Locomotive oil tank calibration method based on deep learning

Abstract

The invention belongs to the technical field of rail transit, and particularly relates to a locomotive fuel tank calibration method based on deep learning. According to the technical scheme, the locomotive fuel tank calibration method based on deep learning comprises the following steps that S1, calibration is conducted before a sensor is applied; s2, constructing a baseline model based on the three-dimensional model; s3, constructing a model for calibrating the oil quantity based on a brand-new oil tank to be used; and S4, constructing a regular calibration model based on the oil tank which is put into use. The invention provides a locomotive fuel tank calibration method, which calibrates the height reading of a sensor based on a machine learning algorithm, realizes fuel tank fuel quantity calibration based on historical data application of deep learning, and combines the two to realize a complete calibration process of a fuel tank.

Description

Technical Field

[0001] The present invention belongs to the field of rail transit technology, and in particular relates to a locomotive fuel tank calibration method based on deep learning. Background Art

[0002] In the process of developing the fuel consumption monitoring system for diesel locomotives, the improvement and optimization of the fuel tank calibration process is the basis for ensuring the accuracy of fuel detection. The realization of this accuracy depends on the precise mapping between fuel height and fuel volume. In order to achieve the accuracy of calibration, it is necessary to focus on improving the following two aspects.

[0003] The first part is to focus on improving the accuracy of the sensor. The models and accuracy of the pressure level meters on the market vary. Based on this premise, we should take the actual situation of the oil tank as an example. If the oil tank is large, the error of every millimeter will affect the calculation of a large amount of oil. Only through the comparison and selection of unified standards can we finally select the sensor with the best performance and the smallest reading error as the final choice.

[0004] The second part is the construction of the oil volume-height correspondence model. Considering the complex premise that the method needs to adapt to the non-standard shapes of different types of fuel tanks, the traditional common method chooses to calibrate the oil volume-height correspondence table exclusive to a fuel tank based on the three-dimensional model of the fuel tank. The advantage of this method is that the correspondence can be switched quickly and accurately. The disadvantage is that the data outside the sampling point still needs to be fitted to realize the estimation, and in actual applications, it is difficult to stop at the calibration point. In addition, considering the liquid characteristics of diesel, the calibration of the three-dimensional model may not be completely consistent with the actual situation.

[0005] This project proposes a method to calibrate the sensor height reading based on machine learning algorithm and realize the fuel tank fuel volume calibration based on historical data application based on deep learning. Combining the two, a complete fuel tank calibration process is realized. Summary of the invention

[0006] In order to solve the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide a locomotive fuel tank calibration method, which calibrates the sensor height reading based on a machine learning algorithm, realizes the tank oil quantity calibration based on the historical data application based on deep learning, and combines the two to realize the complete calibration process of the fuel tank.

[0007] The technical solution adopted by the present invention is:

[0008] A locomotive fuel tank calibration method based on deep learning includes the following steps:

[0009] S1: Sensor calibration before application: The sensor is used to measure the distance from the sensor installation position to the liquid level. Before actual installation and application, design and prepare test fixtures to complete accuracy tests under fixed height and liquid level reduction conditions;

[0010] S2: Construct a baseline model based on the three-dimensional model: According to the three-dimensional structure of the fuel tank and the error range generated in the actual production process, a baseline model is constructed. In the baseline model, a curve is fitted to describe the relationship between volume and height;

[0011] S3: Construct a calibration model for the oil volume based on a brand new fuel tank to be used: conduct actual measurements on a brand new fuel tank, where the fuel tank used in the actual measurement is equipped with diesel from the actual production environment;

[0012] S4: Construct a regular calibration model based on the oil tank that has been put into use: With long-term and multiple uses, the fitting curve needs to be calibrated. The model is directly constructed using an end-to-end approach. The causes are related to oil residue and inner wall adhesion.

[0013] The present invention realizes the pre-application calibration of the sensor based on the machine learning model, ensuring that the height acquisition accuracy is maintained at a high level during the calibration process.

[0014] The present invention effectively utilizes the three-dimensional model of the fuel tank body during modeling to construct a baseline model based on the three-dimensional model. The three-dimensional model is the baseline of the corresponding relationship between the fuel volume and the height conversion under the most ideal conditions. Similarly, the baseline value of the height to fuel volume conversion described by the method can be used to constrain the calculation lower limit based on the algorithm and can also guide the direction of the fitting relationship.

