A data-driven based fuel tank fuel level estimation method

Abstract

The application discloses a kind of based on data-driven storage tank fuel residual amount estimation method, first collection data and construct data set, the data of data set is preprocessed, and training set and test set are divided, after the data of pre-processing are extracted to basic feature, and high-order time series data are extracted based on the basic feature;Fuel residual amount estimation model is constructed, and the model is trained;Finally, based on the fuel residual amount estimation model after training, fuel residual amount is estimated in real time.The scheme of the application requires lower hardware sensor device, only needs to set up rod type liquid level meter and other equipment in the storage tank, without erecting other various sensors, can reduce the cost of fuel residual amount estimation system;Meanwhile, the scheme uses data-driven method, training target is simple, training process is stable, through the analysis of data and feature, the model obtained after training can cover more fuel liquid surface shaking internal law and mode, and the estimation accuracy of fuel residual amount is higher.

Description

Technical Field

[0001] This invention belongs to the field of fuel data processing, specifically relating to a data-driven method for estimating the remaining fuel level in storage tanks. Background Technology

[0002] With the continuous progress and development of my country's aerospace industry, the safety of spacecraft launch, operation, and return has become a major concern. Limited by payload requirements, spacecraft can only carry a limited amount of fuel. Furthermore, to ensure the successful execution of normal space missions, it is necessary to accurately monitor the remaining fuel levels. This helps personnel and systems obtain timely information on the remaining fuel status in the tanks, estimate the rate of fuel consumption and remaining operating time, and prevent accidents caused by fuel depletion.

[0003] Current detection solutions primarily rely on data from liquid level sensors to obtain information on remaining fuel levels. However, the equipment itself undergoes maneuvering during operation, causing some degree of sloshing of the fuel level within the tank. Simultaneously, due to the continuous operation of the equipment, the fuel in the tank is constantly being consumed, resulting in a time-varying downward trend in the liquid level. This trend, combined with the sloshing, creates unavoidable time-varying sloshing interference signals, posing a significant challenge to the current level sensors' ability to accurately detect liquid level and remaining fuel in real time.

[0004] Furthermore, due to the aging and degradation of flow rate sensors in some equipment after long-term use, flow rate measurements become inaccurate, making it impossible to estimate the remaining fuel mass based on the initial liquid level, flow rate, and operating time. Current solutions primarily rely on methods such as mean filtering to handle errors caused by liquid surface sloshing and external disturbances, but their effectiveness is limited, and the estimated liquid level still deviates somewhat from the true value. Therefore, a data-driven method for estimating the remaining fuel mass in storage tanks is urgently needed. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a data-driven method for estimating the remaining fuel level in a storage tank, thereby accurately estimating the liquid level and remaining fuel level under conditions of liquid surface sloshing and fuel consumption.

[0006] The specific technical solution for achieving the objective of this invention is as follows:

[0007] A data-driven method for estimating the remaining fuel level in a fuel tank includes the following steps:

[0008] Step 1: Collect data and build a dataset;

[0009] Step 2: Preprocess the dataset and divide it into training and test sets;

[0010] Step 3: Extract basic features from the preprocessed data, and extract higher-order time series data based on these basic features;

[0011] Step 4: Construct a fuel balance estimation model and train the model;

[0012] Step 5: Estimate the fuel balance in real time based on the trained fuel balance estimation model.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] (1) The solution of the present invention has low requirements for hardware sensor equipment. Only rod level gauges and other equipment need to be installed in the tank. There is no need to install other sensors, such as visual cameras and ultrasonic level gauges, which can further reduce the cost of the fuel balance estimation system.

[0015] (2) The present invention adopts a data-driven approach, with simple training objectives and a stable training process. Through the analysis of data and features, the model obtained after training can cover more internal laws and patterns of fuel liquid surface sloshing, and the estimation accuracy of fuel balance is higher. At the same time, based on the K-fold cross-validation method, the obtained model has better generalization ability.

[0016] (3) In the training of the model, the present invention not only uses the directly extracted basic features, but also uses the high-order time series data features extracted based on the basic features. This can help to find the potential patterns of the data in advance and use them as the input for model training. This is conducive to the rapid convergence of the training process, helps to reduce the model size, improves the modeling efficiency, and makes the training results of the model more accurate.

