data with stage annotations, and determines the acceleration composite term based on acceleration and wind speed change rate to calculate thrust. The aforementioned aerodynamic coefficients are primarily calculated in real-time based on fast access recorder data. Thrust and fuel consumption per unit thrust are calibrated using engine baseline characteristic data, eliminating the need for a complete external aerodynamic database and reducing reliance on external data. Simultaneously, by comparing the engine rotor speed with the baseline characteristic curve during the cruise phase to obtain the degradation coefficient, the engine performance degradation state is quantified and incorporated into the feature system, enabling the prediction model to reflect the engine's actual service condition.

[0021] This invention constructs a candidate feature pool from augmented datasets and physical constraint feature sets, uses distance correlation coefficients for nonlinear screening, conditional mutual information for redundancy removal, and additive noise models for causal direction verification. The resulting core feature set has low redundancy and a unidirectional causal relationship with fuel consumption, clarifying the causal link between aerodynamic characteristics, thrust output, and fuel consumption, and reducing the subjectivity of manual feature selection.

[0022] This invention inputs a core feature set into an encoder composed of a one-dimensional convolutional network and a bidirectional long short-term memory network. A physical baseline fuel consumption is determined using thrust, fuel consumption per unit thrust, the ratio of lift coefficient to drag coefficient, and the decay coefficient. The encoder then learns residual correction terms, which are then superimposed. The physical baseline fuel consumption ensures the basic physical rationality of the prediction results, while the residual correction terms learn bias patterns not covered by the physical model from the data. These two complement each other, improving prediction accuracy. During training, losses are weighted according to flight phases, resulting in higher fitting accuracy during high-fuel-consumption climb and approach phases. When new fast access recorder data arrives, only the parameters of the network layer in the encoder used to output the residual correction terms are fine-tuned. This eliminates the need to retrain all network parameters, enabling online model updates and reducing the computational overhead of model maintenance.

[0023] This invention employs a federated learning strategy, where each airline node trains its model locally using its own data and only uploads the gradients. The central server then aggregates the gradients using a weighted average based on the amount of data and distributes updated parameters. This enables collaborative modeling without the original data leaving the local network, expanding the diversity of aircraft types, airports, and weather scenarios covered by the model training. It also enhances the model's generalization ability under various aircraft types, airports, and changing weather conditions, while protecting the data privacy of all parties.

[0024] The output of this invention not only includes the predicted value of aviation fuel consumption, but also provides the fuel consumption and its proportion in each flight phase, as well as the prediction uncertainty range generated based on deep integration, providing more comprehensive decision reference information for dispatch release and flight plan optimization. Attached Figure Description

[0025] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating an end-to-end prediction method for aviation fuel consumption according to the present invention. Figure 2 This is a structural diagram of an end-to-end aviation fuel consumption prediction system according to the present invention; Figure 3 This is a schematic diagram of the structure of a computer system used to implement the methods, systems, and electronic devices of this application. Detailed Implementation

[0026] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] This invention provides an end-to-end prediction method for aviation fuel consumption, comprising: Step S10: Acquire fast access recorder data and convert it into standardized time-series data. The standardized time-series data includes at least vacuum speed, ground speed, total temperature, timestamp, latitude and longitude location data, altitude data, speed data, flap position data, and landing gear status data. Step S20: Correct the vacuum velocity and ground velocity based on the atmospheric model and wind field data, and use the calculated atmospheric environmental parameters, the corrected vacuum velocity and ground velocity, and all parameters in the standardized time series data as an enhanced dataset. Step S30: The altitude data, speed data, flap position data, and landing gear status data are used as an observation sequence. The flight phases are divided by time-series state decoding and the bank angle is calculated. The time-series data with phase labels is output. Step S40: Calculate drag and lift coefficient, drag coefficient and side force coefficient from the time series data with stage annotations; calculate thrust through force decomposition; determine engine decay coefficient; and generate physical constraint feature set. Step S50: Construct a candidate feature pool from the augmented dataset and the physical constraint feature set, and obtain the core feature set through nonlinear filtering, redundancy removal and causality test; Step S60: Obtain the unit thrust fuel consumption rate, input the core feature set into the encoder composed of a local feature extraction network and a temporal memory network, determine the physical baseline fuel consumption based on the thrust, the unit thrust fuel consumption rate, the ratio of the lift coefficient to the drag coefficient, and the engine decay coefficient, and learn the residual correction term by the encoder, superimpose the physical baseline fuel consumption and the residual correction term, and output the fuel consumption prediction result.

[0029] To more clearly explain the end-to-end prediction method for aviation fuel consumption according to the present invention, the following is in conjunction with... Figure 1 The steps in the embodiments of the present invention are described in detail below: Step S10: Acquire fast access recorder data and convert it into standardized time-series data. The standardized time-series data includes at least vacuum speed, ground speed, total temperature, timestamp, latitude and longitude location data, altitude data, speed data, flap position data, and landing gear status data. In this embodiment, the method for acquiring fast access recorder data and converting it into standardized time-series data is as follows: Step S11: Identify the message format of the fast access recorder data and extract parameter fields according to the field mapping rules of the corresponding format; Step S12: For the extracted data, a combination of median deviation filtering based on sliding window, standard deviation criterion and first-order difference gradient test is used to detect and remove outliers and frozen values. Step S13: Utilize the correlation between parameter channels to reconstruct the missing channels using the Kalman smooth interpolation algorithm, and control the interpolation error within a preset range; Step S14: Resample the engine parameters at the first sampling frequency and the atmospheric data at the second sampling frequency to the same reference frequency, and use a polynomial fitting smoothing filter to reduce noise in the resampled sequence to obtain the standardized time series data.

[0030] In one specific embodiment of the present invention, the first step is to acquire Fast Access Recorder (QAR) data, i.e., QAR data, and convert it into standardized time-series data. Considering the diverse sources of QAR data in real-world operations, such as common ARINC429, ARINC717, A429-FAST, and IOSA, directly processing these heterogeneous raw data is extremely difficult. Therefore, this step aims to transform these raw, noisy, and inconsistently formatted data streams into a clean, complete, time-aligned, and uniformly formatted high-quality dataset through a series of rigorous automated processing steps, laying a solid foundation for subsequent physical feature extraction and accurate prediction model construction.

[0031] Specifically, the data transformation process begins with the adaptive parsing of the original QAR messages. A built-in intelligent message format recognition module automatically identifies the specific format of the QAR data being processed, such as ARINC429 or ARINC717, by analyzing the message's header features, synchronization words, or specific data frame structures. Once the format is successfully identified, the system immediately invokes a pre-defined field mapping rule base corresponding to that format. This rule base defines in detail the physical parameters corresponding to different byte bits or tags in the binary or hexadecimal message. For example, it precisely decodes the encoded value of a specific address into key flight parameters such as vacuum speed, ground speed, timestamps with millisecond precision, latitude and longitude location data, barometric altitude, Mach number, flap angles, and landing gear retraction / extension status. Through this automated parsing and mapping process, the raw, machine-readable bitstream is efficiently converted into a structured sequence of parameters containing preliminary physical meaning.

