A method and apparatus for predicting aviation fuel performance based on few-shot learning and reinforcement
By employing few-shot learning and sample augmentation methods, a semi-supervised learning framework and multi-model stacking learning are constructed, which solves the accuracy and efficiency problems of fuel performance prediction under small-shot conditions, and achieves efficient fuel performance prediction and screening.
Patent Information
- Application Number
- CN202611242866.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing fuel performance prediction methods require a large amount of sample data, and traditional methods lack accuracy and generalization ability under small sample conditions, making it difficult to meet the needs of rapid development and screening of new aviation fuels.
We adopt a method based on few-shot learning and sample augmentation. By constructing a semi-supervised learning framework, we introduce unlabeled samples on the basis of a small number of labeled samples, use unlabeled samples to expand the training sample set, and combine multi-model stacking learning to establish a unified prediction model.
It achieves high-precision synchronous prediction of aviation fuel performance under small sample conditions, improves the prediction effect and efficiency of fuel performance, reduces experimental costs, and is suitable for the rapid development and screening of green and low-carbon fuels and new high-energy fuels.
Smart Images

Figure CN123115646A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of machine learning and energy fuel performance prediction technology, and in particular to a method and apparatus for predicting aviation fuel performance based on few-shot learning and reinforcement. Background Technology
[0002] With the accelerated development of low-carbon and carbon-reduction processes in the aviation industry, market demand for green, low-carbon fuels and high-energy-density fuels continues to grow. The research and development of green, low-carbon fuels and new high-energy fuels has become a key focus for the industry. To ensure the safe application of new aviation fuels in aero engines, their physicochemical properties must meet several key indicators specified in ASTM standards, including density, derived cetane number (DCN), distillation temperature, flash point, freezing point, calorific value, and viscosity. Traditional fuel performance evaluation methods mainly rely on experimental testing methods, such as gas chromatography (GC×GC), spectral analysis, and standard fuel testing procedures. These methods typically require large sample volumes, complex equipment, and long cycles, resulting in high evaluation costs and low efficiency, which is detrimental to the rapid development and screening of new fuels.
[0003] In recent years, machine learning methods have been increasingly applied to fuel property prediction, enabling rapid prediction of fuel performance by establishing a mapping relationship between fuel composition and properties. Currently available fuel performance prediction methods mainly fall into the following categories: (1) Method based on regression model of experimental data This type of method utilizes existing fuel sample data to establish the relationship between fuel composition and performance through linear regression, partial least squares regression, and other methods. However, due to limited training data, the model's generalization ability is poor, and the prediction accuracy drops significantly when the fuel composition exceeds the training range. (2) Prediction methods based on spectra or molecular structures Models can be built based on Fourier transform infrared spectroscopy (FTIR), nuclear magnetic resonance (NMR), or molecular structure information. However, these methods rely on specific testing techniques or molecular-level information, making it difficult to obtain complete structural information for complex, multi-component fuels, thus limiting their application. (3) Prediction methods based on traditional machine learning Random forests and support vector regression (SVR) are commonly used methods, but these methods usually require a large amount of labeled data for training. However, the amount of data that can be obtained simultaneously for fuel composition and physicochemical properties is extremely limited, resulting in insufficient model accuracy. Summary of the Invention
[0004] This specification provides an embodiment of a method and apparatus for predicting aviation fuel performance based on few-sample learning and reinforcement, in order to solve the technical problem of how to improve the effectiveness and efficiency of aviation fuel performance prediction.
[0005] To address the aforementioned technical problems, the embodiments in this specification provide the following technical solutions: This specification provides an embodiment of a method for predicting aviation fuel performance based on few-shot learning and sample augmentation. The method includes: An augmented sample set is obtained by performing one or more sample augmentation operations on a small-scale labeled sample set; the labeled sample is related to aviation fuel. The enhanced sample set is used to train the model to obtain the target model, and the target model is used to predict aviation fuel performance. Each sample augmentation operation includes: constructing an unlabeled sample set that is missing one or more attribute labels; For any missing single-class attribute labels, the attribute labels of that class in the unlabeled sample set are completed through one or more rounds of sample reconstruction operations. The data obtained after all the missing attribute labels are filled in is used as the new labeled sample; Specifically, for the missing single-class attribute labels, the single-round sample reconstruction operation includes: Several basic regression models are trained using a newly constructed training set; the training set includes samples that do not currently lack labels for that type of attribute. For each unlabeled sample lacking the attribute label, the trained basic regression models are used to predict the attribute label of each unlabeled sample, and the prediction results of each basic regression model for the attribute label of each unlabeled sample are evaluated to see if they meet the preset conditions. Complete the attribute labels for unlabeled samples whose prediction results meet the preset conditions.
