Small sample machine learning reinforcement method for engine performance prediction and optimization

Abstract

The invention provides a small sample machine learning reinforcement method for engine performance prediction and optimization, which can be applied to the field of machine learning and the field of energy, and comprises the following steps: firstly, generating a data set with sufficient scale and containing multiple working conditions by using random sampling and a genetic algorithm; preprocessing the data, wherein the preprocessing comprises normalization, working condition classification and uniform grid interpolation of two-dimensional data; and working condition identification is realized by using the classification model. Afterwards, features based on combustion parameters are constructed, the combustion parameters with high prediction precision are screened out, and the combustion parameters are divided into a one-dimensional model and a two-dimensional model to be trained respectively and used for predicting engine performance parameters, emission parameters and key parameters such as in-cylinder pressure, heat release rate and temperature. And finally, according to a classification identification result, predicting the engine performance under a normal combustion condition, and not directly predicting the engine performance under an abnormal combustion condition. According to the method, the accuracy and robustness of engine performance prediction under the small sample condition are effectively improved, and the cost of engine optimization design can be effectively reduced.

Description

Technical Field

[0001] The present invention relates to a small sample machine learning reinforcement method for engine performance prediction and optimization, belonging to the technical field of engine performance optimization. Background Art

[0002] With the rapid development of artificial intelligence, machine learning is increasingly being used in the energy sector, such as engine performance prediction and optimization. However, since there is currently no public engine performance data set, the application of machine learning in engine performance prediction is greatly limited, and model training can only be performed through self-built data sets. The limitation of self-built data sets is that the size of the data set usually does not exceed 1,000, because a large size will result in high computational costs. With a small sample data set of this size, the currently popular machine learning models (such as random forests, support vector machines, neural networks, etc.) have poor prediction accuracy, making them difficult to actually apply to engine performance prediction and operation strategy optimization.

[0003] In addition, the random sampling method is inefficient in generating data sets, which will result in a large number of abnormal combustion conditions, reducing the effectiveness of the data sets used for performance prediction. The method of self-built data sets needs to be further improved. Summary of the invention

[0004] Purpose of the invention: In view of the deficiencies in the prior art, the present invention provides a small sample machine learning enhancement method for engine performance prediction and optimization to solve the problems mentioned in the above background technology.

[0005] Technical solution: A small sample machine learning enhancement method for engine performance prediction and optimization, comprising the following steps:

[0006] S1. Dataset generation: The data set contains input features and output parameters. The Latin hypercube sampling method is used to randomly generate input features containing 100-1000 samples, and the output parameters of the engine under the input features are calculated using the numerical simulation method to construct the initial data set.

[0007] S2. Data preprocessing: Normalize the input parameters in the data set. For the one-dimensional parameters in the output parameters, determine whether the average temperature in the cylinder exceeds 1100K and whether the sound intensity (RI) is higher than 5MW / m 2 Divided into three operating conditions: normal combustion, misfire, and detonation;

[0008] The label value of the normal combustion condition is set to 0, the misfire condition is set to 1, and the knock condition is set to 2. Finally, the data is organized into data set 1 for classification model training. The specific engine output parameters of the normal combustion condition are saved and organized into data set 2 for regression model prediction. For the two-dimensional parameters in the data set, uniform grid interpolation is required according to the crankshaft angle.

[0009] S3, working condition identification: the data set obtained in step S2 contains different working conditions, and the classification model in machine learning is used for training to realize the identification and prediction of different working conditions;

[0010] S4, feature construction: The label values ​​in the data set composed of normal combustion conditions obtained in step S2 include engine performance parameters, combustion parameters and emission parameters, among which the combustion parameters are related to the performance parameters but have no direct impact on the engine performance optimization; firstly, a machine learning model is used to train and predict all combustion parameters, and the determination coefficient (R 2 ) parameters greater than 0.95;

[0011] S5: one-dimensional parameter model training: the training set and the test set are controlled to be the same as step S4, and the combustion parameters and engine operation parameters screened out in step S4 are used as new input features for secondary training to predict engine performance parameters and emission parameters;

