DE102022207540A1 - Method for determining a pressure in a fuel tank - Google Patents

Abstract

The invention relates to a method for determining a pressure (p) in a fuel tank (10) by evaluating an electrical current from a diagnostic pump connected to the fuel tank (10). A data-based model is used to determine this.

Description

The present invention relates to a method for determining a pressure in a fuel tank. The present invention also relates to a computer program which is set up to carry out each step of the method, as well as a machine-readable storage medium on which the computer program is stored. Finally, the invention relates to an electronic control device which is set up to carry out the method.

State of the art

For gasoline engine vehicles that use liquid fuel, some countries require that the fuel tank be diagnosed for leaks. This can be done, for example, using an overprinting process. The overpressure process involves pressurizing the fuel tank, which is connected to an activated carbon filter. If this pressure reaches a certain threshold, the system is sufficiently tight. If this threshold is not reached, there is a leak. In order to make it easier to check this situation, the effective leakage area is calculated based on the calculated pressure signal. If this leakage area is larger than a threshold value, there is a leaky system.

An overprinting process is, for example, from the DE 10 2013 221 794 A1 known. Here, an overpressure is generated in the activated carbon filter and in the fuel tank using a diagnostic pump on an activated carbon filter in the fuel tank. The pump's control current required for this is used to determine the resistance against which the pump is working and from this it is concluded that there is a leak. Alternatively, the time course of the absolute pressure in the fuel tank can be determined. However, a pressure sensor is required for this.

If a pressure sensor is not required for cost reasons, it is possible to determine the pressure using a physical model from the electrical control current of the pump. However, this current fluctuates. Fluctuations can be caused, for example, by changing an electrical supply voltage to the pump. These can occur in the wake of an internal combustion engine of the motor vehicle, for example when electric actuators are controlled. Effects due to internal pump friction that are difficult to reproduce can also lead to fluctuations in the control current.

Disclosure of the invention

A method for determining a pressure in a fuel tank is provided. A diagnostic pump is connected to the fuel tank. In particular, the fuel tank is connected to an activated carbon filter, which is connected to the diagnostic pump. Due to the fluidic connection between the fuel tank and the activated carbon filter, the pressure in the activated carbon filter corresponds to the pressure in the fuel tank. The activated carbon filter is connected in particular to an intake manifold of an internal combustion engine via a tank ventilation valve. In order to prevent pressure equalization in the suction pipe, the tank ventilation valve is closed in the process. Just as with the use of a physical model to determine the pressure, the method provides for evaluating the electrical current with which the diagnostic pump is controlled. However, a data-based model is used instead of a physical model for the determination.

For the data-based model, it is in particular provided that data for the data-based model is collected in a wake of an internal combustion engine that is supplied with fuel from the fuel tank and that the pressure is determined as soon as a stationary state with constant pressure has been reached. For this purpose, more storage capacity is required in an electronic computing device than when using a physical model.

However, in view of the constantly increasing storage volume of modern computer memories, the implementation of this method in the electronic control unit of a motor vehicle is possible. This allows the pressure to be determined more accurately than using a physical model, since the data-based model is less sensitive to fluctuations in the electrical current.

The data-based model preferably uses an electrical current from the diagnostic pump, an electrical voltage and a time or a counter as input variables. By taking the electrical voltage into account in addition to the electrical current, it can be recognized whether a change in current is based on a change in pressure or on a change in voltage. The electrical voltage is in particular the voltage of a battery that supplies the diagnostic pump, or an intermediate circuit voltage. The acquisition of the time or an alternative counter, which in particular starts when a follow-up begins and ends when a termination condition of the method is detected, further makes it possible to correlate the acquired data with time until a steady state is reached.

The data-based model preferably has two submodels. The first submodel has a time-dependent model parameter and the second submodel has a time-independent model parameter. Because the data-based model has two interconnected submodels, it is possible to implement the submodels in different layers, for example when implementing the data-based model in a neural network.

If the data-based model is implemented in a neural network, the first submodel has in particular one non-linear layer or two non-linear layers. The second submodel in particular has a linear layer.

The two model parameters are preferably determined in a training phase of the data-based model. They can then be stored on a machine-readable storage medium in order to make them available to the data-based model when carrying out the method.

