DE102022200059A1 - Method for detecting a pressure anomaly in a fluid supply system - Google Patents

Abstract

The invention relates to a method for detecting a pressure anomaly in a fluid supply system, which has a delivery unit, a delivery line, via which the delivery unit is connected to a fluid tank, and a pressure line, via which the delivery unit is connected to a dosing module for dispensing fluid. comprises obtaining (502) input data (504) comprising a course of a pressure in the fluid supply system over a period of predetermined length, using a determination (506) based on the input data whether there is a pressure anomaly a mesh learning algorithm, and outputting (508) information as to whether there is a pressure anomaly. The invention also relates to a method for operating a fluid supply system and a method for training a machine learning algorithm.

Description

The present invention relates to a method for detecting a pressure anomaly in a fluid supply system, a method for operating a fluid supply system, a method for training a machine learning algorithm, and a computing unit and a computer program for executing it.

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. For this purpose, a dosing module or dosing valve can be used as part of a fluid supply system, which includes 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.

Disclosure of Invention

According to the invention, a method for detecting a pressure anomaly in a fluid supply system, a method for operating a fluid supply system, a method for training a machine learning algorithm, and a computing unit and a computer program for their implementation with the features of the independent patent claims are proposed. Advantageous configurations are the subject of the dependent claims and the following description.

A fluid supply system has a delivery unit, e.g. in the form of a pump or comprising a pump, a delivery line, via which the delivery unit is connected to a fluid tank, and a pressure line, via which the delivery unit is connected to a dosing module for dispensing fluid. Such a fluid supply system can be used in particular in an exhaust aftertreatment system and there, for example, introduce reducing agents or a reducing agent solution into the exhaust system. In addition, a return can be connected to the fluid tank, via which excess fluid can be returned. An orifice or throttle in the return can control the return flow.

In the case of exhaust gas aftertreatment systems, the fluid supply system is usually designed in such a way that it meters the required amount of fluid or reducing agent into the exhaust gas, which is calculated on the basis of the amount of nitrogen oxides emitted at the engine exhaust port. Compliance with this requirement is crucial to ensure that no harmful emissions exit the vehicle's tailpipe. In order to dose correctly, the fluid supply system or the fluid there is pressurized. In order to dose the right amount, this pressure should be kept as stable as possible.

Depending on the design of the fluid supply system, it may happen that air penetrates into the fluid supply system. Air bubbles lead to a pressure drop in the fluid and a pressure regulator that is typically present reacts. In this context, the correct dosing of the fluid is usually interrupted when the pressure drops to or below a predefined limit. In addition, the pressure regulator may produce pressure overshoots in response to air ingress.

These air bubbles can have various consequences for the fluid supply system and its components. Since the dosing of the fluid is typically interrupted during air ingress, if there is too much air in the system, the fluid supply system will not meet the dosing requirements, which would eventually lead to emissions-related penalties. In the worst case, system errors could even be triggered and the vehicle could go into emergency operation. Pressure overshoot can damage fluid delivery system components. Similar to the situation mentioned above, in the worst case this will also trigger the corresponding system errors and the vehicle could go into emergency operation.

In order to prevent these situations, a tank filter in the fluid supply system or there in the fluid tank can be improved accordingly in order to minimize the ingress of air into the system; however, this is time-consuming and costly. Likewise, the pressure regulator could be calibrated accordingly to avoid these pressure peaks. However, this is only a reactive behavior and cannot cover all cases caused by air ingress. A logic in software is also conceivable in order to reset the pressure regulator during the next pressure build-up phase. The disadvantage of this solution, however, is that the update is usually late and there is no validation. In summary, there is no rule-based diagnostic to detect air in the system and thus improve pressure regulator control.

Against this background, a method for detecting a pressure anomaly in a fluid supply system is proposed which uses a machine learning algorithm, in particular an artificial neural network such as a CNN (“Convolutional Neural Network”), ie artificial intelligence. In this way, input data are obtained which include a pressure profile in the fluid supply system, in particular in the pressure line, over a period of time of a predetermined length. Based on the input data, it is then determined whether there is a pressure anomaly using the mesh learning algorithm. Information about this will then be issued.

The input data can also include at least one of the following parameters, for example: a value of an operating variable, e.g. A needle movement of the dosing module can, similar to the pump, be monitored or controlled, e.g. via a duty cycle (in PWM operation) as an operating variable for the dosing module (e.g. duty cycle 0% means: the dosing module is not dosing, 100% means: the dosing module is fully open).

In order to operate a fluid supply system in which the pressure is regulated to a specified setpoint by controlling the delivery unit with an operating variable such as the duty cycle (in PWM operation) as a manipulated variable, this operating variable can be set to a predetermined or earlier (i.e. before detection of the Pressure anomaly used) value can be limited or adjusted when a pressure anomaly has been detected in the manner explained.

