Static-leakage-free ejector injection quantity accurate control method based on physical information neural network

Abstract

The invention aims to provide a physical information neural network-based static-leakage-free ejector injection quantity accurate control method, which belongs to the field of engines, and comprises the following steps of: firstly, constructing a Transform time sequence model, virtually predicting pressure fluctuation of an inlet of an ejector, and replacing a pressure sensor at the inlet of the ejector; and a high-pressure fuel pipe flow model is established, flow is calculated through measured and predicted pressure fluctuations, and the total fuel amount is obtained through integration. And then, constructing a PINN model, and training based on the total amount data of the fuel flowing out of the high-pressure fuel pipe and the injection amount data measured by the experiment so as to realize accurate prediction of the injection amount. And finally, the injection pulse width is corrected in real time through a sectional PID controller, and closed-loop control over the injection amount is achieved. According to the method, only one pressure sensor needs to be installed on the high-pressure fuel pipe, the complexity and cost of multiple sensors are avoided through virtual prediction, and through combination of physical constraint and data driving, the injection quantity prediction precision and the closed-loop control effect under the complex working condition are guaranteed.

Description

Technical Field

[0001] The present invention relates to a method for controlling an injection device, in particular to a method for controlling an injection amount. Background Art

[0002] The fuel injection system is an important part of improving engine performance and reducing emissions. As a widely used technology, the high-pressure fuel injection system has a core component, the injector, which directly affects the engine's combustion efficiency and emission level in terms of injection volume control. Especially under the current policy of energy conservation and emission reduction, accurately controlling the injection volume of the injector has extremely important practical significance.

[0003] Traditional fuel injection system control methods mostly rely on open-loop control or closed-loop control based on feedback variables such as engine speed. However, the static leakage of the needle valve pair introduces an uncontrolled variable, which affects the system's precise control of the injection amount. Therefore, most current injectors use an internal structure without static leakage. Although static leakage-free injectors have the advantage of reducing leakage, they still rely on traditional control algorithms or empirical models for injection amount control, and these methods often cannot fully utilize the physical information in the injector. At the same time, although pure data-driven methods can learn rules from data, it is difficult to explain and ensure physical consistency, which can easily lead to overfitting and model instability. Summary of the invention

[0004] The purpose of the present invention is to provide a method for accurately controlling the injection amount of a static leak-free injector based on a physical information neural network, which can solve the problems that the injection amount of the injector cannot be accurately predicted online in real time and the injection amount cannot be accurately closed-loop controlled.

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

[0006] The present invention provides a method for accurately controlling the injection amount of a static leak-free injector based on a physical information neural network, which is characterized by comprising the following steps:

[0007] (1) A clamp-on pressure sensor is installed at the injector inlet, and an embedded pressure sensor is installed at the center of the high-pressure fuel pipe between the common rail pipe and the injector. The fuel pressure fluctuations of the clamp-on pressure sensor and the embedded pressure sensor are measured under different injection conditions. After obtaining the pressure fluctuation data set, the clamp-on pressure sensor is removed.

[0008] (2) A Transformer time series model is constructed with the pressure fluctuation at the center of the high-pressure fuel pipe as input and the pressure fluctuation at the injector inlet as output. The model is trained based on the pressure fluctuation dataset and performs virtual prediction of the pressure fluctuation during the actual operation of the high-pressure common rail system.

[0009] (3) A high-pressure fuel pipe flow model is established based on the transmission relationship between the center position of the high-pressure fuel pipe and the injector inlet pressure fluctuation. The pressure signal of the center position of the high-pressure fuel pipe actually measured and the injector inlet pressure fluctuation predicted in real time by the Transformer time series model are used as the input of the high-pressure fuel pipe flow model. The flow rate of the high-pressure fuel pipe is calculated, and the amount of fuel flowing out of the high-pressure fuel pipe is obtained by integration. The amount is taken as the sum of the injector injection amount and the reflux amount, and a data set of the amount of fuel flowing out of the high-pressure fuel pipe under different injection conditions is established.

