DE102021205735A1 - Method for controlling a pump, method for training a neural network and fluid supply system - Google Patents

Abstract

The invention relates to a method for controlling a pump with a pump chamber, at least one valve for the pump chamber that can be actively controlled, and with an actuator with which an element delimiting the pump chamber can be moved back and forth, using an artificial neural network (N)-based model (M) of the pump, which receives current values of at least one parameter (T, n, pS, pF, U) characteristic of the operation of the pump as input, a current pressure in the pump chamber is determined, based on switching times (tE,O, tE,S, tA,O, tA,S) for the at least one actively controllable valve for operating the pump are determined on the current pressure (pP) in the pump chamber, and the pump based on the switching times determined (tE,O, tE,S, tA,O, tA,S) is controlled. The invention also relates to a method for training such an artificial neural network (N) and a fluid supply system with such a pump.

Description

The present invention relates to a method for operating a pump, e.g.

Background of the Invention

The so-called SCR method (Selective Catalytic Reduction) can be used in the after-treatment of exhaust gases in motor vehicles, in particular for the reduction of nitrogen oxides (NO _x ). A urea-water solution (HWL) is introduced as a reducing agent solution into the typically oxygen-rich exhaust gas.

A dosing module or dosing valve can be used for this purpose, which comprises a nozzle in order to spray or introduce the urea-water solution into the exhaust gas flow. Upstream of an SCR catalytic converter, the urea-water solution reacts to form ammonia, which then combines with the nitrogen oxides on the SCR catalytic converter, resulting in water and nitrogen.

The metering valve is typically connected to a pump via a pressure line. This pumps the urea-water solution from a reducing agent tank to the dosing module. In addition, a return line can be connected to the reducing agent tank, via which excess urea-water solution can be returned. Such a pump can have a pump chamber and actively controllable valves, for example, as in FIG DE 10 2018 208 112 A1 described.

Disclosure of Invention

According to the invention, a method for operating a pump, a method for training an artificial neural network, a computing unit and a computer program for its implementation, and a fluid supply system with the features of the independent patent claims are proposed. Advantageous configurations are the subject of the dependent claims and the following description.

The invention deals with the operation of a pump which has a pump chamber with at least one, but preferably two valves, of which at least one, but expediently both, can be actively controlled. In addition, an actuator is provided with which an element delimiting the pump chamber can be moved back and forth, ie the volume of the pump chamber can be changed.

The actuator can be an electric motor, for example. The element delimiting the pump chamber can in particular be a membrane which, for example, is coupled via a connecting rod to an eccentric attached to a rotor of the electric motor. This is then a so-called diaphragm pump, as is typically used for the SCR supply systems already mentioned. In principle, however, the element does not have to be a membrane; a piston that (directly) delimits the pump chamber is also conceivable. The two valves are used in particular as an inlet valve and an outlet valve. Instead of an electric motor, a linear actuator can also be considered, e.g. an armature that can be moved by means of an electromagnet, which can be moved up and down against a spring force in the element that delimits the pump chamber (this can also be a membrane).

An actively controllable valve (this applies to both an inlet valve and the outlet valve) is to be understood here as meaning that the opening and closing of the valve can be actively brought about in a targeted manner, e.g. by an electromagnetic actuator with a coil and an armature, so that the valve can be switched by suitably energizing the coil. It goes without saying that other types of actively controllable valves with, for example, multiple coils or other actuators can also be used. In contrast to this, other valves—or valves that are conventionally used in pumps in SCR systems—are valves that open passively or automatically when a certain pressure is applied. With such conventional valves, for example, fluid can be drawn into the pump chamber through the inlet valve during a suction phase of the pump and then pushed out of the pump chamber through the outlet valve during a pumping or delivery phase—when the inlet valve is closed.

A particular advantage of pumps with actively controllable valves is that the pumps can be operated in more or less any way thanks to the individual control of the valves, e.g ). In an SCR supply system in which a fluid such as a urea-water solution is pumped from a fluid tank to a dosing module, this can mean, for example, that the fluid can also be pumped back from the dosing module back into the fluid tank if required can. For this purpose, the valves would only have to be opened and closed in a corresponding manner. This conveying direction, which is opposite to the regular conveying direction, is particularly SCR supply system advantageous because there after switching off the internal combustion engine or the diesel engine, the fluid - or then the urea-water solution - can be pumped back from the dosing module to the fluid tank to prevent freezing, especially in winter. In addition, there is no longer any need for a return, for example.

