Method for observing injection quantity of high-pressure fuel system under multiple injections in real time based on model feature transfer learning

Abstract

The invention aims to provide a model feature transfer learning-based injection quantity real-time observation method for a high-pressure fuel system under multiple injections, and belongs to the field of fuel injection.The method comprises the steps that excitation current signals, injector inlet pressure and injection rate data are collected, and a single-injection data set and a multiple-injection data set are established and subjected to batch processing normalization; constructing a single injection rate prediction model based on a double-layer bidirectional long-short-term memory neural network, and optimizing the performance of the model by using a quantum particle swarm algorithm; characteristic parameters of the single-time injection model are migrated into the multi-time injection model, the migrated parameters are frozen, and training is carried out through relatively few multi-time injection data sets; and performing real-time prediction by using a multi-injection rate prediction model based on the excitation current signal measured in real time and the inlet pressure of the injector to obtain the injection quantity of each injection stage. According to the method, the dependence on a big data set is remarkably reduced, and real-time online accurate observation of the injection quantity of each stage of injection under multiple times of injection can be realized.

Description

Technical Field

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

[0002] Multiple injection technology is an important technology in modern high-pressure fuel systems, but inaccurate injection can lead to incomplete combustion and low efficiency, which in turn causes unstable exhaust temperatures in each cylinder and reduces the conversion efficiency of the catalyst in the exhaust aftertreatment device. The current fuel injection control method is mainly based on open-loop control embedded in the injection quantity calibration map, and the injection pulse width of the required injection quantity is determined by the current speed and load. However, there is a difference between the actual working environment of the engine multi-cylinder and the single-cylinder measurement environment when the calibration map is made, resulting in inconsistencies between the common rail pressure and the fuel pressure during the injector injection process. This makes it difficult to maintain consistency between the actual injection quantity and the expected target value.

[0003] In addition, the impact of pressure fluctuations generated during multiple injections on subsequent injection characteristics has not been fully considered. In the previous injection, the rapid opening and closing of the injector needle valve will generate high-frequency expansion waves and compression waves. These pressure waves are reflected and superimposed in the high-pressure pipeline of the system, thereby affecting the injection pressure of the subsequent injection stage and changing the actual injection rate. Therefore, the injection rate prediction of multiple injections is more complicated than that of single injections. Especially under high pressure and high speed conditions, how to accurately predict the injection rate of each stage and perform feedback control is crucial to the overall performance and emission level of the engine. However, the existing single injection rate prediction method based on mathematical model derivation often cannot meet the accurate prediction requirements in the case of multiple injections. In addition, compared with single injection, the data-driven multiple injection rate prediction model requires a larger data set for training, which is usually exponentially times the amount of data required for a single injection, requiring a lot of measurement time and resources, and increasing R&D costs. Summary of the invention

[0004] The purpose of the present invention is to provide a real-time observation method for injection quantity under multiple injections of a high-pressure fuel system based on model feature transfer learning, which can solve the problem that the injection law and injection quantity under multiple injections cannot be accurately predicted online in real time.

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

[0006] The present invention provides a method for real-time observation of injection quantity under multiple injections of a high-pressure fuel system based on model feature transfer learning, which is characterized by comprising the following steps:

[0007] (1) A current clamp is clamped on the wire of the high-speed solenoid valve of the injector, and a pressure sensor is installed at the connection between the injector and the high-pressure fuel pipe to collect the excitation current signal of the high-speed solenoid valve and the pressure signal at the injector inlet. A single injection instrument is used to collect the injection rate of the injector, and the excitation current, injector inlet pressure and injection rate data sets of single injection, pre-main injection, pre-main-post injection under different working conditions are obtained. The data sets are divided into two categories, one is a single injection data set, and the other is a multiple injection data set. The data in all the data sets are batch processed and normalized;

[0008] (2) A single injection rate prediction model based on a double-layer bidirectional long short-term memory neural network (DL-BiLSTM) was constructed, the model was trained using a single injection dataset, and the quantum particle swarm algorithm was used to optimize the model to improve the prediction performance.