[0032] After obtaining the initial structured parameter sequence, the next step is to rigorously clean and filter the data, as sensor noise, signal transmission interruptions, or transient failures of data recording units often introduce outliers into the data. Therefore, this embodiment employs a robust cleaning strategy combining multiple verifications. A median absolute deviation filtering method based on a sliding window, also known as the Hampel filter, is used to scan the time-series data. The principle is to calculate the median, rather than the mean, of the data within a preset sliding window, such as 5 to 10 sampling points, and then calculate the deviation based on the median. Since the median is far less sensitive to extreme values ​​than the mean, this method can effectively identify and remove outliers that significantly deviate from the local data trend, while simultaneously preventing the outliers themselves from contaminating the detection benchmark.

[0033] As supplementary verification, the system will also apply the traditional 3σ criterion, which identifies data points deviating from the local mean by more than three standard deviations as outliers, further capturing outliers on the edge of the statistical distribution. Furthermore, to identify and handle frozen values ​​caused by sensor malfunctions or signal interruptions—that is, parameters remaining constant within the expected change period—a first-order difference gradient test is introduced. By calculating the rate of change of parameter values ​​at consecutive time points, if it remains zero or close to zero for a prolonged period, the data segment is marked as a frozen value and removed. This triple filtering mechanism, combining robust statistics, classical statistics, and dynamic verification, ensures the comprehensiveness and accuracy of data cleaning.

[0034] Missing values ​​generated during data cleaning or inherent data gaps in the original data can severely affect the continuity of time-series analysis, necessitating scientific reconstruction. This embodiment does not employ simple linear interpolation but instead leverages the strong inherent physical correlations between flight parameters, such as the definite aerodynamic relationships between vacuum velocity, Mach number, total temperature, and static temperature. Based on this, the system uses an advanced Kalman smoothing interpolation algorithm for data reconstruction. This algorithm treats the entire flight parameter system as a dynamic state-space model, using not only data before the missing point for forward prediction but also data after the missing point for backward correction, thus providing the optimal posterior estimate for the missing value. This two-way information fusion significantly improves the accuracy and physical realism of the interpolation. To ensure the quality of the reconstructed data, the system also sets error control thresholds; for example, the relative error generated by interpolation must be controlled within 5%. Reconstructed data points exceeding this threshold are marked or subject to backup strategies, thereby ensuring the high integrity and reliability of the entire dataset.

[0035] To address the issue of inconsistent sampling frequencies across different sensors—for example, engine parameters might be recorded at a high frequency of 8Hz, while atmospheric or navigation data might be recorded at 1Hz—time alignment and standardization of all parameters are necessary. This embodiment resamples all parameter sequences to a reference frequency of 1Hz. For downsampling from high to low frequencies, methods such as sliding window averaging or decimation are used; for upsampling from low to high frequencies, conformal interpolation is employed. After frequency unification, to eliminate residual high-frequency noise from the resampling process and the original data without compromising key dynamic characteristics, the system applies a Savitzky-Gray smoothing filter to all standardized frequency sequences. This filter is essentially a local polynomial fitting-based filtering method. It fits the data with a low-order polynomial within a sliding window, for example, with a window size of 3 to 7 sampling points, thereby smoothing noise while preserving the original signal characteristics, such as rapid changes in climb rate and the establishment of slope during turns. After the above analysis, cleaning, reconstruction and standardization processes, the final output standardized time series data has excellent characteristics of uniform format, reliable quality and time synchronization, providing ideal data input for subsequent fine-grained flight phase division, construction of physical constraint features and high-precision end-to-end fuel consumption prediction.

[0036] Step S20: Correct the vacuum velocity and ground velocity based on the atmospheric model and wind field data, and use the calculated atmospheric environmental parameters, the corrected vacuum velocity and ground velocity, and all parameters in the standardized time series data as an enhanced dataset. In this embodiment, the augmented dataset is constructed as follows: Step S21: Call up the altitude data and total temperature from the standardized time series data, invert the real-time atmospheric environment according to the atmospheric model, and determine the air density and temperature deviation relative to the standard atmosphere under the current flight conditions. Step S22: Using the timestamp and latitude / longitude location data as indexes, perform spatiotemporal interpolation processing on the reanalysis wind field data to decompose it into the downwind component acting on the aircraft's flight path and the crosswind component perpendicular to the flight path. Step S23: Calculate the vacuum speed and ground speed recorded in the standardized time series data using the downwind component and crosswind component to obtain the corrected vacuum speed and ground speed. Step S24: The air density, temperature deviation, downwind component, and crosswind component are added as new parameters, and together with the corrected vacuum velocity and ground velocity and all parameters in the standardized time series data, they constitute the enhanced dataset.

[0037] In this embodiment, the method for constructing the enhanced dataset is as follows. Step S21: Using the altitude data and total temperature from the standardized time-series data, the real-time atmospheric environment is retrieved according to the atmospheric model to determine the air density and temperature deviation relative to the standard atmosphere under the current flight conditions. Specifically, using the altitude data from the standardized time-series data... With total temperature The height data is the input. Given by GNSS altitude. Based on the international standard atmospheric model, from geometric altitude. Determine the standard static pressure corresponding to this layer. With standard static temperature Regarding actual static temperature The solution is obtained by using the relationship between total temperature and static temperature. ,in To approximate the true static temperature using the flight Mach number, an iterative calculation method is employed: using the vacuum velocity or Mach number recorded in the standardized time-series data as an initial value, the static temperature and Mach number are repeatedly updated in conjunction with the sound speed expression corresponding to the current altitude layer until the convergence condition is met, thus obtaining the actual static temperature compatible with the measured total temperature. Based on this, the temperature deviation relative to standard atmosphere is calculated. According to the ideal gas law, the standard static pressure... With the actual static temperature According to the formula Calculate the air density under the current flight conditions. ,in This is the air gas constant. This allows for the inversion of the real-time atmospheric environment, and the resulting air density and temperature deviations directly characterize the environmental benchmark of the aerodynamic forces currently acting on the aircraft.

[0038] Step S22: Using the timestamp and latitude / longitude location data as indexes, interpolation processing is performed on the reanalysis wind field data in both spatiotemporal dimensions to decompose it into the downwind component acting on the aircraft's flight path and the crosswind component perpendicular to the flight path. The reanalysis wind field data uses ERA5 global meteorological reanalysis data with a spatial resolution of 0.25°×0.25°, which provides the horizontal wind components at different isobaric surfaces or altitude levels. , During interpolation, based on the timestamp of the current data frame, the wind field grid points of two adjacent whole-hour intervals before and after that moment are extracted from the ERA5 data, and then processed for each grid point. , The components are interpolated linearly over time to obtain the wind field grid data corresponding to the current moment. Based on the longitude and latitude of the aircraft in the current frame, four adjacent grid points surrounding the aircraft's position are determined in the interpolated grid field. , The components are bilinearly interpolated along the latitudinal and longitudinal directions to obtain the local wind speed vector at that location. The current flight path angle of the aircraft is obtained from the standardized time-series data, and the wind speed vector is... Projected along the track direction, according to the formula Calculate the downwind component ,in The direction angle of the wind speed vector. The track angle is positive, indicating a tailwind, and negative, indicating a headwind. The wind speed vector... Projecting onto a direction perpendicular to the track, according to the formula The crosswind component was calculated. The symbols were agreed upon to represent the left and right crosswind directions. This allowed for the quantification of the influence of atmospheric wind fields on the relationship between the aircraft's ground speed and vacuum speed.