[0006] Preferably, constructing an unlabeled sample set lacking one or more attribute labels includes: Randomly perturb the features of the existing labeled samples other than one or more attribute labels to obtain an unlabeled sample set that does not contain the one or more attribute labels.
[0007] As a preferred option, the evaluation index used to assess whether the prediction results of each basic regression model for the attribute labels of each unlabeled sample meet the preset conditions includes the dispersion of the prediction results.
[0008] Preferably, for any unlabeled sample, if the dispersion of the prediction result of the attribute label of that type for that unlabeled sample is located in a specific distribution area of the dispersion of the prediction results of the attribute label of that type for each unlabeled sample, then the prediction result of the attribute label of that type for that unlabeled sample satisfies the condition.
[0009] Preferably, for any unlabeled sample, the difference between the maximum and minimum values of the predicted results of the attribute label of that unlabeled sample is taken as the dispersion of the predicted results of the attribute label of that unlabeled sample.
[0010] Preferably, the completion of the attribute labels for unlabeled samples whose prediction results meet preset conditions includes: For any unlabeled sample whose prediction results meet the preset conditions, the mean of the prediction results for that unlabeled sample is used as the completion result of the attribute label for that unlabeled sample.
[0011] Preferably, the enhanced sample set is used for model training to obtain the target model, including: The enhanced sample set was used to train each basic regression model; Each trained model is used to make predictions on the test set, and the prediction results of each model on the test set are evaluated. The target model is determined based on the evaluation results of each model's predictions on the test set.
[0012] Preferably, the evaluation metrics used to assess the prediction results of each model on the test set include the coefficient of determination and / or mean absolute error of the prediction results.
[0013] As a preferred option, the missing attribute tags include key physicochemical properties.
[0014] This specification provides an embodiment of an aviation fuel performance prediction device based on few-shot learning and sample augmentation. The device includes: The sample augmentation module is used to obtain an augmented sample set from a small number of labeled samples through one or more sample augmentation operations; the labeled samples are aviation fuel-related samples. The performance prediction module is used to train the model using the enhanced sample set to obtain a target model, and then use the target model to predict aviation fuel performance. Each sample augmentation operation includes: constructing an unlabeled sample set that is missing one or more attribute labels; For any missing single-class attribute labels, the attribute labels of that class in the unlabeled sample set are completed through one or more rounds of sample reconstruction operations. The data obtained after all the missing attribute labels are filled in is used as the new labeled sample; Specifically, for the missing single-class attribute labels, the single-round sample reconstruction operation includes: Several basic regression models are trained using a newly constructed training set; the training set includes samples that do not currently lack labels for that type of attribute. For each unlabeled sample lacking the attribute label, the trained basic regression models are used to predict the attribute label of each unlabeled sample, and the prediction results of each basic regression model for the attribute label of each unlabeled sample are evaluated to see if they meet the preset conditions. Complete the attribute labels for unlabeled samples whose prediction results meet the preset conditions.
[0015] The above-described at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: By constructing a semi-supervised learning framework, unlabeled samples are introduced on top of a small number of labeled samples, and the training sample set is expanded using unlabeled samples, thereby improving data utilization efficiency. Based on this, a unified prediction model is established by combining multi-model stacking learning techniques, achieving high-precision synchronous prediction of aviation fuel performance and effectively improving the prediction effect and efficiency of aviation fuel performance. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments of this specification or the prior art will be briefly described below. Obviously, the drawings used in some embodiments of this application are only described below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the multi-scheme semi-supervised regression process in the first embodiment of this specification. Detailed Implementation
[0018] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments involved in the specific implementation are only a part of the embodiments of this application, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments in the specific implementation without creative effort should fall within the protection scope of this application.