The use of such a machine learning algorithm is possible because it can be trained using a large number of comparison curves with pressures. Such training data can be obtained, for example, from vehicles that are in the field during a validation program and can therefore record and provide the necessary data. In addition to the curves themselves, information is then also required as training data as to whether there is a pressure anomaly in each of these curves. This information can be determined using rule-based functions (i.e. not based on artificial intelligence) or manually.

This makes it possible to identify particularly well and quickly whether or when, for example, an air bubble is detected in the fluid. Such a detection of air or air bubbles in the fluid, based on artificial intelligence or a machine learning algorithm, can in particular also detect such air inlets or air bubbles which, for example, would not have been detected as air bubbles with conventional methods because, for example, a typical rule-based function used for this purpose would not be fulfilled. As has been shown, not only can air bubbles or air ingress in the fluid be detected, but pressure anomalies in general, which may not be air bubbles at all or are not due to such bubbles, but nevertheless show deviations from normal operating conditions.

This last point also shows that this artificial intelligence based air bubble detection feature can also be used as a pressure anomaly detection and the air bubbles is just one of the potential anomalies. If this detection function is embedded in a control unit (e.g. ECU or DCU), an operating variable such as the duty cycle of the delivery unit, i.e. the pump duty cycle, can be adjusted to correct a detected pressure anomaly, as already mentioned. Thus, after the air is removed or vented from the system, the pressure will return to the set point more quickly and without a pressure spike.

A computing unit according to the invention, e.g. a control unit, in particular an engine control unit, of a motor vehicle is set up, in particular in terms of programming, to carry out a method according to the invention.

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 causes particularly low costs, especially if an executing control device is also used for other tasks and is therefore available anyway. Finally, a machine-readable storage medium is provided with a computer program stored thereon as described above. Suitable storage media or data carriers for providing the computer program are, in particular, magnetic, optical and electrical storage devices such as hard drives, flash memories, EEPROMs, DVDs, etc. It is also possible to download a program via computer networks (Internet, intranet, etc.). Such a download can be wired or wired or wireless (e.g. via a WLAN network, a 3G, 4G, 5G or 6G connection, etc.).

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

The invention is illustrated schematically in the drawing using exemplary embodiments and is described below with reference to the drawings.

character list

• 1 shows schematically a fluid supply system in which a method according to the invention can be carried out.
• 2 shows a diagram of a pressure curve in a fluid supply system when air enters without using the invention.
• 3 shows schematically a sequence of a method according to the invention in a preferred embodiment.
• 4 shows a diagram of a pressure curve in a fluid supply system when air enters without using the invention and using a method according to the invention in a preferred embodiment.
• 5 shows a schematic 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, in particular in the form of an SCR supply system, is shown schematically, in which a method according to the invention can be carried out or which can be operated with a method according to the invention. The SCR supply system 100 has a delivery unit 130 designed as a pump with a filter 132 . The pump 130 is set up to deliver fluid or reducing agent 121 (eg a urea-water solution) from a fluid tank 120 via a delivery line 126 and a pressure line 122 to a dosing module or dosing valve 140 . The fluid 121 is then introduced or sprayed into an exhaust system 170 or an exhaust line of an internal combustion engine by means of the metering valve 140 .

Furthermore, a pressure sensor 142 (this can also be integrated into the pump) is provided, which is set up to measure a pressure at least in the pressure line 132—and thus in the fluid. A temperature sensor 123 and filling level sensor 124 are also provided in the fluid tank 120 , for example. A processing unit 150, for example in the form of an exhaust aftertreatment control unit or an engine control unit, is connected to pressure sensor 142 and receives information about the pressure in pressure line 122. Processing unit 150 is also connected to temperature sensor 123 and level sensor 124. In addition, the computing unit 150 is connected to the pump 130 and the metering valve 140 in order to control or operate the SCR supply system 100 .

In addition, the SCR supply system 100 includes, for example, a return line 160, via which the fluid can be returned to the fluid tank 120. In this return line 160, for example, an orifice or throttle 161 is arranged, which offers a local flow resistance. It should be noted, however, that such a return can also be dispensed with in the case of a pump with actively controlled valves.

The computing unit 150 is set up to use relevant data, such as data received from sensors for temperature, pressure and nitrogen oxide content (e.g. sensor 177) in the exhaust gas, to operate the SCR supply system 100 and in particular the pump 130 and to actuate the metering valve 140 in order to feed the urea-water solution 121 (fluid) into the exhaust system 170 in front of an SCR catalytic converter 174 .