By deliberately varying the opening and closing times—or generally switching times or switching times—of the valves, it is also possible, for example, to vary the efficiency of the pump or a desired delivery quantity or a delivery flow.

During regular operation of the pump, the suction and pressure-side valves (or inlet and outlet valves) are not opened and closed within their working range via the applied fluid pressure, but are actively opened and closed with the help of an electrically controllable actuator, for example. In particular, if full delivery is desired, the switching times of the valves (and thus also the activation times of the respective actuators) should be dimensioned in such a way that the valve on the suction side is open during the entire suction phase and the valve on the pressure side during the entire delivery phase.

Passive or hydraulically controlled valves always open and close at the right time due to the principle of operation. The suction valve opens as soon as the pressure in the pump's compression chamber is lower than in the suction path and closes as soon as these conditions reverse. The pressure valve opens as soon as the required delivery pressure is reached in the compression space and closes again as soon as the pressure in the compression space falls below the required delivery pressure.

In the case of actively controlled valves, on the other hand, the valves open and close at the times that are specified by their control. The dead centers of the pump drive, which mechanically define the beginning and end of the suction and pressure or delivery stroke, do not correspond exactly to the beginning and end of the hydraulic suction and delivery phase. This difference depends on a large number of parameters that are characteristic of the operation of the pump and is therefore difficult to apply; any deviation from the ideal switching time can lead to a reduction in the pump delivery rate. This means that the maximum possible flow rate may not be achieved. Even if, for example, the delivery volume is to be reduced in a targeted manner, precise knowledge of the ideal switching time would be advantageous.

In order to determine the ideal switching point in each case, knowledge of the transient pressure profile in the pump or compression chamber of the pump would be necessary. However, this cannot usually be measured in practice for reasons of cost and installation space.

It is therefore proposed to create and train a model based on an (artificial) neural network that is able to calculate the pressure in the pump or pump chamber under many or as many relevant boundary conditions as possible (parameters characteristic of the operation of the pump). To predict the compression space of the pump and to implement this, e.g. in a control unit for the pump. Knowing the transient pressure in the pump chamber modeled with the help of a neural network, the ideal switching times can be determined under various or all boundary conditions, and the pump can therefore always be operated with the best possible delivery rate. In contrast to a physical model, modeling using a neural network requires far fewer resources in terms of computing and storage capacity and can therefore also be implemented on a conventional control device.

The model of the pump receives as input or input values current values of at least one parameter of the pump that is characteristic for the operation of the pump and a current pressure in the pump chamber is determined, i.e. the current pressure in the pump chamber represents the output or an output value of the model or .of the artificial neural network.

Based on the current pressure in the pump chamber, switching times for the at least one actively controllable valve for operating the pump are then determined, in particular opening and closing times for the inlet and outlet valves, in particular based on the pressure conditions across the respective valve. The pump is then controlled or operated based on the determined switching times; this includes in particular the activation of the actively controllable valves and the actuator. In particular, based on the current pressure in the pump chamber, the switching times for operating the pump can also be determined by adjusting or correcting predetermined switching times. For example, standard switching times can be specified for a specific standard pressure in the pump chamber, which can then be adjusted or varied depending on the specific, current pressure. In this sense, one can also speak of a correction function for the switching times.

In this context it should also be mentioned that the control of the valves, for example the application requires a voltage to an electromagnet or solenoid actuator to open the valve. Since a certain amount of time elapses or is necessary from applying the voltage (also referred to as the actuation time) to actually opening (the switching time), appropriate actuation times must be selected depending on the switching times (this applies not only to opening, but also to closing, where the voltage is then removed or switched off).