[0010] (4) According to the ratio of the total area of ​​the spray hole of the injector without static leakage to the area of ​​the OA hole, the proportion of the injection amount and the reflux amount under different injection conditions is calculated to form a database; a physical information neural network PINN model is constructed with the fuel amount flowing out of the high-pressure fuel pipe as input and the injection amount as output. The model adopts a multi-layer feedforward neural network as the framework, and the sum of the physical constraint loss and the regression loss is used as the total loss function of the network model. The data set of the fuel amount flowing out of the high-pressure fuel pipe and the injection amount data under different injection conditions measured by experiments are used for model training;

[0011] (5) Based on the deviation between the actual injection amount predicted in the current cycle and the target injection amount, the injection pulse width of the next cycle is corrected through the segmented PID to achieve closed-loop control of the injection amount.

[0012] The present invention may also include:

[0013] 1. The steps for building the Transformer time series model in step (2) are:

[0014] (21) Divide the data into time segments of the same length, and each time segment is used as a sample; let the input time series P pipe =[P pipe,1 ,P pipe,2 ,...,P pipe,T ], output time series P injector =[P injector,1 ,P injector,2 ,...,P injector,T ], where T represents the time step number, and the pressure fluctuation data is normalized;

[0015] (22) Create position encoding to make the model aware of the information of the time step. The position encoding vector PE is calculated as:

[0016]

[0017] Where i is the dimension index, d model is the embedding dimension;

[0018] (23) Establish the Transformer encoder layer, specifically:

[0019] Capture the long-term temporal dependencies of input time series data through multi-head self-attention layers:

[0020]

[0021] Where Q, K, and V are the query matrix, key matrix, and value matrix corresponding to the input sequence, respectively. k It is the attention dimension;

[0022] A feedforward neural network is established after each attention layer, which is expressed as:

[0023] FFN(x)=max(0,XW 1 +b 1 )W 2 +b 2

[0024] Where W and b are the weight and bias of the feedforward neural network, respectively, and the subscripts 1 and 2 represent the first and second layers in the feedforward network, respectively;

[0025] On this basis, residual connections are added after each layer, and the outputs of the self-attention layer and the feedforward neural network layer are expressed as:

[0026] Output=LayerNorm(X+Attention)

[0027] Output = LayerNorm (X + FFN)

[0028] (24) Construct the loss function of the model, specifically:

[0029]

[0030] In the formula, Characterize the input time series and predict the output injector inlet pressure fluctuation through the Transformer model; i is the actual value of the injector inlet pressure fluctuation. Finally, the model is trained based on the pressure fluctuation dataset.

[0031] 2. The steps for constructing the high-pressure fuel pipe flow model in step (3) are as follows:

[0032] Establish the momentum balance equation in the high-pressure fuel pipe:

[0033]

[0034] Where ρ is the fuel density, u is the flow rate, P is the pressure, D is the pipe diameter, and γ is the wall shear stress;

[0035] Integrate the distance L between the center of the high-pressure fuel pipe and the injector inlet, and then multiply it by the pipe cross-sectional area A to obtain the mass flow rate out of the high-pressure fuel pipe:

[0036]

[0037] In the formula, ξ is a constant related to the speed of sound, and t is time;

[0038] The fuel mass of the high-pressure fuel pipe is expressed as:

[0039] M pipe-out =∫Q pipe-out dt.

[0040] 3. The loss function of the PINN model in step (4) is expressed as:

[0041] Physical constraint loss E 1 It is expressed as:

[0042]

[0043] In the formula, represents the injection amount predicted by the PINN model, κ is the proportion of the injection amount and the recirculation amount recorded in the database;

[0044] Regression loss constraint E 2 It is expressed as:

[0045] The total loss function of the model is: E = λ 1 E 1 +E 2 , where λ 1 is the weight coefficient of the physical constraint loss.

[0046] 4. The specific step of the segmented PID control in step (5) is:

[0047] According to the deviation between the actual injection amount and the target injection amount, different PID control parameters are applied in different stages according to the deviation, as follows:

[0048]

[0049] Where ΔPW(k) is the injection pulse width correction at the kth injection, e(k-1) is the injection amount deviation at the k-1th injection, and σ 1 , σ 2 , σ 3 are the respective deviation thresholds of injection quantity deviation segments of different degrees, is the correction coefficient of the first deviation segment, is the correction coefficient of the second deviation segment, is the correction factor for the third deviation segment;

[0050] The segmented PID outputs the control signal for the kth injection based on the calculated injection pulse width correction value:

[0051] PW(k)=PW(k-1)+ΔPW(k).