[0009] (3) constructing a multiple injection rate prediction model based on a two-layer bidirectional long short-term memory neural network, migrating the model feature parameters such as the first layer weight and bias of the single injection rate prediction model constructed in step (2) to the first layer in the multiple injection rate prediction model, and freezing the migrated model feature parameters. On this basis, the multiple injection rate prediction model is trained using a multiple injection data set;

[0010] (4) Based on the excitation current signal and the pressure signal at the injector inlet measured in real time by the high-pressure fuel system, the multiple injection rate prediction model established in step (3) is used to perform real-time prediction of the injection rate under multiple injections, and the real-time injection amount of each injection stage under multiple injections is obtained by segmented time integration.

[0011] The present invention may also include:

[0012] 1. In the step (1):

[0013] In the single injection data set, the operating conditions include pressure conditions and injection pulse width conditions. The pressure conditions include all pressure conditions with an interval of 10 MPa from the idle pressure to the rated pressure, and the injection pulse width conditions include all injection pulse width conditions with an interval of 0.2 ms from the minimum start-up pulse width to the rated pulse width.

[0014] In the multiple injection data set, the pre-main injection working conditions include pressure conditions, pre-injection pulse width conditions, pre-main injection interval conditions and main injection pulse width conditions, and the pre-main-post-injection working conditions include pressure conditions, pre-injection pulse width conditions, pre-main injection interval conditions, main injection pulse width conditions, main-post-injection interval conditions and post-injection pulse width conditions. The pressure conditions include all pressure conditions between the idle pressure and the rated pressure with an interval of 20MPa, and the pre-injection pulse width conditions include all pressure conditions between the minimum start-injection pulse width and the maximum pre-injection pulse width with an interval of 0.05ms. There are pre-injection pulse width conditions, the main injection pulse width conditions include all main injection pulse width conditions with an interval of 0.2ms between the minimum start-injection pulse width and the rated pulse width, the post-injection pulse width conditions include all post-injection pulse width conditions with an interval of 0.05ms between the minimum start-injection pulse width and the maximum post-injection pulse width, the pre-main injection interval conditions include all interval conditions with an interval of 1ms between the minimum pre-main injection interval and the maximum pre-main injection interval, and the main-post-injection interval conditions include all interval conditions with an interval of 1ms between the minimum main-post-injection interval and the maximum main-post-injection interval.

[0015] 2. The data batch normalization in step (1) is specifically as follows:

[0016]

[0017] In the formula, x * is the processed data; represents the mean of unnormalized data; represents the standard deviation of the data; ε is the margin used to prevent division by zero.

[0018] 3. Step (2) specifically includes:

[0019] (21) Establish a DL-BiLSTM single injection rate prediction model framework, including an input layer, a first BiLSTM layer, a second BiLSTM layer, a fully connected layer, and an output layer; randomly discard a set proportion of neurons between the first BiLSTM layer and the second BiLSTM layer to reduce the model's dependence on specific neurons;

[0020] (22) The data format of the input layer is defined as n driving currents x in the time dimension. a and the pressure x at the ejector inlet b , it is defined that there are m groups of data under different working conditions, and the input matrix of the model is The data format of the output layer is defined as having n injection rates y in the time dimension, and the output matrix is The input matrix and output matrix are expressed as:

[0021]

[0022] (23) The model was trained using a single injection dataset. The input data passed through the first and second BiLSTM layers and then through a fully connected layer for dimensionality reduction and feature extraction, and the predicted injection rate matrix was obtained at the output layer;

[0023] (24) The quantum particle swarm algorithm is used to iteratively optimize the model hyperparameters. The hyperparameters to be optimized are set as the number of units in the first layer BiLSTM and the second layer BiLSTM, the maximum number of iterations, and the initial learning rate. The root mean square error of the prediction results of the validation set is used as the fitness function, and the maximum number of consecutive iterations without improving the global optimal fitness value is used as the termination condition of the optimization. The globally optimal quantum particle position is obtained, and the corresponding optimal hyperparameters are used as the model settings.