[0015] The present invention is based on a brand new oil tank test opportunity and condition to be used, can effectively correct the measurement error, and can effectively compare the approximate range of the error by comparing the difference in oil volume between two schemes.

[0016] The present invention can form a regular inspection mechanism for oil tanks that have been put into use for a period of time. Based on the historical data collected by this mechanism, the correction direction of the altitude-oil volume conversion correspondence relationship as the use time increases can be effectively evaluated.

[0017] As a preferred embodiment of the present invention, in step S1, the accuracy test under the conditions of fixed height and liquid level reduction is specifically as follows:

[0018] Accuracy test under fixed height conditions:

[0019] Inject diesel into the fuel consumption test fixture to the total height, let it stand to stabilize the liquid level, and record the stable liquid level height value;

[0020] The fuel consumption collection board is powered on to read and record the current fuel consumption system output liquid level height and output liquid volume value;

[0021] Compare the measured value of the liquid level tooling with the height value output by the acquisition board; compare the output liquid volume value with the formula to calculate the oil volume value of the oil tank corresponding to the current height liquid level;

[0022] Repeat the test several times;

[0023] When the liquid level is lowered:

[0024] Inject diesel into the fuel consumption test fixture to the total height, let it stand to stabilize the liquid level, and record the stable liquid level height value;

[0025] The fuel consumption collection board is powered on to read and record the current fuel consumption system output liquid level height and output liquid volume value

[0026] The liquid level drops 10mm each time, and the test is repeated several times in total;

[0027] Compare the measured value of the liquid level each time it descends with the height value output by the acquisition board; compare the output liquid volume value with the formula to calculate the oil volume value of the tank corresponding to the current height liquid level

[0028] Additional measurements:

[0029] The lowest measurable height of the sensor;

[0030] Testing of liquid level rise;

[0031] Drift of the sensor when it is suspended in the air.

[0032] As a preferred solution of the present invention, in step S3, the calibration scheme of the calibration model is:

[0033] S31: Setting the oil volume collection unit;

[0034] S32: Design and install a calibrated independent measurement module. A new fuel tank can be equipped with a height sensor to feed the reading back to the module;

[0035] S33: Construct the oil volume-altitude correspondence and optimize the sampling point jump problem.

[0036] As a preferred solution of the present invention, in step S4, the regular calibration model proposes three assumptions:

[0037] Hypothesis 1: As the number of times of oil use and refueling increases, the amount of oil residue will increase and the degree of inner wall adhesion will be serious. Because the inner wall adhesion of oil residue is difficult to quantify, it is assumed that there is a monotonic relationship between the number of times and the change of fitting relationship. Instead of directly establishing a linear relationship to construct a new fitting model, it is retained as a feature.

[0038] Hypothesis 2: It is necessary to clarify whether the amount of oil added each time is the same. If it is different, the amount of oil added will directly affect the adhesion of oil residue to the inner wall. Therefore, there is a second-degree correlation with the oil amount-height fitting curve, which is also retained as a feature;

[0039] Hypothesis 3: The age of the fuel tank is directly listed as a relevant feature.

[0040] As a preferred solution of the present invention, in step S1, the sensor is calibrated before application based on a machine learning algorithm, and the machine learning algorithm is implemented by the following steps:

[0041] S11: Data collection;

[0042] S12: data preprocessing;

[0043] S13: Model selection;

[0044] S14: Model training;

[0045] S15: Model evaluation;

[0046] S16: Model deployment and optimization.

[0047] As a preferred solution of the present invention, step S11 is specifically as follows: collecting data within different measuring ranges under the same liquid environment, and collecting data within the same scale should be repeated fluctuations;

[0048] Step S12 is specifically as follows: processing noise data, abnormal data and readings with large differences from actual readings; selecting data at a certain interval to establish a nonlinear model, and using the model features as the final result of feature selection. Different from traditional machine learning based on at least two or more parameters, by repeatedly fluctuating data in a small range at the same liquid level, relevant features are generated to amplify the relevant features of the input data, and the context of the readings is integrated into the decision-making method of the final machine learning algorithm in the form of features.

[0049] As a preferred embodiment of the present invention, in step S13, the model to be constructed is one of linear regression, polynomial regression, random forest, gradient boosting tree, support vector machine and neural network, and the final model is determined by the performance of the data on each model.