[0019] The first embodiment of this specification (hereinafter referred to as "Embodiment 1") provides a method for predicting aviation fuel performance based on few-shot learning and reinforcement. The executing entity of Embodiment 1 includes, but is not limited to, a terminal, a server, an operating system, or an application; that is, the executing entity can be diverse and can be set, used, or changed as needed. Alternatively, a third-party application can assist the executing entity in executing Embodiment 1. For example, a server can execute the method provided in Embodiment 1, and a corresponding application can be installed on a terminal (which may be held by a user). Data transmission can occur between the terminal or application and the server, thereby assisting the server in executing the method provided in Embodiment 1.
[0020] The aviation fuel performance prediction method based on few-shot learning and reinforcement provided in Example 1 includes: S101: An enhanced sample set is obtained by performing one or more sample enhancement operations on a small-scale labeled sample set; the labeled sample is aviation fuel-related sample. In Example 1, a labeled sample set was constructed. The characteristics of a labeled sample set include: 1. The labeled sample set contains labeled samples, which are samples whose attribute labels of a certain class or certain classes (hereinafter referred to as key attribute labels) are complete or have been assigned values.
[0021] 2. The labeled sample set is small in scale, that is, the labeled sample set contains a small number of labeled samples (for example, the number of labeled samples is less than a certain threshold, which can be set arbitrarily).
[0022] 3. The labeled samples are related to aviation fuel. The labeled samples have one or more features (such as aviation fuel composition features) and key attribute labels. The key attribute labels can include various physicochemical property labels of aviation fuel. These physicochemical property labels can be used to characterize the performance of aviation fuel.
[0023] 4. Labeled samples can be normalized samples to eliminate the impact of dimensional differences on model training.
[0024] refer to Figure 1 In Example 1, in order to train the target model, an augmented sample set is obtained by one or more sample augmentation operations based on a small number of labeled samples.
[0025] Each sample augmentation operation may include: S1011: Construct an unlabeled sample set (denoted as unlabeled sample set S) that is missing one or more attribute labels. The unlabeled sample set S contains unlabeled samples, which are samples with one or more attribute labels (the aforementioned key attribute labels) missing or unassigned. Unlabeled samples are also related to aviation fuel. They generally contain component information but lack key attribute labels, such as physicochemical properties. Unlabeled samples can be normalized to eliminate the impact of dimensional differences on model training. S1013: For the missing single-class attribute label, the attribute label of that class in the unlabeled sample set is completed through one or more rounds of sample reconstruction operation; If the unlabeled samples in the unlabeled sample set S are missing only one type of attribute label (Case 1), then the attribute label of that type of all unlabeled samples is completed through one or more sample reconstruction operations. If the unlabeled samples in the unlabeled sample set S are missing multiple types of attribute labels (Case 2), then the attribute labels of each type are completed sequentially. That is, after completing the missing attribute label of all unlabeled samples through one or more sample reconstruction operations, the missing attribute label of all unlabeled samples is completed through one or more sample reconstruction operations, until all missing attribute labels of each type are completed.
[0026] S1013: The data obtained after all missing attribute labels have been filled in is used as a new labeled sample.
[0027] Once all missing attribute labels for each unlabeled sample in the unlabeled sample set constructed by this sample augmentation operation are filled in, the resulting new sample becomes a new labeled sample and can be used for the next sample augmentation operation.
[0028] The following section further explains how to perform sample reconstruction for missing single-category attribute labels.
[0029] In Example 1, for the missing single-class attribute label, a single-round sample reconstruction operation may include: Several basic regression models are trained using a newly constructed training set; the training set includes samples that do not currently lack labels for that type of attribute. For each unlabeled sample lacking the attribute label, the trained basic regression models are used to predict the attribute label of each unlabeled sample, and the prediction results of each basic regression model for the attribute label of each unlabeled sample are evaluated to see if they meet the preset conditions. Complete the attribute labels for unlabeled samples whose prediction results meet the preset conditions.
[0030] The following example, using a certain type of attribute label (denoted as attribute A), illustrates how to perform a single-round sample reconstruction operation.
[0031] Model training Construct a training set containing samples without missing attribute A, therefore including the aforementioned labeled samples. Furthermore, the training set in the first round of sample reconstruction may only contain the aforementioned labeled samples.