[0017] (4) The present invention provides a visualization result of the estimated fuel balance in the storage tank, which can clearly and accurately present the real-time liquid level sensor measurement value, real-time flow rate estimation value, storage tank filling ratio, and predicted fuel balance result in the storage tank, so as to facilitate the user to monitor the fuel in the storage tank in real time and help subsequent task execution and operation.

[0018] The present invention will be further described below with reference to specific embodiments. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the steps in the data-driven tank fuel balance estimation method of the present invention.

[0020] Figure 2 This is a schematic diagram illustrating the principle of measuring the remaining fuel level in the storage tank according to the present invention.

[0021] Figure 3 This is a schematic diagram of the original dataset format in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the dataset format after feature extraction in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of a visualization interface for estimating the remaining fuel level in the storage tank, as shown in an embodiment of the present invention. Detailed Implementation

[0024] A data-driven method for estimating the remaining fuel level in a fuel tank includes the following steps:

[0025] Step 1: Collect data and build a dataset:

[0026] Step 1-1: As the liquid level gradually decreases during fuel consumption, fuel sloshing may occur under different filling ratios during data acquisition. The frequency and amplitude of the sloshing signal will vary slightly depending on the filling ratio. To address this, liquid level data are collected under different filling ratios and sloshing scenarios within the tank. The filling ratio δ represents the percentage of the remaining fuel mass in the tank compared to the mass of fuel when the tank is fully loaded, i.e.:

[0027]

[0028] in, It represents the current mass of fuel remaining in the tank. This refers to the full load mass of fuel in the storage tank;

[0029] The raw data collected for the shaking scene needs to cover different filling ratio states, and the liquid level sensor measurement value and the estimated filling ratio value at the current moment should be collected respectively;

[0030] Steps 1-2: Determine the gold standard value of fuel balance at different times based on the true value of flow rate and time. :

[0031]

[0032] Where s is the flow velocity, in kg / s, and t is the measurement time, in seconds;

[0033] Steps 1-3: Based on the data from Steps 1-1 and 1-2, construct a dataset, which includes liquid level sensor measurements, filling ratio estimates, and fuel balance values ​​at different time points.

[0034] Step 2: Preprocess the dataset and divide it into training and test sets:

[0035] Step 2-1: Perform data smoothing on the centralized liquid level sensor measurement values, filling ratio estimates, and fuel balance gold standard values ​​using the mean filtering method.

[0036]

[0037] in, This represents the liquid level measurement, filling ratio estimate, or fuel balance gold standard value at time k, and N represents the number of samples.

[0038] Step 2-2: Normalize the smoothed liquid level sensor measurement value, fill ratio estimate value, and fuel balance gold standard value respectively:

[0039]

[0040] in, This represents the minimum value among the measured liquid level, estimated fill ratio, or gold standard value for fuel balance. This indicates the maximum value among the measured liquid level, estimated fill ratio, or gold standard value for fuel balance. This represents the normalized data;

[0041] Steps 2-3: Divide the processed dataset into training and test sets.

[0042] Step 3: Extract basic features from the preprocessed data, and extract higher-order time series data based on these basic features:

[0043] Step 3-1: Use the time history, liquid level sensor measurement value, and filling ratio estimate in the preprocessed dataset as basic features;

[0044] Step 3-2: Based on the basic features, extract the higher-order time series features, specifically the first-order and multiple-order changes of the time history, liquid level sensor measurement values, and filling ratio estimation values.

[0045] For example, the first-order change, second-order change, and third-order change of the liquid level sensor measurement value, where the first-order change refers to the derivative of the preprocessed liquid level sensor measurement value, i.e.

[0046]

[0047] Similarly, the second-order change in the liquid level sensor measurement refers to the second derivative of the preprocessed liquid level sensor measurement, i.e.

[0048]

[0049] The same applies to the third-order changes in the liquid level sensor readings.

[0050] Step 4: Construct a fuel surplus estimation model and train the model:

[0051] Step 4-1: Select MAE and RMSE as performance metrics:

[0052]

[0053]

[0054] in, The gold standard value for residual fuel mass. It is the estimated or predicted value, where N is the total sample size;

[0055] Step 4-2: Based on the performance metrics in Step 4-1, use the basic features and high-order time series data determined in Step 3 as the input values ​​of the model, and the fuel balance gold standard value as the output value of the model. Select a deep learning architecture or method, not limited to long short-term memory neural network (LSTM), nonlinear auto-regressive model (NARX), or convolutional neural network (CNN), to train the model.