[0012] S6: Two-dimensional parameter model training: Consistent with step S5, on the same training set and test set, the combustion parameters, engine operating parameters and crankshaft angle values ​​selected in step S4 are used together as new input feature machine learning model training to predict the engine cylinder pressure, heat release rate and temperature parameters;

[0013] S7: Engine performance prediction: For given engine operating parameters, the classification model trained in step S3 is first used for identification. If it is a normal combustion condition, the prediction model trained in steps S5 and S6 is used to predict the specific performance parameters; if it is an abnormal combustion condition, the engine performance parameters are not specifically predicted.

[0014] Preferably, step S1 specifically comprises:

[0015] Based on the initial data set, the genetic algorithm is used to actively generate a new data set until the determination coefficient (R 2 ) is greater than 0.95;

[0016] Among them, the input features are the operating parameters of the engine, including intake temperature, intake pressure, injection amount, injection timing and injection pressure; the output parameters are divided into one-dimensional parameters and two-dimensional parameters. The one-dimensional parameters are the engine's performance parameters, combustion parameters and emission parameters; the two-dimensional parameters are the changes in the engine's combustion parameters with the crankshaft angle.

[0017] Preferably, step S2 specifically comprises:

[0018] In step S2, the formula for normalization processing is: , where X is the normalized value, X i Represents the original value, is the average value, X max is the maximum value, X min is the minimum value;

[0019] In step S2, the engine performance parameters are specifically fuel consumption, combustion noise, efficiency, indicated mean pressure (IMEP), and total output power; the combustion parameters are specifically the maximum explosion pressure in the cylinder, the pressure rise rate, the crankshaft angle corresponding to 10% of the burned mass rate (CA10), the crankshaft angle corresponding to 50% of the burned mass rate (CA50), and the crankshaft angle corresponding to 90% of the burned mass rate (CA90); the emission parameters are specifically nitrogen oxide (NOx) emissions, soot emissions, unburned hydrocarbons (HC), and carbon monoxide (CO) emissions;

[0020] The two-dimensional parameters of the engine specifically refer to the changes in the engine cylinder pressure, heat release rate, and temperature parameters with the crankshaft angle;

[0021] Uniform grid interpolation specifically refers to the interpolation calculation of the cylinder pressure, heat release rate and temperature at the same crankshaft angle interval. The interpolation method is linear interpolation and its formula is:

[0022]

[0023] Among them, x0 and x1 are the horizontal coordinates of the two data points of the original data, y0 and y1 are the vertical coordinates corresponding to the two data points, x is the horizontal coordinate of the interpolation point, and y is the vertical coordinate.

[0024] Preferably, step S4 is specifically:

[0025] In step S4, the determination coefficient R 2 , the formula expression is:

[0026]

[0027] The public expression of MSE is:

[0028]

[0029] In the formula, R 2 The value of is between 0 and 1. The closer it is to 1, the better the model prediction performance. MSE is the mean square error. Its value is between 0 and positive infinity. The closer it is to 0, the better the model prediction performance. N represents the number of data points. Represents the predicted value, yi represents the true value, Represents the mean of the true values.

[0030] Preferably, in step S3 and step S4, the machine learning model is selected as a regression model, including but not limited to support vector machine, random forest, extreme gradient boosting, neural network and Gaussian process regression.

[0031] Preferably, in step S3 and step S4, the machine learning model is selected as an automatic machine learning model, including but not limited to Autogluon, Auto-sklearn, TPOT and H2O models.

[0032] Therefore, the machine learning model is selected as the support vector machine (SVM) for modeling. The optimization goal of SVM is to find a function To make the error between the predicted value and the true value within the set range, the expression of the loss function is:

[0033] L(y,f(x))=max(0,|yf(x)|-∈)

[0034] Among them, K(x i ,x) is the kernel function. Common kernel functions include linear kernel, radial basis function kernel and polynomial kernel. b is the bias term, y is the true value, f(x) is the predicted value, ∈ is the insensitive bandwidth, which is used to control the tolerance of the model to small errors.