[0032] It should be noted that during the model training process, a semi-supervised learning method is used to train multiple basic regression models in the stacked ensemble learning. These basic regression models include, but are not limited to, XGBoost (ExtremeGradient Boosting), Random Forest (RF), Support Vector Regression (SVR), and Gaussian Process Regression (GPR). In Example 1, the selection of basic regression models is not limited.
[0033] After the model is trained, it can be tested using a test set containing labeled samples (the test set can be built separately). Once the model meets the predetermined conditions, it can be used for subsequent attribute prediction.
[0034] Attribute prediction After training each basic regression model, the trained basic regression models are used to predict attribute A for each unlabeled sample that is missing attribute A. Each basic regression model will output a predicted value.
[0035] For example, if there are four basic regression models: XGBoost, RF, SVR, and GPR, then for any unlabeled sample, each of these four models will output a predicted value for attribute A of the unlabeled sample. Thus, four predicted values for attribute A of the unlabeled sample will be obtained. These four predicted values can be used as the prediction results of each basic regression model for attribute A of the unlabeled sample.
[0036] The prediction results of each basic regression model for attribute A of any unlabeled sample will be referred to as the prediction results of that unlabeled sample.
[0037] Predictive Evaluation Based on the obtained prediction results, the evaluation criteria for whether the prediction results of each unlabeled sample meet the preset conditions are assessed. Preferably, the evaluation index used to assess whether the prediction results of each unlabeled sample meet the preset conditions includes the dispersion of the prediction results. That is, for any unlabeled sample, the dispersion (denoted as RL) of the prediction results of that unlabeled sample can be used to assess whether the prediction results of that unlabeled sample meet the preset conditions.
[0038] For any unlabeled sample, the dispersion can be the difference between the maximum and minimum predicted values (referred to as the maximum and minimum values) among the predicted results for that unlabeled sample. For example, if there are four basic regression models: XGBoost, RF, SVR, and GPR, then four predicted values for attribute A of the unlabeled sample (from these four models respectively) will be obtained, denoted as y1, y2, y3, and y4. The dispersion (RL) of the predicted results for this unlabeled sample is represented as follows: RL=Y i max -Y i min ; i=1,2,3,….,4,Y i max Let Y be the maximum value among y1, y2, y3, and y4. i min It is the minimum value among y1, y2, y3, and y4.
[0039] The dispersion represents the maximum deviation between the prediction results of each basic regression model for attribute A of the unlabeled sample. The smaller the dispersion, the higher the consistency of the prediction of attribute A of the unlabeled sample among the basic regression models, and the higher the prediction credibility or confidence.
[0040] For any unlabeled sample, if the dispersion of the predicted result of the unlabeled sample lies within a specific distribution region of the dispersion of the predicted results of all unlabeled samples, then the predicted result of the unlabeled sample satisfies the condition. Specifically, assuming there are n unlabeled samples lacking the attribute label in this round of sample reconstruction, the predicted result of any unlabeled sample will have a dispersion calculated (i.e., the dispersion of the predicted result of each base regression model for attribute A of the unlabeled sample), thus generating n dispersions. If the dispersion of the predicted result of an unlabeled sample lies within the top 10% of these n dispersions (belonging to a specific distribution region), then the predicted result of the unlabeled sample is deemed to satisfy the condition.
[0041] Generally, this specific distribution area is a specific region defined by sorting the dispersion of the prediction results of each unlabeled sample. If the dispersion of the prediction results of the unlabeled sample is located within this specific distribution area, it means that the dispersion of the unlabeled sample is relatively small and the confidence level is relatively high.
[0042] Attribute completion For any unlabeled sample, if the prediction result of the unlabeled sample meets the condition, then attribute A of the unlabeled sample is completed.
[0043] Preferably, completing attribute A of the unlabeled sample may include: using the mean of the prediction results of the unlabeled sample as the completed result (or completed value; the completed attribute label is called a pseudo label) of attribute A of the unlabeled sample.
[0044] As explained above, this round of sample reconstruction operation can complete the missing attribute A in a portion of the unlabeled samples (i.e., unlabeled samples whose prediction results meet the conditions, which may be one or more). If there are still unlabeled samples missing attribute A, the next round of sample reconstruction operation is performed as above. Furthermore, in each round of sample reconstruction operation, the unlabeled samples whose attribute A has been completed in that round are added to the training set, thereby constructing a new training set for the next round of sample reconstruction operation.