[0052] The advantages of the present invention are:

[0053] (1) The present invention organically combines physical models and data-driven methods, avoiding the error accumulation caused by traditional methods that rely on experience adjustment;

[0054] (2) The present invention only requires one pressure sensor to be installed on the high-pressure fuel pipe to virtually predict the pressure fluctuation at the injector inlet, which effectively avoids the need to add additional sensors in actual working conditions and reduces complexity and cost;

[0055] (3) The present invention performs segmented PID injection pulse width correction according to the injection quantity deviation threshold, thereby ensuring that the system can be adjusted quickly and enhancing the stability of injection quantity control. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the present invention;

[0057] Figure 2 This is a schematic diagram of the test device connection. DETAILED DESCRIPTION

[0058] The present invention is described in more detail below with reference to the accompanying drawings:

[0059] Combination Figure 1-2 The present invention provides a method for accurately controlling the injection amount of a static leak-free injector based on a physical information neural network, comprising the following steps:

[0060] Step S1, such as Figure 2 As shown, a clamp-type pressure sensor is installed at the injector inlet, and an embedded pressure sensor is installed at the center of the high-pressure fuel pipe between the common rail pipe and the injector. The fuel pressure fluctuations at the above two locations are measured under different injection conditions. After the pressure fluctuation data set is obtained, the clamp-type pressure sensor is removed.

[0061] Step S2, constructing a Transformer time series model with the pressure fluctuation at the center of the high-pressure fuel pipe as input and the pressure fluctuation at the injector inlet as output, and training based on the pressure fluctuation data set. The Transformer time series model is used for virtual prediction of pressure fluctuation during actual operation of the high-pressure common rail system, replacing the clamp-on pressure sensor.

[0062] Step S3, according to the Euler partial differential equation of generalized mass and momentum, a high-pressure fuel pipe flow model is established based on the transmission relationship between the center position of the high-pressure fuel pipe and the injector inlet pressure fluctuation, the pressure signal actually measured at the center position of the high-pressure fuel pipe and the injector inlet pressure fluctuation predicted in real time by the Transformer time series model are used as inputs of the high-pressure fuel pipe flow model, the flow of the high-pressure fuel pipe is calculated, and the amount of fuel flowing out of the high-pressure fuel pipe is obtained by integration, which is taken as the sum of the injector injection amount and the reflux amount, and a data set of the amount of fuel flowing out of the high-pressure fuel pipe under different injection conditions is established.

[0063] Step S4, according to the ratio of the total area of ​​the spray hole of the non-static leak injector to the area of ​​the OA hole, calculate the proportion of the injection amount and the reflux amount under different injection conditions to form a database. Construct a PINN model with the amount of fuel flowing out of the high-pressure fuel pipe as input and the injection amount as output. The model uses a multi-layer feedforward neural network as the framework, and uses the sum of the physical constraint loss and the regression loss as the total loss function of the network model. The data set of the amount of fuel flowing out of the high-pressure fuel pipe and the injection amount data under different injection conditions measured by experiments are used as model training, so that the PINN model obtained after the final training can accurately predict the injection amount.

[0064] Step S5, in the actual working process, based on the deviation between the actual injection amount predicted in the current cycle and the target injection amount, the injection pulse width of the next cycle is corrected through the segmented PID to achieve closed-loop control of the injection amount.

[0065] The specific steps of building the Transformer time series model in step S2 are:

[0066] Step (a) divides the data into time segments of the same length, and each time segment is used as a sample. Suppose the input time series P pipe =[P pipe,1 ,P pipe,2 ,...,P pipe,T ], output time series P injector =[P injector,1 ,P injector,2 ,...,P injector,T ], where T represents the time step. The pressure fluctuation data is normalized to improve the stability of the model.

[0067] Step (b) creates a position encoding to make the model aware of the time step information. The position encoding vector PE is calculated as:

[0068]

[0069] Where i is the dimension index, d model is the embedding dimension.

[0070] Step (c) builds the Transformer encoder layer, specifically:

[0071] Capture the long-term temporal dependencies of input time series data through multi-head self-attention layers:

[0072]

[0073] Where Q, K, and V are the query matrix, key matrix, and value matrix corresponding to the input sequence, respectively. k It is the attention dimension.

[0074] A feedforward neural network is established after each attention layer to improve the nonlinear representation ability of the model, which is expressed as:

[0075] FFN(x)=max(0,XW 1 +b 1 )W 2 +b 2 (3)

[0076] Where W and b are the weight and bias of the feedforward neural network, respectively, and the subscripts 1 and 2 represent the first and second layers in the feedforward network, respectively.