The data-based model is trained in particular using a test vehicle that has a pressure sensor that is fluidly connected to the fuel tank. This can be arranged in particular in the fuel tank, in the activated carbon filter or in a line between the fuel tank and the activated carbon filter. In addition, the fuel tank of the test vehicle contains a hole that can be specifically opened or closed using a panel in order to be able to set a defined leakage area in individual tests when the engine is running.

When determining the time-dependent model parameter, a loss function is preferably used. This has, as an offset, a sum of the amounts of differences between successive weights of the first submodel. This makes it possible to filter out any oscillations that occur in the input data during the training phase.

If sufficient computing capacity is available, it is further preferred that the data-based model is combined with a physical model in a steady state. The hybrid model obtained in this way enables the pressure to be determined even more precisely.

The computer program is set up to carry out every step of the method, especially if it runs on a computing device or on an electronic control device. For this purpose it is stored on the machine-readable storage medium.

By installing the computer program on a conventional electronic control device, the electronic control device is obtained, which is set up to use the method to determine the pressure in an activated carbon filter of a fuel tank.

Brief description of the drawings

• 1 zeigt schematisch einen Aktivkohlefilter, dessen Innendruck mittels Ausführungsbeispielen des erfindungsgemäßen Verfahrens ermittelt werden kann.
• 2 zeigt schematisch Massenströme in dem Aktivkohlefilter und einem damit verbundenen Kraftstofftank gemäß 1.
• 3 zeigt schematisch ein datenbasiertes Modell, das in einem Ausführungsbeispiel des erfindungsgemäßen Verfahrens verwendet wird.
• 4 zeigt schematisch ein neuronales Netzwerk, das in einem Verfahren gemäß der Erfindung verwendet wird.

Exemplary embodiments of the invention are shown in the drawings and are explained in more detail in the following description.

• 1 shows schematically an activated carbon filter, the internal pressure of which can be determined using exemplary embodiments of the method according to the invention.
• 2 shows schematically mass flows in the activated carbon filter and a fuel tank connected to it 1 .
• 3 shows schematically a data-based model that is used in an exemplary embodiment of the method according to the invention.
• 4 shows schematically a neural network used in a method according to the invention.

Embodiments of the invention

1 shows a fuel tank 10 of a motor vehicle, which is fluidly connected to an activated carbon filter 12 by means of a first line 11. The activated carbon filter 12 contains activated carbon 13. By means of a second line 14, it is fluidly connected to an intake pipe, not shown, of an internal combustion engine of the motor vehicle. A tank ventilation valve 15 is arranged in the second line 14. Fuel vapors 21 gas out of fuel 20, which is stored in the fuel tank 10. These reach the activated carbon filter 12 via the first line 11 and are adsorbed there by the activated carbon 13. When the tank ventilation valve 15 is opened, the activated carbon 13 can be flushed by directing the adsorbed fuel vapors 21 through the second line 14 and the intake manifold into the internal combustion engine and burning them there. If the tank ventilation valve 15 is closed, the pressure in the fuel tank 10 and in the activated carbon filter 12 match. In order to be able to determine this pressure in the wake of the internal combustion engine, a diagnostic pump 16 is connected to the activated carbon filter 12. The diagnostic pump 16 is supplied with an electrical voltage U by a battery 17. An electrical current I, which is used to control the Diagnostic pump 16 flows is reported to an electronic control unit 18.

In 2 Mass flows in the fuel tank 10 and in the activated carbon filter 12 are shown. There is a pressure p in the activated carbon filter 12, which corresponds to the pressure in the fuel tank 10. Using the diagnostic pump 16, a mass flow ṁ _in is introduced into the activated carbon filter 12. This causes air to enter the activated carbon filter 12. Since the air volume V in the fuel tank 10, the first line 11 and the activated carbon filter 12, which has a temperature T, remains unchanged, this creates an overpressure in the activated carbon filter 12 and thus also in the fuel tank 10. One out of the Fuel 20 in the fuel tank 10 outgassing mass flow ṁ _ex of fuel vapors 21 is thereby reduced. Due to a leak in the fuel tank 10, a mixture of fuel vapors 21 and air can leave the fuel tank 10 into the environment as a mass flow ṁ _out .