[0039] Step S23, utilizing the downwind component crosswind component The corrected vacuum velocity is obtained by calculating the vacuum velocity and ground velocity recorded in the standardized time series data. With ground speed During the calculation, the ground velocity with higher accuracy from the standardized time-series data is used. Based on the ground speed The data is provided by GNSS measurements and corrected for consistency based on the wind speed triangle principle. The ground speed vector is then... The local wind vector obtained in the preceding steps Perform vector subtraction according to the formula The corrected vacuum velocity was calculated. The corrected ground velocity was also calculated. Take as Alternatively, Savitzky-Golay filtering can be used. The result is obtained after smoothing and noise reduction. This correction forces a physical compatibility between vacuum speed, ground speed, and wind vector, eliminating inconsistencies caused by sensor errors or missing wind field data. The corrected vacuum speed better represents the aircraft's true speed relative to the air, providing reliable basic parameters for subsequent calculations of physical characteristics such as lift coefficient and drag coefficient; the corrected ground speed serves as the basis for trajectory estimation and phase division.

[0040] Step S24, the air density Temperature deviation Downwind weight and crosswind component As a new parameter, it is related to the corrected vacuum velocity. Corrected ground speed The enhanced dataset, together with all parameters in the standardized time-series data, constitutes the data set. Each frame of this enhanced dataset carries original flight state parameters, atmospheric physical property parameters, and wind field environmental parameters. The original flight state parameters include, but are not limited to, altitude data, total temperature, indicated airspeed, Mach number, heading angle, pitch angle, roll angle, low-pressure rotor speed of the engine, high-pressure rotor speed of the engine, exhaust temperature, fuel flow rate, flap position data, slat position data, landing gear status data, timestamp, and latitude and longitude position data. This enhanced dataset provides a unified, complete, and physically consistent data foundation for subsequent steps involving fine-tuning of flight phases and bank angles based on hidden Markov models, as well as the generation of physical constraint features.

[0041] Step S30: The altitude data, speed data, flap position data, and landing gear status data are used as an observation sequence. The flight phases are divided by time-series state decoding and the bank angle is calculated. Time-series data with phase labels are output. The time-series state decoding adopts a hidden Markov model. In this embodiment, step S30 includes: Step S31: The altitude data, speed data, flap position data, and landing gear status data are used as the observation sequence and input into the Hidden Markov Model for state decoding. The model identifies and labels each flight stage, including taxiing out, takeoff run, initial climb, climb, acceleration, cruise, descent, approach, go-around, and taxiing in. Step S32: Extract latitude and longitude location data from the enhanced dataset, estimate the turning radius using the differential curvature method, and calculate the real-time slope angle by combining the corrected vacuum speed in the enhanced dataset; Step S33: Merge flight phases with durations below a preset threshold into adjacent flight phases, append the phase labeling results and the slope angle to the augmented dataset, and output the time-series data with phase labels.

[0042] In this embodiment, the specific implementation of this step is as follows. Step S31: The altitude data, speed data, flap position data, and landing gear status data in the enhanced dataset are used as a four-dimensional observation sequence and input into the Hidden Markov Model for state decoding. Ten flight phases are identified and labeled: taxiing out, takeoff run, initial climb, climb, acceleration, cruise, descent, approach, go-around, and taxiing in.

[0043] The Hidden Markov Model (HMM) requires training before use, with ten hidden states corresponding to the ten flight phases mentioned above. The observation probability density function is represented by a Gaussian mixture model, where each hidden state is composed of a weighted combination of several Gaussian components to fit the distribution of observation parameters within that phase. Training samples are taken from QAR data of flights labeled with flight phases. The state transition probability matrix, observation probability parameters, and initial state distribution of the model are estimated iteratively using the expectation-maximization algorithm. The constraint on the state transition probability matrix is ​​that non-zero probabilities are only allowed for transitions in directions reasonable to the actual flight process, prohibiting physically impossible transition paths such as a direct transition from cruise to takeoff roll, to enhance the physical plausibility of the decoding results.

[0044] During decoding, the complete four-dimensional observation sequence of the current flight is input into a trained Hidden Markov Model (HMM), and the Viterbi algorithm is used to solve for the most probable state path, obtaining the flight phase label corresponding to each time frame. The parameters in the four-dimensional observation sequence are obtained as follows: the altitude data is taken from the altitude data in the augmented dataset, which is given by GNSS altitude; the velocity data is taken from the corrected ground speed in the augmented dataset. Or corrected vacuum speed The flap position data is obtained from flap angle sensors or discrete flap position signals; the landing gear status data is obtained from landing gear compression switches or proximity sensors. After obtaining preliminary stage annotations through Viterbi decoding, checks are performed according to the defined rules for each stage, including: during the takeoff roll phase, the landing gear status data must simultaneously indicate ground compression and the speed data must be greater than a threshold; during the initial climb phase, the landing gear status data must indicate a continuous increase in altitude after takeoff and the flap position data must indicate takeoff configuration; during the cruise phase, the altitude data fluctuation must be less than a preset range and the speed data must be stable; during the approach phase, the flap position data must indicate deployment to landing configuration and the landing gear status data must indicate deployment. Annotated frames inconsistent with the defined rules are corrected to ensure that the stage division results strictly match the flight physical state.

[0045] Step S32: Extract latitude and longitude location data from the enhanced dataset, and estimate the turning radius using the differential curvature method. The real-time slope angle is calculated by combining the corrected vacuum velocity from the enhanced dataset. .

[0046] Estimating turning radius using the differential curvature method The specific process is as follows: Longitude and latitude coordinates of three temporally consecutive adjacent frames are extracted from the augmented dataset and projected onto the local tangent plane rectangular coordinate system to obtain a planar position coordinate sequence. A second-order difference is performed on the planar position coordinate sequence to obtain the discrete curvature of the trajectory curve. curvature The calculation formula is In the formula , These are the first-order differences of the coordinates of the tangent plane. , These are the second-order differences of the coordinates of the tangent plane, respectively. (Based on curvature) Taking the reciprocal gives the turning radius. Regarding the turning radius Apply median filtering to remove abnormal spikes caused by positioning noise, so that the estimation results remain smooth and continuous.

[0047] Obtain turning radius Then, according to the formula Calculate real-time slope angle In the formula It is the acceleration due to gravity. The corrected vacuum velocity in the enhanced dataset, The turning radius is estimated using the differential curvature method. This formula is derived from the force balance relationship during uniform turning flight and characterizes the physical constraints between vacuum speed, turning radius, and bank angle.

[0048] Step S33: Merge flight phases with durations below a preset threshold into adjacent flight phases, append the phase labeling results and the slope angle to the augmented dataset, and output the time-series data with phase labels.

[0049] The preset threshold is set to 30 seconds. The stage labels obtained from Viterbi decoding in step S31 are scanned. When the continuous duration of a certain flight stage is less than 30 seconds, it is considered a short-duration abnormal stage. During merging and repair, the types of the adjacent stages before and after the abnormal stage are checked: if the types of the adjacent stages are the same, the label of the abnormal stage is directly replaced with the same stage type; if the types of the adjacent stages are different, the abnormal stage is split and merged into the adjacent stages according to the principle of physical rationality. Through merging and repair, stage jumps caused by instantaneous sensor disturbances or blurred decoding boundaries are eliminated, ensuring the overall physical rationality of the stage sequence.