[0035] In addition to finding the regression function with the smallest loss function, the goal of SVM is to keep the model simple, that is, the norm of the weights is small, which can be achieved by minimizing the following function:

[0036]

[0037] Among them, ||w|| 2 is the complexity term of the model, C is the regularization parameter used to control the complexity of the model, ξ i and is the slack variable.

[0038] Beneficial effect: The present invention provides a small sample machine learning enhancement method for engine parameter prediction, which can be used to enhance the prediction accuracy of the machine learning model for the one-dimensional performance parameters and the two-dimensional parameters that vary with time of the engine, improve the generalization ability of the machine learning model, and make the small sample machine learning model have a better prediction accuracy in the coefficient of determination (R 2 ) and mean square error (MSE) performance have been greatly improved. It can be applied to multi-objective optimization of engine performance and rapid evaluation of engine cycle changes, which can significantly reduce the time cost of engine optimization design. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0040] Figure 1 It is a flow chart of the present invention.

[0041] Figure 2 Schematic diagram of the data preprocessing process of the present invention.

[0042] Figure 3 It is a principle diagram of the characteristic construction method of the present invention.

[0043] Figure 4 The MSE comparison diagram of the cylinder pressure prediction value and the CFD calculated value before and after the characteristic construction of the present invention is shown. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] In the description of the present invention, it is necessary to understand that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0046] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0047] Taking the performance prediction and optimization of methanol-hydrogen engine as an example, Figure 1 This is a flow chart of a small sample machine learning enhancement method for engine performance prediction and optimization proposed by the present invention, and the specific steps are as follows:

[0048] S1: Dataset generation: Use the Latin hypercube random sampling method to randomly generate an initial data set containing 200 samples within the specified input feature range. Based on this data set, use the genetic algorithm to generate a new generation of data sets containing 40 samples through a series of processes such as screening, sorting, crossover, and mutation. Repeat the iteration 20 times to finally generate a data set containing 1000 samples. The range of input features is shown in Table 1:

[0049] Table 1 Range of input features

[0050]

[0051] S2: Data preprocessing: Normalize the input features in the data set. The output features are divided into one-dimensional parameters and two-dimensional parameters. The one-dimensional parameters distinguish whether the engine misfires according to whether the maximum average temperature in the engine cylinder is greater than 1100K, and the combustion noise (RI) exceeds 5MW / m 2 To determine whether knock occurs, the label value of misfire is 1, knock is 2, and normal combustion condition is 0. Secondly, the cylinder pressure in the data set is uniformly grid interpolated, and the crankshaft angle is interpolated from -143 to 130°CA at an interval of 1 degree;

[0052] S3: Working condition identification: The data set obtained in step S2 contains label values ​​of different working conditions. First, the data set is randomly divided, where the ratio of training set to test set is 4:1. Then, the LightGBM model in machine learning is used for training to realize the identification and prediction of different working conditions. Since there is a sample imbalance problem in the data set, that is, the number of samples of misfire, detonation and normal combustion conditions is unbalanced, the synthetic minority oversampling technology (SMOTE) is used to process the data set, and the final classification model is saved. The prediction results of the classification model are shown in Table 2:

[0053] Table 2 Recognition results of classification model

[0054]

[0055] The results on the diagonal of the table are all correct classifications. If the classification model is not used for preliminary screening, the result of step S6 will be very different from the numerical simulation result.