To determine the change in pressure p over time t using a conventional physical model, the pressure change p(t) can be calculated using the general gas law according to Formula 1: $$\dot{\text{p}}\left(\text{t}\right)=\frac{\text{R}\cdot \text{T}}{\text{v}}\cdot \left({\dot{\text{m}}}_{\text{in}}\left(\text{t}\right)-{\dot{\text{m}}}_{\text{out}}\left(\text{t}\right)+{\dot{\text{m}}}_{\text{ex}}\left(\text{t}\right)\right)$$

The mass flow ṁ _in (t) is directly proportional to the electrical current I(t) of the control of the diagnostic pump 16. The escaping mass flow ṁ _out (t) is proportional to the pressure p in the fuel tank 10, in the first line 11 and in the activated carbon filter 12 as well to the area A of all leaks in the fuel tank 10. Using two proportionality factors k ₁ and k ₂ , formula 2 results: $$\dot{\text{p}}\left(\text{t}\right)=\frac{\text{R}\cdot {\text{T}}_{\text{in}}}{{\text{v}}_{\text{in}}}\cdot \left({\text{k}}_1\cdot \text{I}\left(\text{t}\right)-\text{A}\cdot {\text{k}}_2\cdot \text{p}\left(\text{t}\right)\cdot \sqrt{\frac{2}{\text{R}\cdot \text{T}}}+{\dot{\text{m}}}_{\text{ex}}\left(\text{t}\right)\right)$$

Given the fact that the temperature T in the fuel tank 10 and the leakage area A are unknown and the volume flow ṁ _ex of the outgassing fuel vapors 21 is generally negligible, formula 3 can be formulated by introducing two new proportionality factors k ₃ and k ₄ : $$\dot{\text{p}}\left(\text{t}\right)={\text{k}}_3\cdot \text{I}\left(\text{t}\right)-{\text{k}}_4\cdot \text{p}\left(\text{t}\right)$$

The pressure change p(t) is negligible in the stationary state. This results in formula 4 for calculating the pressure p(t) by introducing a fifth proportionality factor k ₅ : $$\text{p}\left(\text{t}\right)={\text{k}}_5\cdot \text{I}\left(\text{t}\right)$$

In one embodiment of the method according to the invention it is provided to determine the pressure p using a data-based model f _θ . This is in 3 shown. According to Formula 5, the data-based model f _θ uses the electrical current I, the electrical voltage U and the time t as input variables: $$\text{p}\left(\text{t}\right)={\text{f}}_{\mathrm{\theta }}\left(\text{I}\left(\text{t}\right),\text{U}\left(\text{t}\right),\text{t}\right)+\mathrm{\sigma }\left(\text{t}\right)$$

The value σ(t) represents model error and can be ignored if it is sufficiently small.

According to Formula 6, the data-based model f _θ has two submodels h _θh(t) and g _θg : $${\text{f}}_{\mathrm{\theta }}={\text{H}}_{\mathrm{\theta }\text{H}\left(\text{t}\right)}\circ {\text{G}}_{\mathrm{\theta }\text{G}}$$

The first submodel h _θh(t) is time dependent. According to Formula 7, it has a weight w(t) and an offset b(t) for each element x of an output vector: $${\text{H}}_{\mathrm{\theta }\text{H}\left(\text{t}\right)}\left(\text{x}\right)=\text{w}\left(\text{t}\right)\cdot \text{x}+\text{b}\left(\text{t}\right)$$

In the next time step t + 1, formula 8 results: $${\text{H}}_{\mathrm{\theta }\text{H}\left(\text{t}+1\right)}\left(\text{x}\right)=\text{w}\left(\text{t}+1\right)\cdot \text{x}+\text{b}\left(\text{t}+1\right)$$

c bezeichnet hierbei eine Konstante. Da der Betrag der Differenz zweier aufeinanderfolgender Gewichte |w(t + 1) - w(t)| gegen 0 geht, können potenzielle Oszillationen von Eingabewerten bei der Ausgabe herausgefiltert werden. Hierzu wird beim Trainieren des Untermodells h_θh(t) eine Verlustfunktion $L\left(\mathrm{\theta }\text{h}\right)$

eingeführt, die gemäß Formel 10 mit einem zusätzlichen Offset beaufschlagt wird, um eine zu registrierende Verlustfunktion $L_{\text{reg}}\left(\mathrm{\theta }\text{h}\right)$

zu erhalten: $$L_{\text{reg}}\left(\mathrm{\theta }\text{h}\right)=L\left(\mathrm{\theta }\text{h}\right)+{\displaystyle \sum _{\text{t}}\left|\text{w}\left(\text{t}+1\right)-\text{w}\left(\text{t}\right)\right|}$$