[0077] On this basis, a residual connection is added after each layer to prevent the gradient from disappearing and improve the training stability. The outputs of the self-attention layer and the feedforward neural network layer are expressed as:

[0078] Output=LayerNorm(X+Attention) (4)

[0079] Output = LayerNorm (X + FFN) (5)

[0080] Step (d) constructs the loss function of the model to evaluate the prediction error of the model, specifically:

[0081]

[0082] In the formula, Characterize the input time series and predict the output injector inlet pressure fluctuation through the Transformer model; i is the actual value of the injector inlet pressure fluctuation. Finally, the model is trained based on the pressure fluctuation dataset.

[0083] The high-pressure fuel pipe flow model is constructed in step S3 as follows:

[0084] Establish the momentum balance equation in the high-pressure fuel pipe:

[0085]

[0086] Where ρ is the fuel density, u is the flow rate, P is the pressure, D is the pipe diameter, and γ is the wall shear stress.

[0087] Integrate the distance L between the center of the high-pressure fuel pipe and the injector inlet, and then multiply it by the pipe cross-sectional area A to obtain the mass flow rate out of the high-pressure fuel pipe:

[0088]

[0089] Where ξ is a constant related to the speed of sound, and t is time.

[0090] The fuel mass of the high-pressure fuel pipe is expressed as:

[0091] M pipe-out =∫Q pipe-out dt (9)

[0092] The loss function of the PINN model in step S4 is expressed as:

[0093] Physical constraint loss E 1 It is expressed as:

[0094]

[0095] In the formula, represents the injection amount predicted by the PINN model, and κ is the proportion of the injection amount and the recirculation amount recorded in the database.

[0096] Regression loss constraint E 2 It is expressed as:

[0097] In summary, the total loss function of the model is: E = λ 1 E 1 +E 2 (12)

[0098] In the formula, λ 1 is the weight coefficient of the physical constraint loss.

[0099] The segmented PID control in step S5 is specifically as follows:

[0100] According to the deviation between the actual injection amount and the target injection amount, different PID control parameters are applied in different stages according to the deviation. The specific formula is as follows:

[0101]

[0102] Where ΔPW(k) is the injection pulse width correction at the kth injection, e(k-1) is the injection amount deviation at the k-1th injection, and σ 1 , σ 2 , σ3 are the respective deviation thresholds of injection quantity deviation segments of different degrees, is the correction coefficient of the first deviation segment, is the correction coefficient of the second deviation segment, is the correction coefficient for the third deviation segment.

[0103] The segmented PID outputs the control signal for the kth injection based on the calculated injection pulse width correction value:

[0104] PW(k)=PW(k-1)+ΔPW(k) (14).

Claims

1. A method for precise control of injection quantity of a static leak-free injector based on a physical information neural network, characterized by: The following steps are involved: (1) A clamp-on pressure sensor is installed at the injector inlet, and an embedded pressure sensor is installed at the center of the high-pressure fuel pipe between the common rail pipe and the injector. The fuel pressure fluctuations of the clamp-on pressure sensor and the embedded pressure sensor are measured under different injection conditions. After obtaining the pressure fluctuation data set, the clamp-on pressure sensor is removed. (2) A Transformer time series model is constructed with the pressure fluctuation at the center of the high-pressure fuel pipe as input and the pressure fluctuation at the injector inlet as output. The model is trained based on the pressure fluctuation dataset and performs virtual prediction of the pressure fluctuation during the actual operation of the high-pressure common rail system. (3) A high-pressure fuel pipe flow model is established based on the transmission relationship between the center position of the high-pressure fuel pipe and the injector inlet pressure fluctuation. The pressure signal of the center position of the high-pressure fuel pipe actually measured and the injector inlet pressure fluctuation predicted in real time by the Transformer time series model are used as the input of the high-pressure fuel pipe flow model. The flow rate of the high-pressure fuel pipe is calculated, and the amount of fuel flowing out of the high-pressure fuel pipe is obtained by integration. The amount is taken as the sum of the injector injection amount and the reflux amount, and a data set of the amount of fuel flowing out of the high-pressure fuel pipe under different injection conditions is established. (4) According to the ratio of the total area of ​​the spray hole of the injector without static leakage to the area of ​​the OA hole, the proportion of the injection amount and the reflux amount under different injection conditions is calculated to form a database; a physical information neural network PINN model is constructed with the fuel amount flowing out of the high-pressure fuel pipe as input and the injection amount as output. The model adopts a multi-layer feedforward neural network as the framework, and the sum of the physical constraint loss and the regression loss is used as the total loss function of the network model. The data set of the fuel amount flowing out of the high-pressure fuel pipe and the injection amount data under different injection conditions measured by experiments are used for model training; (5) Based on the deviation between the actual injection amount predicted in the current cycle and the target injection amount, the injection pulse width of the next cycle is corrected through the segmented PID to achieve closed-loop control of the injection amount.

