CN-121980227-A - Self-adaptive prediction method, system and equipment for aviation oil consumption by integrating physical model and multitask learning

CN121980227ACN 121980227 ACN121980227 ACN 121980227ACN-121980227-A

Abstract

The invention belongs to the field of aviation data analysis, and relates to an aviation oil consumption self-adaptive prediction method, an aviation oil consumption self-adaptive prediction system and aviation oil consumption self-adaptive prediction equipment integrating a physical model and multi-task learning, so as to solve the problems of sensitivity to abnormal data, insufficient robustness and easiness in long-term running performance degradation. The method comprises the steps of obtaining and preprocessing data of a quick access recorder of an aircraft, analyzing a flight stage, calculating atmospheric parameters, constructing full-dimensional data, constructing an input characteristic reflecting real flight working conditions, constructing a multi-task prediction model comprising a main task for predicting the consumption of aviation oil and an auxiliary task for identifying abnormal states, sharing a characteristic extraction layer by the multi-task prediction model and the multi-task prediction model, reducing the influence of the abnormal data on the main task by utilizing an auxiliary task identification result through a reverse constraint mechanism, continuously monitoring the prediction result, and automatically triggering model rollback or characteristic reconstruction when performance drift is detected. The method can remarkably improve the stability, reliability and physical consistency of the prediction of the aviation oil consumption, and ensure the long-term prediction precision under complex working conditions.

Inventors

YANG SHI
LIU WEIREN
LI JIAXIN
YANG LEI
WANG ZHONGXING
SU JIANFEI
ZENG RUIQI
ZHANG ZIHONG

Assignees

南航科技(广东横琴)有限公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (10)