[0050] After the repair is completed, the stage annotation results of each time frame and the slope angle calculated in step S32 are combined. As a new field, it is appended to the corresponding time frame of the augmented dataset to form time-series data with stage annotations. This time-series data retains all parameters of the augmented dataset, and adds flight stage labels and real-time slope angle information, providing accurate stage boundaries and flight dynamic parameters for subsequent physical constraint feature generation and stage-adaptive weighted training.

[0051] Step S40: Calculate drag and lift coefficient, drag coefficient and side force coefficient from the time series data with stage annotations; calculate thrust through force decomposition; determine engine decay coefficient; and generate physical constraint feature set. In this embodiment, step S40 includes: Step S41: Extract acceleration and track angle from the time series data with stage annotations, decompose the forces on the aircraft along the track direction, and calculate the drag. Step S42: Obtain the reference area of ​​the aircraft wing and the mass of the aircraft. Combine the drag, the air density in the enhanced data set and the corrected vacuum speed to calculate the lift coefficient, drag coefficient and side force coefficient respectively, and perform low-pass filtering smoothing. Step S43: Based on the acceleration and the rate of change of wind speed, determine the acceleration composite term, and add the product of the aircraft mass and the acceleration composite term, the gravity component along the flight path, the drag, and the wind component along the thrust direction determined based on the tailwind component and the crosswind component to calculate the thrust. Step S44: Identify the cruise phase from the time-series data with phase labels, and extract the engine rotor speed corresponding to the cruise phase; Step S45: Compare the engine rotor speed with the reference characteristic curve of the corresponding engine model to obtain the engine degradation coefficient; Step S46: Combine the lift coefficient, drag coefficient, lateral force coefficient, thrust, and engine decay coefficient to form the physical constraint feature set.

[0052] In this embodiment, the generation process of the physical constraint feature set is as follows. Step S41: Extract acceleration and track angle from the time-series data with stage annotations, decompose the forces acting on the aircraft along the track direction, and calculate the drag. Specifically, the drag is calculated from the corrected ground speed in the enhanced dataset. Taking the first derivative with respect to time yields the acceleration along the trajectory. The track angle is calculated from the attitude angle or velocity vector direction of the augmented dataset. To perform force decomposition, a baseline estimate of the thrust must first be obtained. This is based on the engine low-pressure rotor speed recorded in the time-series data with stage annotations. The height data in the augmented dataset ,Mach number and total temperature It calls the reference thrust characteristic database provided by the engine manufacturer for the corresponding aircraft model and uses multidimensional linear interpolation to obtain the reference thrust that the new engine should produce under the current flight conditions. Based on this, the equilibrium equations are established for the forces acting along the flight path according to Newton's second law. Ignoring the effects of the angle of attack and thrust installation angle, the resultant force along the flight path consists of the thrust, aerodynamic drag, and the component of gravity along the flight path, satisfying the following relationship: ,in For the mass of the aircraft, This is the acceleration due to gravity. From this, the drag force can be calculated. The thrust reference value required for this back-calculation process comes from the engine manufacturer's steady-state characteristic data, does not rely on external aerodynamic databases, and the back-calculation error can be controlled within 3%.

[0053] Step S42: Obtain the reference area of ​​the aircraft wing. and aircraft mass Combined with the aforementioned resistance Air density in the enhanced dataset and corrected vacuum speed The lift coefficient, drag coefficient, and side force coefficient were calculated separately, and then smoothed using low-pass filtering. Lift coefficient Derived from the relationship between lift and gravity balance under level flight or small track angle conditions, according to the formula Calculation; drag coefficient From the resistance already obtained According to the formula Calculate the lateral force coefficient. It is necessary to obtain the lateral force acting on the aircraft. This lateral force was obtained through lateral force analysis along a direction perpendicular to the flight path: lateral acceleration was extracted from the enhanced dataset. It can be calculated using the formula Calculation, where The rate of change of heading angle, This is the corrected ground speed; crosswind component is also considered. The effect of the time rate of change on the inertial force is used to obtain the expression for the lateral force. Then Substitute into the formula This yields the lateral force coefficient. The above calculations... , , The sequences were subjected to a low-pass filter with a cutoff frequency of 0.1 Hz to remove high-frequency jitter caused by aeroelastic vibration and measurement noise, so that the coefficient curves were smooth and conformed to quasi-steady aerodynamic characteristics.

[0054] Step S43: Determine the acceleration composition term based on the acceleration and the rate of change of wind speed. The mass of the aircraft The acceleration composition term Gravity component along the flight path The resistance And the component of the wind thrust direction determined based on the downwind component and the crosswind component. Add them together to calculate the thrust. Specifically, the acceleration composition term. It is obtained by vector synthesis of the differentials of ground speed and wind speed, and its calculation formula is as follows: ,in This is the derivative of the downwind component with respect to time. The component of the wind acting along the thrust direction. The rate of change of the wind speed vector projected onto the thrust direction is determined, and is approximately composed of the differential terms of the downwind component. Based on this, a dual-axis thrust model is used to calculate the real-time thrust, and its expression is as follows: This model comprehensively considers inertial terms, gravitational components, aerodynamic drag, and wind acceleration effects, achieving a thrust calculation accuracy better than 2%.

[0055] Step S44: Identify frames labeled as cruise from the time-series data with stage labels, and extract the engine rotor speed corresponding to that cruise stage. The extracted rotor speed can be the low-pressure rotor speed. or engine pressure ratio This embodiment uses These are the baseline parameters. To ensure data stability, altitude data during the cruise phase, when the flight status is stable, are selected. For continuous data segments with speed fluctuations less than a preset range, calculate the value within that segment. The average value is used as the measured cruising speed. .

[0056] Step S45: Compare the engine rotor speed with the reference characteristic curve of the corresponding engine model to obtain the engine degradation coefficient. The aforementioned baseline characteristic curves are provided by the engine manufacturer for the new engine under given flight conditions (altitude data). ,Mach number Total temperature () and the corresponding reference low-pressure rotor speed at a given thrust level The relational data. During comparison, the average altitude, average Mach number, and average total temperature of the cruise phase are used as query conditions, and the thrust calculated in step S43 is used as the comparison data. As the thrust input, the reference speed that the new engine should possess under the same operating conditions is obtained by interpolation on the reference characteristic curve. Then follow the formula Calculate the engine degradation coefficient and... Value constraints within the range The internal parameters characterize the degree of performance degradation that occurs in an engine due to cumulative use over time. This yields... It reflects the current level of engine thrust output relative to the new state and is an important correction factor for subsequent physical baseline predictions.

[0057] Step S46, the lift coefficient drag coefficient Lateral force coefficient The thrust and the engine degradation coefficient The data are merged to form the physical constraint feature set. This physical constraint feature set, together with the time-series data with stage annotations, constitutes the input for subsequent feature selection and end-to-end prediction. Each time frame contains the above five parameters derived from the physical mechanism, realizing the embedding of physical constraints for aerodynamic-thrust-engine health status.

