High-pressure fuel system injection quantity long-term online prediction method based on real-time training optimization of kernel extreme learning machine

Abstract

The invention aims to provide a high-pressure fuel system injection quantity long-term online prediction method based on kernel extreme learning machine real-time training optimization, and belongs to the field of fuel injection.The method comprises the steps that firstly, an injection quantity data set and an engine power data set are established, and a kernel extreme learning machine V1 model is trained based on the injection quantity data set; according to different working conditions, a V1 model is used for injection quantity prediction, and the injection quantity is supplemented to a power data set; and training the V2 model based on the new power data set, and optimizing the power prediction capability of the V2 model. In the long-term operation process of the engine, rail pressure and injection pulse width are monitored in real time, the injection amount is predicted online through the V1 model, and power is predicted through the V2 model in combination with real-time data; according to the deviation between the predicted power and the actual power, the injection pulse width is dynamically adjusted based on the MPC algorithm, the injection amount data set is updated, and the V1 model is retrained in real time. According to the invention, high-precision injection quantity long-term online prediction can be realized, the real-time training optimization capability is realized, and shutdown retraining is not needed.

Description

Technical Field

[0001] The present invention relates to a fuel system control method, in particular to a fuel injection prediction method. Background Art

[0002] The dynamic injection stability of the high-pressure fuel system directly affects the power, economy and emissions of the engine. The MAP diagram of the cycle injection amount changing with the pressure and injection pulse width is embedded in the electronic control system. Then, the injection amount is corrected by looking up the table according to the actual working conditions of the engine. This control strategy is suitable for different engines. However, there is a difference between the actual working environment of the multi-cylinder engine and the single-cylinder measurement environment when the calibration map is made, resulting in inconsistency between the common rail pressure and the fuel pressure during the injector injection process, which makes it difficult to maintain a high degree of consistency between the actual injection amount and the expected injection amount.

[0003] In the current injection quantity control method based on model prediction, the premise of accurate control is an accurate injection quantity prediction model. Although the models based on feature recognition or data-driven have improved the accuracy of prediction to a certain extent, they also have obvious limitations. These models usually rely on a large amount of historical data for training, and the historical data fails to cover the changes in future system parameters. During the long-term operation of the engine, the state and environmental factors of the engine will change, and the model does not take into account the impact of changes in key structures and driving parameters caused by the wear of system components and the change of dimensional tolerances over time on the injection quantity, which makes the existing injection quantity prediction method unsatisfactory in the long-term prediction effect of practical applications. Summary of the invention

[0004] The purpose of the present invention is to provide a long-term online prediction method for the injection quantity of a high-pressure fuel system based on real-time training optimization of a nuclear extreme learning machine, which can solve the problem that the performance of the engine changes after long-term operation, resulting in the inability to accurately predict the injection quantity online in real time.

[0005] The object of the present invention is achieved in that:

[0006] The present invention provides a long-term online prediction method for injection quantity of a high-pressure fuel system based on real-time training optimization of a nuclear extreme learning machine, which is characterized by comprising the following steps:

[0007] (1) Experiment on the injector: Use a single injection instrument to measure the injection volume of the injector under different working conditions, and establish an injection volume data set of rail pressure, injection pulse width and injection volume; Experiment on the engine: Measure and record the rail pressure, injection pulse width, air-fuel ratio, intake pressure, intake temperature, and power under different working conditions, and establish an engine power data set;

[0008] (2) Taking rail pressure and injection pulse width as input and injection quantity as output, the nuclear extreme learning machine V1 model is constructed and the V1 model is trained based on the injection quantity dataset;

[0009] (3) Based on the rail pressure and injection pulse width conditions at different speeds and loads in the engine power data set, the nuclear extreme learning machine V1 model is used to predict the injection amount, and the predicted injection amount is added to the engine power data set;

[0010] (4) using the injection amount, air-fuel ratio, intake pressure, and intake temperature as inputs and power as output, constructing a core extreme learning machine V2 model, training the V2 model based on the new engine power data set in step (3), and optimizing the model using a genetic algorithm;

[0011] (5) During the operation of the engine, the injection amount is predicted online through the V1 model based on the real-time rail pressure and injection pulse width of the current cycle, and the power is predicted online through the V2 model based on the injection amount predicted by the V1 model and the real-time air-fuel ratio, intake pressure, and intake temperature;

[0012] (6) According to the deviation between the predicted power and the actual power, the injection pulse width is adjusted based on the MPC algorithm until the deviation between the actual power and the predicted power is less than 5%. The adjusted injection pulse width is then recorded and updated in the injection quantity data set as the injection characteristic after component aging. The V1 model is then retrained in real time based on the updated injection quantity data set to obtain the kernel extreme learning machine V 1_new model, thereby achieving long-term online prediction of injection quantity.

[0013] The present invention may also include:

[0014] 1. The conditions of the injection quantity data set and the engine power data set in step (1) are specifically:

[0015] In the injection quantity data set, the operating conditions include rail pressure conditions and injection pulse width conditions. The rail pressure conditions include all rail pressure conditions with an interval of 10 MPa between the idle rail pressure and the rated rail pressure. The injection pulse width conditions include all injection pulse width conditions with an interval of 0.1 ms between the minimum injection start pulse width and the rated pulse width.