[0056] S4: Feature construction: The combustion parameters of normal combustion conditions obtained in step S2 are first trained using the support vector machine model, and the training set and the test set are randomly divided into a ratio of 4:1. The Bayesian optimization method is used for hyperparameter optimization, and the number of experiments is 200. The R of different combustion parameters on the test set is calculated. 2 And MSE, as shown in Table 3:

[0057] Table 3 Prediction accuracy of combustion parameters on the test set

[0058]

[0059] S5: One-dimensional parameter model training: During model training, the training set and test set are kept consistent with those in step S4, and the Pmax, PPRR, IMEP, CA50 and engine operating parameters selected in step S4 are used as new input features for secondary training to predict engine fuel consumption, combustion noise, emissions and other parameters, such as Figure 2 Calculate the R of different combustion parameters on the test set 2 and MSE, as shown in Table 4. The table compares the prediction results of the unprocessed AutoGluon automatic machine learning model with the AutoGluon model after using the method of the present invention. It can be clearly seen that the prediction performance of the method of the present invention on a small sample data set is R 2 Both the performance and MSE have been significantly improved.

[0060] Table 4 Comparison of results between common machine learning model and the method of the present invention on the test set

[0061]

[0062] S6: 2D parameter model training: Keep the division of training set and test set consistent with step S5, and use the combustion parameters, engine operation parameters and crankshaft angle values ​​selected in step S4 as new input feature machine learning model training to predict engine cylinder pressure, heat release rate, temperature and other parameters. Finally, all crankshaft angle results are spliced ​​into the final result to achieve the prediction of cylinder pressure curve. The MSE distribution of the model on the test set is as follows: Figure 3 As shown, it can be found that the feature construction method proposed in this paper has an obvious improvement effect on the prediction of the engine cylinder pressure curve.

[0063] S7: Engine performance prediction: Based on the classification model and regression model established in steps S3 and S5, for a specified engine operating condition, the classification model is first used to identify the condition. If it is a misfire or knock case, its value can be set to a larger value, such as 999. If it is a normal combustion condition, the regression model is continued to be called to predict its engine performance parameters.

[0064] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0065] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A small sample machine learning enhancement method for engine performance prediction and optimization, characterized in that: The following steps are involved: S1. Dataset generation: The data set contains input features and output parameters. The Latin hypercube sampling method is used to randomly generate input features containing 100-1000 samples, and the output parameters of the engine under the input features are calculated using the numerical simulation method to construct the initial data set. S2. Data preprocessing: Normalize the input parameters in the data set. For the one-dimensional parameters in the output parameters, determine whether the average temperature in the cylinder exceeds 1100K and whether the sound intensity (RI) is higher than 5MW / m 2 Divided into three operating conditions: normal combustion, misfire, and detonation; The label value of the normal combustion condition is set to 0, the misfire condition is set to 1, and the knock condition is set to 2. Finally, the data is organized into data set 1 for classification model training. The specific engine output parameters of the normal combustion condition are saved and organized into data set 2 for regression model prediction. For the two-dimensional parameters in the data set, uniform grid interpolation is required according to the crankshaft angle. S3, working condition identification: the data set obtained in step S2 contains different working conditions, and the classification model in machine learning is used for training to realize the identification and prediction of different working conditions; S4, feature construction: The label values ​​in the data set composed of normal combustion conditions obtained in step S2 include engine performance parameters, combustion parameters and emission parameters, among which the combustion parameters are related to the performance parameters but have no direct impact on the engine performance optimization; firstly, a machine learning model is used to train and predict all combustion parameters, and the determination coefficient (R 2 ) parameters greater than 0.95; S5: one-dimensional parameter model training: the training set and the test set are controlled to be the same as step S4, and the combustion parameters and engine operation parameters screened out in step S4 are used as new input features for secondary training to predict engine performance parameters and emission parameters; S6: Two-dimensional parameter model training: Consistent with step S5, on the same training set and test set, the combustion parameters, engine operating parameters and crankshaft angle values ​​selected in step S4 are used together as new input feature machine learning model training to predict the engine cylinder pressure, heat release rate and temperature parameters; S7: Engine performance prediction: For given engine operating parameters, the classification model trained in step S3 is first used for identification. If it is a normal combustion condition, the prediction model trained in steps S5 and S6 is used to predict specific performance parameters; If the combustion condition is abnormal, the engine performance parameters are not specifically predicted.