[0045] It is evident that each round of sample reconstruction operation not only completes the attribute A of at least some unlabeled samples, but also updates the training set for the next round of sample reconstruction (provided that the next round of sample reconstruction is required). In this way, after one or more rounds of sample reconstruction operations, the attribute A of all unlabeled samples in the unlabeled sample set S can be completed.
[0046] After completing attribute A for all unlabeled samples, if another type of attribute label (denoted as attribute B, i.e., case two mentioned above) is still missing, one or more rounds of sample reconstruction operations can be performed on attribute B to complete attribute B for all unlabeled samples in the unlabeled sample set S. Thus, each round or more of sample reconstruction operations can complete one type of attribute label for unlabeled samples until all missing attribute labels for all unlabeled samples in the unlabeled sample set S are completed. After all attribute labels for each unlabeled sample in the unlabeled sample set S are completed, they become new data (i.e., new samples). These new samples can serve as new labeled samples, thereby increasing the number of labeled samples.
[0047] As described in Example 1, in each sample augmentation operation, the unlabeled samples in the unlabeled sample set S are reconstructed into new labeled samples, thereby expanding the labeled samples. Such sample augmentation operations can be performed multiple times to continuously expand the labeled samples until the size of the labeled samples reaches the target size, which forms the augmented sample set.
[0048] Preferably, in each sample augmentation operation, constructing an unlabeled sample set lacking one or more attribute labels can include: randomly perturbing the features (e.g., component features) of the existing labeled samples other than one or more attribute labels (i.e., key attribute labels) (e.g., random perturbation within ±5% of the original value), resulting in an unlabeled sample set that does not contain the one or more attribute labels. That is, after random perturbation, the original sample's one or more attribute labels are not retained, thus obtaining an unlabeled sample set lacking these attribute labels, which can then be used for sample augmentation to obtain new labeled samples. For example, given M1 labeled samples, after M2 random perturbations, a large number of M1*M2 unlabeled samples can be quickly generated.
[0049] S103: Use the enhanced sample set to train the model and obtain the target model, and use the target model to predict aviation fuel performance.
[0050] After constructing the augmented sample set, it can be used as the training set for model training to obtain the target model for predicting aviation fuel performance. The augmented sample set significantly expands upon the initial labeled samples, enhancing the data support capability during the target model training process and enabling the model to fully learn the complex mapping relationship between various characteristics of aviation fuel and key physicochemical properties.
[0051] Preferably, the model is trained using an augmented sample set to obtain the target model, including: The basic regression models in stacked learning are trained using augmented sample sets; whereby the basic regression models used here can be the same as those used in the sample augmentation operation, and the basic regression models here can be the basic regression models after the sample augmentation operation is completed. Each trained model is used to predict on the test set, and the prediction results of each model on the test set are evaluated. Each basic regression model can be trained on the augmented sample set to obtain a model. Therefore, if there are multiple basic regression models, multiple models will be obtained by training on the augmented sample set. The target model is determined based on the evaluation results of each model's predictions on the test set.
[0052] As a preferred approach, the evaluation metrics used to assess the prediction results of each model on the test set include the coefficient of determination (denoted as R) of the prediction results. 2 ) and / or mean absolute error (denoted as MAE).
[0053] The coefficient of determination and mean absolute error are explained below.
[0054] Coefficient of determination Let the actual value be The predicted value is The number of samples in the prediction set is n, and the average value of the true values is .
[0055] ; in, The value of i ranges from 1 to n.
[0056] meaning: R 2 =1: The prediction is completely accurate; R 2 The closer the model is to 1, the better the model fit. R 2 =0: The model performs comparably to predictions made directly using the average of the true values; R 2 <0: The model performs worse than predicting using the average of the true values.
[0057] Mean Absolute Error ; meaning: MAE represents the average absolute error between the predicted result and the actual value; The closer the MAE is to 0, the more accurate the prediction result. MAE is measured in the same units as the predicted performance.
[0058] In Example 1, both labeled and unlabeled samples can have aviation fuel component characteristics. Based on this, a multi-scheme semi-supervised regression model including sample reconstruction and sample enhancement is used to predict the physicochemical properties of unlabeled samples, expand samples, enhance samples, and train the target model that integrates multiple basic regression models. This establishes a unified, complex, and accurate mapping from aviation fuel component characteristics to various key physicochemical properties.