The ratio of the successive weights w(t) and w(t + 1) then flows into Formula 9: $${\text{H}}_{\mathrm{\theta }\text{H}\left(\text{t}+1\right)}\left(\text{x}\right)=\frac{\text{w}\left(\text{t}+1\right)}{\text{w}\left(\text{t}\right)}\cdot {\text{H}}_{\mathrm{\theta }\text{H}\left(\text{t}\right)}\left(\text{x}\right)+\text{c}$$

c denotes a constant. Since the magnitude of the difference between two consecutive weights |w(t + 1) - w(t)| approaches 0, potential oscillations of input values can be filtered out in the output. For this purpose, h _θh(t) becomes a loss function when training the submodel $L\left(\mathrm{\theta }\text{H}\right)$

introduced, which is subjected to an additional offset according to Formula 10 in order to create a loss function to be registered $L_{\text{reg}}\left(\mathrm{\theta }\text{H}\right)$

to obtain: $$L_{\text{reg}}\left(\mathrm{\theta }\text{H}\right)=L\left(\mathrm{\theta }\text{H}\right)+{\displaystyle \sum _{\text{t}}\left|\text{w}\left(\text{t}+1\right)-\text{w}\left(\text{t}\right)\right|}$$

The second submodel g _θg is time-independent and uses the electric current I and the voltage U as input variables.

The two submodels h _θh(t) and g _θg are implemented in a neural network 30. This is in 4 shown. In the exemplary embodiment shown, this has a nonlinear layer 31 for the first submodel h _θh(t) and a linear layer 32 for the second submodel g _θg . In another exemplary embodiment, not shown, two nonlinear layers 31 can be provided for the first submodel h _θh(t) .

In a further exemplary embodiment of the method according to the invention, the pressure p is determined using a hybrid model. This results from combining the physical model according to Formula 4 and the data-based model according to Formula 6 as Formula 11: $$\text{p}\left(\text{t}\right)={\text{k}}_5\cdot \text{I}\left(\text{t}\right)+{\text{f}}_{\mathrm{\theta }}\left(\text{I}\left(\text{t}\right),\text{U}\left(\text{t}\right),\text{t}\right)$$

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• DE 102013221794 A1 [0003]

Claims (10)

Method for determining a pressure (p) in a fuel tank (10) by evaluating an electrical current (I) of a diagnostic pump (16) connected to the fuel tank (10), characterized in that a data-based model is used for the determination. Procedure according to Claim 1 , characterized in that the data-based model uses the electrical current (I) of the diagnostic pump (16), an electrical voltage (U) and a time (t) or a counter as input variables. Procedure according to Claim 1 or 2 , characterized in that the data-based model has two submodels (h _θh(t) , g _θg ), whereby the first submodel (h _θh (t)) has a time-dependent model parameter (θh(t)) and the second submodel (g _θg ) has a time-independent model parameter (θg). Procedure according to Claim 3 , characterized in that the model parameters (θh(t), θg) are determined in a training phase of the data-based model. Procedure according to Claim 4 , characterized in that when determining the time-dependent model parameter (θh(t)) a loss function $\left(L_{\text{reg}}\right)$

is used, which has as an offset of a sum of the amounts of differences between time-successive weights (w(t), w(t+1)) of the first submodel (h _θh (t)). Procedure according to one of the Claims 3 until 5 , characterized in that the data-based model is implemented in a neural network (30), in which the first submodel (h _θh (t)) has a nonlinear layer (31) or two nonlinear layers and the second submodel (g _θg ) has one linear layer (32). Procedure according to one of the Claims 1 until 6 , characterized in that the data-based model is combined with a physical model in a stationary state. Computer program that is set up to carry out each step of a procedure according to one of the Claims 1 until 7 to carry out. Machine-readable storage medium on which a computer program can be written Claim 8 is stored. Electronic control device (18), which is set up to use a method according to one of the Claims 1 until 7 to determine a pressure (p) in a fuel tank (10).

Images