In 2 FIG. 12 is a diagram with a profile of a pressure p (in bar) of fluid in a pressure line of a fluid supply system such as that from FIG 1 shown over time t (in s). A target value for the pressure is denoted here by P _target and is 9 bar, for example. As mentioned, air can enter the system, for example. The air bubbles lead to a drop in pressure, as can be seen shortly before time t=200s, and a pressure controller reacts by increasing the flow rate or flow rate of the pump (eg 130). Provision can also be made for dosing to be interrupted when the pressure falls to or falls below a predefined limit value, here for example 1.5 or 2 bar. The intervention of the pressure regulator generates a pressure overshoot, the peak value of which can, in certain cases, be above the component limits (here of 13 bar, for example). In order to avoid such situations, such pressure anomalies should be recognized at an early stage in order to be able to react accordingly, in particular to prevent such controller interventions.

According to a preferred embodiment of the invention, pressure anomalies are detected based on the pressure history using a machine learning algorithm, e.g., a CNN. For training data for training such a machine learning algorithm, curves with pressures should be used where it is known that there are air bubbles or air in the fluid (or generally: pressure anomalies); this can be determined, for example, with conventional, rule-based functions. In particular, curves can be used in which air bubbles definitely occur, as well as—possibly for comparison—courses in which—at least according to conventional determination—no air bubbles occur. In addition, the times at which the air bubbles appeared should be known for the curves.

Pressure value curves (so-called time windows) are expediently considered for the pressure anomaly detection. These windows can either overlap (ie multiple pressure values are part of different histories) or non-overlap. A length of several seconds has proven useful for such time windows, for example in the range from 10 to 100 s. A suitable time window can have a length of between 30 and 35 s, for example 32 s. Particularly good learning results could be achieved when a pressure drop is approximately in the middle of a time window with a length of 32 s. Time windows without pressure anomalies accordingly show no pressure drop. The values for the length of the time window can, for example, also be adjusted as a function of a (typical) amount of air and a flow rate within the fluid supply system.

Other parameters or characteristics such as values of the operating quantity of the delivery unit (e.g. pump duty cycle or duty cycle) as well as needle movement of the dosing module can also be taken into account to improve the model; however, there will usually be a compromise between the complexity of the model, its size on the computing unit and the added value for predictability.

For the training, a rule-based function can be used to generate the desired output data of the machine learning algorithm, i.e. the desired output information whether a trace shows a pressure anomaly or not.

• - der Druck im Fluid bzw.im Fluidversorgungssystem steigt über 10 bar,
• - das Fluidversorgungssystem dosiert länger als 15s (um Verwechslungen mit Druckstabilisierung nach dem Druckaufbau zu vermeiden),
• - der Tastgrad der Pumpe liegt über 65%, und
• - das Maximum des Tastgrads der Nadelbewegung des Dosiermoduls liegt unter 50%.

An example rule-based function for detecting an air bubble can be defined as follows:

• - the pressure in the fluid or in the fluid supply system rises above 10 bar,
• - the fluid supply system doses longer than 15s (to avoid confusion with pressure stabilization after pressure build-up),
• - the duty cycle of the pump is over 65%, and
• - the maximum duty cycle of the needle movement of the dosing module is below 50%.

Classic machine learning models such as decision trees achieve an accuracy of e.g. 85%; however, the problem of a time dependency can occur here. This means that the air bubble can appear anywhere in the considered time window, which leads to different pressure profiles. For example, if an air bubble never appears at the end of the window in the training data, it may not be detected in reality.

In order to counteract this dependency and also to take into account that the pressure anomalies or pressure changes can be very dynamic, the use of so-called "convolutional neural networks" (CNN) can be considered. This deep learning architecture is able to handle the temporal and spatial dependency through so-called automatic feature extraction.

As such, CNNs are typically designed to process images that can be viewed as multidimensional matrices. In the present case, the pressure profile is a one-dimensional time series. It has been shown that it is better to transform such one-dimensional vectors into matrices. For example, a technique called “Grammian Angular Field” can be used for this. This includes, for example, normalizing the data (a time series) to the range [-1; +1], finding polar coordinates of the normalized data, and then computing the matrix using the polar coordinates.

For additional validation, a model can be used to re-examine the trajectories determined by the rule-based function to contain no air bubbles. For this purpose, the curves can in turn be divided into (e.g. 32 s long, non-overlapping) time windows and transferred to the model. By using the machine learning algorithm (CNN), air bubbles or pressure anomalies can then be detected in certain profiles that were not detected with rule-based functions.

As already mentioned, the machine learning algorithm can be embedded in a control unit in a particularly simple manner. As a rule, a compromise has to be found between the size (i.e. required memory space) of the machine learning algorithm, its required computing power (in the control unit) and the performance (e.g. accuracy) of the detection of the pressure anomaly.