[0024] 4. The function form of parameter migration in step (3) is specifically:

[0025] θ T =argm θ inL T (f(X T θ frozen ,θ train ),Y T )

[0026] In the formula, θ T represents the model parameters trained on the multiple injection dataset, L T is the loss function of the multiple injection rate prediction model, θ frozen The first layer BiLSTM weights and biases are migrated and frozen from the single injection rate prediction model, θ train are the remaining parameters to be trained, X T and Y T They are the input data and output data of the multiple injection rate prediction model respectively.

[0027] The advantages of the present invention are:

[0028] (1) The construction process of the multiple injection rate prediction model of the present invention does not require a large amount of data sets. A high-precision prediction model can be obtained by only transferring learning model features and using a relatively small number of multiple injection data sets for training;

[0029] (2) The present invention can predict the injection rate and injection amount of each injection stage under multiple injections in real time online only based on the injector inlet pressure and excitation current, thereby realizing online real-time and accurate observation of injection information under multiple injections. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of the present invention;

[0031] Figure 2 The single injection rate prediction model structure and optimization schematic diagram;

[0032] Figure 3 Schematic diagram of the structure and migration process of the multiple injection rate prediction model. DETAILED DESCRIPTION

[0033] The present invention is described in more detail below with reference to the accompanying drawings:

[0034] Combination Figure 1-3 The present invention provides a method for real-time observation of injection quantity under multiple injections of a high-pressure fuel system based on model feature transfer learning, comprising the following steps:

[0035] Step S1, clamping a current clamp on the wire of the high-speed solenoid valve of the injector, installing a pressure sensor at the connection between the injector and the high-pressure fuel pipe, collecting the excitation current signal of the high-speed solenoid valve and the pressure signal at the injector inlet, and using a single injection instrument to collect the injection rate of the injector, obtaining data sets of the excitation current, injector inlet pressure and injection rate of single injection, pre-main injection, pre-main-post injection under different working conditions, dividing the data sets into two categories, one is a single injection data set, and the other is a multiple injection data set, and performing batch processing and normalization on the data in all the data sets;

[0036] Step S2, constructing a single injection rate prediction model based on a double-layer bidirectional long short-term memory neural network (DL-BiLSTM), training the model using a single injection data set, and optimizing the model using a quantum particle swarm algorithm to improve the prediction performance;

[0037] Step S3, such as Figure 3 As shown, a multiple injection rate prediction model based on a double-layer bidirectional long short-term memory neural network is constructed, and the model feature parameters such as the first layer weight and bias of the single injection rate prediction model constructed in step S2 are migrated to the first layer in the multiple injection rate prediction model, and the migrated model feature parameters are frozen. On this basis, the multiple injection rate prediction model is trained using a multiple injection data set;

[0038] Step S4, based on the excitation current signal measured in real time by the high-pressure fuel system and the pressure signal at the injector inlet, the multiple injection rate prediction model established in step S3 is used to perform real-time prediction of the injection rate under multiple injections, and the real-time injection amount of each injection stage under multiple injections is obtained by segmented time integration.

[0039] The data set working conditions in step S1 are specifically:

[0040] In the single injection data set, the operating conditions include pressure and injection pulse width, wherein the pressure conditions include all pressure conditions with an interval of 10 MPa between the idle pressure and the rated pressure, and the injection pulse width conditions include all injection pulse width conditions with an interval of 0.2 ms between the minimum injection start pulse width and the rated pulse width;

[0041] In the multiple injection data set, the pre-main injection working conditions include pressure, pre-injection pulse width, pre-main injection interval and main injection pulse width, and the pre-main-post injection working conditions include pressure, pre-injection pulse width, pre-main injection interval, main injection pulse width, main-post injection interval and post injection pulse width, wherein the pressure condition includes all pressure conditions between the idle pressure and the rated pressure with an interval of 20MPa, the pre-injection pulse width condition includes all pre-injection pulse width conditions between the minimum start-injection pulse width and the maximum pre-injection pulse width with an interval of 0.05ms, and the main The spray pulse width conditions include all main spray pulse width conditions with an interval of 0.2ms between the minimum start-injection pulse width and the rated pulse width. The post-injection pulse width conditions include all post-injection pulse width conditions with an interval of 0.05ms between the minimum start-injection pulse width and the maximum post-injection pulse width. The pre-main spray interval conditions include all interval conditions with an interval of 1ms between the minimum pre-main spray interval and the maximum pre-main spray interval. The main-post-injection interval conditions include all interval conditions with an interval of 1ms between the minimum main-post-injection interval and the maximum main-post-injection interval.