[0058] Step S50: Construct a candidate feature pool from the augmented dataset and the physical constraint feature set, and obtain the core feature set through nonlinear filtering, redundancy removal and causality test; In this embodiment, step S50 includes: Step S51: Extract the original flight parameters and atmospheric environment parameters from the augmented dataset, extract the physical constraint parameters from the physical constraint feature set, and calculate the time-series statistics of the original flight parameters, atmospheric environment parameters, and physical constraint parameters within the sliding window. Then, merge the original flight parameters, atmospheric environment parameters, physical constraint parameters, and time-series statistics to form a candidate feature pool. Step S52: Calculate the distance correlation coefficient between each feature in the candidate feature pool and fuel consumption, and retain features whose distance correlation coefficient is higher than a preset correlation threshold; Step S53: Calculate the conditional mutual information for the retained features and remove features whose conditional mutual information is lower than a preset redundancy threshold. Step S54: After removing redundant features, use an additive noise model to perform causal direction verification, retain the features that have a unidirectional causal relationship with fuel consumption, and form the core feature set.

[0059] In this embodiment, the process of obtaining the core feature set is as follows. In step S51, the original flight parameters are taken from the measurement parameters directly related to the flight state in the enhanced dataset, including the corrected ground speed. Corrected vacuum speed Altitude data ,Mach number Heading angle Pitch angle Roll angle flap position data slat position Landing gear status data Engine low-pressure rotor speed Engine high-pressure rotor speed Exhaust temperature Indicating airspeed Vertical velocity Total temperature Fuel flow Throttle lever angle Angle of attack Sideslip angle It has a total of 20 dimensions. Among them, height data... The altitude is given by GNSS, and the velocity data is incorporated in various forms such as corrected ground speed, corrected vacuum speed, and indicated airspeed. Flange position data is also included. Landing gear status data The data are taken from the flap angle sensor and the landing gear compression switch, respectively. The atmospheric environmental parameters are taken from environmental parameters obtained through atmospheric compensation in the enhanced dataset, including air density. Temperature deviation Downwind weight Crosswind component The total dimensions are 4. The physical constraint parameters are taken from the physical constraint feature set, including lift coefficient, drag coefficient, side force coefficient, thrust, and engine decay coefficient, totaling 5 dimensions.

[0060] For each of the 29 parameters mentioned above, eight time-series statistics are calculated within a sliding window: mean, standard deviation, maximum value, minimum value, peak-to-peak value, skewness, kurtosis, and slope. The length of the sliding window is set to 10 seconds, corresponding to the 1 Hz sampling frequency of the augmented dataset, meaning that every 10 consecutive data frames constitute one window, and the window slides along the time axis in 1-second steps. Each parameter generates eight corresponding statistics within each window, thus expanding the 29 original features into 232 time-series statistical features. The 20 original flight parameters, the 4 atmospheric environment parameters, the 5 physical constraint parameters, and the 232 time-series statistical features are merged to form a candidate feature pool with 261 dimensions. This candidate feature pool comprehensively covers potential information from multiple levels, including flight state, atmospheric environment, physical mechanisms, and time-series dynamic characteristics.

[0061] In step S52, the distance correlation coefficient The calculation process is as follows: for features and fuel consumption Calculate the distance matrix for each sample separately, perform bicentering on the distance matrix, and then apply the formula... Calculate, where Let the distance covariance be... and They are respectively and The distance variance. The distance correlation coefficient can capture the nonlinear correlation between variables, and its value is equal to 0 if and only if the two variables are independent. The preset correlation threshold is set to 0.65. The distance correlation coefficient between each of the 261-dimensional features in the candidate feature pool and fuel consumption is calculated, and the coefficients are retained. By identifying the characteristics and removing those that are not sufficiently correlated with fuel consumption, a preliminary set of features is obtained.

[0062] In step S53, conditional mutual information The calculation formula is ,in For the features to be evaluated, Fuel consumption For the feature subset that has been selected into the core feature set, This represents the probability distribution. The K-nearest neighbor entropy estimation method is used to estimate the probability density of the continuous variable during calculation, and the number of nearest neighbors... Set to 3. Conditional mutual information measures the information within a known subset of selected features. Under the condition, characteristics fuel consumption Additional information contribution. The preset redundancy threshold is set to 0.1, when When, it indicates the characteristics In the existing feature subset Based on The incremental information is limited and is considered redundant, so it is removed. The removal process employs a greedy forward selection strategy, successively selecting from the candidate features. The feature that is the largest and not lower than the threshold is added to the core subset until no feature meets the condition. The final retained features constitute the feature set after redundancy removal.

[0063] In step S54, the additive noise model The verification process is as follows: for features With fuel consumption Assume two causal directions respectively, namely and Regarding the causal direction of the hypothesis. Establish a regression model ,in Gaussian process regression was used for fitting. This is the residual term. If the direction is true, then the residual... With causal variables Statistically independent; conversely, if the residuals and If a dependency exists, the direction is rejected. The independence of the residuals is tested using a statistical test based on the Hilbert-Schmidt independence criterion, with a significance level set at 0.05. If only the direction... Direction through independence test If it fails, then confirm. yes If a feature is identified as a unidirectional cause of fuel consumption, it is retained. If both directions pass the test, indicating a bidirectional fuzzy causal relationship between the feature and fuel consumption, the feature is not directly included in the core feature set. It is only retained if physical prior verification confirms it as a unidirectional cause of fuel consumption; otherwise, it is discarded. After the above causal direction test, the final retained features constitute the core feature set, which has a dimension of 16 and a feature redundancy controlled within 10%. This core feature set accurately characterizes the intrinsic causal link between aerodynamics, thrust, and fuel consumption, providing low-redundancy and highly interpretable input for the end-to-end prediction model.

[0064] Step S60: Obtain the unit thrust fuel consumption rate, input the core feature set into the encoder composed of a local feature extraction network and a temporal memory network, determine the physical baseline fuel consumption based on the thrust, the unit thrust fuel consumption rate, the ratio of the lift coefficient to the drag coefficient, and the engine decay coefficient, and learn the residual correction term by the encoder, superimpose the physical baseline fuel consumption and the residual correction term, and output the fuel consumption prediction result.

[0065] In this embodiment, the fuel consumption per unit thrust is obtained, and the core feature set is input into an encoder composed of a local feature extraction network and a temporal memory network. The physical baseline fuel consumption is determined based on the thrust, the fuel consumption per unit thrust, the ratio of the lift coefficient to the drag coefficient, and the engine degradation coefficient. The method is as follows: Step S61: The local feature extraction network is composed of multiple cascaded convolutional layers with the number of convolutional kernels increasing layer by layer; the temporal memory network is a bidirectional long short-term memory layer. Step S62: Input the core feature set into the encoder, and extract temporal features by the local feature extraction network and the temporal memory network; Step S63: Multiply the thrust by the fuel consumption rate per unit thrust to obtain the baseline fuel consumption rate; multiply the ratio of the lift coefficient to the drag coefficient and the engine degradation coefficient by the correction factor; determine the fuel consumption correction coefficient according to the correction factor and a preset ratio; divide the baseline fuel consumption rate by the fuel consumption correction coefficient to obtain the physical baseline fuel consumption.

