Adaptive prediction method, system and equipment for fuel consumption of aircraft based on fusion of physical model and multi-task learning

Abstract

The application belongs to the field of aviation data analysis, and relates to a fuel consumption adaptive prediction method, system and equipment fusing a physical model and multi-task learning, so as to solve the problems of sensitivity to abnormal data, insufficient robustness and easy degradation of long-term operation performance. The application comprises the following steps: acquiring and preprocessing aircraft rapid access recorder data; analyzing flight stages and calculating atmospheric parameters to construct full-dimensional data; constructing input features reflecting real flight working conditions; constructing a multi-task prediction model comprising a main task of fuel consumption prediction and an auxiliary task of abnormal state identification, both sharing a feature extraction layer, and reducing the influence of abnormal data on the main task by using the auxiliary task identification result through a reverse constraint mechanism; continuously monitoring the prediction result, and automatically triggering model rollback or feature reconstruction when performance drift is detected. The application can significantly improve the stability, reliability and physical consistency of fuel consumption prediction, and ensure long-term prediction accuracy under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of aviation data analysis technology, specifically to an adaptive prediction method, system, and device for fuel consumption that integrates physical models and multi-task learning. Background Technology

[0002] Accurate fuel consumption forecasting is crucial for improving airline operational efficiency, optimizing fuel management strategies, and ensuring flight safety. Current technologies typically utilize flight data collected by the Quick Access Recorder (QAR) installed on the aircraft to build predictive models using a data-driven approach, thereby estimating fuel consumption for different flight phases or the entire flight segment.

[0003] Currently, mainstream methods for predicting aviation fuel consumption mainly rely on time integration and statistical analysis of fuel flow parameters in QAR data, or on combining aircraft aerodynamic models to construct functional relationships between fuel consumption and flight attitude and environmental parameters. These methods provide a certain technical foundation for aircraft fuel management.

[0004] However, in practical engineering applications, the aforementioned existing technical solutions still have the following shortcomings that urgently need to be addressed:

[0005] First, it is highly sensitive to QAR outlier data, resulting in insufficient stability of prediction results. In actual flight, QAR data often introduces outliers due to factors such as sensor jitter, communication delays, or sudden environmental changes. Existing technologies typically employ independent preprocessing steps to remove outliers, but when outlier data is not fully identified, subsequent prediction models are directly affected by erroneous data, leading to significant deviations in prediction results and poor stability.

[0006] Secondly, outlier handling and prediction models operate independently, lacking a collaborative mechanism. Existing technologies generally employ a sequential structure of "handling outliers first, then making predictions," which results in a lack of effective information exchange between the outlier identification process and the prediction model. The prediction model lacks the ability to proactively suppress the impact of potential outlier data, and its robustness is insufficient when facing complex flight conditions or data quality fluctuations.

[0007] Secondly, over-reliance on fixed physical models or static parameters leads to insufficient model adaptability. Some physical model-based methods depend on pre-set aerodynamic or thrust model parameters, which are difficult to reflect in real time changes in flight status caused by factors such as load variations, weather conditions, and engine performance degradation. Therefore, when actual operating conditions deviate from model assumptions, prediction results are prone to systematic biases.

[0008] Finally, there is a lack of self-diagnosis and automatic correction capabilities for model performance. After a single training iteration, the performance of existing predictive models gradually declines over time due to changes in data distribution or error accumulation—a phenomenon known as "model drift." Current technologies generally rely on manual retraining or parameter tuning based on human experience, lacking real-time monitoring and automatic correction mechanisms for model performance degradation. This impacts the long-term reliability and engineering application value of the predictive system.

[0009] Therefore, how to effectively suppress abnormal data interference, enhance model robustness, and achieve long-term self-sustaining of model performance are technical problems that urgently need to be solved in the field of jet fuel consumption prediction. Summary of the Invention

[0010] To address the aforementioned problems in existing technologies, namely their sensitivity to abnormal data, insufficient robustness, and tendency to degrade performance over long-term operation, this invention provides an adaptive prediction method, system, and device for aviation fuel consumption that integrates physical models and multi-task learning.

[0011] In a first aspect, this invention proposes an adaptive prediction method for aviation fuel consumption that integrates physical models and multi-task learning, comprising the following steps:

[0012] Acquire and preprocess the aircraft's fast access recorder data to obtain preprocessed flight data;

[0013] Based on the preprocessed flight data, the flight phase is analyzed and the corresponding atmospheric environmental parameters are calculated. The analysis and calculation results are then correlated with the flight data to obtain full-dimensional data containing flight status and environmental parameters.

[0014] Based on the aircraft aerodynamic model and engine thrust model, feature construction is performed on the full-dimensional data to generate aerodynamic force characteristics and thrust characteristics that reflect flight conditions, which are used as input feature parameters.

[0015] A multi-task prediction model is constructed, including a shared feature extraction layer, a main task branch for fuel consumption prediction, and an auxiliary task branch for abnormal state identification. The shared feature extraction layer performs deep extraction on the input feature parameters to generate a shared feature representation. The main task branch outputs the predicted fuel consumption value based on the shared feature representation. The auxiliary task branch outputs the abnormal state identification result based on the shared feature representation.

[0016] During the training of the multi-task prediction model, the abnormal state identification auxiliary task branch applies a reverse constraint to the shared feature extraction layer.

[0017] The trained model is used to predict the fuel consumption of the target flight mission, and the prediction results are output.

[0018] The prediction results are continuously monitored. When a drift in prediction performance is detected and exceeds preset conditions, the model rollback or feature reconstruction operation is automatically triggered to restore prediction performance.

[0019] Furthermore, the aircraft's fast access recorder data is acquired and preprocessed to obtain preprocessed flight data, including:

[0020] The fast access recorder data is cleaned by combining physical rule thresholding and statistical model anomaly detection methods to remove outliers.

[0021] The cleaned data is time-aligned based on the timestamp.

[0022] For randomly missing data and consecutively missing data, interpolation and imputation by the mean of similar data are used to handle missing values, respectively.

[0023] The data, after time alignment and missing value processing, is standardized to eliminate the dimensional differences between different parameters, resulting in standardized and time-consistent preprocessed flight data.

[0024] Furthermore, based on the preprocessed flight data, the flight phase is analyzed, and the corresponding atmospheric environmental parameters are calculated, including:

[0025] Using preprocessed flight data such as flight altitude, airspeed, landing gear status, engine speed, and throttle position as input, and employing threshold judgment combined with time sequence continuity verification logic, the flight process is divided into multiple flight phases including takeoff, climb, cruise, descent, approach, and landing.

[0026] Based on the international standard atmospheric model, using flight altitude as input, and according to the different atmospheric layers at the flight altitude, the corresponding temperature, pressure and density function relationships are used to calculate the atmospheric environmental parameters at the corresponding altitude.

[0027] The analyzed flight phase information and calculated atmospheric environmental parameters are correlated and fused with the preprocessed flight data according to the timestamp to obtain full-dimensional data containing flight status and environmental parameters.

[0028] Furthermore, aerodynamic force characteristics and thrust characteristics reflecting flight conditions are generated, including:

[0029] Based on the aircraft aerodynamic model, combined with the atmospheric density, vacuum velocity and reference area in the full-dimensional data, the lift characteristics are calculated by the lift coefficient function related to the angle of attack and Mach number; the drag characteristics are calculated by the drag coefficient function combining the parasitic drag coefficient and the induced drag factor.

[0030] Based on the engine thrust model, combined with the engine rotor speed, atmospheric pressure and atmospheric temperature in the full-dimensional data, the net thrust characteristics of the engine are calculated through the functional relationship between the engine characteristic coefficient and the environmental attenuation coefficient.

[0031] Based on the lift characteristics, drag characteristics and engine net thrust characteristics, derivative operating condition characteristics including lift-to-drag ratio, thrust-to-drag difference, aerodynamic force change rate and operating condition adaptability deviation are constructed.

[0032] The constructed features are verified by physical laws and filtered for data discrimination to obtain the final core feature set, which serves as the input feature parameters for the model.

[0033] Furthermore, a multi-task prediction model is constructed, including:

[0034] A shared feature extraction layer is constructed, which combines a fully connected layer and an attention mechanism, to encode the input feature parameters of the multi-task prediction model and output a shared feature representation.

[0035] A main task branch for predicting aviation fuel consumption is constructed. The main task branch is a multi-layer fully connected regression network that takes the shared feature representation as input and outputs continuous predicted values ​​of aviation fuel consumption.

[0036] An auxiliary task branch for abnormal state recognition is constructed. The auxiliary task branch is a classification network containing a fully connected layer and a classification function. The shared feature representation is used as input, and the probability distribution of different abnormal state categories is output.

[0037] A reverse constraint interaction layer is constructed to perform weighted fusion of the abnormal state identification results output by the auxiliary task branch and the intermediate layer features of the main task branch. The prediction error of the main task branch is transformed into a penalty term to adjust the feature weights of the auxiliary task branch in reverse, thereby realizing bidirectional reverse constraints between the main task and the auxiliary task.

[0038] Furthermore, the shared feature extraction layer is subjected to a reverse constraint through the abnormal state recognition auxiliary task branch, including:

[0039] During the forward inference process of model training, both the predicted fuel consumption value and the abnormal state identification result are obtained simultaneously.

[0040] A composite loss function is constructed, which includes the main task loss, the auxiliary task loss, and the cooperative constraint loss. The cooperative constraint loss is used to ensure that the predicted value of aviation fuel consumption under abnormal conditions and the predicted value under normal conditions meet the preset physical consistency relationship.

[0041] Backpropagation of gradients is performed based on the composite loss function, so that the parameters of the shared feature extraction layer are optimized by the losses of the main task and the auxiliary task simultaneously, and the main task and the auxiliary task are mutually constrained by the gradient chain rule.

[0042] During training, the weights of each loss term in the composite loss function are dynamically adjusted based on changes in the prediction accuracy of the main task and the recognition accuracy of the auxiliary task.

[0043] Furthermore, the prediction results will be continuously monitored, including:

[0044] Construct a three-dimensional monitoring system that includes core indicators of the main task, related indicators of auxiliary tasks, and pre-input data quality indicators, and set sliding statistical windows and performance thresholds for each indicator;

[0045] A combination of timed triggering and threshold triggering is used to calculate and update various monitoring indicators within the sliding window in real time;

[0046] When a monitoring indicator is detected to be not in line with a preset threshold, the system combines the status of the preceding indicators and the trend of indicator changes to determine whether the performance drift is data-driven drift or model aging drift, and quantifies the severity of the drift.