[0042] The data batch normalization in step S1 is specifically as follows:

[0043]

[0044] In the formula, x * is the processed data; represents the mean of unnormalized data; represents the standard deviation of the data; ε is the margin used to prevent division by zero.

[0045] like Figure 2 As shown, step S2 specifically includes:

[0046] Step S201, establishing a DL-BiLSTM single injection rate prediction model framework, including an input layer, a first BiLSTM layer, a second BiLSTM layer, a fully connected layer, and an output layer; randomly discarding a specific proportion of neurons between the first BiLSTM layer and the second BiLSTM layer to reduce the model's dependence on specific neurons;

[0047] Step S202: define the data format of the input layer as n driving currents x in the time dimension. a and the pressure x at the ejector inlet b , it is defined that there are m groups of data under different working conditions, and the input matrix of the model is The data format of the output layer is defined as having n injection rates y in the time dimension, and the output matrix is The input matrix and output matrix are expressed as:

[0048]

[0049] Step S203, the model is trained using a single injection data set, the input data passes through the first layer BiLSTM and the second layer BiLSTM, and then is subjected to dimensionality reduction and feature extraction through a fully connected layer, and a predicted injection rate matrix is ​​obtained at the output layer;

[0050] Step S204, using quantum particle swarm algorithm to iteratively optimize the model hyperparameters, setting the hyperparameters to be optimized as the number of units of the first layer BiLSTM and the second layer BiLSTM, the maximum number of iterations and the initial learning rate; taking the root mean square error of the prediction results of the validation set as the fitness function, and taking the maximum number of consecutive iterations without improving the global best fitness value as the termination condition of the optimization, the global optimal quantum particle position is obtained, and the corresponding optimal hyperparameters are used as the model settings.

[0051] The function form of parameter migration in step S3 is specifically:

[0052]

[0053] In the formula, θ T represents the model parameters trained on the multiple injection dataset, L T is the loss function of the multiple injection rate prediction model, θ frozen The first layer BiLSTM weights and biases are migrated and frozen from the single injection rate prediction model, θ train are the remaining parameters to be trained, X T and Y T They are the input data and output data of the multiple injection rate prediction model respectively.

Claims

1. A real-time observation method for injection quantity under multiple injections of a high-pressure fuel system based on model feature transfer learning, characterized by: The following steps are involved: (1) A current clamp is clamped on the wire of the high-speed solenoid valve of the injector, and a pressure sensor is installed at the connection between the injector and the high-pressure fuel pipe to collect the excitation current signal of the high-speed solenoid valve and the pressure signal at the injector inlet. A single injection instrument is used to collect the injection rate of the injector, and the excitation current, injector inlet pressure and injection rate data sets of single injection, pre-main injection, pre-main-post injection under different working conditions are obtained. The data sets are divided into two categories, one is a single injection data set, and the other is a multiple injection data set. The data in all the data sets are batch processed and normalized; (2) A single injection rate prediction model based on a double-layer bidirectional long short-term memory neural network (DL-BiLSTM) was constructed, the model was trained using a single injection dataset, and the quantum particle swarm algorithm was used to optimize the model to improve the prediction performance. (3) constructing a multiple injection rate prediction model based on a two-layer bidirectional long short-term memory neural network, migrating the model feature parameters such as the first layer weight and bias of the single injection rate prediction model constructed in step (2) to the first layer in the multiple injection rate prediction model, and freezing the migrated model feature parameters. On this basis, the multiple injection rate prediction model is trained using a multiple injection data set; (4) Based on the excitation current signal and the pressure signal at the injector inlet measured in real time by the high-pressure fuel system, the multiple injection rate prediction model established in step (3) is used to perform real-time prediction of the injection rate under multiple injections, and the real-time injection amount of each injection stage under multiple injections is obtained by segmented time integration.