In 3 a sequence of a method according to the invention in a preferred embodiment for training a machine learning algorithm is briefly presented in summary. In a step 300, training data 302 is obtained; in a step 304, the machine learning algorithm 306 is adapted such that its outputs correspond to the results of the training data 302 as far as possible. In step 308, the trained machine learning algorithm is then output or made available for further use.

In 4a is a diagram with a profile of a pressure p (in bar) of fluid in a fluid or SCR supply system similar to that in FIG 2 shown, with the pressure (upper trace) on no manner according to the invention should be adjusted to a setpoint p _setpoint . In addition, a duty cycle T is shown as the operating variable of the pump (lower curve).

As air enters the system, the pressure drops and the pressure regulator responds by increasing the pump duty cycle to compensate for the pressure drop. As soon as the air is purged from the pump or system, the pressure rises sharply due to the high duty cycle and overshoots before stabilizing at the set pressure.

In 4b is a chart similar to the one from 8a shown, but using a preferred embodiment of a method according to the invention.

In 5 a corresponding sequence of a preferred embodiment—an application—is shown schematically. In a step 500, the pressure is regulated to the desired value p _desired ; this is regular operation.

According to step 502, the machine learning algorithm receives (e.g. continuously) input data 504, e.g. a pressure profile over a specific period of time, e.g. the last 32 s. With the machine learning algorithm, step 506, a pressure anomaly, e.g. an air bubble, can be detected , here for example at time t ₁ .

In a preferred step 508, information about this can then be output, for example.

Then, in step 510, the duty cycle can be set or limited to a fixed value or to the value just before the occurrence of the pressure anomaly, such as used at time t<to. The pressure drop cannot be completely avoided, as it is due to the air in the system, but after the pressure anomaly there is no pressure peak and the system stabilizes much more quickly at the target pressure.

Claims (13)

Method for detecting a pressure anomaly in a fluid supply system (100), which has a delivery unit (130), a delivery line (126) via which the delivery unit (110) is connected to a fluid tank (120), and a pressure line (122), via which the delivery unit (130) is connected to a dosing module (140) for dispensing fluid (121), comprising: Obtaining (502) input data (504), comprising a profile of a pressure (p) in the fluid supply system (100) over a period of predetermined length, determining (506), based on the input data, whether there is a pressure anomaly using a mesh learning algorithm, and outputting (508) information as to whether there is a pressure anomaly. procedure after claim 1 , wherein the input data further comprise at least one of the following parameters: a value of an operating variable of the delivery unit (130) during the period of time, and an operating variable of a needle movement of the dosing module (140) during the period of time. procedure after claim 1 or 2 , wherein detecting the pressure anomaly includes detecting air bubbles in the fluid (121) of the fluid supply system (100). Method for operating a fluid supply system, which has a delivery unit (130), a delivery line (126) via which the delivery unit (130) is connected to a fluid tank (120), and a pressure line (122) via which the delivery unit (130 ) is connected to a dosing module (140), comprising: regulating (500) a pressure (p) in the fluid supply system (100) to a predetermined desired value (p _desired ) by controlling the delivery unit (130) with an operating variable of Delivery unit as manipulated variable, and limiting or setting (510) the operating variable (T') for controlling the delivery unit (130) to a predetermined or earlier value if a pressure anomaly has been detected in the fluid supply system using a method according to one of the preceding claims is. Method for training a machine learning algorithm, which is provided for detecting a pressure anomaly in a fluid supply system, which has a delivery unit (110), a delivery line (130) via which the delivery unit (110) is connected to a fluid tank (120), and a pressure line (132) via which the delivery unit (110) is connected to a dosing module (136) for dispensing fluid, comprising: Obtaining (300) training data (302), comprising: a plurality of profiles of a pressure in each case in a fluid supply system over a period of predetermined length, and information as to whether there is a pressure anomaly in each of these profiles, adjusting (304) the machine learning algorithm (306) based on the training data such that the machine learning algorithm detects whether a pressure anomaly is present, and providing (308) the trained machine learning algorithm. procedure after claim 5 , wherein the information about whether there is a pressure anomaly in each of these curves is determined using rule-based functions. Method according to one of the preceding claims, wherein the predetermined length of the time period is between 10 s and 100 s, preferably between 30 and 35 s. Method according to one of the preceding claims, wherein an artificial neural network, in particular a CNN, is used as the machine learning algorithm. A method according to any one of the preceding claims, wherein a history of a pressure is transformed into a matrix for use in the machine learning algorithm. Method according to one of the preceding claims, wherein the fluid supply system (100) has an SCR supply system, and wherein the fluid (121) is or has a urea-water solution. Arithmetic unit which is set up to carry out all method steps of a method according to one of the preceding claims. Computer program that causes a computing unit 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. Machine-readable storage medium with a computer program stored on it claim 12 .

Images