[0047] Furthermore, automatically triggering model rollback or feature reconstruction operations includes:

[0048] Establish a historical model version management system to standardize the storage and recording of model versions that meet the preservation conditions;

[0049] When a model is determined to be aging drift and the preset rollback trigger conditions are met, candidate versions with better performance than the current model are selected from historical model versions, and the rollback version is selected according to the principle of optimal performance.

[0050] Load the selected rollback version to replace the current online model, and validate the rollback model using a preset number of normal flight data. Once the validation is successful, resume the prediction process.

[0051] When the model fails validation or performance drift is frequently triggered after rollback, it is determined that the existing feature system is invalid, triggering a feature reconstruction operation. Feature dimensions are added or adjusted and the feature construction operation is re-executed until the core monitoring indicators of the main task of the reconstructed model meet the preset requirements.

[0052] In a second aspect, the present invention proposes an adaptive fuel consumption prediction system integrating a physical model and multi-task learning, for implementing the adaptive fuel consumption prediction method integrating a physical model and multi-task learning described in the first aspect. The system includes:

[0053] The data acquisition and processing module is configured to acquire and preprocess the aircraft's fast access recorder data to obtain preprocessed flight data.

[0054] The full-dimensional data analysis and calculation module is configured to analyze the flight phase based on the preprocessed flight data, calculate the corresponding atmospheric environmental parameters, and associate the analysis and calculation results with the flight data to obtain full-dimensional data containing flight status and environmental parameters.

[0055] The input feature parameter acquisition module is configured to construct features from the full-dimensional data based on the aircraft aerodynamic model and engine thrust model, generating aerodynamic force features and thrust features that reflect flight conditions, which are used as input feature parameters.

[0056] The model building module is configured to build a multi-task prediction model, including a shared feature extraction layer, a main task branch for fuel consumption prediction, and an auxiliary task branch for abnormal state identification. The shared feature extraction layer performs deep extraction on the input feature parameters to generate a shared feature representation. The main task branch outputs the predicted fuel consumption value based on the shared feature representation. The auxiliary task branch outputs the abnormal state identification result based on the shared feature representation.

[0057] The training module is configured to apply a reverse constraint to the shared feature extraction layer through the abnormal state recognition auxiliary task branch during the training of the multi-task prediction model.

[0058] The prediction module is configured to use the trained model to predict the fuel consumption of the target flight mission and output the prediction results.

[0059] The monitoring module is configured to continuously monitor the prediction results. When a drift in prediction performance is detected and exceeds preset conditions, it will automatically trigger model rollback or feature reconstruction operations to restore prediction performance.

[0060] In a third aspect, the present invention provides an apparatus comprising:

[0061] At least one processor;

[0062] and a memory communicatively connected to at least one of the processors;

[0063] The memory stores instructions that can be executed by the processor to implement, in accordance with the first aspect, an adaptive prediction method for fuel consumption that integrates a physical model and multi-task learning.

[0064] The beneficial effects of this invention are:

[0065] Existing technologies typically employ a sequential approach of anomaly handling followed by model prediction, with these two processes operating independently. This leads to the prediction model being highly sensitive to incompletely removed anomaly data. This invention addresses this by constructing a multi-task model that parallels the primary task of fuel consumption prediction with an anomaly identification auxiliary task, and utilizes a reverse constraint mechanism for collaborative optimization during training. This structure enables the model to actively learn and identify anomaly data features while simultaneously learning fuel consumption patterns, thereby generating more generalizable feature representations that are less sensitive to anomalies at the shared feature extraction layer. This fundamentally enhances the model's ability to actively suppress data noise and interference, ensuring stable and reliable prediction results even under complex flight conditions or fluctuating data quality.

[0066] Traditional data-driven methods, when constructing features, may rely solely on the statistical correlation of data, sometimes leading to predictions that deviate from the actual physical laws governing aircraft operation. This invention, in the feature construction stage, deeply integrates aircraft aerodynamic models and engine thrust models, mapping raw flight parameters into features with clear physical meanings, such as lift, drag, thrust, and lift-to-drag ratio. This approach embeds the principles of aeronautical flight dynamics as prior knowledge into the model, constraining its learning direction and ensuring that the prediction process not only depends on data patterns but also follows physical laws. This makes the prediction results more consistent with actual flight conditions, significantly improving its interpretability and reliability in engineering applications.

[0067] Existing technologies lack the ability to automatically monitor and repair performance degradation after model deployment, typically relying on manual intervention, which is slow and costly. This invention designs a closed-loop system comprising "real-time monitoring - drift determination - model rollback / feature reconstruction." This system can diagnose the health status of the model in real time through a quantitative multi-dimensional indicator system. Once model aging drift is detected, it can automatically trigger a rollback to a higher-performing historical version or perform feature reconstruction, thereby autonomously restoring performance. This adaptive and self-correcting capability transforms the static prediction model into a dynamic, self-maintaining intelligent system, effectively solving the performance degradation problem of long-term model operation, ensuring the stability and accuracy of the prediction system throughout its entire lifecycle, and reducing operational costs.

[0068] Compared to the fragmented technical aspects of existing technologies, this invention provides a complete end-to-end solution from data access to long-term stable model operation. This makes fuel consumption prediction not only theoretically feasible but also highly robust, reliable, and easy to maintain in engineering practice. It can provide more accurate and reliable data-driven decision support for airlines' core businesses such as refined fuel management and flight operation optimization, demonstrating significant practical application value. Attached Figure Description

[0069] 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:

[0070] Figure 1 This is a flowchart of an adaptive prediction method for aviation fuel consumption that integrates physical models and multi-task learning in the first embodiment of the present invention;

[0071] Figure 2 This is a structural block diagram of an adaptive prediction system for aviation fuel consumption that integrates physical models and multi-task learning, as described in the second embodiment of the present invention.

[0072] 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

[0073] 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.

[0074] 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.

[0075] The first embodiment of the present invention provides an adaptive prediction method for aviation fuel consumption that integrates physical models and multi-task learning, comprising the following steps:

[0076] Step S10: Acquire the aircraft's fast access recorder data and preprocess it to obtain preprocessed flight data;

[0077] Step S20: Based on the preprocessed flight data, analyze the flight phase and calculate the corresponding atmospheric environmental parameters. Associate the analysis and calculation results with the flight data to obtain full-dimensional data including flight status and environmental parameters.

[0078] Step S30: Based on the aircraft aerodynamic model and engine thrust model, feature construction is performed on the full-dimensional data to generate aerodynamic force characteristics and thrust characteristics that reflect flight conditions, which are used as input feature parameters.

[0079] Step S40: Construct a multi-task prediction model, including a shared feature extraction layer, a main task branch for fuel consumption prediction, and an auxiliary task branch for abnormal state identification; the shared feature extraction layer performs deep extraction on the input feature parameters to generate a shared feature representation; the main task branch outputs the predicted fuel consumption value based on the shared feature representation; the auxiliary task branch outputs the abnormal state identification result based on the shared feature representation.

[0080] Step S50: During the training of the multi-task prediction model, an inverse constraint is applied to the shared feature extraction layer through the abnormal state recognition auxiliary task branch.

[0081] Step S60: Use the trained model to predict the fuel consumption of the target flight mission and output the prediction results.

[0082] Step S70: Continuously monitor the prediction results. When a drift in prediction performance is detected and exceeds the preset conditions, automatically trigger model rollback or feature reconstruction operations to restore prediction performance.

[0083] To more clearly illustrate the adaptive prediction method for aviation fuel consumption that integrates physical models and multi-task learning according to the present invention, the following will be combined with... Figure 1 The steps in the embodiments of the present invention are described in detail below:

[0084] Step S10: Acquire the aircraft's fast access recorder data and preprocess it to obtain preprocessed flight data;

[0085] In this embodiment, the preprocessing steps include:

[0086] Step S11: Combine the physical rule threshold method and the statistical model anomaly detection method to clean the data of the fast access recorder and remove outliers.

[0087] Step S12: Based on the timestamp, perform time alignment on the cleaned data;

[0088] Step S13: For randomly missing data and consecutively missing data, interpolation and imputation by the mean of similar data are used to handle missing values, respectively.

[0089] Step S14: Standardize the data after time alignment and missing value processing to eliminate the dimensional differences between different parameters and obtain standardized, time-consistent preprocessed flight data.

[0090] In one specific implementation, the adaptive prediction method for aviation fuel consumption that integrates physical models and multi-task learning, as described in this invention, first performs step S10, which involves acquiring and preprocessing the aircraft's Quick Access Recorder (QAR) data to obtain high-quality preprocessed flight data. This step is fundamental to ensuring the accuracy and stability of subsequent model predictions. Its core objective is to transform the raw, multi-source QAR data collected from the aircraft, which may contain noise and missing values, into a structured, anomaly-free, time-aligned, and standardized data format to meet the needs of subsequent physical model mapping and deep learning model training. This preprocessing process can be specifically broken down into a series of refined operations such as data cleaning, time alignment, missing value handling, and standardization.

[0091] Specifically, the preprocessing process begins with data cleaning in step S11. Full QAR parameter data for the specified flight is acquired. This data covers flight status parameters such as altitude, airspeed, and landing gear status, as well as engine operating parameters such as high-pressure rotor speed (N2) and throttle position. For this raw data, this embodiment employs a combined strategy of physical rule thresholding and statistical model anomaly detection to remove outliers. Physical rule thresholding is primarily used to remove extreme outliers, i.e., data points that clearly violate the physical laws of aircraft operation or exceed equipment design limits. For example, the system sets valid ranges for each core parameter based on the flight manual and relevant aviation industry standards for the specific aircraft type (e.g., Boeing 737NG or Airbus A320). For instance, flight altitude cannot be negative, and engine speed should not exceed 110% of the maximum permissible speed. Any data point exceeding this range will be directly identified as an anomaly and removed. However, physical thresholding alone cannot identify hidden outliers that are within a reasonable range but do not conform to the current flight conditions. To address this, this embodiment further introduces a statistical model anomaly detection method, such as using isolated forests or models based on multivariate Gaussian distributions, to identify data distribution anomalies in multidimensional space. For example, when an aircraft is stationary on the ground, its airspeed should be close to zero, and the engine speed should be at idle or off. If a frame of data shows an airspeed of 50 knots, although this value is normal during flight, it is a latent anomaly under this specific condition. The statistical model can effectively capture such anomalous data that is inconsistent with the context and remove or mark it.