2. The small sample machine learning enhancement method for engine performance prediction and optimization according to claim 1 is characterized in that: Step S1 is specifically as follows: Based on the initial data set, the genetic algorithm is used to actively generate a new data set until the determination coefficient (R 2 ) is greater than 0.95; Among them, the input features are the operating parameters of the engine, including intake temperature, intake pressure, injection amount, injection timing and injection pressure; the output parameters are divided into one-dimensional parameters and two-dimensional parameters. The one-dimensional parameters are the engine's performance parameters, combustion parameters and emission parameters; the two-dimensional parameters are the changes in the engine's combustion parameters with the crankshaft angle.

3. The small sample machine learning enhancement method for engine performance prediction and optimization according to claim 2 is characterized in that: Step S2 is specifically as follows: In step S2, the formula for normalization processing is: , where X is the normalized value, X i Represents the original value, is the average value, X max is the maximum value, X min is the minimum value; In step S2, the engine performance parameters are specifically fuel consumption, combustion noise, efficiency, indicated mean pressure (IMEP), and total output power; the combustion parameters are specifically the maximum explosion pressure in the cylinder, the pressure rise rate, the crankshaft angle corresponding to 10% of the burned mass rate (CA10), the crankshaft angle corresponding to 50% of the burned mass rate (CA50), and the crankshaft angle corresponding to 90% of the burned mass rate (CA90); the emission parameters are specifically nitrogen oxide (NOx) emissions, soot emissions, unburned hydrocarbons (HC), and carbon monoxide (CO) emissions; The two-dimensional parameters of the engine specifically refer to the changes in the engine cylinder pressure, heat release rate, and temperature parameters with the crankshaft angle; Uniform grid interpolation specifically refers to the interpolation calculation of the cylinder pressure, heat release rate and temperature at the same crankshaft angle interval. The interpolation method is linear interpolation and its formula is: Among them, x0 and x1 are the horizontal coordinates of the two data points of the original data, y0 and y1 are the vertical coordinates corresponding to the two data points, x is the horizontal coordinate of the interpolation point, and y is the vertical coordinate.

4. The small sample machine learning enhancement method for engine performance prediction and optimization according to claim 3 is characterized in that: Step S4 is specifically as follows: In step S4, the determination coefficient R 2 , the formula expression is: The public expression of MSE is: In the formula, R 2 The value of is between 0 and 1. The closer it is to 1, the better the model prediction performance. MSE is the mean square error. Its value is between 0 and positive infinity. The closer it is to 0, the better the model prediction performance. N represents the number of data points. Represents the predicted value, y i represents the true value, Represents the mean of the true values.

5. The small sample machine learning enhancement method for engine performance prediction and optimization according to claim 4 is characterized in that: In step S3 and step S4, the machine learning model is selected as a regression model, including but not limited to support vector machine, random forest, extreme gradient boosting, neural network and Gaussian process regression.

6. The small sample machine learning enhancement method for engine performance prediction and optimization according to claim 4 is characterized in that: In step S3 and step S4, the machine learning model is selected as an automatic machine learning model, including but not limited to Autogluon, Auto-sklearn, TPOT and H2O models.

7. The small sample machine learning enhancement method for engine performance prediction and optimization according to claim 5, characterized in that: Therefore, the machine learning model is selected as the support vector machine (SVM) for modeling. The optimization goal of SVM is to find a function Make the error between the predicted value and the true value within the set range, and the expression of the loss function is: L(y,f(x))=max(0,|yf(x)|-∈) Among them, K(x i ,x) is the kernel function. Common kernel functions include linear kernel, radial basis function kernel and polynomial kernel. b is the bias term, y is the true value, f(x) is the predicted value, ∈ is the insensitive bandwidth, which is used to control the tolerance of the model to small errors. In addition to finding the regression function with the smallest loss function, the goal of SVM is to keep the model simple, that is, the norm of the weights is small, which can be achieved by minimizing the following function: Among them, ||w|| 2 is the complexity term of the model, C is the regularization parameter used to control the complexity of the model, ξ i and is the slack variable.

Images