[0050] As a preferred solution of the present invention, step S14 is specifically as follows: dividing the data set based on the preprocessed data, dividing the data into a training set and a validation set according to a ratio of 8:2, and then performing a 3- or 5-fold cross-validation on the training set to select the optimal model, and finally completing model evaluation on the selected model on the training set.

[0051] As a preferred embodiment of the present invention, in step S15, the evaluation indicators of model evaluation are set to the following four values: mean square error, root mean square error, mean absolute error, and determination coefficient; all evaluation indicators are calculated separately on the validation set and the test set.

[0052] As a preferred solution of the present invention, in step S16, the trained and finally selected model can be deployed in the actual application process. A monitoring module should be set to record data in real time, and spot check to supplement the actual readings, so as to accumulate a certain amount of data for optimization in the later stage.

[0053] The beneficial effects of the present invention are:

[0054] 1. The present invention realizes the pre-application calibration of sensors based on machine learning models, ensuring that the height acquisition accuracy during the calibration process is maintained at a high level.

[0055] 2. The present invention effectively utilizes the three-dimensional model of the fuel tank body during modeling to construct a baseline model based on the three-dimensional model. The three-dimensional model is the baseline of the corresponding relationship between the fuel volume and the height conversion under the most ideal conditions. Similarly, the baseline value of the height to fuel volume conversion described by the method can be used to constrain the calculation lower limit based on the algorithm and can also guide the direction of the fitting relationship.

[0056] 3. The present invention is based on a brand new oil tank test opportunity and conditions to be used, which can effectively correct the measurement error, and can effectively compare the approximate range of the error by comparing the difference in oil volume between the two schemes.

[0057] 4. The present invention can form a regular inspection mechanism for oil tanks that have been put into use for a period of time. Based on the historical data collected by this mechanism, the correction direction of the altitude-oil volume conversion correspondence relationship can be effectively evaluated as the use time increases. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flow chart of the calibration use of the present invention;

[0059] Figure 2 It is a flow chart of the hypothetical scenario of oil level height conversion. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0061] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0062] like Figure 1 and Figure 2 As shown, a locomotive fuel tank calibration method based on deep learning in this embodiment includes the following steps:

[0063] S1: The first thing to be completed is the sensor calibration before application. The calibration process is as follows:

[0064] The sensor is used to measure the distance from the sensor installation position to the liquid level. Before actual installation and application, it is necessary to design and prepare test fixtures to complete accuracy tests under fixed height and liquid level reduction conditions.

[0065] Fixed Height:

[0066] 1) Inject diesel into the fuel consumption test fixture to the total height (low liquid level: below 1 / 4; middle liquid level: 1 / 4 to 3 / 4; high liquid level: above 3 / 4), let it stand to stabilize the liquid level, and record the stable liquid level height value;

[0067] 2) The fuel consumption collection board is powered on to read and record the current fuel consumption system output liquid level height and output liquid volume value;

[0068] 3) Compare the liquid level tooling measurement value and the height value output by the acquisition board; compare the output liquid volume value with the formula to calculate the oil volume value of the oil tank corresponding to the current height liquid level;

[0069] 4) Repeat the above steps 10 times.

[0070] When the liquid level is lowered:

[0071] 1) Inject diesel into the fuel consumption test fixture to the total height (low liquid level: below 1 / 4; middle liquid level: 1 / 4 to 3 / 4; high liquid level: above 3 / 4), let it stand to stabilize the liquid level, and record the stable liquid level height value;

[0072] 2) The fuel consumption collection board is powered on to read and record the current fuel consumption system output liquid level height and output liquid volume value;

[0073] 3) The liquid level drops 10 mm each time, for a total of 10 drops (the number of drops is determined based on the actual height of the container);

[0074] 4) Compare the measured value of the liquid level height tooling each time it descends with the height value output by the acquisition board; compare the output liquid volume value with the formula to calculate the oil volume value in the tank corresponding to the current height liquid level.

[0075] Additional measurements:

[0076] 1) The lowest measurable height of the sensor;

[0077] 2) Testing of liquid level rise;

[0078] 3) Drift of the sensor when it is suspended (not submerged in liquid).

[0079] Sensor pre-application calibration based on machine learning algorithm. The algorithm is implemented in the following steps:

[0080] S11: Data collection: Collect data within different measuring ranges under the same liquid environment. The data should be collected repeatedly within the same scale to avoid the situation where the ADC conversion process is less sensitive to slight changes.