[0092] After data cleaning, step S12 is executed to time-align the cleaned data. Since different parameters in the QAR system may be collected by different sensors, their data acquisition frequencies and timestamps may have slight deviations; the acquisition frequency is typically between 1Hz and 16Hz. For example, the timestamp for altitude data may lag behind airspeed data by 0.1 seconds. To ensure that all parameters accurately reflect the aircraft's overall state at any given moment, multi-source parameters must be unified onto the same reference time axis. This embodiment preferably uses the Global Positioning System (GPS) timestamp as the reference due to its high accuracy and global uniformity. The alignment process uses linear interpolation. The system first constructs a standard reference time axis based on the minimum sampling interval of all parameters (e.g., 0.0625 seconds, corresponding to 16Hz) or a preset uniform time interval (e.g., 0.1 seconds). For any parameter, if its original timestamp does not fall on a node of the reference time axis, or if there is no sampled value at a certain reference time node, the system performs linear interpolation based on the two nearest valid data points to calculate the corresponding value of the parameter at the reference time node. This step ensures that all parameter sequences are resampled and aligned to a unified time axis, thereby guaranteeing the consistency of the data over time and laying the foundation for subsequent feature construction and model analysis based on time series.

[0093] Subsequently, the system executes step S13 to handle missing values ​​in the time-aligned data. QAR data missing values ​​can be categorized into two types: random missing values ​​and continuous missing values. Random missing values ​​are typically caused by the loss of a single or several frames of data during transmission, with a short missing length (e.g., less than 3 frames). For this type of situation, this embodiment uses linear interpolation for filling. For example, if the altitude at t=4s is 3000 feet, the altitude at t=6s is 3020 feet, and data at t=5s is missing, the system will interpolate to calculate the altitude at t=5s as 3010 feet. This method is very effective for scenarios where flight parameters change quasi-linearly over a short period. The other type is continuous missing values, which are usually caused by temporary sensor malfunctions or signal interruptions resulting in a longer period of data gaps (e.g., more than 5 frames or lasting several seconds). In this case, simple linear interpolation would introduce a large error. Therefore, this embodiment uses a strategy of filling with the average of similar flight data. Specifically, when engine N2 data for a flight is continuously missing during a certain flight phase (such as the climb phase) between t=100s and 120s, the system will retrieve several flights (e.g., 10 flights) with the same aircraft type, flying the same route, and with similar flight conditions from the historical database. It will then extract the N2 data of these flights within the same time period of the corresponding flight phase, calculate their frame-by-frame average, and form an N2 mean curve. This curve will be used as filler to complete the missing data for the current flight. If a sufficient number of similar flights cannot be found or the missing time is too long, to avoid introducing unreliable data, the system will mark this data segment as invalid and remove it from subsequent model training.

[0094] Finally, step S14 is executed to standardize the data that has undergone the above processing and is free of anomalies, has consistent timing, and is complete without missing data. Because the dimensions and numerical ranges of different flight parameters vary greatly (for example, altitude is measured in feet and can reach tens of thousands; while engine speed is measured as a percentage and is usually within 100), directly inputting it into the model can lead to an imbalance in the weights of different features, affecting the convergence speed and final performance of the model training. This embodiment uses the Z-score standardization method to eliminate the dimensional differences between parameters. For each parameter sequence, this method first calculates its mean across the entire training dataset. and standard deviation Then, for each data point X in the sequence, apply the formula... Perform the conversion. These are values ​​after Z-score standardization. For example, if the mean of an airspeed dataset during the cruise phase of a flight is 280 knots and the standard deviation is 15 knots, then an airspeed value of 310 knots will become (310-280) / 15=2 after standardization. Through this process, all parameters are transformed into a dimensionless distribution with a mean of 0 and a standard deviation of 1. It is worth noting that the mean calculated during the training phase... and standard deviation This data will be saved and used in subsequent predictions of new data. and Standardization is performed to ensure the consistency of data distribution. At this point, the entire preprocessing process is complete, and the system outputs standardized, complete, and time-series consistent high-quality preprocessed QAR data, providing reliable data input for subsequent steps S20's flight status analysis and environmental parameter calculations.

[0095] Step S20: Based on the preprocessed flight data, analyze the flight phase and calculate the corresponding atmospheric environmental parameters. Associate the analysis and calculation results with the flight data to obtain full-dimensional data including flight status and environmental parameters.

[0096] In this embodiment, the flight phase is analyzed based on the preprocessed flight data, and the corresponding atmospheric environmental parameters are calculated, including:

[0097] Step S21: Using the flight altitude, airspeed, landing gear status, engine speed and throttle position in the preprocessed flight data as input, the flight process is divided into multiple flight phases including takeoff, climb, cruise, descent, approach and landing by using threshold judgment combined with time sequence continuity verification logic.

[0098] Step S22: Based on the international standard atmospheric model, using flight altitude as input, and according to the different atmospheric layers at the flight altitude, the corresponding temperature, pressure and density function relationships are used to calculate the atmospheric environmental parameters at the corresponding altitude.

[0099] Step S23: The parsed flight phase information and calculated atmospheric environmental parameters are correlated and fused with the preprocessed flight data according to the timestamp to obtain full-dimensional data containing flight status and environmental parameters.

[0100] After completing the initial preprocessing of the Quick Access Recorder (QAR) data, the present invention proceeds to step S20. Its core task is to further analyze the specific flight stage at each moment during the entire flight based on the preprocessed flight data, and simultaneously calculate the atmospheric environmental parameters corresponding to the flight altitude. Finally, these analyzed and calculated high-dimensional contextual information are accurately correlated back into the original data stream, thereby generating a full-dimensional dataset containing flight state and environmental parameters, providing necessary and rich input for subsequent feature construction based on the physical model. In a specific implementation, this step includes three closely interconnected parts: flight stage analysis, atmospheric environmental parameter calculation, and data fusion supplementation.

[0101] First, the system performs flight phase analysis. This process utilizes a multi-parameter fusion judgment logic, taking the preprocessed QAR core parameters obtained in step S10 as input, specifically including flight altitude, calibrated airspeed, landing gear status, engine high-pressure rotor speed (N2), and throttle position. The core of this logic is a robust identification strategy combining "threshold judgment and temporal continuity verification," aiming to accurately divide the continuous flight process into six core phases: takeoff roll, takeoff climb, cruise, descent, approach, and landing roll. In a specific implementation, the specific judgment logic for each phase is as follows:

[0102] (1) Takeoff roll: The judgment conditions are that the landing gear is in the down position (logic value is 1), the altitude is not greater than the ground elevation plus 50 feet, the airspeed increases from 0 to the landing gear lift speed VR (typical value is 150-170 knots), and the engine speed N2 is not less than 95%. This stage continues in time until the airspeed reaches VR and the aircraft lifts the landing gear and begins to ascend.

[0103] (2) Takeoff Climb: The criteria for this stage are that the altitude continuously increases from 50 feet above ground level to 1,000 feet, the landing gear status changes from down (logic value 1) to retracted (logic value 0), the engine speed N2 is not less than 90%, and the airspeed is not less than the safe takeoff speed V2 (typically 150-170 knots). This stage continues until the aircraft's altitude stabilizes and reaches the cruise altitude range.

[0104] (3) Cruise: The criteria for judgment are that the altitude is within the model-compatible range of 30,000 feet to 45,000 feet, and the vertical altitude fluctuation within 30 seconds is no more than ±500 feet, while the airspeed is stable in the range of 280-340 knots and the engine speed N2 is maintained between 85% and 90%. This stage needs to last for more than 10 minutes in terms of duration to exclude short-term level flight transitions.

[0105] (4) Descent: The judgment condition is that the altitude continues to decrease from the cruising altitude at a rate of not less than 500 feet per minute, while the airspeed is in the range of 250-300 knots and the landing gear is in the retracted state (logic value 0). This phase continues in time until the altitude drops to above 1,000 feet.

[0106] (5) Approach: The criteria for judgment are an altitude of no more than 1,000 feet, a rate of descent of no more than 300 feet per minute, landing gear in the down position (logic value 1), and airspeed of no more than 180 knots. This phase continues in sequence until the aircraft touches down.

[0107] (6) Landing roll: The judgment condition is that after the aircraft touches down, that is, the altitude is no greater than the ground elevation plus 20 feet, the landing gear is in the down position (logic value 1), the airspeed continuously decreases from the landing speed to 0, and the engine speed N2 is no higher than 70%. This stage continues in time until the aircraft comes to a complete stop.

[0108] To avoid misjudgment of stages due to instantaneous fluctuations or noise in single-frame data, this invention introduces a temporal continuity verification mechanism. This means that all parameter combinations and threshold conditions for each stage must continuously satisfy a minimum time window, such as 3 seconds (corresponding to 3 to 48 consecutive data frames based on the QAR data acquisition frequency). Only then is the entry into the corresponding flight stage confirmed. Through this detailed and rigorous logic, the system can accurately label each timestamp in the preprocessed QAR dataset with the flight stage.

[0109] Simultaneously, atmospheric environmental parameters are calculated in parallel. This process is explicitly based on the International Standard Atmospheric Model (ISAM) as defined by the International Civil Aviation Organization (ICAO), which provides a recognized and standardized environmental benchmark for aircraft performance analysis. The core input for the calculation is the "flight altitude" parameter from the preprocessed QAR data. Since the altitude in QAR data is usually in feet, the system first converts it to geometric altitude H in meters (m) required for the ISA model calculation using a conversion formula (1 ft = 0.3048 m). The ISA model vertically divides the atmosphere into different layers. Civil transport aircraft mainly operate in the troposphere (0 to 11 km) and the lower stratosphere (11 to 20 km). The ISA model defaults to the following sea-level benchmark conditions: temperature T0 = 288.15 K (15 °C), air pressure... =101325 Pa, density =1.225kg / m 3 Therefore, the calculation process is layered: when the flight altitude H is within the troposphere (0 to 11 km), the model assumes that the atmospheric temperature T decreases linearly with altitude, with a decrease rate of α = 0.0065 K / m. The system uses the formula... Calculate the temperature, where The standard temperature at sea level is 288.15 K; for example, when the flight altitude H = 3000 m, the temperature T = 288.15 - 0.0065 × 3000 = 268.65 K (-4.5℃).