[0066] The encoder learns a residual correction term, which is then superimposed on the physical baseline fuel consumption to output a predicted fuel consumption result. The method is as follows: Step S64: The encoder outputs the residual correction term based on the timing characteristics; Step S65: Add the physical baseline fuel consumption to the residual correction term to obtain the predicted fuel consumption value; The encoder is obtained through training, and a joint loss function is used during training. The joint loss function includes a first loss term that represents the deviation between the predicted fuel consumption and the actual fuel consumption, a second loss term that constrains the residual correction term to approach zero, and a model parameter regularization term. The loss is weighted according to the flight phase, with the loss weight of the climb phase and the approach phase being greater than the loss weight of the taxiing, taxiing, and cruise phases. Step S66: When new fast access recorder data arrives, fix the parameters of the local feature extraction network and the temporal memory network, and only fine-tune the parameters of the network layer in the encoder used to output the residual correction term for online model update; Step S67: Output the fuel consumption prediction result, which includes the fuel consumption of each flight phase and its proportion in the total fuel consumption, as well as the prediction uncertainty interval generated based on deep integration.

[0067] Furthermore, a federated learning strategy is employed when updating the model online, and the federated learning strategy includes: Each airline node trains the encoder locally using its own fast access recorder data, and extracts the model gradient after training is complete. The extracted model gradients are uploaded to the central server, but the original fast access recorder data is not uploaded. The central server aggregates the received model gradients by weighted average based on the amount of data to obtain the global gradient, updates the encoder parameters using the global gradient, and distributes the updated parameters to each airline node.

[0068] In this embodiment, the specific implementation of this step is as follows. Step S61: Construct an encoder composed of a local feature extraction network and a temporal memory network. The local feature extraction network consists of multiple cascaded convolutional layers with the number of convolutional kernels increasing layer by layer. Specifically, it includes three one-dimensional convolutional layers with the number of convolutional kernels set to 64, 128, and 256 respectively. The kernel size is 3×1, the stride is 1, and the activation function is a modified linear unit. The first convolutional layer extracts local patterns from the input core feature set, and the second and third convolutional layers abstract higher-level feature representations layer by layer based on this. A batch normalization layer is set after each convolutional layer to accelerate training convergence and suppress overfitting. The temporal memory network is a bidirectional long short-term memory layer with 128 hidden neurons. The bidirectional long short-term memory layer receives the feature sequence output by the local feature extraction network, performs recursive operations along the forward and backward time axes respectively, and concatenates the hidden states in the two directions to obtain a temporal feature representation that simultaneously contains forward and backward dependencies.

[0069] Step S62: The core feature set is input into the encoder, where the local feature extraction network and the temporal memory network extract temporal features. The core feature set is a 16-dimensional feature sequence retained after the causal-correlation hybrid feature selection in step S50, arranged into a two-dimensional matrix according to time frames, with each row corresponding to a time frame and each column corresponding to a feature dimension. This matrix is ​​first input into the local feature extraction network, processed layer by layer by three one-dimensional convolutional layers to output a high-dimensional feature map; then, the high-dimensional feature map is input into the bidirectional long short-term memory layer, which performs sequence modeling along the time dimension, and finally outputs the temporal feature vector corresponding to each time frame. The time-series feature vector It integrates short-term change patterns within the local receptive field with long-term temporal dependencies.

[0070] Step S63, based on the thrust Fuel consumption per unit thrust The lift coefficient With the drag coefficient The ratio and the engine degradation coefficient Determine physical baseline fuel consumption Specifically, according to the formula... Calculate the physical baseline fuel consumption. Where, The real-time thrust calculated in step S43; The fuel consumption rate per unit thrust is determined by interpolation based on the new engine baseline fuel consumption characteristics provided by the engine manufacturer and combined with the current flight conditions. The lift-to-drag ratio is the ratio of the lift coefficient to the drag coefficient, which characterizes the aerodynamic efficiency of the aircraft. The engine degradation coefficient obtained in step S45 has a value range of [value range missing]. The physical meaning of this formula is: This constitutes the baseline fuel consumption rate. A correction factor is established to reflect the combined impact of aerodynamic efficiency and engine health on fuel consumption. A fuel consumption correction coefficient is formed based on the correction factor and a preset ratio. The baseline fuel consumption rate is then corrected by this coefficient to obtain the physical baseline fuel consumption. This physical baseline fuel consumption provides a theoretical estimate of aviation fuel consumption based on the physical mechanisms of aerodynamics, thrust, and engine health.

[0071] Step S64, the encoder outputs the residual correction term based on the timing characteristics. Specifically, a fully connected layer is set after the bidirectional long short-term memory layer, and the input of the fully connected layer is the temporal feature vector output in step S62. The output is the fuel consumption relative to the physical baseline. Same-dimensional residual correction term The activation function of the fully connected layer is a linear function, enabling it to output positive and negative values ​​for bidirectional correction of the physical baseline fuel consumption. The residual correction term... Its role is to capture high-order nonlinear fuel consumption factors that cannot be covered by physical baseline formulas, such as transient wind field changes, pilot operation differences, configuration transition effects, and other data-driven patterns.

[0072] Step S65, the physical baseline fuel consumption With the residual correction term Add them together to get the predicted fuel consumption value. This structure achieves a deep integration of physical mechanisms and data-driven approaches: the physical baseline ensures the physical consistency and extrapolation stability of the prediction results under extreme conditions, while the residual correction term learns the complex patterns of residual bias from historical data, and the two complement each other.

[0073] The training process of the encoder is as follows. A joint loss function is used during training. It consists of three items: In the formula, the first loss term... The predicted value of aviation fuel consumption Compared to real-world fuel consumption Mean absolute error between, weights Set to 0.7, dominating overall prediction accuracy; second loss term Constrain the residual correction term Approaching 0, weight Setting it to 0.2 reflects a preference for the physical baseline, encouraging the model to rely on the physical baseline as much as possible, and only correcting for systematic biases in the physical baseline through the residual correction term; the third loss term For model parameters Regularization term, weight The value is set to 0.1 to prevent model overfitting. During training, the loss is also weighted according to the flight phases, with weights for the climb and approach phases. Set to 2.0, loss weights for the slide-out and slide-in phases. Set to 0.5, the loss weight during the cruise phase. The value is set to 1.0 to enable the model to achieve higher prediction accuracy during periods of high fuel consumption and drastic changes.

[0074] Step S66: When new fast access recorder data arrives, the parameters of the local feature extraction network and the temporal memory network are fixed, and only the parameters of the fully connected layer in the encoder used to output the residual correction term are fine-tuned for online model updates. Specifically, each incremental update requires at least 100 new data entries, and the update cycle does not exceed 24 hours. During fine-tuning, only the weight parameters of the fully connected layer are updated via backpropagation; the parameters of the three one-dimensional convolutional layers and the bidirectional long short-term memory layer remain frozen. This strategy keeps the model's low-level feature extraction capability stable, while only the top-level correction logic adaptively adjusts to factors such as engine performance degradation and seasonal weather changes, achieving rapid adaptation while maintaining model stability.