[0081] S12: Data preprocessing: Process noise data, abnormal data, and readings with large differences from actual readings; select data at a certain interval to establish a nonlinear model, and use the model features as the final result of feature selection. Unlike traditional machine learning based on at least two or more parameters, the source data only has one type of sensor reading, so it is necessary to generate relevant features by repeatedly fluctuating data in a small range at the same liquid level, which is used to amplify the relevant features of the input data and integrate the context of the readings into the decision-making method of the final machine learning algorithm in the form of features. This method is set for low sensitivity during ADC conversion; the input and output are normalized, which can effectively improve the training speed and performance of the machine learning model.

[0082] S13: Model selection: There are many types of machine learning, and you should choose the model that is ultimately suitable for the current problem. The models to be built include but are not limited to the following: linear regression, polynomial regression, random forest, gradient boosting tree, support vector machine, neural network, etc.

[0083] S14: Model training: Divide the data set based on the preprocessed data into a training set and a validation set in an 8:2 ratio. Perform a 3- or 5-fold cross-validation on the training set to select the optimal model. The final selected model is evaluated on the training set.

[0084] S15: Model evaluation: To achieve consistency in model evaluation, evaluation indicators should be strictly set, which can well evaluate the accuracy, stability and generalization ability of the model. The following four values ​​can be set for calculation indicators: mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). All evaluation indicators should be calculated separately on the validation set and the test set. The performance on the validation set shows the generalization performance, and the performance on the test set shows the final performance.

[0085] S16: Model deployment and optimization: The trained and finally selected model should be deployed in the actual application process. A monitoring module should be set up to record data in real time, and spot check to supplement the actual readings, so as to accumulate a certain amount of data for optimization in the later stage.

[0086] S2: It is necessary to build a baseline model based on the 3D model. The calibration process is as follows:

[0087] According to the three-dimensional structure of the fuel tank, a baseline model is constructed. The model is the most ideal baseline case. In the baseline model, a curve can be fitted to describe the relationship between volume and height.

[0088] Implementation requirements: A three-dimensional model of the fuel tank is required, as well as the error range generated during the actual production process. This baseline model can be directly used to fit the oil volume-height correspondence. The advantage is that it reduces human errors compared to actual measurements and is easier to model. The disadvantage is that the error with the actual correspondence will always exist, and will continue to change with the frequent use of the fuel tank, and the direction of change is not constant.

[0089] S3: A calibration model for the oil volume based on a brand new tank to be used needs to be constructed. As the result of step S2, a baseline model has been constructed based on the three-dimensional structure, and actual measurements are still required on a brand new tank. The tank used for the actual measurement needs to be equipped with diesel from the actual production environment (maintaining a state where there may be oil residue, without the need for filter residue treatment).

[0090] Calibration scheme:

[0091] S31: Setting the oil volume collection unit; for example, taking every increase of 50 liters of diesel as an example, reading the current oil volume level height;

[0092] S32: Design and install calibrated independent measurement modules. New fuel tanks can be equipped with height sensors, which feed the readings back to the module for easy reading;

[0093] S33: Construct the oil volume-altitude correspondence and optimize the sampling point jump problem.

[0094] After the calibration model of the new fuel tank is completed, it can be used directly as a fitting relationship, or it can replace the three-dimensional model and become a new warning baseline model. Similar to the baseline model, this model also does not consider the changes in the fitting relationship after multiple and long-term use.

[0095] S4: It is necessary to build a regular calibration model based on the oil tank that has been put into use. To solve the problems mentioned in steps S2 and S3, the fitting curve needs to be calibrated with long-term and multiple uses. The model is directly constructed using an end-to-end approach. The cause is obviously related to oil residue and inner wall adhesion. The regular calibration model proposes three assumptions, as shown in the attached figure:

[0096] Hypothesis 1: As the number of times of oil use and refueling increases, the amount of oil residue will increase and the degree of inner wall adhesion will be serious. Because the inner wall adhesion of oil residue is difficult to quantify, it is assumed that there is a monotonic relationship between the number of times and the change in the fitting relationship. Instead of directly establishing a linear relationship to construct a new fitting model, it is retained as a feature.

[0097] Hypothesis 2: It is necessary to clarify whether the amount of refueling is the same each time. If it is different, the amount of refueling will directly affect the adhesion of oil residue to the inner wall. Therefore, there is a second-degree correlation between the oil amount-height fitting curve, which is also retained as a feature.