[0110] Subsequently, the relationships were derived based on the hydrostatic equation and the ideal gas law. Calculate atmospheric pressure ,as well as Calculate atmospheric density ,in These are, respectively, the standard atmospheric pressure at sea level, the standard air density at sea level, the standard gravitational acceleration, the average molar mass of dry air, and the universal gas constant. Where g is the gravitational acceleration (9.80665 m / s²). 2 M is the molar mass of dry air (0.0289644 kg / mol), R is the universal gas constant (8.31432 J / (mol·K)), and the exponent is approximately -5.25588. For example, at H = 3000 m, the pressure P ≈ 101325 × (268.65 / 288.15) -5.25588 ≈69700 Pa. Density ρ≈69700×0.0289644 / (8.31432×268.65)≈0.903 kg / m³ 3 ;

[0111] When the flight altitude H enters the lower stratosphere (11km to 20km), the model assumes this layer is isothermal with a constant temperature. At this point, the pressure P and density ρ are determined according to... and Perform calculations, where and These are the altitude of the tropopause and standard atmospheric pressure, respectively. Through this layered and precise calculation, the system matches a set of environmental parameters, including atmospheric temperature, pressure, and density, to the flight altitude at each timestamp.

[0112] in =11000m, The standard atmospheric pressure at an altitude of 11 km is 22632.06 Pa, and exp is the natural exponential function. For example, when the flight altitude H = 12000 m, the pressure P ≈ 22632.06 × exp(-9.80665 × 0.0289644 × (12000 - 11000) / (8.31432 × 216.65)) ≈ 22595 Pa. The density ρ ≈ 22595 × 0.0289644 / (8.31432 × 216.65) ≈ 0.363 kg / m³ 3 .

[0113] Finally, data fusion and supplementation are performed. The system precisely matches and correlates the "flight phase labels" (such as "cruise," "descent," etc.) generated in the first two steps and the "atmospheric environmental parameters" (temperature, pressure, density) calculated frame by frame, strictly according to the timestamps. This information is then added as new data columns to the preprocessed QAR data entries for the corresponding time series. After this step, the original dataset is successfully expanded into a full-dimensional preprocessed QAR dataset. Each row of this dataset not only records the aircraft's original flight parameters at that moment but also includes its macroscopic flight state (flight phase) and external physical environment (atmospheric parameters). This provides complete and high-quality data support for the subsequent step S30 to construct features with clear physical meaning and ultimately improve the accuracy and interpretability of the fuel consumption prediction model.

[0114] Step S30: Based on the aircraft aerodynamic model and engine thrust model, feature construction is performed on the full-dimensional data to generate aerodynamic force characteristics and thrust characteristics that reflect flight conditions, which are used as input feature parameters.

[0115] In this embodiment, the aerodynamic force characteristics and thrust characteristics reflecting the flight conditions are generated, including:

[0116] Step S31: Based on the aircraft aerodynamic model, combined with the atmospheric density, vacuum velocity and reference area in the full-dimensional data, the lift characteristics are calculated using the lift coefficient function related to the angle of attack and Mach number; the drag characteristics are calculated using the drag coefficient function combining the parasitic drag coefficient and the induced drag factor.

[0117] Step S32: Based on the engine thrust model, combined with the engine rotor speed, atmospheric pressure and atmospheric temperature in the full-dimensional data, the net thrust characteristics of the engine are calculated through the functional relationship between the engine characteristic coefficient and the environmental attenuation coefficient.

[0118] Step S33: Based on the lift characteristics, drag characteristics and engine net thrust characteristics, construct derived operating condition characteristics including lift-to-drag ratio, thrust-to-drag difference, aerodynamic force change rate and operating condition adaptability deviation.

[0119] Step S34: Verify the physical laws and filter the data discrimination of the constructed features to obtain the final core feature set, which is used as the input feature parameters of the model.

[0120] After acquiring comprehensive data including flight status and environmental parameters, the present invention proceeds to step S30, a core technical point, which involves deep feature construction and constraint mapping of the comprehensive data based on the aircraft physical model. This aims to transform the original, physically discrete QAR parameters into aerodynamic force and thrust characteristics that directly and quantitatively reflect the aircraft's actual flight conditions and stress states, ultimately forming a set of high-quality input feature parameters for subsequent multi-task prediction models. The core logic of this step lies in introducing classical aircraft dynamics theory as a strong constraint to achieve a nonlinear mapping from the original data space to a feature space with clear physical meaning and high operational condition identification. This effectively overcomes the black-box problem that may exist in purely data-driven methods, ensuring the physical consistency and interpretability of model predictions. In this embodiment, this process revolves around a dual-physical model constraint framework of "aerodynamic model + thrust model" and follows a systematic process of "basic feature calculation - derived feature construction - feature screening and verification."

[0121] Specifically, the feature construction process first involves quantifying the basic physical characteristics based on a pre-defined aircraft aerodynamic model and engine thrust model. In terms of aerodynamic feature calculation, this embodiment employs a six-degree-of-freedom nonlinear aerodynamic model adapted for civil aviation fixed-wing transport aircraft. This model uses flight state parameters (such as angle of attack α and Mach number Ma) from the full-dimensional data and atmospheric environment parameters (such as atmospheric density ρ) calculated in step S20 as core inputs. The system first calculates the aircraft lift (L), which follows a core formula. Where V is the flight vacuum speed obtained from QAR data and converted to a unit (e.g., from knots to meters per second), and S is the reference area from the specific aircraft manual (e.g., Airbus A320 reference area S=124.6m). 2 The aircraft reference area was obtained from [reference data]. The key lift coefficient is... It is a bivariate function related to the angle of attack α and the Mach number Ma. Its specific form is obtained by polynomial fitting of aerodynamic characteristic test data of a specific aircraft model. For example, it can be expressed as: Among them, the coefficients (such as the zero lift coefficient) Slope of the lift line (etc.) are all known or preset model-related constants. and These are the current (or instantaneous) flight Mach number and the reference (or initial) Mach number, respectively.

[0122] The model is applicable to the following ranges: angle of attack α∈[-5°,20°] (the angle of attack range for conventional civil aviation flight; exceeding this range will result in a stall risk zone), sideslip angle β∈[-10°,10°] (the range for conventional flight with no sideslip / small sideslip), and Mach number Ma∈[0.2,0.86] (the range for subsonic to high subsonic flight of civil aviation transport aircraft).

[0123] In a specific calculation example, for an Airbus A320 aircraft, during the cruise phase (H=12000m), the atmospheric density ρ=0.363kg / m³. 3 The flight vacuum velocity V = 250 knots (approximately 128.6 m / s), and the reference area S = 124.6 m². 2 Angle of attack α = 3°, Mach number Ma = 0.78. Assume zero lift coefficient. =0.25, slope of the lift line =0.09 / °, Mach number influence coefficient =0.15, reference Mach number =0.5, then the lift coefficient =0.25 + 0.09 × 3 0.15×(0.78 0.5) = 0.478. Substituting into the formula, the lift L≈0.5×0.363×128.62×124.6×0.478≈180000N (approximately 18 tons of force). This result is consistent with the physical law of lift and gravity balancing during the cruise phase.

[0124] Similarly, the system calculates aircraft drag D, and its core formula is: The drag coefficient in this formula consists of two parts: a parasitic drag coefficient that is essentially independent of the angle of attack. and the induced drag coefficient generated by lift. The induced drag factor K is also a model-specific constant. Through the above calculations, the system generates corresponding instantaneous lift and drag characteristic values ​​for each timestamp in the full-dimensional data.

[0125] Following the calculation example from the cruise phase above, let the parasitic drag coefficient of the A320 be... =0.018, induced drag factor K=0.045, lift coefficient has been calculated. =0.478. Therefore, the drag coefficient = 0.018 + 0.045 × 0.4782 ≈ 0.0283. Substituting into the formula, the drag D ≈ 0.5 × 0.363 × 128.62 × 124.6 × 0.0283 ≈ 10500 N (approximately 1.05 tons of force). This result conforms to the physical law of balancing drag and engine thrust during the cruise phase.

[0126] Meanwhile, in terms of thrust characteristic calculation, this embodiment adopts a thrust calculation model based on component-level characteristics, adapted to mainstream civil aviation twin-shaft turbofan engines. This model uses core engine parameters from full-dimensional data (such as low-pressure rotor speed N1 and high-pressure rotor speed N2), atmospheric environmental parameters calculated in step S20 (such as atmospheric pressure P and atmospheric temperature T), and flight Mach number Ma as inputs. Engine net thrust. The calculation formula can be simplified to The first term of the formula represents the total thrust produced by the engine, where , , These are the thrust characteristic index coefficients for the low-pressure rotor speed (N1) and the high-pressure rotor speed (N2) of the engine, respectively. and The two items are correction factors used to quantify the impact of high-altitude, low-pressure, and low-temperature environments on engine performance degradation (wherein...). and (The standard atmospheric parameters are at sea level); the second term of the formula represents the ram drag generated by the intake air during high-speed aircraft flight, where K2 is the intake drag coefficient. This represents the inlet area of ​​the air intake. Using this formula, the system can accurately estimate the net thrust characteristics of each engine based on real-time engine operating parameters and flight environment.

[0127] In a specific calculation example, using the cruise phase parameters mentioned above (P=22595Pa, T=216.65K), assuming the engine is a CFM56-5B with cruise speeds of N1=75% and N2=88%, and setting engine characteristic coefficients K1=0.85, a=0.6, b=0.9; intake drag coefficient K2=0.025, and inlet area... =1.4m 2 Substituting into the formula, the net thrust of a single engine is approximately Fn ≈ 4233 N. 210N = 4023N. For a twin-engine aircraft, the total thrust is approximately 8046N, which basically matches the calculated cruise drag (10500N). This deviation is within a reasonable range in the simplified engineering model. Using this formula, the system can accurately estimate the net thrust characteristics of each engine based on real-time engine operating parameters and flight environment.

[0128] After calculating the three primary fundamental characteristics—lift, drag, and net thrust—the system further performs temporal and physical dimension correlation processing on these fundamental characteristics in step S33, combining the context information of the flight phase, to construct a series of derived operating condition characteristics with higher recognizability and richer physical connotations. These derived characteristics mainly include: the "lift-to-drag ratio" (L / D) used to reflect aerodynamic efficiency; and "thrust-drag balance" characteristics used to characterize the acceleration or deceleration trend of the aircraft, such as the thrust-drag difference (L / D). ) and thrust-to-drag ratio ( The features used to quantify the severity of flight attitude adjustments include "aerodynamic force change rate" characteristics, such as lift change rate (dL / dt) and drag change rate (dD / dt), which can be obtained by calculating the difference between the basic features between adjacent data frames and then dividing by the time interval; and "condition adaptability deviation" characteristics used to evaluate flight performance economy, such as the deviation between the current lift coefficient and the optimal lift coefficient for this flight phase. And the current thrust and the thrust required to maintain the current flight status ( These derived features combine the instantaneous stress state of the aircraft with dynamic changing trends and performance optimization objectives, providing a more comprehensive characterization of flight conditions.