[0056] The input values ​​include, but are not limited to, basic features such as data time history, liquid level sensor measurements, and filling ratio estimates, as well as higher-order temporal features such as first-, second-, and third-order changes in liquid level sensor measurements. Taking convolutional neural network training as an example, it generally has an input layer, hidden layers, and an output layer, with one input layer and one output layer, and 1 to n hidden layers. Each layer consists of neurons. Here, the input layer has no fewer than 6 neurons, corresponding to no fewer than 6 basic and higher-order temporal features, namely, data time history, liquid level sensor measurements, filling ratio estimates, and first-, second-, and third-order changes in liquid level sensor measurements. In actual training, the no fewer than 6 features from the previous time step are used as input, and the current fuel balance gold standard value is used as the output value for training. That is, the no fewer than 6 features from time step k-1 are used as input, and the fuel balance gold standard value from time step k is used as the output value for training. Finally, the fuel balance model is obtained after training.

[0057] By utilizing high-order time-series data features extracted from basic features, we can identify potential patterns in the data in advance and use them as input for model training. This facilitates rapid convergence of the training process, helps reduce model size, improves modeling efficiency, and makes the training results of the model more accurate.

[0058] Step 4-3: Use the K-fold cross-validation method to select the model. Select the model with the best test performance among all trained models as the final prediction model, or select the average of all model results as the final model output.

[0059] Step 5: Estimate the fuel balance in real time based on the trained fuel balance estimation model:

[0060] Based on the model trained in step 4, the remaining fuel in the tank is estimated, and the results are visualized to present the liquid level sensor measurement value, real-time flow rate estimate, tank filling ratio, feature data, and real-time estimated fuel remaining value in real time.

[0061] It also displays the fuel balance curve.

[0062] Example

[0063] Combination Figure 1 A data-driven method for estimating the remaining fuel level in a fuel tank includes the following steps:

[0064] Step 1: Collect data and build a dataset:

[0065] Step 1-1, Combining Figure 2 Because the liquid level gradually decreases during fuel consumption, fuel sloshing may occur at different filler ratios during data acquisition, and the frequency and amplitude of the sloshing signal will vary slightly depending on the filler ratio. To address this, liquid level data were collected under different filler ratios and sloshing scenarios within the tank. The filler ratio δ represents the percentage of the remaining fuel mass in the tank compared to the mass of fuel when the tank is fully loaded, i.e.:

[0066]

[0067] in, It represents the current mass of fuel remaining in the tank. This refers to the full load mass of fuel in the storage tank;

[0068] The raw data collected for the shaking scene needs to cover different filling ratio states, and the liquid level sensor measurement value and the estimated filling ratio value at the current moment should be collected respectively;

[0069] The system can calculate the variance of the liquid level sensor measurements over a 5-second time period. If this variance exceeds a set threshold, it is considered that there is swaying interference within this time period. The remaining mass at the initial moment is then taken, and the filling ratio at that remaining mass is calculated and rounded up. For example, if the calculated filling ratio is 67.2%, then the filling ratio at that moment is taken as 70%; if the calculated filling ratio is 64.3%, then the filling ratio at that moment is taken as 65%.

[0070] Steps 1-2: Determine the gold standard value of fuel balance at different times based on the true value of flow rate and time. :

[0071]

[0072] Where s is the flow velocity, in kg / s, and t is the measurement time, in seconds;

[0073] Steps 1-3: Based on the data from Steps 1-1 and 1-2, construct a dataset, which includes liquid level sensor measurements, fill ratio estimates, and fuel balance values ​​at different time points. Figure 3 As shown.

[0074] Step 2: Preprocess the dataset and divide it into training and test sets:

[0075] Step 2-1: Perform data smoothing on the centralized liquid level sensor measurement values, filling ratio estimates, and fuel balance gold standard values ​​using the mean filtering method.

[0076]

[0077] in, This represents the liquid level measurement, filling ratio estimate, or fuel balance gold standard value at time k. In this embodiment, the number of samples is 10.

[0078] Step 2-2: Normalize the smoothed liquid level sensor measurement value, fill ratio estimate value, and fuel balance gold standard value respectively:

[0079]

[0080] in, This represents the minimum value among the measured liquid level, estimated fill ratio, or gold standard value for fuel balance. This indicates the maximum value among the measured liquid level, estimated fill ratio, or gold standard value for fuel balance. This represents the normalized data;

[0081] Steps 2-3: Divide the processed dataset into a training set and a test set. In this embodiment, 80% of the preprocessed dataset is used as the training set and 20% is used as the test set, that is, the number of training set samples is 8000 and the number of test set samples is 2000.