[0059] After training the target model, it can be used to predict aviation fuel performance, including the physicochemical properties of new fuel samples. In practical applications, the fuel composition information of the sample to be predicted (which may contain fuel composition information but generally does not contain key physicochemical properties, thus belonging to unlabeled samples) is written into the model input file format. The target model script is run, and the target model predicts the physicochemical properties of the input fuel sample through its constructed complex mapping relationship between fuel composition and various physicochemical properties (including key physicochemical properties), and outputs it to the specified data file.
[0060] In Example 1, practical verification was also conducted, including comparing the prediction results of the target model for the physicochemical properties of actual fuels with those of actual aviation fuels not used in model training. The prediction accuracy was evaluated using the coefficient of determination, mean absolute error, and mean relative error as indicators. Verification showed that the target model obtained in Example 1 has high prediction accuracy for all aviation fuel properties specified by ASTM, and the prediction errors all meet the tolerance range specified in the relevant standards. This demonstrates the feasibility and effectiveness of Example 1, indicating that the method provided in Example 1 has good accuracy and engineering application value.
[0061] Example 1 can achieve the following beneficial effects: In the field of sustainable aviation fuel research, the high cost of experiments makes it difficult to obtain large quantities of fuel composition and physicochemical properties, resulting in a lack of high-quality data samples. Limited sample data leads to low accuracy and poor generalization ability in existing prediction models, and makes it difficult to achieve unified predictions of multiple key physicochemical properties.
[0062] Example 1 constructs a semi-supervised learning framework, introducing unlabeled samples on top of a small number of labeled samples. A base regression model is trained to predict the attributes of the unlabeled samples and select high-confidence samples. This data reconstruction mechanism reconstructs the unlabeled samples into labeled samples, expanding the training set for the subsequent target model and improving data utilization efficiency. Based on the expanded and enhanced sample set, Example 1 trains a multi-model fusion stacked ensemble prediction model as the target model, effectively improving the prediction performance and efficiency of aviation fuel.
[0063] Example 1 combines semi-supervised and multi-model collaborative prediction mechanisms to perform semi-supervised label expansion. Specifically, it includes constructing a small-scale labeled sample base, training stacked models (i.e., each basic regression model), sample reconstruction, sample augmentation, and other machine learning schemes. Example 1 constructs a small-sample learning framework based on a combination of semi-supervised learning and stacked ensemble learning. It trains initial basic regression models using a training set including labeled samples and predicts the attributes of unlabeled samples, supplementing the corresponding attribute labels (pseudo-labels) to achieve collaborative utilization of labeled and unlabeled samples. Example 1 effectively solves the problems of missing sample information and insufficient sample size. The final sample data volume is larger, and the trained model can fully learn the complex mapping relationship between input variables (fuel components) and output variables (physicochemical properties), ensuring that training accuracy and prediction quality meet requirements.
[0064] Example 1 fully mines the latent information in unlabeled samples through a semi-supervised learning mechanism, and expands the distribution range of labeled sample data used to train the target model by combining sample reconstruction and sample augmentation mechanisms. This enables the final trained target model to learn more complete data features, thereby significantly improving prediction accuracy and generalization ability. Simultaneously, by integrating the prediction results of multiple basic regression models through ensemble learning, the error of a single model is reduced, and prediction stability is improved. Example 1 can achieve simultaneous prediction of multiple fuel physicochemical properties within a single model framework, thereby effectively reducing experimental testing costs and improving fuel screening efficiency.
[0065] Example 1 achieves high-precision prediction of multiple key physicochemical properties under the condition of a limited number of fuel composition and performance data samples. Example 1 uses fuel composition data (i.e., labeled samples) as input and constructs a modeling framework combining semi-supervised learning and stacked ensemble learning. It jointly utilizes limited labeled and unlabeled data (i.e., unlabeled samples) and expands the training sample size by combining sample reconstruction and sample augmentation techniques, thereby establishing a stable mapping relationship between fuel composition and physicochemical properties. The sample reconstruction, sample augmentation, model training, and performance prediction stages of Example 1 are interconnected, forming a complete data-driven prediction process.