[0129] Finally, in step S34, the system rigorously screens and validates the validity of all constructed basic and derived features to form the core feature set ultimately input into the prediction model. This screening process employs a dual mechanism combining physical significance verification and data discrimination verification. Physical significance verification aims to eliminate feature values ​​that clearly violate physical laws due to data noise or model simplification. For example, if the calculated lift-to-drag ratio (L / D) exceeds the theoretical limit of the aircraft's aerodynamic efficiency (e.g., greater than 40), the data point will be deemed invalid. Data discrimination verification uses statistical analysis methods, such as analysis of variance, to evaluate the discriminative power of each feature across different flight phases (e.g., takeoff, cruise, and approach). Only features that can significantly distinguish between different operating conditions (e.g., whose ANOVA F-statistic is greater than a preset threshold of 10) are retained. Through this series of meticulous construction, derivation, and screening processes, the system ultimately outputs a core feature set that is appropriately dimensional, has clear physical meaning, and is highly recognizable for different flight conditions. For example, for the A320 model, it can generate a feature set that includes lift, drag, lift-to-drag ratio, lift rate of change, etc. and The combination of aerodynamic features such as deviations and thrust features such as engine net thrust, thrust-to-drag ratio, and thrust-to-demand thrust deviation will serve as high-quality input feature parameters for the next stage of the multi-task prediction model.

[0130] Step S40: Construct a multi-task prediction model, including a shared feature extraction layer, a main task branch for fuel consumption prediction, and an auxiliary task branch for abnormal state identification; the shared feature extraction layer performs deep extraction on the input feature parameters to generate a shared feature representation; the main task branch outputs the predicted fuel consumption value based on the shared feature representation; the auxiliary task branch outputs the abnormal state identification result based on the shared feature representation.

[0131] In this embodiment, a multi-task prediction model is constructed, including:

[0132] Step S41: Construct a shared feature extraction layer that includes a combination of a fully connected layer and an attention mechanism to encode the input feature parameters of the multi-task prediction model and output a shared feature representation.

[0133] Step S42: Construct the main task branch for predicting fuel consumption. The main task branch is a multi-layer fully connected regression network that takes the shared feature representation as input and outputs continuous predicted values ​​of fuel consumption.

[0134] Step S43: Construct an auxiliary task branch for abnormal state recognition. The auxiliary task branch is a classification network containing a fully connected layer and a classification function. It takes the shared feature representation as input and outputs the probability distribution of different abnormal state categories.

[0135] Step S44: Construct a reverse constraint interaction layer, which weights and fuses the abnormal state identification results output by the auxiliary task branch with the intermediate layer features of the main task branch, and converts the prediction error of the main task branch into a penalty term to reversely adjust the feature weights of the auxiliary task branch, thereby realizing bidirectional reverse constraints between the main task and the auxiliary task.

[0136] After completing the feature construction based on the physical model, the present invention proceeds to step S40 to construct an innovative multi-task prediction model. The core objective of this model is to utilize the high-quality, physically meaningful input feature parameters generated in step S30 to not only accurately predict fuel consumption but also simultaneously identify abnormal states during flight. Through a collaborative constraint mechanism between the two, it fundamentally solves the problems of sensitivity to abnormal data and insufficient robustness in existing prediction models. The model constructed in this embodiment adopts a dual-tower structure of serial sharing and parallel tasks. This structure can be subdivided into a shared feature extraction layer, a task-specific prediction layer, and a crucial reverse constraint interaction layer, thereby achieving deep collaboration and mutual promotion between the main task (fuel consumption prediction) and the auxiliary task (abnormal state identification).

[0137] Specifically, model construction begins in step S41, which involves building a shared feature extraction layer. The core function of this layer is to perform deep nonlinear encoding on the input fused features (e.g., a 12-dimensional feature vector) containing multi-dimensional information such as aerodynamics, thrust, and operating conditions. This extracts a general high-dimensional feature representation that simultaneously considers the variation patterns of fuel consumption and the inherent characteristics of abnormal states, thus avoiding the feature learning bias that may occur in single-task models. In this embodiment, the shared layer adopts a composite structure combining a fully connected layer and an attention mechanism, comprising three deep network layers. The first layer is a fully connected layer (FC1), whose input dimension matches the feature dimension output from S30 (e.g., 12 dimensions), and the output dimension is expanded to 64 dimensions. A LeakyReLU activation function (preferably with a slope of 0.01) is used to alleviate the gradient vanishing problem, while a Dropout layer (preferably with a dropout rate of 0.2) is introduced to enhance the model's generalization ability and prevent overfitting. Following this is a multi-head attention mechanism layer, with 8 heads and 64 input and output dimensions. The introduction of this layer is a technical innovation of this invention. Its core function is to dynamically capture the temporal and intrinsic physical correlations between input features. For example, the model can learn the hysteresis response relationship between engine thrust changes and fuel consumption, or the high correlation between abnormal fluctuations in lift-to-drag ratio and certain equipment failure states through this mechanism. The last layer is a fully connected layer (FC2), which further compresses the 64-dimensional weighted feature vector output by the attention layer to 32 dimensions. It also uses the LeakyReLU activation function and introduces a batch normalization layer to accelerate model convergence and improve training stability. This layer ultimately outputs a 32-dimensional general feature vector, denoted as... This serves as the unified input for all subsequent task branches.

[0138] Based on this shared feature representation The system then constructs two task-specific prediction branches in parallel. First, in step S42, the main task branch for fuel consumption prediction is constructed. The goal of this branch is to perform a regression task, namely, to accurately predict the average fuel consumption rate (unit: kg / s) within a short future time window (e.g., 10 seconds). Its network structure is a three-layer fully connected regression network. The first layer (FC3) integrates 32-dimensional... The first layer (FC4) maps the 8-dimensional features to 16 dimensions, and the second layer (FC5) further maps them to 8 dimensions, both using LeakyReLU as the activation function. The final layer (FC5) serves as the output layer, mapping the 8-dimensional features to 1 dimension without setting any activation function, so as to directly output continuous and unbounded fuel consumption predictions. Next, in step S43, an abnormal state recognition auxiliary task branch is constructed. The goal of this branch is to perform a multi-classification task, such as identifying four predefined flight states: thrust anomaly, aerodynamic imbalance, fuel leak, and normal state, thereby providing real-time operating condition constraints for fuel consumption prediction. Its network structure is a two-layer fully connected classification network. The first layer (FC6) maps the 32-dimensional features to 16 dimensions. The first layer maps the 16-dimensional features to 16 dimensions, employs the LeakyReLU activation function, and adds a Dropout layer (with a dropout rate of 0.2). The second layer (FC7) is the output layer, which maps the 16-dimensional features to a dimension equal to the number of anomaly categories (e.g., 4 dimensions) and employs the Softmax activation function. The Softmax function transforms the output values ​​into a probability distribution, where each element represents the probability that the current flight state belongs to the corresponding anomaly category.

[0139] Finally, and most importantly, the core innovation of this invention's model structure lies in step S44, which constructs a unique reverse constraint interaction layer. This layer establishes a dynamic, bidirectional constraint channel between the main task and the auxiliary task, thereby achieving closed-loop collaboration in prediction, identification, and correction. This completely breaks the limitation of traditional multi-task models where tasks only share low-level features without deep interaction. This interaction layer contains two reverse constraint modules. The first module implements the reverse constraint from anomaly identification to fuel consumption prediction. Specifically, it sets the anomaly category probability vector p (e.g., a 4-dimensional vector) output by the auxiliary task branch as the constraint. The features of the main task branch and the intermediate layer features (e.g., the 8-dimensional output features of FC4) are dynamically weighted and fused. The fusion formula can be designed as follows: ,in It is an adjustable constraint strength coefficient (e.g., an empirical value of 0.3).

[0140] The physical logic of this mechanism is as follows: when the auxiliary task identifies a high-probability abnormal state (e.g., a significantly increased probability p3 of fuel leakage), this abnormal signal amplifies the feature weights related to the abnormality in the main task network through multiplicative gating, thereby intelligently guiding the prediction logic of the main task to correct towards a pattern consistent with the abnormal condition (e.g., predicting a higher fuel consumption value). The second module implements a reverse constraint from "fuel consumption prediction error" to "abnormal state identification." This module calculates the absolute error E between the main task's predicted value and the actual value in real time. When the error E exceeds a preset threshold (e.g., 0.5 kg / s), the error signal is converted into a penalty term. Through gradient backpropagation, this penalty term adjusts the internal weights of the attention mechanism in the shared feature extraction layer, thereby forcing the model to pay more attention to input features that can cause significant deviations in fuel consumption prediction. This is equivalent to using the prediction performance of the main task as a supervisory signal to optimize the feature sensitivity of the auxiliary task, enabling it to more accurately identify the fundamental conditions causing abnormal fuel consumption. Through the collaborative work of these two reverse constraint modules, the model as a whole forms a dynamic and adaptive collaborative learning system, which significantly improves the accuracy and robustness of fuel consumption prediction under complex and abnormal operating conditions.

[0141] Step S50: During the training of the multi-task prediction model, an inverse constraint is applied to the shared feature extraction layer through the abnormal state recognition auxiliary task branch.

[0142] This embodiment specifically includes:

[0143] Step S51: During the forward inference process of model training, the predicted value of aviation fuel consumption and the result of abnormal state identification are obtained simultaneously.

[0144] Step S52: Construct a composite loss function that includes main task loss, auxiliary task loss and cooperative constraint loss, wherein the cooperative constraint loss is used to ensure that the predicted value of aviation fuel consumption under abnormal conditions and the predicted value under normal conditions meet the preset physical consistency relationship.

[0145] Step S53: Backpropagation of gradients is performed based on the composite loss function, so that the parameters of the shared feature extraction layer are optimized by the losses of the main task and the auxiliary task at the same time, and the main task and the auxiliary task are mutually constrained by the gradient chain rule.

[0146] Step S54: During the training process, dynamically adjust the weights of each loss term in the composite loss function based on the changes in the prediction accuracy of the main task and the recognition accuracy of the auxiliary task.