The at least one parameter that is characteristic of the operation of the pump and whose values are used as input for the model is preferably selected from: a temperature of a fluid to be delivered by means of the pump, an ambient pressure, a speed or frequency of the pump (depending on this, e.g depending on the type of pump, with an electric motor or linear actuator), a pressure on a suction side of the pump (suction pressure) and a pressure on a delivery side of the pump (delivery pressure). Likewise, flow rate tolerances, pressure pulsations and pressure losses within the pump (each one or more) come into consideration. These parameters also influence the ideal switching times of the valves. For example, a higher suction pressure means that the inlet valve can open earlier, since the pressure in the compression chamber (which is initially reduced) falls below the suction pressure earlier, so that fluid can be sucked in. For a more detailed description of such relationships, reference is also made to the figures and the associated description at this point.

A coil temperature and/or a control voltage of a magnetic actuator used for active control of the at least one valve typically have no influence on the pressure in the pump chamber; however, they are preferably used for the calculation or determination of the control times.

The neural network models the relationship between the input variables such as pump speed, fluid temperature, etc. and the desired, current pressure in the pump chamber as a kind of virtual sensor signal. Various types of neural networks are suitable for modeling, e.g. "feedforward networks" such as the so-called "multilayer perceptron" or multi-layer perceptron (MLP), or also recurrent neural networks such as the so-called "long short-term memory" ( LSTM).

The neural network is trained or taught in advance, i.e. before it is used to operate the pump. The training of such an artificial neural network of a model of a pump as described above is also the subject of the present invention. For this purpose, a pressure in the pump chamber is determined in each case for a plurality of sets of reference values of the at least one parameter that is characteristic of the operation of the pump by means of the model. Depending on a respective deviation of the pressure determined by means of the model or neural network from a reference pressure corresponding to the reference values, the model or there the weights in the neurons of the neural network are adjusted.

So-called supervised learning, for example, can be considered as a learning method, in which the neural network is given an input pattern and the output that the neural network produces in its current state is compared with the value that it should actually output. By comparing the target and actual output, conclusions can be drawn about the changes to be made to the network configuration. The so-called delta rule (also perceptron learning rule) can be applied to single-layer perceptrons. Multilayer perceptrons are typically trained using backpropagation (or error feedback), which is a generalization of the delta rule.

This learning or training can be based, for example, on measurement data or transient simulation data, i.e. the multiple sets of reference values of the at least one parameter that is characteristic of the operation of the pump and the corresponding reference pressures are determined using test measurements and/or simulations. Especially when using simulations, the underlying simulation models should be validated and sufficiently accurate and able to depict the necessary physical effects (e.g. influence of temperature-dependent fluid properties on the pressure in the compression chamber).

A computing unit according to the invention, e.g. a control unit of a motor vehicle, a pump control unit or an exhaust gas aftertreatment control unit, is set up, in particular in terms of programming, to carry out a method according to the invention. If necessary, a special or general computer or PC can also be used for the training.

The subject matter of the invention is also a fluid supply system, in particular an SCR supply system, with a pump explained above and a computing unit according to the invention, in particular as a control unit.

Also the implementation of a method according to the invention in the form of a computer program or computer program product with program code for carrying out all Method steps is advantageous because this results in particularly low costs, in particular if an executing control device is also used for other tasks and is therefore available in any case. Suitable data carriers for providing the computer program are, in particular, magnetic, optical and electrical memories, such as hard drives, flash memories, EEPROMs, DVDs, etc. It is also possible to download a program via computer networks (Internet, intranet, etc.).

Further advantages and refinements of the invention result from the description and the attached drawing.

The invention is shown schematically in the drawing using an exemplary embodiment and is described below with reference to the drawing.

character list

• 1 shows schematically a fluid supply system with a pump in which a method according to the invention can be carried out.
• 2 shows schematically a pump in which a method according to the invention can be carried out.
• 3 shows pressure and stroke curves when controlling a pump to explain the invention.
• 4 shows schematically a sequence of a method according to the invention in a preferred embodiment.
• 5 shows schematically a sequence of a method according to the invention in a further preferred embodiment.

embodiment(s) of the invention

In 1 a fluid supply system 100 embodied as an SCR supply system is shown schematically and by way of example, in which or with a pump present there a method according to the invention can be carried out. The SCR supply system 100 includes a pump or delivery pump 210 with a pump chamber 220, two actively controllable valves 221 and 222 for the pump chamber 220 and a filter 230. These components together form an example of a delivery unit 200, which is available as a structural unit, for example can be asked.