[0098] Assumption 3: The age of the fuel tank can be directly listed as a relevant feature.

[0099] Step S4 builds a calibration model based on the fuel tank that has been put into use. This model will build a deep learning algorithm and compare several common recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), etc. in the feature extraction stage. Convolutional neural networks (CNN) can also be used. Although this algorithm is generally used in image processing scenarios, it is possible to consider using CNN structures to extract high-dimensional features of data for early processing of encoding and decoding. At the same time, transformer model structures can also be considered, such as BERT or GPT variant frameworks. Although such frameworks are mostly used in the field of natural language processing, they are also applicable to time series data from a data logic perspective. This type of general framework is applied in the early data feature extraction process. After the later encoding and decoding, the fitting parameters of steps S2 and S3 are fused to form a new vector, and the final model is obtained using loss constraints.

[0100] Example:

[0101] The first step is to collect sensor readings at low, medium and high liquid levels based on sensors with certain accuracy, and to have multiple people take standardized actual readings at the same time. The average of the readings of multiple people is taken as the true value, and a machine learning algorithm is constructed to fit the collected readings to the true value. Unlike traditional machine learning based on at least two or more parameters, this source data only has one type of sensor reading. Therefore, it is necessary to generate relevant features by repeatedly fluctuating data in a small range at the same liquid level to amplify the relevant features of the input data and integrate the context of the readings into the final decision-making method of the machine learning algorithm in the form of features. After the machine learning algorithm is constructed and selected, whenever there is feedback from the actual collection results of the sensor, the feedback value is not just one, but also includes the fluctuation value within the range. The liquid level height value h close to the actual reading will be given through the machine learning algorithm.

[0102] The second step is to calibrate the three-dimensional model of the fuel tank and build a baseline model. After obtaining the baseline model, the piecewise function f1(h) can be constructed based on the structural characteristics. The height h in the first step can be input into the piecewise function to obtain the oil height v1 after table lookup.

[0103] The third step is to build a calibration model for a new oil tank to be used. Assume that the oil tank is not allowed to be repeatedly added or drained during the calibration process, which will cause unnecessary waste, so there is only one chance to calibrate the record. Therefore, by adding a unit amount of oil in a quantitative manner, and combining the value h of the sensor in the first step, a new piecewise function f2(h) is fitted. The lowest measurement error e under the ideal drawing and actual environment can be evaluated by comparing f1(h) and f2(h).

[0104] The fourth step is to consider the impact of oil quality on the readings for fuel tanks that have been in use for a period of time. A deep learning model is constructed to comprehensively consider the impact of time-dimensional information such as the number of refuelings, the amount of refueling, and the length of use on the oil residue and inner wall adhesion in the fuel tank. At the same time, the f2(h) reading can also be used as a reference. The features are introduced into the final encoding and decoding stage to supplement the information, improve the accuracy, and even predict future usage.

[0105] The present invention is not limited to the above-mentioned optional implementation modes. Anyone can derive other various forms of products under the inspiration of the present invention. However, no matter what changes are made in the shape or structure, all technical solutions that fall within the scope defined by the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A locomotive fuel tank calibration method based on deep learning, characterized in that: The following steps are involved: S1: Sensor calibration before application: The sensor is used to measure the distance from the sensor installation position to the liquid level. Before actual installation and application, design and prepare test fixtures to complete accuracy tests under fixed height and liquid level reduction conditions; S2: Construct a baseline model based on the three-dimensional model: According to the three-dimensional structure of the fuel tank and the error range generated in the actual production process, a baseline model is constructed. In the baseline model, a curve is fitted to describe the relationship between volume and height; S3: Construct a fuel volume calibration model based on a brand new fuel tank to be used: conduct actual measurements on a brand new fuel tank equipped with diesel from an actual production environment; S4: Construct a regular calibration model based on the oil tank that has been put into use: With long-term and multiple uses, the fitting curve needs to be calibrated. The model is directly constructed using an end-to-end approach. The causes are related to oil residue and inner wall adhesion.