[0066] Example 1 enables high-precision, multi-performance prediction of aviation fuel properties or performance under small sample or small-scale sample conditions. It can be used for predicting the physicochemical properties and performance evaluation of sustainable aviation fuels and traditional jet fuels. It has the function of achieving high-precision fuel performance prediction and rapid pre-screening based on small sample data.
[0067] In Example 1, during the sample augmentation process, unlabeled samples with high prediction confidence are selected for attribute completion and used as new labeled samples for subsequent sample augmentation. This strategy of gradually expanding the data samples effectively suppresses the model shift problem caused by abrupt changes in data distribution and the accumulation of abnormal samples, thereby avoiding numerical instability or overfitting during the training process.
[0068] In Example 1, new unlabeled samples are generated through random perturbation, and sample augmentation operations are continued to expand the scale of labeled samples, significantly broadening the sample distribution space and enhancing the model's robustness to input feature perturbations, thereby effectively improving prediction accuracy and generalization ability. Based on this, a unified target model is trained by combining the augmented sample set with a multi-model stacking learning method, achieving high-precision simultaneous prediction of multiple key physicochemical properties of aviation fuel.
[0069] The second embodiment of this specification provides an aviation fuel performance prediction device based on few-shot learning and sample augmentation, corresponding to the method provided in the embodiments. The device includes: The sample augmentation module is used to obtain an augmented sample set from a small number of labeled samples through one or more sample augmentation operations; the labeled samples are aviation fuel-related samples. The performance prediction module is used to train the model using the enhanced sample set to obtain a target model, and then use the target model to predict aviation fuel performance. Each sample augmentation operation includes: constructing an unlabeled sample set that is missing one or more attribute labels; For any missing single-class attribute labels, the attribute labels of that class in the unlabeled sample set are completed through one or more rounds of sample reconstruction operations. The data obtained after all the missing attribute labels are filled in is used as the new labeled sample; Specifically, for the missing single-class attribute labels, the single-round sample reconstruction operation includes: Several basic regression models are trained using a newly constructed training set; the training set includes samples that do not currently lack labels for that type of attribute. For each unlabeled sample lacking the attribute label, the trained basic regression models are used to predict the attribute label of each unlabeled sample, and the prediction results of each basic regression model for the attribute label of each unlabeled sample are evaluated to see if they meet the preset conditions. Complete the attribute labels for unlabeled samples whose prediction results meet the preset conditions.
[0070] Preferably, constructing an unlabeled sample set lacking one or more attribute labels includes: Randomly perturb the features of the existing labeled samples other than one or more attribute labels to obtain an unlabeled sample set that does not contain the one or more attribute labels.
[0071] As a preferred option, the evaluation index used to assess whether the prediction results of each basic regression model for the attribute labels of each unlabeled sample meet the preset conditions includes the dispersion of the prediction results.
[0072] Preferably, for any unlabeled sample, if the dispersion of the prediction result of the attribute label of that type for that unlabeled sample is located in a specific distribution area of the dispersion of the prediction results of the attribute label of that type for each unlabeled sample, then the prediction result of the attribute label of that type for that unlabeled sample satisfies the condition.
[0073] Preferably, for any unlabeled sample, the difference between the maximum and minimum values of the predicted results of the attribute label of that unlabeled sample is taken as the dispersion of the predicted results of the attribute label of that unlabeled sample.
[0074] Preferably, the completion of the attribute labels for unlabeled samples whose prediction results meet preset conditions includes: For any unlabeled sample whose prediction results meet the preset conditions, the mean of the prediction results for that unlabeled sample is used as the completion result of the attribute label for that unlabeled sample.
[0075] Preferably, the enhanced sample set is used for model training to obtain the target model, including: The enhanced sample set was used to train each basic regression model; Each trained model is used to make predictions on the test set, and the prediction results of each model on the test set are evaluated. The target model is determined based on the evaluation results of each model's predictions on the test set.
[0076] Preferably, the evaluation metrics used to assess the prediction results of each model on the test set include the coefficient of determination and / or mean absolute error of the prediction results.
[0077] As a preferred option, the missing attribute tags include key physicochemical properties.
[0078] The above embodiments can achieve the same technical effect. The contents of each embodiment can be referred to or supplemented by each other and used in combination.
[0079] The above description is merely an embodiment of this specification and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.