1. The self-adaptive prediction method for the aviation oil consumption by integrating the physical model and the multi-task learning is characterized by comprising the following steps of: Acquiring data of a quick access recorder of the aircraft and preprocessing the data to obtain preprocessed flight data; analyzing a flight phase based on the preprocessed flight data, calculating corresponding atmospheric environment parameters, and associating analysis and calculation results to the flight data to obtain full-dimension data containing flight states and environment parameters; based on an aircraft aerodynamic model and an engine thrust model, carrying out feature construction on the full-dimensional data to generate aerodynamic stress features and thrust features reflecting flight conditions as input feature parameters; The method comprises the steps of constructing a multi-task prediction model, wherein the multi-task prediction model comprises a shared feature extraction layer, an aviation oil consumption prediction main task branch and an abnormal state identification auxiliary task branch, the shared feature extraction layer is used for carrying out deep extraction on input feature parameters to generate shared feature representation, the main task branch is used for representing and outputting an aviation oil consumption prediction value based on the shared feature, and the auxiliary task branch is used for representing and outputting an abnormal state identification result based on the shared feature; In the multi-task prediction model training process, applying reverse constraint to the shared feature extraction layer through auxiliary task branches; predicting the aviation oil consumption of the target flight mission by using the trained model, and outputting a prediction result; And continuously monitoring the prediction result, and automatically triggering model rollback or feature reconstruction operation when the prediction performance is detected to drift and exceed a preset condition so as to recover the prediction performance.
2. The method for adaptively predicting the consumption of aviation oil by combining a physical model and multitask learning according to claim 1, wherein the steps of obtaining data of a fast access recorder of an aircraft and preprocessing the data to obtain preprocessed flight data comprise: carrying out data cleaning on the data of the quick access recorder by combining a physical rule threshold method and a statistical model anomaly detection method, and removing anomaly values; Time alignment is carried out on the cleaned data by taking the time stamp as a reference; for random missing data and continuous missing data, respectively adopting an interpolation method and a similar data average filling method to carry out missing value processing; And carrying out standardization processing on the data subjected to time alignment and missing value processing, and eliminating dimension differences among different parameters to obtain standardized and time sequence-consistent preprocessed flight data.
3. The method for adaptively predicting the consumption of aviation oil by combining a physical model and multi-task learning according to claim 1, wherein analyzing the flight phase based on the preprocessed flight data and calculating the corresponding atmospheric environment parameters comprises: Taking the flight height, airspeed, landing gear state, engine rotating speed and throttle lever position in the preprocessed flight data as inputs, adopting a logic of combining threshold judgment with time sequence continuity verification, and dividing the flight process into a plurality of flight phases including take-off, climbing, cruising, descending, approaching and landing; based on an international standard atmosphere model, taking flying height as input, and calculating to obtain atmospheric environment parameters of corresponding heights by adopting corresponding temperature, pressure and density function relations according to different atmospheres where the flying height is located; and carrying out association fusion on the flight phase information obtained by analysis and the atmospheric environment parameters obtained by calculation and the preprocessed flight data according to the time stamp to obtain full-dimension data containing the flight state and the environment parameters.
4. The method for adaptively predicting the consumption of aviation oil by combining a physical model and multitask learning according to claim 1, wherein generating aerodynamic stress characteristics and thrust characteristics reflecting the flight conditions comprises: Based on an aircraft aerodynamic model, combining the air density, the vacuum velocity and the reference area in the full-dimensional data, calculating lift characteristics through lift coefficient functions related to attack angles and Mach numbers; based on an engine thrust model, combining the engine rotor rotating speed, the atmospheric pressure and the atmospheric temperature in the full-dimensional data, and calculating the engine net thrust characteristic through the functional relation between the engine characteristic coefficient and the environment attenuation coefficient; on the basis of the lift force characteristic, the resistance characteristic and the engine net thrust characteristic, derivative working condition characteristics including lift-drag ratio, thrust resistance difference, aerodynamic stress change rate and working condition adaptability deviation are constructed; And carrying out physical rule verification and data differentiation screening on the constructed features to obtain a final core feature set which is used as a model input feature parameter.
5. The method for adaptively predicting the consumption of aviation oil by fusing a physical model and multitasking learning of claim 1, wherein constructing the multitasking prediction model comprises: constructing a shared feature extraction layer comprising a full connection layer and attention mechanism combined structure, encoding input feature parameters of the multi-task prediction model, and outputting a shared feature representation; Constructing a main task branch of the aviation oil consumption prediction, wherein the main task branch is a multi-layer full-connection regression network, and the sharing characteristic is used as input to output a continuous prediction value of the aviation oil consumption; Constructing an abnormal state identification auxiliary task branch, wherein the auxiliary task branch is a classification network comprising a full connection layer and a classification function, and the shared characteristic is used as input to output probability distribution of different abnormal state categories; and constructing a reverse constraint interaction layer, carrying out weighted fusion on the abnormal state identification result output by the auxiliary task branch and the middle layer characteristic of the main task branch, converting the prediction error of the main task branch into a punishment item, and reversely adjusting the characteristic weight of the auxiliary task branch to realize bidirectional reverse constraint between the main task and the auxiliary task.
6. An adaptive prediction method for fuel consumption in combination with a physical model and multitasking learning as claimed in claim 1 or 5, wherein applying a reverse constraint to the shared feature extraction layer by assisting task branches comprises: in the forward reasoning process of model training, simultaneously obtaining a predicted value of aviation oil consumption and an abnormal state identification result; Constructing a composite loss function comprising main task loss, auxiliary task loss and cooperative constraint loss, wherein the cooperative constraint loss is used for ensuring that a predicted value of aviation oil consumption in an abnormal state and a predicted value in a normal state meet a preset physical consistency relation; Performing gradient counter propagation based on the composite loss function, so that the parameters of the shared feature extraction layer are simultaneously optimized by the loss of the main task and the auxiliary task, and the main task and the auxiliary task are mutually constrained by a gradient chain rule; and dynamically adjusting the weight of each loss term in the composite loss function according to the change of the prediction precision of the main task and the recognition accuracy of the auxiliary task in the training process.
7. The method for adaptively predicting the consumption of aviation oil by combining a physical model and multitask learning according to claim 1, wherein the method for continuously monitoring the predicted result comprises the following steps: Constructing a three-dimensional monitoring system comprising a main task core index, an auxiliary task associated index and an input data quality preposed index, and setting a sliding statistical window and a performance threshold of each index; calculating and updating each monitoring index in the sliding window in real time by adopting a mode of combining timing trigger and threshold trigger; when the monitoring index is detected not to meet the preset threshold, the type of the performance drift is judged as data-driven drift or model aging drift by combining the state of the front index and the change trend of the index, and the severity of the drift is quantified.
8. The method for adaptively predicting the consumption of aviation oil by combining a physical model and multi-task learning according to claim 1 or 7, wherein the automatic triggering of the model rollback or the feature reconstruction operation comprises the following steps: establishing a history model version management system, and carrying out normalized storage and recording on model versions meeting storage conditions; When the model aging drift is judged and a preset rollback triggering condition is reached, candidate versions with the performance superior to that of the current model are screened out from the historical model versions, and the rollback versions are selected according to the principle of optimal performance; Loading the selected rollback version to replace the current online model, verifying the model after rollback by using the preset number of normal flight data, and recovering the prediction flow after verification; And when the model verification after rollback is not passed or the performance drift is frequently triggered, judging that the existing characteristic system fails, triggering characteristic reconstruction operation, adding or adjusting characteristic dimension and re-executing characteristic construction operation until the main task core monitoring index of the model after reconstruction meets the preset requirement.
9. An adaptive prediction system for aviation oil consumption by combining a physical model and multi-task learning, for implementing the adaptive prediction method for aviation oil consumption by combining a physical model and multi-task learning according to any one of claims 1 to 8, characterized in that the system comprises: the data acquisition and processing module is configured to acquire the data of the quick access recorder of the aircraft and perform preprocessing to obtain preprocessed flight data; the full-dimensional data analysis and calculation module is configured to analyze a flight phase based on the preprocessed flight data, calculate corresponding atmospheric environment parameters, and correlate analysis and calculation results into the flight data to obtain full-dimensional data comprising flight states and environment parameters; the input characteristic parameter acquisition module is configured to perform characteristic construction on the full-dimension data based on an aircraft aerodynamic model and an engine thrust model, and generate aerodynamic stress characteristics and thrust characteristics reflecting flight conditions as input characteristic parameters; The model construction module is configured to construct a multi-task prediction model and comprises a shared feature extraction layer, an aviation oil consumption prediction main task branch and an abnormal state identification auxiliary task branch, wherein the shared feature extraction layer is used for carrying out deep extraction on input feature parameters to generate shared feature representation; a training module configured to impose a reverse constraint on the shared feature extraction layer through auxiliary task branches during the multi-task prediction model training process; The prediction module is configured to predict the aviation oil consumption of the target flight mission by utilizing the trained model and output a prediction result; And the monitoring module is configured to continuously monitor the prediction result, and automatically trigger model rollback or feature reconstruction operation to restore the prediction performance when the prediction performance is detected to drift and exceed a preset condition.
10. An apparatus, comprising: At least one processor; and a memory communicatively coupled to at least one of the processors; Wherein the memory stores instructions executable by the processor for execution by the processor to implement an adaptive prediction of fuel consumption by fusion of a physical model with multi-task learning as claimed in any one of claims 1 to 8.