In the case of a regular conveying direction, valve 221 serves as an inlet valve, while valve 222 serves as an outlet valve. In addition, the pump 210 has a conveying element 225—or an element delimiting the pump chamber 220—in order to increase and decrease the volume of the pump chamber 220 . The conveying element 225 can be a membrane, for example, as will be explained in more detail below.

The pump 210 is now set up to deliver reducing agent 121 (or a reducing agent solution) as the fluid to be delivered from a fluid tank 120 via a pressure line 122 to a dosing module or dosing valve 130 . There, the reducing agent 121 is then sprayed into an exhaust line 170 of an internal combustion engine.

Furthermore, purely by way of example, a pressure sensor 140 is provided (this can also be accommodated in the delivery unit 200), which is set up to measure a pressure at least in the pressure line 122; it is then the delivery pressure. It goes without saying that further sensors for other parameters such as further pressures and temperatures can be provided. A computing unit 150, embodied, for example, as an exhaust gas aftertreatment control device, is connected to pressure sensor 140 and receives information from it about the pressure in pressure line 122. Exhaust gas aftertreatment control device 150 is also connected to delivery unit 200, there in particular to pump 210, and to dosing module 130. to be able to control them. This also includes activation of the actively controllable valves 221 and 222.

In addition, the SCR supply system 100 includes, for example, a return 160 through which the reducing agent can be returned from the system to the fluid tank 120 . In this return 160, for example, an orifice or throttle 161 is arranged, which offers a local flow resistance. However, it should be noted that such a return can also be omitted in a pump with actively controlled valves.

The exhaust gas aftertreatment control unit is set up to use relevant data, such as data received from the engine control unit or from sensors for temperature, pressure and nitrogen oxide content in the exhaust gas, to coordinate the system's actuators in order to inject the urea-water solution into the exhaust system in front of the exhaust tract in accordance with the operating strategy Introduce SCR catalytic converter. Furthermore, for example, an on-board diagnosis (OBD) monitors the components and assemblies of the exhaust gas aftertreatment system that are relevant for compliance with the exhaust gas limit values.

In 2 is a schematic of a pump 210 in which a method according to the invention can be carried out, in more detail than in FIG 1 shown in a sectional view. In addition to the pump chamber 220, the pump 210 has two actively controllable valves 221 and 222 for the pump chamber 220, in particular an element 225 designed as a membrane, which delimits the pump chamber 220.

In addition, an electric motor 240 is provided, on the rotor 245 of which a connecting rod 250 is attached, for example by means of an eccentric (cf. angle φ), which is also connected to the membrane. In this way, an up and down movement of the membrane 225 can be achieved by a rotational movement of the runner 250 .

The two valves 221 and 222 have here, for example, electromagnetic actuators or magnetic actuators, each with a coil 223 and an armature 224, by means of which a suitable element can be actuated in order to release a flow, i.e. to open or block the valve. so to close the valve.

In 3 shows pressure and stroke curves when controlling a pump, as explained above, for a more detailed explanation of the invention. In the diagram above, a pressure p is plotted against a time t. A suction pressure is shown with ps and a delivery pressure with p _f . The pressure curves p ₁ , p ₂ and p ₃ each indicate the curve of the pressure in the pump or compression space, specifically for different boundary conditions or parameters such as suction pressure. In the diagram below, a stroke h of the pump drive or actuator (and thus also of the membrane) is shown.

At time t _OT the top dead center (TDC) of the pump, ie the mechanical end of the delivery stroke and the mechanical beginning of the suction stroke, is reached. It should be mentioned here that the stroke h from the point of view of the electric motor in the direction of the membrane (cf. 2 ) is defined. However, the suction phase only begins at time t _E,O , when the pressure in the compression chamber falls below the pressure ps on the suction side of the pump. In the figure, this point in time t _E,O is only drawn in for the pressure profile P ₁ ; however, it can be seen that the other pressure curves p ₂ and p ₃ result in other—in this case later—points in time. The suction or inlet valve must be opened at this point in time - for optimum pumping conditions. The ideal switching time for opening therefore differs depending on the pressure in the compression chamber and thus depending on the boundary conditions.