[0082] Step 3: Extract basic features from the preprocessed data, and extract higher-order time series data based on these basic features:

[0083] Step 3-1: Use the time history, liquid level sensor measurement value, and filling ratio estimate in the preprocessed dataset as basic features;

[0084] Step 3-2: Based on the basic features, extract the higher-order time series features, specifically the first-order and multiple-order changes of the time history, liquid level sensor measurement values, and filling ratio estimation values.

[0085] For example, the first-order change, second-order change, and third-order change of the liquid level sensor measurement value, where the first-order change refers to the derivative of the preprocessed liquid level sensor measurement value, i.e.

[0086]

[0087] Similarly, the second-order change in the liquid level sensor measurement refers to the second derivative of the preprocessed liquid level sensor measurement, i.e.

[0088]

[0089] The same applies to the third-order changes in the liquid level sensor readings.

[0090] In this embodiment, the dataset after feature extraction is in the form of: Figure 4 As shown.

[0091] Step 4: Construct a fuel surplus estimation model and train the model:

[0092] Step 4-1: Select MAE and RMSE as performance metrics:

[0093]

[0094]

[0095] in, The gold standard value for residual fuel mass. It is the estimated or predicted value, where N is the total sample size;

[0096] Step 4-2: Based on the performance metrics in Step 4-1, use the basic features and high-order time series data determined in Step 3 as the input values ​​of the model, and the fuel balance gold standard value as the output value of the model. Select a deep learning architecture or method, not limited to long short-term memory neural network (LSTM), nonlinear auto-regressive model (NARX), or convolutional neural network (CNN), to train the model.

[0097] By utilizing high-order time-series data features extracted from basic features, we can identify potential patterns in the data in advance and use them as input for model training. This facilitates rapid convergence of the training process, helps reduce model size, improves modeling efficiency, and makes the training results of the model more accurate.

[0098] Step 4-3: Use the K-fold cross-validation method to select the model. Select the model with the best test performance among all trained models as the final prediction model, or select the average of all model results as the final model output.

[0099] Step 5: Estimate the fuel balance in real time based on the trained fuel balance estimation model:

[0100] Based on the model trained in step 4, the remaining fuel in the tank is estimated, and the results are visualized to present the liquid level sensor measurement value, real-time flow rate estimate, tank filling ratio, feature data, and real-time estimated fuel remaining value in real time.

[0101] It also displays the fuel balance curve.

[0102] like Figure 5 As shown, the input of this embodiment is the real-time liquid level sensor measurement value. After the residual estimation model inference, the residual estimation result and intermediate process variables can be obtained. It can be used to present the user with parameters such as the real-time liquid level sensor measurement value, real-time flow rate estimation value, tank filling ratio, main features and the model inference and prediction of fuel residual results obtained in step 4, and to give a schematic diagram of the fuel residual curve.

[0103] Module 1 is the residual monitoring process and result value, which allows users to understand the process data in real time and helps with subsequent task operations. This module includes, but is not limited to, intermediate or result data such as liquid level sensor measurement value, real-time flow rate estimate, tank filling ratio estimate, fuel residual estimate result, fuel residual estimate error, real-time first-order characteristic value, real-time second-order characteristic value, and real-time third-order characteristic value. Module 2 is the image presentation module, which allows users to intuitively understand the changing trends of measured value and estimated value curves, estimation error curves, high-order time series characteristic curves, and spectrum analysis curves in real time.

[0104] This embodiment of the solution has low requirements for hardware sensor equipment, requiring only rod-type level gauges and eliminating the need for other sensors such as visual cameras and ultrasonic level gauges, thus further reducing the cost of the fuel balance estimation system. Secondly, this solution adopts a data-driven approach with simple training objectives and a stable training process. Through the analysis of data and features, the trained model can cover more of the inherent laws and patterns of fuel level sloshing, resulting in higher accuracy in fuel balance estimation. Furthermore, the K-fold cross-validation method enhances the generalization ability of the obtained model. Finally, this invention provides a tank fuel balance estimation software interface that clearly and accurately presents users with real-time liquid level sensor measurements, real-time flow rate estimates, tank filling ratio, and predicted fuel balance results, facilitating real-time monitoring of fuel in the tank and aiding in subsequent task execution and operation.