2. The method for real-time observation of injection quantity under multiple injections of a high-pressure fuel system based on model feature transfer learning according to claim 1 is characterized by: In the step (1): In the single injection data set, the operating conditions include pressure conditions and injection pulse width conditions. The pressure conditions include all pressure conditions with an interval of 10 MPa from the idle pressure to the rated pressure, and the injection pulse width conditions include all injection pulse width conditions with an interval of 0.2 ms from the minimum start-up pulse width to the rated pulse width. In the multiple injection data set, the pre-main injection working conditions include pressure conditions, pre-injection pulse width conditions, pre-main injection interval conditions and main injection pulse width conditions, and the pre-main-post-injection working conditions include pressure conditions, pre-injection pulse width conditions, pre-main injection interval conditions, main injection pulse width conditions, main-post-injection interval conditions and post-injection pulse width conditions. The pressure conditions include all pressure conditions between the idle pressure and the rated pressure with an interval of 20MPa, and the pre-injection pulse width conditions include all pressure conditions between the minimum start-injection pulse width and the maximum pre-injection pulse width with an interval of 0.05ms. There are pre-injection pulse width conditions, the main injection pulse width conditions include all main injection pulse width conditions with an interval of 0.2ms between the minimum start-injection pulse width and the rated pulse width, the post-injection pulse width conditions include all post-injection pulse width conditions with an interval of 0.05ms between the minimum start-injection pulse width and the maximum post-injection pulse width, the pre-main injection interval conditions include all interval conditions with an interval of 1ms between the minimum pre-main injection interval and the maximum pre-main injection interval, and the main-post-injection interval conditions include all interval conditions with an interval of 1ms between the minimum main-post-injection interval and the maximum main-post-injection interval.

3. The method for real-time observation of injection quantity under multiple injections of a high-pressure fuel system based on model feature transfer learning according to claim 1 is characterized by: The data batch normalization in step (1) is specifically as follows: In the formula, x * is the processed data; represents the mean of unnormalized data; represents the standard deviation of the data; ε is the remainder used to prevent division by zero.

4. The method for real-time observation of injection quantity under multiple injections of a high-pressure fuel system based on model feature transfer learning according to claim 1 is characterized by: Step (2) specifically includes: (21) Establish a DL-BiLSTM single injection rate prediction model framework, including an input layer, a first BiLSTM layer, a second BiLSTM layer, a fully connected layer, and an output layer; randomly discard a set proportion of neurons between the first BiLSTM layer and the second BiLSTM layer to reduce the model's dependence on specific neurons; (22) The data format of the input layer is defined as n driving currents x in the time dimension. a and the pressure x at the ejector inlet b , it is defined that there are m groups of data under different working conditions, and the input matrix of the model is The data format of the output layer is defined as having n injection rates y in the time dimension, and the output matrix is The input matrix and output matrix are expressed as: (23) The model was trained using a single injection dataset. The input data passed through the first and second BiLSTM layers and then through a fully connected layer for dimensionality reduction and feature extraction, and the predicted injection rate matrix was obtained at the output layer; (24) The quantum particle swarm algorithm is used to iteratively optimize the model hyperparameters. The hyperparameters to be optimized are set as the number of units in the first layer BiLSTM and the second layer BiLSTM, the maximum number of iterations, and the initial learning rate. The root mean square error of the prediction results of the validation set is used as the fitness function, and the maximum number of consecutive iterations without improving the global optimal fitness value is used as the termination condition of the optimization. The globally optimal quantum particle position is obtained, and the corresponding optimal hyperparameters are used as the model settings.

5. The method for real-time observation of injection quantity under multiple injections of a high-pressure fuel system based on model feature transfer learning according to claim 1 is characterized by: The function form of parameter migration in step (3) is specifically: In the formula, θ T represents the model parameters trained on the multiple injection dataset, L T is the loss function of the multiple injection rate prediction model, θ frozen The first layer BiLSTM weights and biases are migrated and frozen from the single injection rate prediction model. train are the remaining parameters to be trained, X T and Y T They are the input data and output data of the multiple injection rate prediction model respectively.

Images