At time t _UT the bottom dead center (UT) of the pump, ie the mechanical end of the suction stroke and the mechanical beginning of the delivery or compression stroke, is reached. However, the suction phase only ends at time t _E,S , when the pressure (again shown here for p ₁ ) in the compression space exceeds the pressure ps on the suction side. This point in time is also different for the three cases shown. The inlet valve is to be closed at time t _E,S .

The delivery phase begins at time t _A,O , when the pressure in the compression space (shown here again for p ₁ ) exceeds the pressure p _F on the delivery side. Only at this point in time may the pressure or outlet valve open; in particular, both valves must be closed between times t _E,S and t _A,O . The time t _A,O is also different for the three cases.

At time t _A,S , when the pressure in the compression chamber (again shown here for p ₁ ) falls below the pressure p _F on the delivery side, the end of the delivery phase is finally reached, at which point the pressure or outlet valve must close. Again, there are differences between the three cases; the time t _A,S is also not identical to the now recurring time t _OT , the top dead center of the pump.

Since the ideal switching times of the valves change when the operating boundary conditions of the pump change, these would traditionally have to be applied for the entire range of operating boundary conditions. Operating boundary conditions that affect the switching times are, for example, fluid temperature, ambient pressure, control voltage, coil temperature (when using a magnetic actuator), pump speed or the suction and delivery pressure or possibly other parameters, as already described above. Based on 3 For example, the influence of suction and discharge pressure is clearly visible: If the discharge pressure p _F were higher, the point in time t _A,O , i.e. the opening of the outlet valve, would be delayed (intersection of p _F with p ₁ ).

An application that takes all these boundary conditions into account is correspondingly complex. It is therefore proposed, as mentioned, to determine the switching times using a model, e.g. in the control unit. Since a physical model is usually not an option due to the resources available in the control unit, a model based on an (artificial) neural network is to be created and trained in such a way that it is able to work under all relevant operating boundary conditions of the pump to predict the pressure in the compression chamber as a virtual sensor and thus to determine the optimal control times in each case.

In 4 a sequence of a method according to the invention is shown schematically in a preferred embodiment, namely the control or operation of the pump using the model M based on an artificial neural network N.

For this purpose, for example, a temperature T of the fluid to be delivered by the pump, a speed n of the pump, a pressure ps on a suction side of the pump, and a pressure p _F on a delivery side of the pump are recorded as current values (these values are e.g measured by means of suitable sensors) and fed to the model M or its artificial neural network N as an input or input values.

The current pressure pp in the pump chamber or compression chamber of the pump is then determined there as the output or initial value. From this or based on it, for example depending on the pressures p _S and p _F as in 3 shown, the switching times for the valves are determined, in particular opening time t _E,O and closing time t _E,S of the intake valve and opening time t _A,O and closing time t _A,S of the exhaust valve.

Control times are then determined from this, taking into account a control voltage u of a magnetic actuator used for active control of the at least one valve and a coil temperature Ts of the magnetic actuator. The valves are then controlled accordingly. The difference between the activation time and the opening time (or switching time in general), the so-called activation offset, depends on the coil temperature and activation voltage. This can e.g. be calibrated.

In 5 a sequence of a method according to the invention is shown schematically in a further preferred embodiment, specifically the training or learning of the artificial neural network N, as is the case, for example, according to FIG 4 is used. First of all, for a set P _R,1 of reference values, the parameter, such as is used in connection with 4 are mentioned (temperature T etc.) by means of the model or the neural network N, a pressure pp is determined in the pump chamber.

This pressure pp thus obtained is then compared with a reference pressure p _R which, for example, was determined during test measurements on the test bench for the parameter set P _R,1 . A weight G of a neuron of the neural network N, for example, is then adjusted as a function of a deviation from p _P to p _R ; the weight G' is obtained. With this adapted neural network, the process is then repeated with a different set of parameters P _R,2 .