[0147] After constructing the multi-task prediction model containing the bidirectional reverse constraint mechanism, the present invention then performs the key step S50, which is to perform collaborative training on the model and apply anomaly feature constraints. The core of this is to effectively transmit the constraint effect of the anomaly state identification auxiliary task on the main task of fuel consumption prediction from the model structure level to every link of parameter optimization through a carefully designed training process and loss function. This allows the model to learn how to accurately predict fuel consumption while internally learning to identify and suppress the interference caused by abnormal data features.

[0148] Specifically, in each iteration of model training, the forward inference process described in step S51 is executed first. The system feeds a batch of input feature parameters extracted from historical QAR data and processed in step S30 into the multi-task prediction model. The data stream first passes through a shared feature extraction layer, and after deep encoding with fully connected and attention mechanisms, generates a shared feature representation that can meet the needs of both tasks. Subsequently, this shared feature representation is fed in parallel into the main task branch for fuel consumption prediction and the auxiliary task branch for abnormal state recognition, thereby simultaneously obtaining preliminary fuel consumption prediction values ​​and the recognition probabilities of each abnormal state. At the same time, within the main task branch, the constraint module located in the reverse constraint interaction layer uses the abnormal probabilities output by the auxiliary task branch in real time to dynamically weight the intermediate layer features of the main task branch, generating a corrected feature that incorporates abnormal state information, and finally outputs the fuel consumption prediction value based on this corrected feature.

[0149] Next, in step S52, the system constructs and calculates a composite loss function based on the output of forward inference and data labels, namely, the actual fuel consumption value and the actual abnormal state labels. This composite loss function is the core of achieving dual-task collaborative optimization. It consists of three key weighted components: the main task loss, the auxiliary task loss, and the collaborative constraint loss. The main task loss preferably adopts a modified Huber loss function. This loss function is equivalent to mean squared error (MSE) when the prediction error is small, for example, less than a preset threshold δ=0.8kg / s, to ensure prediction accuracy. When the error is large, it becomes a linearly increasing penalty, similar to L1 loss, thereby effectively reducing the excessive influence of extreme abnormal fuel consumption data points on the model training process and enhancing the robustness of the model. The auxiliary task loss employs weighted cross-entropy loss, aiming to address the class imbalance problem caused by the far greater number of normal state samples than various abnormal state samples in actual flight data. By assigning higher loss weights to anomalous categories with smaller sample sizes (such as fuel leaks, thrust anomalies, etc.) (for example, weights are calculated based on "total number of samples / (number of categories × number of samples in that category)," e.g., a normal state weight of 1 and an anomalous state weight of 5.7), the model is ensured to fully learn the ability to identify these crucial but scarce anomalous states during training. Crucially, the collaborative constraint loss uses a clear mathematical form to enforce physical consistency between the outputs of the two tasks.

[0150] Subsequently, in step S53, the system performs backpropagation of gradients based on the calculated total composite loss using a modern optimizer such as Adam (e.g., setting a learning rate of 1e-4 and a decay rate of 0.9) to update all learnable parameters of the model. During this process, the bidirectional back-constraint mechanism designed in this invention is implemented through the gradient chain rule. On one hand, the gradient of the auxiliary task loss not only updates the parameters of its own branch but also feeds back to the shared feature extraction layer through the total loss. This means that if the abnormal state is not accurately identified, the parameters of the shared layer (especially the weights of the attention mechanism) will be adjusted to enhance the ability to capture input features highly correlated with abnormal states, thereby indirectly improving the prediction accuracy of the main task when facing abnormal conditions. On the other hand, the gradients of the main task loss and the collaborative constraint loss are also fed back to the shared layer for parameter optimization. Furthermore, when the prediction error of the main task is too large, the back-constraint interaction layer will convert this error signal into a direct adjustment of the weights of the auxiliary task's attention layer, thereby prompting the auxiliary task to more accurately identify key conditions that could lead to deviations in fuel consumption prediction. In this way, the parameter updates of the shared feature extraction layer are simultaneously supervised and optimized by two tasks, and a closed loop of mutual correction and co-evolution is formed between the two task branches through gradient flow.

[0151] More specifically, in step S50, the system trains and optimizes the dual-task deep collaborative prediction model, which includes a reverse constraint interaction layer, constructed in step S40. The core design of this embodiment lies in an innovative composite loss function and a complete reverse constraint implementation mechanism.

[0152] The composite loss function adopts the "main task loss" +Auxiliary task losses +Cooperative constraint loss The composite loss function, "", balances the training priorities of the two tasks through weight allocation, ensuring the prediction accuracy of the main task while leveraging the constraint effect of the auxiliary task. Its specific formula is as follows: ;in, As loss weights, in a specific implementation, their empirical values ​​can be set to 0.6, 0.3, and 0.1, and optimized through cross-validation. The definitions of each loss term are as follows:

[0153] Main task loss: For fuel consumption prediction, an improved Huber loss is used to balance prediction accuracy and robustness to outliers. The formula is:

[0154] when ;

[0155] when ;

[0156] in, For predicted values, For truth labels, The threshold value is 0.8 kg / s, which can be determined empirically based on a large amount of flight data. The loss function uses mean squared error (MSE) to ensure accuracy when the error is small, and switches to L1 loss when the error is large to reduce the interference of extreme values ​​on model training.

[0157] Auxiliary task loss: For anomaly state identification, weighted cross-entropy loss is used to effectively address the imbalance problem of flight anomaly state samples. The formula is: ;

[0158] in, For real labels, This represents the probability of the corresponding class predicted by the model. This refers to the category weights. For example, when the normal state samples account for 85% and the abnormal state samples account for 15%, the normal state weight can be calculated as "total number of samples / (number of categories × number of samples in that category)". Abnormal state weights This ensures that the contribution of each type of loss is balanced.

[0159] Collaborative constraint loss: Construct consistency constraints between the primary and secondary tasks, transforming physical laws into quantifiable loss terms. The formula is:

[0160] ;

[0161] in, To assist in identifying the main task's predicted value when the auxiliary task is in a "normal state," The predicted values ​​are used to identify abnormal conditions such as "fuel leak / thrust anomaly". The minimum difference threshold can be empirically set to 0.3 kg / s. The reverse constraint is implemented through a closed-loop process of "forward inference → loss calculation → gradient backpropagation → parameter adjustment," with the specific steps as follows:

[0162] Step 1: Forward Inference. The fused features constructed in step S30 are input into the model, passed through a shared feature extraction layer and two task heads, to obtain the predicted fuel consumption value. And the probability of predicting abnormal states, P.

[0163] Step 2: Calculate the composite loss. Based on the true labels, calculate sequentially... And according to weight The total loss is obtained by weighted summation.

[0164] Step 3: Gradient backpropagation and parameter tuning. The Adam optimizer (in one embodiment, the learning rate can be set to 1e-4 and the decay rate to 0.9) is used to... Perform backpropagation. Use the gradient chain rule to achieve mutual constraints between the two tasks:

[0165] Constraints of auxiliary tasks on the main task: When When it is large, its gradient passes through Backpropagation focuses on adjusting the attention weights of the shared feature extraction layer, making the model pay more attention to features related to anomalies (such as thrust fluctuations), thereby optimizing the input of the main task and improving the accuracy of fuel consumption prediction.

[0166] Constraints of the primary task on the auxiliary task: When Larger and When the error gradient of the main task is greater than 0 (the prediction result does not conform to the laws of physics), the error gradient of the main task will adjust the parameters of the auxiliary task head in the opposite direction, prompting the auxiliary task to more accurately identify the real operating conditions that cause abnormal fuel consumption, thus forming a collaborative optimization closed loop.

[0167] Step 4: Adaptive Adjustment of Constraint Strength. To avoid task imbalance caused by fixed weights, a dynamic weight adjustment mechanism is introduced. For example, during training, if the prediction accuracy of the main task (R...) changes... 2 If it is below 0.8, then increase. (e.g., down to 0.8), decrease (If the accuracy is reduced to 0.1), prioritize ensuring the accuracy of the primary task; if the accuracy (Acc) of the auxiliary task is lower than 0.85, then increase the accuracy. (e.g., down to 0.5), decrease (e.g., up to 0.4), strengthen auxiliary task training to ensure the effectiveness of reverse constraints.

[0168] Through the training mechanism described above, which includes collaborative constraint loss and reverse dynamic constraints, the model constructed in this invention not only achieves deep collaboration between two tasks but also achieves a significant performance improvement. Compared with traditional single-task models, the method described in this invention has improved the accuracy of fuel consumption prediction and the reliability (accuracy) of abnormal state identification to a certain extent.

[0169] Step S60: Use the trained model to predict the fuel consumption of the target flight mission and output the prediction results.

[0170] Specifically, when it is necessary to predict fuel consumption for a specific flight mission (whether it is an ongoing real-time flight or a completed historical flight), the system first acquires the raw QAR data stream or data file corresponding to the mission. This newly acquired data undergoes the same data preprocessing and feature engineering pipeline as during model training. First, the data undergoes data cleaning, GPS timestamp-based time alignment, interpolation or imputation for random or consecutive missing values, and Z-score standardization in step S10, generating a uniformly formatted and time-ordered preprocessed QAR dataset. Subsequently, this data is fed into the flight state and environmental parameter parsing module in step S20. Through multi-parameter fusion logic, each data point is precisely labeled with its corresponding flight phase (e.g., cruise, descent), and key environmental parameters such as atmospheric temperature, pressure, and density at the corresponding flight altitude are calculated based on the International Standard Atmosphere (ISA) model, thus forming comprehensive data encompassing "raw parameters - flight state - atmospheric environment." Following this, this complete comprehensive data is transferred to the feature construction and physical constraint mapping module in step S30. Here, the system calls the same aircraft aerodynamic model and engine thrust model as during training, calculates the data, and generates basic features including lift, drag, and net engine thrust, as well as a series of derived operating condition features such as lift-to-drag ratio and thrust-to-drag difference. After physical law verification and data discrimination filtering, a core feature vector with dimensions that perfectly match the model input layer is finally formed.

[0171] After completing the rigorous data preparation process described above, the system inputs the generated core feature vector sequence into the trained and deployed multi-task prediction model in chronological order, performing a pure forward inference calculation. During this inference process, the model's network weights remain fixed, without any gradient calculations or parameter updates. The input features first flow through a shared feature extraction layer, where they are encoded into a high-dimensional shared feature representation. Subsequently, this shared feature representation is simultaneously sent to two parallel task branches. The main task branch for fuel consumption prediction, based on this shared representation, processes it through its internal fully connected regression network, ultimately generating a continuous numerical value at the output layer. This value represents the model's prediction of the average fuel consumption rate within a predetermined time window (e.g., the next 10 seconds) from the current moment, typically expressed in kilograms per second (kg / s). Simultaneously, the auxiliary task branch for anomaly identification, based on the same shared representation, processes it through its classification network and Softmax function, outputting a probability distribution vector. Each element of this vector corresponds to the probability of a preset flight state (e.g., normal, thrust anomaly, aerodynamic imbalance, fuel leak, etc.).