Claims
1. A method for predicting aviation fuel performance based on few-shot learning and sample augmentation, characterized in that, The method includes: An augmented sample set is obtained by performing one or more sample augmentation operations on a small-scale labeled sample set; the labeled sample is related to aviation fuel. The enhanced sample set is used to train the model to obtain the target model, and the target model is used to predict aviation fuel performance. Each sample augmentation operation includes: constructing an unlabeled sample set that is missing one or more attribute labels; For any missing single-class attribute labels, the attribute labels of that class in the unlabeled sample set are completed through one or more rounds of sample reconstruction operations. The data obtained after all the missing attribute labels are filled in is used as the new labeled sample; Specifically, for the missing single-class attribute labels, the single-round sample reconstruction operation includes: Several basic regression models are trained using a newly constructed training set; the training set includes samples that do not currently lack labels for that type of attribute. For each unlabeled sample lacking the attribute label, the trained basic regression models are used to predict the attribute label of each unlabeled sample, and the prediction results of each basic regression model for the attribute label of each unlabeled sample are evaluated to see if they meet the preset conditions. Complete the attribute labels for unlabeled samples whose prediction results meet the preset conditions.
2. The method as described in claim 1, characterized in that, Constructing an unlabeled sample set missing one or more attribute labels includes: Randomly perturb the features of the existing labeled samples other than one or more attribute labels to obtain an unlabeled sample set that does not contain the one or more attribute labels.
3. The method as described in claim 1, characterized in that, The evaluation metrics used to assess whether the prediction results of each basic regression model for the attribute labels of each unlabeled sample meet the preset conditions include the dispersion of the prediction results.
4. The method as described in claim 3, characterized in that, For any unlabeled sample, if the dispersion of the prediction result of the attribute label of that unlabeled sample is located in a specific distribution area of the dispersion of the prediction results of the attribute label of each unlabeled sample, then the prediction result of the attribute label of that unlabeled sample satisfies the condition.
5. The method as described in claim 3, characterized in that, For any unlabeled sample, the difference between the maximum and minimum values of the predicted results of the attribute label of that unlabeled sample is taken as the dispersion of the predicted results of the attribute label of that unlabeled sample.
6. The method according to any one of claims 3 to 5, characterized in that, The completion of attribute labels for unlabeled samples whose prediction results meet preset conditions includes: For any unlabeled sample whose prediction results meet the preset conditions, the mean of the prediction results for that unlabeled sample is used as the completion result of the attribute label for that unlabeled sample.
7. The method as described in claim 1, characterized in that, The enhanced sample set is used to train the model to obtain the target model, including: The enhanced sample set was used to train each basic regression model; Each trained model is used to make predictions on the test set, and the prediction results of each model on the test set are evaluated. The target model is determined based on the evaluation results of each model's predictions on the test set.
8. The method as described in claim 7, characterized in that, The evaluation metrics used to assess the prediction results of each model on the test set include the coefficient of determination and / or mean absolute error of the prediction results.
9. The method as described in claim 1, characterized in that, The missing attribute tags include key physicochemical properties.
10. An aviation fuel performance prediction device based on few-shot learning and sample augmentation, characterized in that, The device includes: The sample augmentation module is used to obtain an augmented sample set from a small number of labeled samples through one or more sample augmentation operations; the labeled samples are aviation fuel-related samples. The performance prediction module is used to train the model using the enhanced sample set to obtain a target model, and then use the target model to predict aviation fuel performance. Each sample augmentation operation includes: constructing an unlabeled sample set that is missing one or more attribute labels; For any missing single-class attribute labels, the attribute labels of that class in the unlabeled sample set are completed through one or more rounds of sample reconstruction operations. The data obtained after all the missing attribute labels are filled in is used as the new labeled sample; Specifically, for the missing single-class attribute labels, the single-round sample reconstruction operation includes: Several basic regression models are trained using a newly constructed training set; the training set includes samples that do not currently lack labels for that type of attribute. For each unlabeled sample lacking the attribute label, the trained basic regression models are used to predict the attribute label of each unlabeled sample, and the prediction results of each basic regression model for the attribute label of each unlabeled sample are evaluated to see if they meet the preset conditions. Complete the attribute labels for unlabeled samples whose prediction results meet the preset conditions.