[0075] Step S67: Output the fuel consumption prediction result. The fuel consumption prediction result includes: a second-level fuel flow curve, that is, the fuel flow curve corresponding to each time frame. Values; fuel consumption and its percentage in each flight phase are obtained by integrating the second-level fuel flow rate in each flight phase over time, and the percentage of fuel consumption in each phase to the total consumption is calculated; based on the prediction uncertainty interval generated by deep integration, an integrated model is formed by training multiple independently initialized encoders, each model outputs a prediction value, and the prediction mean and ±2σ confidence interval are calculated to provide a reliable uncertainty reference for dispatch release.

[0076] Furthermore, the online update model also includes a federated learning strategy. The specific implementation of the federated learning strategy is as follows: each airline node trains the encoder locally using its own Fast Access Recorder (FAR) data, following the same training process as described above. After training, the model gradient is extracted. The extracted model gradient is uploaded to the central server, but the original FAR data is not uploaded, thus protecting the data privacy of each airline. The central server aggregates the received model gradients using a weighted average based on data volume, with an aggregation period of 7 days, to obtain the global gradient. The encoder parameters are then updated using the global gradient, and the updated parameters are distributed to each airline node. This federated learning strategy enables the model to collaboratively optimize using multi-airline data, improving generalization ability while meeting the requirements of data anonymization and privacy protection.

[0077] like Figure 2 As shown, a second embodiment of the present invention provides an end-to-end fuel consumption prediction system for executing a first embodiment of an end-to-end fuel consumption prediction method. The system includes: The data preprocessing module is used to acquire fast access recorder data and convert it into standardized time-series data. The standardized time-series data includes at least vacuum speed, ground speed, total temperature, timestamp, latitude and longitude location data, altitude data, speed data, flap position data, and landing gear status data. The atmospheric environment compensation module is used to correct the vacuum velocity and ground velocity based on the atmospheric model and wind field data. The calculated atmospheric environment parameters, the corrected vacuum velocity and ground velocity, and all parameters in the standardized time series data are used together as an enhanced dataset. The flight phase segmentation module is used to take altitude data, speed data, flap position data and landing gear status data as observation sequences, divide the flight phases through time-series state decoding and calculate the bank angle, and output time-series data with phase labels. The physical feature generation module is used to back-calculate drag and lift coefficient, drag coefficient and side force coefficient from the time series data with stage annotation, calculate thrust through force decomposition, determine engine decay coefficient, and generate physical constraint feature set; The feature selection module is used to construct a candidate feature pool from the augmented dataset and the physical constraint feature set, and obtain the core feature set through nonlinear filtering, redundancy removal and causality test; The end-to-end prediction module is used to obtain the fuel consumption rate per unit thrust. The core feature set is input into an encoder composed of a local feature extraction network and a temporal memory network. The physical baseline fuel consumption is determined based on the thrust, the fuel consumption rate per unit thrust, the ratio of the lift coefficient to the drag coefficient, and the engine decay coefficient. The encoder learns the residual correction term, and the physical baseline fuel consumption and the residual correction term are superimposed to output the fuel consumption prediction result.

[0078] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the system described above can be found in the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0079] It should be noted that the end-to-end aviation fuel consumption prediction system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0080] A device according to a third embodiment of the present invention includes: At least one processor; and a memory communicatively connected to at least one of the processors; The memory stores instructions that can be executed by the processor to implement the aforementioned end-to-end prediction method for aviation fuel consumption.

[0081] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, which are executed by the computer to implement the above-described end-to-end prediction method for aviation fuel consumption.

[0082] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the storage device and processing device described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0083] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system for implementing embodiments of the systems, methods, and electronic devices of this application. Figure 3 The server shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0084] like Figure 3As shown, the computer system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in Read Only Memory (ROM) 302 or programs loaded from storage section 308 into Random Access Memory (RAM) 303. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0085] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a cathode ray tube (CRT), liquid crystal display (LCD), and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0086] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0087] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0088] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0089] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0090] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0091] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for end-to-end prediction of aviation fuel consumption, characterized in that, include: Acquire fast access recorder data and convert it into standardized time-series data, wherein the standardized time-series data includes at least vacuum speed, ground speed, total temperature, timestamp, latitude and longitude location data, altitude data, speed data, flap position data and landing gear status data; The vacuum velocity and ground velocity are corrected based on atmospheric models and wind field data. The calculated atmospheric environmental parameters, the corrected vacuum velocity and ground velocity, and all parameters in the standardized time series data are used together as an enhanced dataset. Altitude data, speed data, flap position data, and landing gear status data are used as observation sequences. Flight phases are divided by time-series state decoding and bank angle is calculated. Time-series data with phase labels are output. The drag is calculated from the time-series data with stage annotations, and the lift coefficient, drag coefficient, and side force coefficient are calculated. The thrust is calculated through force decomposition, and the engine decay coefficient is determined to generate a set of physical constraint features. A candidate feature pool is constructed from the augmented dataset and the physical constraint feature set, and the core feature set is obtained through nonlinear filtering, redundancy removal and causality test. The unit thrust fuel consumption rate is obtained, and the core feature set is input into an encoder composed of a local feature extraction network and a temporal memory network. The physical baseline fuel consumption is determined based on the thrust, the unit thrust fuel consumption rate, the ratio of the lift coefficient to the drag coefficient, and the engine decay coefficient. The encoder learns the residual correction term, and the physical baseline fuel consumption and the residual correction term are superimposed to output the fuel consumption prediction result.

2. The end-to-end prediction method for aviation fuel consumption according to claim 1, characterized in that, The method for acquiring fast access logger data and converting it into standardized time-series data is as follows: Identify the message format of the fast access logger data and extract parameter fields according to the field mapping rules of the corresponding format; The extracted data is processed by a combination of median deviation filtering based on sliding window, standard deviation criterion and first difference gradient test to detect and remove outliers and frozen values. By utilizing the correlation between parameter channels, the missing channels are reconstructed using the Kalman smooth interpolation algorithm, and the interpolation error is controlled within a preset range. The engine parameters at the first sampling frequency and the atmospheric data at the second sampling frequency are resampled to the same reference frequency. The resampled sequence is then denoised using a polynomial fitting smoothing filter to obtain the standardized time series data.

3. The end-to-end prediction method for aviation fuel consumption according to claim 1, characterized in that, The augmented dataset is constructed as follows: By calling up the altitude data and total temperature from the standardized time-series data, the real-time atmospheric environment is retrieved according to the atmospheric model to determine the air density and temperature deviation relative to the standard atmosphere under the current flight conditions. Using the timestamp and latitude / longitude location data as indexes, the reanalysis wind field data is interpolated in the spatiotemporal dimensions to decompose the tailwind component acting on the aircraft's flight path and the crosswind component perpendicular to the flight path. The vacuum velocity and ground velocity recorded in the standardized time series data are calculated using the downwind component and crosswind component to obtain the corrected vacuum velocity and ground velocity. The air density, temperature deviation, downwind component, and crosswind component are added as new parameters, and together with the corrected vacuum velocity and ground velocity, as well as all parameters in the standardized time series data, constitute the enhanced dataset.