[0016] In the engine power data set, the operating conditions include speed conditions and load conditions. The speed conditions include all speed conditions between idle speed and rated speed with an interval of 10% of the rated speed. The load conditions include all load conditions between no load and full load with an interval of 10%.

[0017] 2. The output functions of the kernel extreme learning machine V1 and V2 models are:

[0018]

[0019] Where x is the input matrix of rail pressure and injection pulse width, n is the dimension of the input matrix, I is the unit matrix, C is the regularization coefficient, β is the weight value between the hidden layer and the injection output matrix, and H is the output matrix of the hidden layer; g(x,x i ) is the Gaussian kernel function, expressed as:

[0020]

[0021] Where γ is the kernel parameter.

[0022] 3. The MPC algorithm in step (6) adjusts the injection pulse width as follows:

[0023] Calculate the predicted power P for the current cycle predict (t) and actual power P actual Deviation of (t):

[0024] P error (t) = P predict (t)-P actual (t)

[0025] The injection pulse width of the next cycle is adjusted by feedback with dynamic characteristics:

[0026] W(t+1)=W(t)+λ1·P error (t)+λ2·∫P error (t)dt+λ3·(P error (t)-P error (t-1))

[0027] Where W(t) is the injection pulse width in the current cycle, and λ1, λ2, and λ3 are feedback adjustment coefficients.

[0028] The advantages of the present invention are:

[0029] (1) The present invention can train and optimize the injection quantity prediction model in real time based on the deviation between the predicted engine power and the real-time power, and realize high-precision long-term online prediction of the injection quantity of the high-pressure fuel system by continuously dynamically learning and updating the model parameters.

[0030] (2) Compared with the traditional injection quantity prediction model, the nuclear extreme learning machine prediction model constructed in the present invention has real-time training and optimization capabilities, and does not require the engine to be shut down to retrain the model, so that the high-pressure fuel system can respond to changes in injection quantity in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the present invention;

[0032] Figure 2Schematic diagram of the training and optimization process of the kernel extreme learning machine V2 model;

[0033] Figure 3 Schematic diagram of the real-time retraining process of the kernel extreme learning machine V1 model. DETAILED DESCRIPTION

[0034] The present invention is described in more detail below with reference to the accompanying drawings:

[0035] Combination Figure 1-3 The present invention provides a long-term online prediction method for injection quantity of a high-pressure fuel system based on real-time training optimization of a nuclear extreme learning machine, comprising the following steps:

[0036] Step S1, conducting an experiment on a high-pressure fuel system test bench, using a single injection instrument to measure the injection amount of the injector under different working conditions, and establishing an injection amount data set of rail pressure, injection pulse width and injection amount; conducting an experiment on an engine test bench, measuring and recording the rail pressure, injection pulse width, air-fuel ratio, intake pressure, intake temperature, and power under different working conditions, and establishing an engine power data set;

[0037] Step S2, taking the rail pressure and the injection pulse width as input and the injection amount as output, constructing a core extreme learning machine V1 model, and training the V1 model based on the injection amount data set;

[0038] Step S3, according to the rail pressure and injection pulse width conditions at different speeds and loads in the engine power data set, the nuclear extreme learning machine V1 model is used to predict the injection amount, and the predicted injection amount is added to the engine power data set;

[0039] Step S4, using injection quantity, air-fuel ratio, intake pressure, and intake temperature as inputs and power as output, constructs a core extreme learning machine V2 model, and trains the V2 model based on the new engine power data set in step S3. Use a genetic algorithm to optimize the model, such as Figure 2 As shown;

[0040] Step S5, during the operation of the engine, based on the real-time rail pressure and injection pulse width of the current cycle, the injection amount is predicted online through the V1 model, and based on the injection amount predicted by the V1 model and the real-time air-fuel ratio, intake pressure, and intake temperature, the power is predicted online through the V2 model;

[0041] Step S6, as Figure 3 As shown in FIG. 1 , according to the deviation between the predicted power and the actual power, the injection pulse width is adjusted based on the MPC algorithm until the deviation between the actual power and the predicted power is less than 5%. After that, the adjusted injection pulse width is recorded and updated in the injection quantity data set as the injection characteristic after component aging. Then, the V1 model is retrained in real time based on the updated injection quantity data set to obtain the kernel extreme learning machine V 1_newmodel, thereby achieving long-term online prediction of injection quantity.

[0042] The data set working conditions in step S1 are specifically:

[0043] In the injection quantity data set, the operating conditions include rail pressure and injection pulse width, wherein the rail pressure conditions include all rail pressure conditions with an interval of 10 MPa between the idle rail pressure and the rated rail pressure, and the injection pulse width conditions include all injection pulse width conditions with an interval of 0.1 ms between the minimum start-up pulse width and the rated pulse width;

[0044] In the engine power data set, the operating conditions include speed and load, wherein the speed condition includes all speed conditions between idle speed and rated speed with an interval of 10% of the rated speed, and the load condition includes all load conditions between no load and full load with an interval of 10%.