2. According to claim 1, a method for precise control of injection quantity of a static leak-free injector based on a physical information neural network is characterized by: The steps for building the Transformer time series model in step (2) are: (21) Divide the data into time segments of the same length, and each time segment is used as a sample; let the input time series P pipe =[P pipe,1 ,P pipe,2 ,...,P pipe,T ], output time series P injector =[P injector,1 ,P injector,2 ,…,P injector,T ], where T represents the time step number, and the pressure fluctuation data is normalized; (22) Create position encoding to make the model aware of the information of the time step. The position encoding vector PE is calculated as: Where i is the dimension index, d model is the embedding dimension; (23) Establish the Transformer encoder layer, specifically: Capture the long-term temporal dependencies of input time series data through multi-head self-attention layers: Where Q, K, and V are the query matrix, key matrix, and value matrix corresponding to the input sequence, respectively. k It is the attention dimension; A feedforward neural network is established after each attention layer, which is expressed as: FFN(x)=max(0,XW1+b1)W2+b2 Where W and b are the weight and bias of the feedforward neural network, respectively, and the subscripts 1 and 2 represent the first and second layers in the feedforward network, respectively; On this basis, residual connections are added after each layer, and the outputs of the self-attention layer and the feedforward neural network layer are expressed as: Output=LayerNorm(X+Attention) Output = LayerNorm (X + FFN) (24) Construct the loss function of the model, specifically: In the formula, Characterize the input time series and predict the output injector inlet pressure fluctuation through the Transformer model; i is the actual value of the injector inlet pressure fluctuation. Finally, the model is trained based on the pressure fluctuation dataset.

3. The method for precise control of injection quantity of a static leak-free injector based on a physical information neural network according to claim 1, characterized in that: The steps for constructing the high-pressure fuel pipe flow model in step (3) are specifically as follows: Establish the momentum balance equation in the high-pressure fuel pipe: Where ρ is the fuel density, u is the flow rate, P is the pressure, D is the pipe diameter, and γ is the wall shear stress; Integrate the distance L between the center of the high-pressure fuel pipe and the injector inlet, and then multiply it by the pipe cross-sectional area A to obtain the mass flow rate out of the high-pressure fuel pipe: In the formula, ξ is a constant related to the speed of sound, and t is time; The fuel mass of the high-pressure fuel pipe is expressed as: M pipe-out =∫Q pipe-out dt。 4. The method for precise control of injection quantity of static leak-free injector based on physical information neural network according to claim 1, characterized in that: The loss function of the PINN model in step (4) is expressed as: The physical constraint loss E1 is expressed as: In the formula, represents the injection amount predicted by the PINN model, κ is the proportion of the injection amount and the recirculation amount recorded in the database; The regression loss constraint E2 is expressed as: The total loss function of the model is: E=λ1E1+E2, where λ1 is the weight coefficient of the physical constraint loss.

5. The method for precise control of injection quantity of static leak-free injector based on physical information neural network according to claim 1, characterized in that: The specific step of the segmented PID control in step (5) is: According to the deviation between the actual injection amount and the target injection amount, different PID control parameters are applied in different stages according to the deviation, as follows: Where ΔPW(k) is the injection pulse width correction at the kth injection, e(k-1) is the injection quantity deviation at the k-1th injection, σ1, σ2, σ3 are the respective deviation thresholds of different degrees of injection quantity deviation segments, is the correction coefficient of the first deviation segment, is the correction coefficient of the second deviation segment, is the correction factor for the third deviation segment; The segmented PID outputs the control signal for the kth injection based on the calculated injection pulse width correction value: PW(k)=PW(k-1)+ΔPW(k).

Images