[0172] Ultimately, the system integrates and encapsulates the outputs of the two task branches to form structured prediction results for external output. A typical output data frame includes a precise timestamp, the fuel consumption prediction value given by the main task, the complete abnormal state probability distribution given by the auxiliary task, and the most likely flight state label (i.e., the category with the highest probability) determined based on this probability distribution. This information can be pushed to the airline's flight operation management system or fuel management system in real time. For example, the fuel management system can use continuous fuel consumption prediction values ​​to compare with the fuel consumption curve in the flight plan in real time, achieving accurate monitoring of flight fuel status and immediate early warning of abnormal consumption; while when the probability of an abnormal state output by the auxiliary task branch (such as "fuel leak") continuously exceeds the warning threshold, the system can automatically generate alarm information, providing critical data support for flight safety and operational decisions. For batch prediction of historical flight data, the system generates a detailed analysis report covering the entire flight mission, including a comparison of second-by-second prediction values ​​and actual values, providing a powerful and reliable technical tool for airline fleet performance evaluation, fuel-saving strategy optimization, and pilot behavior analysis.

[0173] Step S70: Continuously monitor the prediction results. When a drift in prediction performance is detected and exceeds the preset conditions, automatically trigger model rollback or feature reconstruction operations to restore prediction performance.

[0174] In this embodiment, the prediction results are continuously monitored, including:

[0175] Step S71: Construct a three-dimensional monitoring system that includes core indicators of the main task, related indicators of auxiliary tasks, and pre-input data quality indicators, and set the sliding statistical window and performance threshold for each indicator.

[0176] Step S72: Using a combination of timed triggering and threshold triggering, the various monitoring indicators within the sliding window are calculated and updated in real time.

[0177] Step S73: When the monitoring indicator is detected to be not in line with the preset threshold, the performance drift type is determined to be data-driven drift or model aging drift by combining the status of the preceding indicators and the trend of indicator changes, and the severity of the drift is quantified.

[0178] Automatically trigger model rollback or feature reconstruction operations, including:

[0179] Step S74: Establish a historical model version management system to standardize the storage and recording of model versions that meet the preservation conditions;

[0180] Step S75: When it is determined that the model is aging drift and the preset rollback trigger condition is met, candidate versions with better performance than the current model are selected from the historical model versions, and the rollback version is selected according to the principle of optimal performance.

[0181] Step S76: Load the selected rollback version to replace the current online model, and use a preset number of normal flight data to verify the rollback model. After successful verification, resume the prediction process.

[0182] Step S77: When the model verification fails or performance drift is frequently triggered after rollback, it is determined that the existing feature system is invalid, triggering feature reconstruction operation, adding or adjusting feature dimensions and re-executing feature construction operation until the core monitoring indicators of the main task of the reconstructed model meet the preset requirements.

[0183] After the multi-task prediction model is deployed for long-term stable engineering applications via step S60, to address the performance degradation issues caused by changes in operating conditions, data distribution evolution, or model aging in existing technologies, this invention further implements a crucial step S70: a closed-loop, automated prediction performance monitoring and model self-diagnosis and self-healing mechanism. This step ensures that the fuel consumption prediction system maintains high accuracy and reliability throughout its entire lifecycle without manual intervention. First, in step S71, the system constructs a comprehensive and three-dimensional real-time performance monitoring system. This system abandons single-dimensional error monitoring and instead systematically evaluates the model's health status from three levels: firstly, focusing on the core indicators of the main task of fuel consumption prediction, specifically selecting the coefficient of determination. Mean Absolute Percentage Error (MAPE) and Maximum Absolute Error Three items, and set strict performance baseline thresholds for them (e.g., It should be no less than 0.85, and MAPE should be no more than 5%. (Not exceeding 1.0 kg / s); secondly, focus on the correlation indicators of auxiliary tasks, including the accuracy of abnormal state identification (Acc) and the anomaly-consistency coefficient. Its threshold (e.g., Acc ≥ 0.85), The first measure is used to indirectly verify the effectiveness of the shared feature extraction layer; the second measure is to monitor the quality of input data in advance, such as the feature missing rate and the abnormal rate of atmospheric parameter calculation. The thresholds (e.g., ≤1% and ≤0.5%) are designed to isolate performance fluctuations caused by data source problems as soon as possible. To ensure statistical effectiveness, all indicators are calculated over a configurable sliding time window. For example, the window for core and related indicators can be set to "the last 100 flights" or "the last 72 hours," while data quality indicators use a shorter "the last 24 hours" window for faster response.

[0184] Based on this monitoring system, step S72 employs a dual-mode driving mechanism combining "timed triggering and threshold triggering" to perform real-time calculations and updates. After processing data for each flight, the system periodically triggers a routine monitoring of all indicators, updating the statistical values ​​within the sliding window and comparing them with the thresholds. Simultaneously, the system also establishes highly sensitive threshold triggers; for example, if the absolute error of fuel consumption prediction in any frame of data instantaneously exceeds a more stringent emergency threshold (e.g., ...), ... When the threshold is 1.5 times (i.e., 1.5 kg / s), a special diagnosis is immediately triggered to capture sudden data anomalies or model failures. Once any monitoring activity detects that an indicator does not meet the preset threshold, the system proceeds to step S73 to perform precise performance drift diagnosis and classification. This diagnostic logic first checks the input data quality pre-indicators. If the indicator fails to meet the standard, the performance degradation is determined to be "data-driven drift," and the data quality tracing process is initiated; if the data quality is not a problem, it is determined to be "model aging drift." Based on this, the system combines the number and duration of the non-compliant indicators to quantify the degree of drift as "slight drift" (e.g., only one core indicator is slightly below the threshold and lasts for two statistical windows) or "severe drift" (e.g., multiple core or related indicators are significantly below the threshold). To avoid misjudgments caused by random data fluctuations, the system further introduces trend verification, that is, calculating the slope of the key indicator's change in the most recent statistical windows. Only when the indicator shows a continuous and significant downward trend (e.g., the slope is less than -0.02) will the drift be finally confirmed.

[0185] After a model is diagnosed with aging-related drift, the system automatically activates the model's self-healing procedure. This procedure is based on the standardized historical model version management system established in step S74. This system automatically archives model versions according to preset rules (such as monthly regular saving, saving during performance optimization, and emergency saving before self-healing operations). Each version fully includes the model weight file (e.g., .pth format), structure configuration file (e.g., .yaml format), performance baseline metrics, and training data range, and adopts a hierarchical strategy of "local hot storage + cloud cold storage" to balance rollback speed and storage cost. Next, in step S75, the system initiates model rollback based on preset rollback trigger conditions (e.g., severe drift occurs, or slight drift persists for more than three statistical windows without self-recovery). It first filters from the numerous historical versions stored locally that meet the baseline requirements and are superior to the current faulty model. Then, based on the principle of "optimal performance first, recent time second," it automatically selects the best version as the primary rollback target.

[0186] Step S76 is the specific execution stage of the model rollback. The system first saves the complete state of the current faulty model for post-event analysis. Then, it loads the weights and configuration of the selected rollback version from the repository to replace the currently running model. After replacement, service is not immediately restored. Instead, a small batch of recent, verified normal flight data (e.g., 5 flights) is used to quickly verify and infer the rolled-back model. Only when the verification results show that its core monitoring indicators (such as R² and MAPE) meet the preset thresholds is the system confirmed as a successful rollback and seamlessly resumes its normal prediction service process. After the rollback, the system also incorporates the new data accumulated during this event into subsequent incremental training plans, aiming to train a new version with better performance based on the rollback.

[0187] Finally, to address extreme situations, this invention also includes a final self-healing mechanism: feature reconstruction in step S77. When the model rollback operation fails to validate after trying all available versions, or when the system frequently triggers severe drift within a short period (e.g., within three consecutive months), the system determines that this is not simply model parameter aging, but rather that the underlying feature system can no longer effectively represent the current flight conditions. At this point, the system triggers a feature reconstruction alarm and automatically executes a preset feature reconstruction process. This process returns to step S30, where, based on the original feature construction framework, new original parameters (e.g., engine exhaust temperature) are introduced, or the combination and weighting of existing features are adjusted to generate a completely new feature system. Subsequently, the system uses recent flight data (e.g., 100 flights) to rigorously test the model trained on the new features. Only when the core monitoring indicators of the new model reach or exceed the initial optimal performance standards (e.g., ...) will the system be able to successfully reconstruct the model. Only when the system reaches a certain threshold will it be approved for deployment, thus completing a deep self-healing process from the root cause of the feature, ensuring the ultimate adaptability of the entire prediction system in the face of unknown changes.

[0188] See Figure 2 The second embodiment of the present invention provides an adaptive fuel consumption prediction system that integrates physical models and multi-task learning, used to implement an adaptive fuel consumption prediction method that integrates physical models and multi-task learning. The system includes:

[0189] The data acquisition and processing module is configured to acquire and preprocess the aircraft's fast access recorder data to obtain preprocessed flight data.

[0190] The full-dimensional data analysis and calculation module is configured to analyze the flight phase based on the preprocessed flight data, calculate the corresponding atmospheric environmental parameters, and associate the analysis and calculation results with the flight data to obtain full-dimensional data containing flight status and environmental parameters.

[0191] The input feature parameter acquisition module is configured to construct features from the full-dimensional data based on the aircraft aerodynamic model and engine thrust model, generating aerodynamic force features and thrust features that reflect flight conditions, which are used as input feature parameters.

[0192] The model building module is configured to build a multi-task prediction model, including a shared feature extraction layer, a main task branch for fuel consumption prediction, and an auxiliary task branch for abnormal state identification. The shared feature extraction layer performs deep extraction on the input feature parameters to generate a shared feature representation. The main task branch outputs the predicted fuel consumption value based on the shared feature representation. The auxiliary task branch outputs the abnormal state identification result based on the shared feature representation.

[0193] The training module is configured to apply a reverse constraint to the shared feature extraction layer through the abnormal state recognition auxiliary task branch during the training of the multi-task prediction model.