2. The locomotive fuel tank calibration method based on deep learning according to claim 1 is characterized in that: In step S1, the accuracy test under the condition of fixed height and lowered liquid level is specifically as follows: Accuracy test under fixed height conditions: Inject diesel into the fuel consumption test fixture to the total height, let it stand to stabilize the liquid level, and record the stable liquid level height value; The fuel consumption collection board is powered on to read and record the current fuel consumption system output liquid level height and output liquid volume value; Compare the measured value of the liquid level tooling with the height value output by the acquisition board; compare the output liquid volume value with the formula to calculate the oil volume value of the oil tank corresponding to the current height liquid level; Repeat the test several times; When the liquid level is lowered: Inject diesel into the fuel consumption test fixture to the total height, let it stand to stabilize the liquid level, and record the stable liquid level height value; The fuel consumption collection board is powered on to read and record the current fuel consumption system output liquid level height and output liquid volume value The liquid level drops 10mm each time, and the test is repeated several times in total; Compare the measured value of the liquid level each time it descends with the height value output by the acquisition board; compare the output liquid volume value with the formula to calculate the oil volume value of the tank corresponding to the current height liquid level Additional measurements: The lowest measurable height of the sensor; Testing of liquid level rise; Drift of the sensor when it is suspended in the air.

3. The locomotive fuel tank calibration method based on deep learning according to claim 1, characterized in that: In step S3, the calibration scheme of the calibration model is: S31: Setting the oil volume collection unit; S32: Design and install a calibrated independent measurement module. A new fuel tank can be equipped with a height sensor to feed the reading back to the module; S33: Construct the oil volume-altitude correspondence and optimize the sampling point jump problem.

4. The locomotive fuel tank calibration method based on deep learning according to claim 1 is characterized in that: In step S4, the regular calibration model proposes three assumptions: Hypothesis 1: As the number of times of oil use and refueling increases, the amount of oil residue will increase and the degree of inner wall adhesion will be serious. Because the inner wall adhesion of oil residue is difficult to quantify, it is assumed that there is a monotonic relationship between the number of times and the change of fitting relationship. Instead of directly establishing a linear relationship to construct a new fitting model, it is retained as a feature. Hypothesis 2: It is necessary to clarify whether the amount of oil added each time is the same. If it is different, the amount of oil added will directly affect the adhesion of oil residue to the inner wall. Therefore, there is a second-degree correlation with the oil amount-height fitting curve, which is also retained as a feature; Hypothesis 3: The age of the fuel tank is directly listed as a relevant feature.

5. The locomotive fuel tank calibration method based on deep learning according to claim 1, characterized in that: In step S1, the sensor pre-application calibration is based on a machine learning algorithm, which is implemented by the following steps: S11: Data collection; S12: data preprocessing; S13: Model selection; S14: Model training; S15: Model evaluation; S16: Model deployment and optimization.

6. The locomotive fuel tank calibration method based on deep learning according to claim 5 is characterized in that: Step S11 is specifically as follows: under the condition of the same liquid environment, data within different measuring ranges should be collected by repeatedly fluctuating within the same scale; Step S12 is specifically as follows: processing noise data, abnormal data and readings with large differences from actual readings; selecting data at a certain interval to establish a nonlinear model, and using the model features as the final result of feature selection. Different from traditional machine learning based on at least two or more parameters, by repeatedly fluctuating data in a small range at the same liquid level, relevant features are generated to amplify the relevant features of the input data, and the context of the readings is integrated into the decision-making method of the final machine learning algorithm in the form of features.

7. The locomotive fuel tank calibration method based on deep learning according to claim 5 is characterized in that: In step S13, the model to be constructed is one of linear regression, polynomial regression, random forest, gradient boosting tree, support vector machine and neural network, and the final model is determined by the performance of the data on each model.

8. The locomotive fuel tank calibration method based on deep learning according to claim 5, characterized in that: Step S14 is specifically as follows: based on the preprocessed data, the data set is divided into a training set and a validation set according to an 8:2 ratio, and then the training set is cross-validated in a 3-fold or 5-fold manner to select the optimal model. The final selected model is evaluated on the training set.

9. The locomotive fuel tank calibration method based on deep learning according to claim 5, characterized in that: In step S15, the evaluation indicators of model evaluation are set to the following four values: mean square error, root mean square error, mean absolute error, and determination coefficient; all evaluation indicators are calculated separately on the validation set and the test set.

10. The locomotive fuel tank calibration method based on deep learning according to claim 5, characterized in that: In step S16, the trained and finally selected model is deployed in the actual application process. A monitoring module should be set to record data in real time, and spot check to supplement the actual readings, so as to accumulate a certain amount of data for optimization in the later stage.

Images