[0105] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A data-driven method for estimating the remaining fuel level in a storage tank, characterized in that, Includes the following steps: Step 1: Collect data and build a dataset: Step 1-1: Collect liquid level data in the storage tank under different filling ratios and sloshing scenarios. The filling ratio represents the percentage of the remaining fuel mass in the storage tank relative to the mass of fuel when the tank is fully loaded, i.e.: ； in, It represents the current mass of fuel remaining in the tank. This refers to the full load mass of fuel in the storage tank; The raw data collected for the shaking scene needs to cover different filling ratio states, and the liquid level sensor measurement value and the estimated filling ratio value at the current moment should be collected respectively; Steps 1-2: Calculate the gold standard value of fuel balance at different times based on the true value of flow velocity and time. : ； Where s is the flow velocity, in kg / s, and t is the measurement time, in seconds; Steps 1-3: Based on the data from Steps 1-1 and 1-2, construct a dataset, which includes liquid level sensor measurements, filling ratio estimates, and fuel balance values ​​at different time points. Step 2: Preprocess the dataset and divide it into training and test sets; Step 3: Extract basic features from the preprocessed data, and extract higher-order time series data based on these basic features; Step 4: Construct a fuel balance estimation model and train the model; Step 5: Estimate the fuel balance in real time based on the trained fuel balance estimation model.

2. The data-driven method for estimating tank fuel balance according to claim 1, characterized in that, The data preprocessing in step 2 specifically includes: Step 2-1: Perform data smoothing on the centralized liquid level sensor measurement value, filling ratio estimate value, and fuel balance gold standard value in the dataset respectively; Step 2-2: Normalize the smoothed liquid level sensor measurement value, filling ratio estimate value, and fuel balance gold standard value respectively; Steps 2-3: Divide the processed dataset into training and test sets.

3. The data-driven method for estimating tank fuel balance according to claim 2, characterized in that, The data smoothing process in step 2-1 uses the mean filtering method: ； in, This represents the measured liquid level, estimated filling ratio, or gold standard value of fuel balance at time k, and N represents the number of samples.

4. The data-driven method for estimating tank fuel balance according to claim 3, characterized in that, The normalization process in step 2-2 is specifically as follows: ； in, This represents the minimum value among the measured liquid level, estimated fill ratio, or gold standard value for fuel balance. This indicates the maximum value among the measured liquid level, estimated fill ratio, or gold standard value for fuel balance. This represents the normalized data.

5. The data-driven method for estimating tank fuel balance according to claim 1, characterized in that, The feature extraction of the preprocessed dataset in step 3 specifically involves: Step 3-1: Use the time history, liquid level sensor measurement value, and filling ratio estimate in the preprocessed dataset as basic features; Step 3-2: Based on the basic features, extract the higher-order temporal features respectively.

6. The data-driven method for estimating tank fuel balance according to claim 5, characterized in that, The higher-order time series features are first-order to multiple-order changes in time history, liquid level sensor measurements, and filling ratio estimates.

7. The data-driven method for estimating tank fuel balance according to claim 1, characterized in that, The model building and training in step 4 are specifically as follows: Step 4-1: Select MAE and RMSE as performance metrics: ； ； in, The gold standard value for residual fuel mass. It is the estimated or predicted value, where N is the total sample size; Step 4-2: Based on the performance metrics in Step 4-1, use the basic features and high-order time series data determined in Step 3 as the input values ​​of the model, and the fuel balance gold standard value as the output value of the model. Select a deep learning architecture or method, not limited to long short-term memory neural network, nonlinear autoregressive neural network or convolutional neural network, for model training. Step 4-3: Use the K-fold cross-validation method to select the model. Select the model with the best test performance among all trained models as the final prediction model, or select the average of the results of all models as the final model output.

8. The data-driven method for estimating tank fuel balance according to claim 1, characterized in that, The real-time estimation of remaining fuel in step 5 specifically involves: Based on the model trained in step 4, the remaining fuel in the tank is estimated, and the results are visualized to present the liquid level sensor measurement value, real-time flow rate estimate, tank filling ratio, feature data, and real-time estimated fuel remaining value in real time. It also displays the fuel balance curve.

Images