Description

Self-adaptive prediction method, system and equipment for aviation oil consumption by integrating physical model and multitask learning Technical Field The invention relates to the technical field of aviation data analysis, in particular to an aviation oil consumption self-adaptive prediction method, an aviation oil consumption self-adaptive prediction system and aviation oil consumption self-adaptive prediction equipment integrating a physical model and multi-task learning. Background The accurate prediction of aviation fuel consumption is a key link for improving the operation efficiency of an airline company, optimizing a fuel management strategy and guaranteeing the flight safety. In the prior art, a prediction model is usually built in a data-driven manner by using flight data collected by a Quick Access Recorder (QAR) installed on an aircraft, so as to realize estimation of fuel consumption of different flight phases or the whole flight segment. At present, the mainstream aviation fuel consumption prediction method mainly relies on time integration and statistical analysis of fuel flow parameters in QAR data, or combines an aircraft aerodynamic model to construct a functional relationship between fuel consumption and flight attitude and environmental parameters. These methods provide a certain technical basis for aircraft fuel management. However, in practical engineering applications, the above prior art solutions still have the following drawbacks to be solved: first, it is highly sensitive to QAR anomaly data and the predicted outcome is not stable enough. In actual flight, QAR data often introduces outliers due to sensor jitter, communication delay, or environmental abrupt changes. In the prior art, an independent preprocessing step is generally adopted to remove the abnormality, but when abnormal data is not completely identified, a subsequent prediction model can be directly influenced by error data, so that a prediction result has larger deviation and poor stability. Secondly, outlier processing and a prediction model are mutually independent and lack a cooperative mechanism. The prior art generally adopts a serial structure of firstly processing the abnormality and then carrying out prediction, which causes the lack of effective information interaction between the abnormality identification process and the prediction model. The prediction model does not have the capability of actively suppressing the influence of potential abnormal data, and has insufficient robustness when facing complex flight conditions or data quality fluctuation. Again, depending on fixed physical models or static parameters too much, the model is not sufficiently adaptable. The method based on the physical model depends on preset aerodynamic or thrust model parameters, and the parameters are difficult to reflect the flight state changes caused by factors such as load changes, meteorological conditions, engine performance attenuation and the like in real time. Therefore, when the actual operating conditions deviate from the model assumptions, systematic deviations in the predicted results are likely to occur. Finally, the ability to self-diagnose and automatically correct the performance of the model is lacking. The performance of the existing predictive model gradually decreases over time due to data distribution changes or error accumulation, i.e., model drift, after one-time training is completed. In the prior art, manual retraining or parameter adjustment of a model is generally carried out by relying on manual experience, and a real-time monitoring and automatic correction mechanism for model performance degradation is lacked, so that the long-term reliability and engineering application value of a prediction system are affected. Therefore, how to effectively suppress abnormal data interference, enhance model robustness and realize long-term self-maintenance of model performance is a technical problem to be solved in the current field of aviation oil consumption prediction. Disclosure of Invention In order to solve the problems in the prior art, namely the problems of sensitivity to abnormal data, insufficient robustness and easy degradation of long-term operation performance in the prior art, the invention provides a self-adaptive prediction method, a self-adaptive prediction system and self-adaptive prediction equipment for aviation oil consumption, which are used for fusing a physical model and multi-task learning. The invention provides a self-adaptive prediction method for aviation oil consumption by combining a physical model and multi-task learning, which comprises the following steps: Acquiring data of a quick access recorder of the aircraft and preprocessing the data to obtain preprocessed flight data; analyzing a flight phase based on the preprocessed flight data, calculating corresponding atmospheric environment parameters, and associating analysis and calculation results to the flight data to obtain full-