[0045] The output functions of the kernel extreme learning machine V1 and V2 models are:

[0046]

[0047] Where x is the input matrix of rail pressure and injection pulse width, n is the dimension of the input matrix, I is the unit matrix, C is the regularization coefficient, β is the weight value between the hidden layer and the injection output matrix, and H is the output matrix of the hidden layer; g(x,x i ) is the Gaussian kernel function, expressed as:

[0048]

[0049] Where γ is the kernel parameter.

[0050] The MPC algorithm in step S6 adjusts the injection pulse width as follows:

[0051] Calculate the predicted power P for the current cycle predict (t) and actual power P actual Deviation of (t):

[0052] P error (t) = P predict (t)-P actual (t) (3)

[0053] The injection pulse width of the next cycle is adjusted by feedback with dynamic characteristics:

[0054] W(t+1)=W(t)+λ1·P error (t)+λ2·∫P error (t)dt+λ3·(P error (t)-P error (t-1)) (4)

[0055] Where W(t) is the injection pulse width in the current cycle, and λ1, λ2, and λ3 are feedback adjustment coefficients.

Claims

1. A long-term online prediction method for injection quantity of a high-pressure fuel system based on real-time training optimization of a nuclear extreme learning machine, characterized by: The following steps are involved: (1) Experiment on the injector: Use a single injection instrument to measure the injection volume of the injector under different working conditions, and establish an injection volume data set of rail pressure, injection pulse width and injection volume; Experiment on the engine: Measure and record the rail pressure, injection pulse width, air-fuel ratio, intake pressure, intake temperature, and power under different working conditions, and establish an engine power data set; (2) Taking rail pressure and injection pulse width as input and injection quantity as output, the nuclear extreme learning machine V1 model is constructed and the V1 model is trained based on the injection quantity dataset; (3) Based on the rail pressure and injection pulse width conditions at different speeds and loads in the engine power data set, the nuclear extreme learning machine V1 model is used to predict the injection amount, and the predicted injection amount is added to the engine power data set; (4) using the injection amount, air-fuel ratio, intake pressure, and intake temperature as inputs and power as output, constructing a core extreme learning machine V2 model, training the V2 model based on the new engine power data set in step (3), and optimizing the model using a genetic algorithm; (5) During the operation of the engine, the injection amount is predicted online through the V1 model based on the real-time rail pressure and injection pulse width of the current cycle, and the power is predicted online through the V2 model based on the injection amount predicted by the V1 model and the real-time air-fuel ratio, intake pressure, and intake temperature; (6) According to the deviation between the predicted power and the actual power, the injection pulse width is adjusted based on the MPC algorithm until the deviation between the actual power and the predicted power is less than 5%. The adjusted injection pulse width is then recorded and updated in the injection quantity data set as the injection characteristic after component aging. The V1 model is then retrained in real time based on the updated injection quantity data set to obtain the kernel extreme learning machine V 1_new model, thereby achieving long-term online prediction of injection quantity.

2. The long-term online prediction method for injection quantity of a high-pressure fuel system based on real-time training optimization of a nuclear extreme learning machine according to claim 1 is characterized by: The conditions of the injection quantity data set and the engine power data set in step (1) are specifically: In the injection quantity data set, the operating conditions include rail pressure conditions and injection pulse width conditions. The rail pressure conditions include all rail pressure conditions with an interval of 10 MPa between the idle rail pressure and the rated rail pressure. The injection pulse width conditions include all injection pulse width conditions with an interval of 0.1 ms between the minimum injection start pulse width and the rated pulse width. In the engine power data set, the operating conditions include speed conditions and load conditions. The speed conditions include all speed conditions between idle speed and rated speed with an interval of 10% of the rated speed. The load conditions include all load conditions between no load and full load with an interval of 10%.

3. The method for long-term online prediction of injection quantity of a high-pressure fuel system based on real-time training optimization of a nuclear extreme learning machine according to claim 1 is characterized by: The output functions of the kernel extreme learning machine V1 and V2 models are: Where x is the input matrix of rail pressure and injection pulse width, n is the dimension of the input matrix, I is the unit matrix, C is the regularization coefficient, β is the weight value between the hidden layer and the injection output matrix, and H is the output matrix of the hidden layer; g(x,x i ) is the Gaussian kernel function, expressed as: Where γ is the kernel parameter.

4. The long-term online prediction method for injection quantity of a high-pressure fuel system based on real-time training optimization of a nuclear extreme learning machine according to claim 1 is characterized by: The MPC algorithm in step (6) adjusts the injection pulse width as follows: Calculate the predicted power P for the current cycle predict (t) and actual power P actual Deviation of (t): P error (t)=P predict (t)-P actual (t) The injection pulse width of the next cycle is adjusted by feedback with dynamic characteristics: W(t+1)=W(t)+λ1·P error (t)+λ2·∫P error (t)dt+λ3·(P error (t)-P error (t-1)) Where W(t) is the injection pulse width in the current cycle, and λ1, λ2, and λ3 are feedback adjustment coefficients.

Images