4. The method of claim 1, wherein, Altitude data, velocity data, flap position data, and landing gear status data are used as an observation sequence. Flight phases are divided through time-series state decoding, and bank angles are calculated. Time-series data with phase labels are output. The method is as follows: The temporal state decoding adopts a hidden Markov model. The altitude data, speed data, flap position data, and landing gear state data are used as the observation sequence and input into the hidden Markov model for state decoding. The model identifies and labels each flight stage, including taxiing out, takeoff run, initial climb, climb, acceleration, cruise, descent, approach, go-around, and taxiing in. Latitude and longitude location data are extracted from the enhanced dataset, the turning radius is estimated using the differential curvature method, and the real-time slope angle is calculated by combining the corrected vacuum speed in the enhanced dataset. Flight phases with durations below a preset threshold are merged into adjacent flight phases. The phase labeling results and the slope angle are appended to the augmented dataset, and the time-series data with phase labels is output.

5. The method of claim 3, wherein, The drag, lift coefficient, drag coefficient, and side force coefficient are calculated from the time-series data with stage annotations. Thrust is calculated through force decomposition, and the engine decay coefficient is determined. A set of physical constraint features is generated. The method is as follows: Acceleration and track angle are extracted from the time-series data with stage annotations. The forces acting on the aircraft are decomposed along the track direction, and the drag is calculated in reverse. The reference area of ​​the aircraft wing and the mass of the aircraft are obtained. Combined with the drag, the air density in the enhanced dataset and the corrected vacuum velocity, the lift coefficient, drag coefficient and side force coefficient are calculated respectively, and low-pass filtering is performed for smoothing. The acceleration composition term is determined based on the acceleration and the rate of change of wind speed. The thrust is calculated by adding the product of the aircraft mass and the acceleration composition term, the gravity component along the flight path, the drag, and the wind component along the thrust direction determined based on the tailwind component and the crosswind component. Identify the cruise phase from the time-series data with phase annotations, and extract the engine rotor speed corresponding to the cruise phase; The engine rotor speed is compared with the reference characteristic curve of the corresponding engine model to obtain the engine degradation coefficient; The lift coefficient, drag coefficient, lateral force coefficient, thrust, and engine decay coefficient are combined to form the physical constraint feature set.

6. The method of claim 1, wherein, A candidate feature pool is constructed from the augmented dataset and the physical constraint feature set. Through nonlinear filtering, redundancy removal, and causality testing, a core feature set is obtained. The method is as follows: The original flight parameters and atmospheric environment parameters are extracted from the augmented dataset, and the physical constraint parameters are extracted from the physical constraint feature set. The time-series statistics of the original flight parameters, atmospheric environment parameters and physical constraint parameters within the sliding window are calculated respectively. The original flight parameters, atmospheric environment parameters, physical constraint parameters and time-series statistics are merged to form a candidate feature pool. Calculate the distance correlation coefficient between each feature in the candidate feature pool and fuel consumption, and retain features whose distance correlation coefficient is higher than a preset correlation threshold; Calculate the conditional mutual information for the retained features, and remove features whose conditional mutual information is lower than a preset redundancy threshold; After removing redundant features, an additive noise model is used to test the causal direction. Features with a unidirectional causal relationship with fuel consumption are retained, forming the core feature set.

7. The end-to-end prediction method for aviation fuel consumption according to claim 1, characterized in that, To obtain the fuel consumption rate per unit thrust, the core feature set is input into an encoder composed of a local feature extraction network and a temporal memory network. Based on the thrust, the fuel consumption rate per unit thrust, the ratio of the lift coefficient to the drag coefficient, and the engine degradation coefficient, the physical baseline fuel consumption is determined. The method is as follows: The local feature extraction network is composed of multiple convolutional layers cascaded together, with the number of convolutional kernels increasing layer by layer; the temporal memory network is a bidirectional long short-term memory layer. The core feature set is input into the encoder, and the local feature extraction network and the temporal memory network extract temporal features. The reference fuel consumption rate is obtained by multiplying the thrust by the fuel consumption rate per unit thrust. The correction factor is obtained by multiplying the ratio of the lift coefficient to the drag coefficient and the engine degradation coefficient. The fuel consumption correction coefficient is determined according to the correction factor and a preset ratio. The reference fuel consumption rate is divided by the fuel consumption correction coefficient to obtain the physical baseline fuel consumption.

8. The method of claim 7, wherein, The encoder learns a residual correction term, which is then superimposed on the physical baseline fuel consumption to output a predicted fuel consumption result. The method is as follows: The encoder outputs the residual correction term based on the temporal characteristics; The predicted fuel consumption is obtained by adding the physical baseline fuel consumption to the residual correction term. The encoder is obtained through training, and a joint loss function is used during training. The joint loss function includes a first loss term that represents the deviation between the predicted fuel consumption and the actual fuel consumption, a second loss term that constrains the residual correction term to approach zero, and a model parameter regularization term. The loss is weighted according to the flight phase, with the loss weight of the climb phase and the approach phase being greater than the loss weight of the taxiing, taxiing, and cruise phases. When new fast access recorder data arrives, the parameters of the local feature extraction network and the temporal memory network are fixed, and the parameters of the network layer in the encoder used to output the residual correction term are fine-tuned only to update the model online. The predicted fuel consumption results are output, including the fuel consumption of each flight phase and its proportion in the total fuel consumption, as well as the prediction uncertainty range generated based on deep integration.

9. The method of claim 8, wherein, A federated learning strategy is used when updating the model online. The federated learning strategy includes: Each airline node trains the encoder locally using its own fast access recorder data, and extracts the model gradient after training is complete. The extracted model gradients are uploaded to the central server, but the original fast access recorder data is not uploaded. The central server aggregates the received model gradients by weighted average based on the amount of data to obtain the global gradient, updates the encoder parameters using the global gradient, and distributes the updated parameters to each airline node.

10. An end-to-end aviation fuel consumption prediction system, used to execute the end-to-end aviation fuel consumption prediction method according to any one of claims 1-9, characterized in that, The system includes: The data preprocessing module is used to acquire fast access recorder data and convert it into standardized time-series data. The standardized time-series data includes at least vacuum speed, ground speed, total temperature, timestamp, latitude and longitude location data, altitude data, speed data, flap position data, and landing gear status data. The atmospheric environment compensation module is used to correct the vacuum velocity and ground velocity based on the atmospheric model and wind field data. The calculated atmospheric environment parameters, the corrected vacuum velocity and ground velocity, and all parameters in the standardized time series data are used together as an enhanced dataset. The flight phase segmentation module is used to take altitude data, speed data, flap position data and landing gear status data as observation sequences, divide the flight phases through time-series state decoding and calculate the bank angle, and output time-series data with phase labels. The physical feature generation module is used to back-calculate drag and lift coefficient, drag coefficient and side force coefficient from the time series data with stage annotation, calculate thrust through force decomposition, determine engine decay coefficient, and generate physical constraint feature set; The feature selection module is used to construct a candidate feature pool from the augmented dataset and the physical constraint feature set, and obtain the core feature set through nonlinear filtering, redundancy removal and causality test; The end-to-end prediction module is used to obtain the fuel consumption rate per unit thrust. The core feature set is input into an encoder composed of a local feature extraction network and a temporal memory network. The physical baseline fuel consumption is determined based on the thrust, the fuel consumption rate per unit thrust, the ratio of the lift coefficient to the drag coefficient, and the engine decay coefficient. The encoder learns the residual correction term, and the physical baseline fuel consumption and the residual correction term are superimposed to output the fuel consumption prediction result.

Images