[0194] The prediction module is configured to use the trained model to predict the fuel consumption of the target flight mission and output the prediction results.

[0195] The monitoring module is configured to continuously monitor the prediction results. When a drift in prediction performance is detected and exceeds preset conditions, it automatically triggers model rollback or feature reconstruction operations to restore prediction performance. Those skilled in the art will understand that, for ease of description and brevity, the specific working process and related explanations of the method described above can be found in the corresponding processes of the foregoing system embodiments, and will not be repeated here.

[0196] It should be noted that the above embodiment of the adaptive prediction system for aviation fuel consumption integrating physical models and multi-task learning 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.

[0197] A device according to a third embodiment of the present invention includes:

[0198] At least one processor;

[0199] and a memory communicatively connected to at least one of the processors;

[0200] The memory stores instructions that can be executed by the processor to implement the aforementioned adaptive prediction method for fuel consumption that integrates a physical model and multi-task learning.

[0201] 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 adaptive prediction method for fuel consumption that integrates a physical model and multi-task learning.

[0202] 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.

[0203] 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.

[0204] 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.

[0205] 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.

[0206] 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.

[0207] 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).

[0208] 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, can 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.

[0209] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0210] 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.

[0211] 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. An adaptive prediction method for aviation fuel consumption that integrates physical models and multi-task learning, characterized in that, Includes the following steps: Acquire and preprocess the aircraft's fast access recorder data to obtain preprocessed flight data; Based on the preprocessed flight data, the flight phase is analyzed and the corresponding atmospheric environmental parameters are calculated. The analysis and calculation results are then correlated with the flight data to obtain full-dimensional data containing flight status and environmental parameters. Based on the aircraft aerodynamic model and engine thrust model, feature construction is performed on the full-dimensional data to generate aerodynamic force characteristics and thrust characteristics that reflect flight conditions, which are used as input feature parameters. A multi-task prediction model is constructed, including a shared feature extraction layer, a main task branch for fuel consumption prediction, and an auxiliary task branch for abnormal state identification. The shared feature extraction layer performs deep extraction on the input feature parameters to generate a shared feature representation. The main task branch outputs the predicted fuel consumption value based on the shared feature representation. The auxiliary task branch outputs the abnormal state identification result based on the shared feature representation. Among them, a reverse constraint interaction layer is constructed, which weights and fuses the abnormal state identification results output by the auxiliary task branch with the intermediate layer features of the main task branch, and converts the prediction error of the main task branch into a penalty term to adjust the feature weights of the auxiliary task branch in reverse, thereby realizing bidirectional reverse constraint between the main task and the auxiliary task. A composite loss function is constructed, which includes the main task loss, the auxiliary task loss, and the cooperative constraint loss. The cooperative constraint loss is used to ensure that the predicted value of aviation fuel consumption under abnormal conditions and the predicted value under normal conditions meet the preset physical consistency relationship. Backpropagation of gradients is performed based on the composite loss function, so that the parameters of the shared feature extraction layer are optimized by the losses of the main task and the auxiliary task simultaneously, and the main task and the auxiliary task are mutually constrained by the gradient chain rule. During the training of the multi-task prediction model, an inverse constraint is applied to the shared feature extraction layer through an auxiliary task branch; The trained model is used to predict the fuel consumption of the target flight mission, and the prediction results are output. The prediction results are continuously monitored. When a drift in prediction performance is detected and exceeds preset conditions, the model rollback or feature reconstruction operation is automatically triggered to restore prediction performance.

2. The adaptive prediction method for aviation fuel consumption integrating physical models and multi-task learning according to claim 1, characterized in that, Acquire and preprocess the aircraft's fast access recorder data to obtain preprocessed flight data, including: The fast access recorder data is cleaned by combining physical rule thresholding and statistical model anomaly detection methods to remove outliers. The cleaned data is time-aligned based on the timestamp. For randomly missing data and consecutively missing data, interpolation and imputation by the mean of similar data are used to handle missing values, respectively. The data, after time alignment and missing value processing, is standardized to eliminate the dimensional differences between different parameters, resulting in standardized and time-consistent preprocessed flight data.

3. The adaptive prediction method for aviation fuel consumption integrating physical models and multi-task learning according to claim 1, characterized in that, Based on the preprocessed flight data, the flight phase is analyzed, and the corresponding atmospheric environmental parameters are calculated, including: Using preprocessed flight data such as flight altitude, airspeed, landing gear status, engine speed, and throttle position as input, and employing threshold judgment combined with time sequence continuity verification logic, the flight process is divided into multiple flight phases including takeoff, climb, cruise, descent, approach, and landing. Based on the international standard atmospheric model, using flight altitude as input, and according to the different atmospheric layers at the flight altitude, the corresponding temperature, pressure and density function relationships are used to calculate the atmospheric environmental parameters at the corresponding altitude. The analyzed flight phase information and calculated atmospheric environmental parameters are correlated and fused with the preprocessed flight data according to the timestamp to obtain full-dimensional data containing flight status and environmental parameters.

4. The adaptive prediction method for aviation fuel consumption integrating physical models and multi-task learning according to claim 1, characterized in that, Generate aerodynamic force and thrust characteristics that reflect flight conditions, including: Based on the aircraft aerodynamic model, combined with the atmospheric density, vacuum velocity and reference area in the full-dimensional data, the lift characteristics are calculated by the lift coefficient function related to the angle of attack and Mach number; the drag characteristics are calculated by the drag coefficient function combining the parasitic drag coefficient and the induced drag factor. Based on the engine thrust model, combined with the engine rotor speed, atmospheric pressure and atmospheric temperature in the full-dimensional data, the net thrust characteristics of the engine are calculated through the functional relationship between the engine characteristic coefficient and the environmental attenuation coefficient. Based on the lift characteristics, drag characteristics and engine net thrust characteristics, derivative operating condition characteristics including lift-to-drag ratio, thrust-to-drag difference, aerodynamic force change rate and operating condition adaptability deviation are constructed. The constructed features are verified by physical laws and filtered for data discrimination to obtain the final core feature set, which serves as the input feature parameters for the model.

5. The adaptive prediction method for aviation fuel consumption integrating physical models and multi-task learning according to claim 1, characterized in that, Constructing a multi-task prediction model includes: A shared feature extraction layer is constructed, which combines a fully connected layer and an attention mechanism, to encode the input feature parameters of the multi-task prediction model and output a shared feature representation. A main task branch for predicting aviation fuel consumption is constructed. The main task branch is a multi-layer fully connected regression network that takes the shared feature representation as input and outputs continuous predicted values ​​of aviation fuel consumption. An auxiliary task branch for abnormal state recognition is constructed. The auxiliary task branch is a classification network containing a fully connected layer and a classification function. The shared feature representation is used as input, and the probability distribution of different abnormal state categories is output.

6. The adaptive prediction method for aviation fuel consumption integrating physical models and multi-task learning according to claim 1 or 5, characterized in that, Applying inverse constraints to the shared feature extraction layer through an auxiliary task branch includes: During the forward inference process of model training, both the predicted fuel consumption value and the abnormal state identification result are obtained simultaneously. During training, the weights of each loss term in the composite loss function are dynamically adjusted based on changes in the prediction accuracy of the main task and the recognition accuracy of the auxiliary task.

7. The adaptive prediction method for aviation fuel consumption integrating physical models and multi-task learning according to claim 1, characterized in that, Continuous monitoring of the prediction results includes: Construct a three-dimensional monitoring system that includes core indicators of the main task, related indicators of auxiliary tasks, and pre-input data quality indicators, and set sliding statistical windows and performance thresholds for each indicator; A combination of timed triggering and threshold triggering is used to calculate and update various monitoring indicators within the sliding window in real time; When a monitoring indicator is detected to be not in line with a preset threshold, the system combines the status of the preceding indicators and the trend of indicator changes to determine whether the performance drift is data-driven drift or model aging drift, and quantifies the severity of the drift.

8. The adaptive prediction method for aviation fuel consumption integrating physical models and multi-task learning according to claim 1 or 7, characterized in that, Automatically trigger model rollback or feature reconstruction operations, including: Establish a historical model version management system to standardize the storage and recording of model versions that meet the preservation conditions; When a model is determined to be aging drift and the preset rollback trigger conditions are met, candidate versions with better performance than the current model are selected from historical model versions, and the rollback version is selected according to the principle of optimal performance. Load the selected rollback version to replace the current online model, and validate the rollback model using a preset number of normal flight data. Once the validation is successful, resume the prediction process. When the model fails validation or performance drift is frequently triggered after rollback, it is determined that the existing feature system is invalid, triggering a feature reconstruction operation. Feature dimensions are added or adjusted and the feature construction operation is re-executed until the core monitoring indicators of the main task of the reconstructed model meet the preset requirements.

9. A fuel consumption adaptive prediction system integrating physical models and multi-task learning, used to implement the fuel consumption adaptive prediction method integrating physical models and multi-task learning as described in any one of claims 1-8, characterized in that, The system includes: The data acquisition and processing module is configured to acquire and preprocess the aircraft's fast access recorder data to obtain preprocessed flight data. The full-dimensional data analysis and calculation module is configured to analyze the flight phase based on the preprocessed flight data, calculate the corresponding atmospheric environmental parameters, and associate the analysis and calculation results with the flight data to obtain full-dimensional data containing flight status and environmental parameters. The input feature parameter acquisition module is configured to construct features from the full-dimensional data based on the aircraft aerodynamic model and engine thrust model, generating aerodynamic force features and thrust features that reflect flight conditions, which are used as input feature parameters. The model building module is configured to build a multi-task prediction model, including a shared feature extraction layer, a main task branch for fuel consumption prediction, and an auxiliary task branch for abnormal state identification. The shared feature extraction layer performs deep extraction on the input feature parameters to generate a shared feature representation. The main task branch outputs the predicted fuel consumption value based on the shared feature representation. The auxiliary task branch outputs the abnormal state identification result based on the shared feature representation. The training module is configured to apply a reverse constraint to the shared feature extraction layer through an auxiliary task branch during the training of the multi-task prediction model. The prediction module is configured to use the trained model to predict the fuel consumption of the target flight mission and output the prediction results. The monitoring module is configured to continuously monitor the prediction results. When a drift in prediction performance is detected and exceeds preset conditions, it will automatically trigger model rollback or feature reconstruction operations to restore prediction performance.

10. A device, characterized in that, include: 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 adaptive prediction method for fuel consumption that integrates physical models and multi-task learning as described in any one of claims 1-8.

Images