This can, for example, be repeated with a large number of existing parameter sets and corresponding reference pressures until the deviations between the specific pressure pp and the reference pressure p _R fall below a predetermined threshold value or are as small as possible. The neural network thus trained can then be used to operate the pump as described in relation to FIG 4 was explained.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited 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.

Patent Literature Cited

• DE 102018208112 A1 [0004]

Claims (14)

Method for controlling a pump (210) with a pump chamber (220), at least one valve (221, 222) for the pump chamber (220), which can be actively controlled, and with an actuator with which an element delimiting the pump chamber (220). (225) can be moved back and forth, using a model (M) of the pump based on an artificial neural network (N) which has current values of at least one parameter (T, n, P _S , p _F , U) receives as an input, a current pressure in the pump chamber is determined, based on the current pressure (pp) in the pump chamber switching times (t _E,O , t _E,S , t _A,O , t _A,S ) for the at least one actively controllable valve is determined for operating the pump, and the pump is controlled based on the determined switching times (t _E,O , t _E,S , t _A,O , t _A,S ). procedure after claim 1 , wherein the pump has at least two actively controllable valves, one of which is used as an inlet valve (221) and one as an outlet valve (222). procedure after claim 1 or 2 , wherein the switching times are selected from: an opening time (t _E,O ) of an intake valve, a closing time (t _E,S ) of the intake valve, an opening time (t _A,O ) of an exhaust valve and a closing time (t _A,S ) of the exhaust valve. Method according to one of the preceding claims, wherein based on the current pressure (p _P ) in the pump chamber, the switching times (t _E,O , t _E,S , t _A,O , t _A,S ) for the at least one actively controllable valve Operation of the pump can be determined by adjusting predetermined switching times (t _E,O , t _E,S , t _A,O , t _A,S ). Method according to one of the preceding claims, in which a fluid (121) is conveyed in a fluid supply system (100), in particular an SCR supply system, by means of the pump. Method according to one of the preceding claims, wherein based on the determined switching times (t _E,O , t _E,S , t _A,O , t _A,S ) control times (t ') are determined, according to which the at least one actively controllable valve is controlled. procedure after claim 6 , wherein the control times (t') are determined taking into account a coil temperature (T _S ) and/or a control voltage (u) of a magnetic actuator used for active control of the at least one valve. Method for training an artificial neural network of a model (M) of a pump with a pump chamber (220), at least one valve (221, 222) for the pump chamber (220), which can be actively controlled, and with an actuator with which a Pump chamber (220) delimiting element (225) is movable back and forth, and using which model (M), the current values of at least one characteristic of the operation of the pump parameter (T, n, p _S , p _F , U) as input, a current pressure (p _P ) _in the pump chamber can be determined, with a _pressure (pp ) is determined in the pump chamber, and the model is adapted depending on a respective deviation of the pressure (pp) determined by means of the model from a reference pressure (p _R ) corresponding to the reference values. procedure after claim 8 , wherein the multiple sets (P _R,1 P _R,2 ) of the reference values of the at least one parameter characteristic of the operation of the pump and the corresponding reference pressures (p _R ) are determined on the basis of test measurements and/or simulations. Method according to one of the preceding claims, wherein the at least one parameter characteristic of the operation of the pump is selected from: a temperature (T) of a fluid (121) to be delivered by means of the pump, an ambient pressure, a speed (n) or frequency of the pump , a pressure (ps) on a suction side of the pump, a pressure (p _F ) on a delivery side of the pump, a delivery amount tolerance, a pressure pulsation, and a pressure loss inside the pump. Arithmetic unit (150) which is set up to carry out all method steps of a method according to one of the preceding claims. Fluid supply system (100), in particular SCR supply system, with a pump (210) with a pump chamber (220), at least one valve (221, 222) for the pump chamber (220), which is actively controllable, and with an actuator, with which an element (225) delimiting the pump chamber (220) can be moved back and forth, and a computing unit (150). claim 11 . Computer program that causes a computing unit (150) to carry out all the method steps of a method according to one of Claims 1 until 10 to be performed when it is executed on the computing unit (150). Machine-readable storage medium with a computer program stored on it Claim 13 .

Images