Method for determining characteristic variables of solenoid valves and method for training pattern recognition methods based on artificial intelligence

Abstract

The invention relates to a method for determining a variable characterizing the opening and / or closing of a flow-through opening of an electromagnetic valve, for which an electromagnetic coil is energized in order to lift a armature for opening a flow-through opening for a fluid, wherein during the operation of the electromagnetic valve a curve of the current (I) in the electromagnetic coil is determined, and wherein using an artificial intelligence-based pattern recognition method (430) the characteristic variable is determined with the aid of at least one section of the curve or a curve derived therefrom using a neural network. The invention also relates to a method for applying and training an artificial intelligence-based pattern recognition method (430).

Description

Technical Field

[0001] The present invention relates to a method for determining variables characterizing the flow opening of a solenoid valve that is open and / or closed, a method for training an artificial intelligence-based pattern recognition method for determining such characteristic variables, and a computing unit and computer program for implementing them. Background Technology

[0002] A solenoid valve consists of a solenoid coil and an armature, which is lifted or attracted by energizing the coil—typically achieved through pulse width modulation (PWM) control to apply a specific or average voltage, thereby opening a flow passage. Typical applications of solenoid valves include the precise metering of liquid or gaseous working materials (e.g., fuel injectors, working material metering valves, control valves in hydraulic presses). There, solenoid valves are typically used to open (or close) or close a time-precisely defined flow passage for the working material. Summary of the Invention

[0003] According to the present invention, a method for determining variables characterizing the flow opening of a solenoid valve and / or the closing of a solenoid valve, a method for training an artificial intelligence-based pattern recognition method, and a computing unit and computer program for implementing them are provided. Advantageous designs are the subject of the following description.

[0004] This invention relates to a solenoid valve having an electromagnetic coil and an armature; in such a solenoid valve, the electromagnetic coil is typically energized by applying a voltage or a time-defined control voltage to lift the armature. This opens or opens the flow opening of the solenoid valve for fluids (especially liquids or gases). When energization is terminated, the armature may fall back with a delay—or in particular be pushed back by a spring—and then close or seal the flow opening again. In this case, it can also be described very generally as the opening and closing of the solenoid valve.

[0005] These solenoid valves can be used in a variety of applications, such as in vehicles. There, they can be specifically configured to introduce fuel as a fluid into the cylinders of an internal combustion engine, or to introduce a reducing agent or reducing agent solution as a fluid into the exhaust system (e.g., when using a so-called SCR system for "selective catalytic reduction," i.e., reducing harmful emissions). Similarly, independent of the vehicle, these solenoid valves can be used to selectively introduce other fluids into chambers or spaces. Other examples were mentioned at the beginning.

[0006] Accurately knowing when a solenoid valve opens and / or closes is not only important, but especially crucial when used in vehicles, as it allows a defined amount of working medium (fluid) to be introduced into a target cavity. Furthermore, the minimum requirement for operating a valve is knowing whether the solenoid valve is open or closed.

[0007] Identifying whether and when a solenoid valve used as an injection valve opens and closes is crucial for determining the injection quantity of fluids such as diesel, gasoline, reducing agents, hydraulic oil, and compressed air. For this purpose, expensive needle motion sensors or evaluations of the control process of electronically controlled valves are typically used, for example. In the case of a solenoid valve, this can be achieved using the current profile during the energization or activation of the solenoid coil, which, during solenoid valve operation, i.e., during injection, pulls the armature from the valve seat by the generated magnetic field, and thus also pulls the (valve) needle, thereby achieving injection, i.e., opening the flow passage. The figure shows a typical current profile with pull-in and hold phases and valve closure.

[0008] The mechanical movement of the armature or needle, and possibly other components within the solenoid valve, induces an inductive reaction characterizing the opening and closing motion. This inductive reaction can be analyzed by measuring the current with high time resolution. Particularly preferred here is the so-called "jet pulse start / end" (BIP / EIP) method, which determines the timing of the solenoid valve opening or closing under a given control by analyzing the curvature change of the current curve, i.e., the measured current curve (corresponding to the second time derivative). The start and end of the needle or armature movement thus induce measurable curvature changes in the current curve.

[0009] The absence or drift (deviation) of characteristic BIP or EIP features can then be used for diagnostic purposes, typically indicating a defect or accelerated wear in the solenoid valve. By understanding the actual measured opening time of the solenoid valve (i.e., the time interval between opening and closing), the actual amount of fluid injected can be calculated and corrected, for example, in a closed-loop control circuit, by referencing the requested amount of fluid to be injected for the next injection. Typically, the operation of the solenoid valve can then be performed and / or adjusted using these variables characterizing opening and / or closing. Diagnosis can also be performed, or, if indicated, emergency measures can be taken to prevent danger or damage from a malfunctioning valve.

[0010] To determine the change in curvature, as mentioned, the second-order time derivative of the current curve or current signal can be used in particular. While the change in curvature allows for very precise determination of the relevant time points, this approach is practically susceptible to interference; that is, it is sensitive to noise in the measured signal. In cyclic control scenarios, such as using a solenoid valve designed for a 12V control voltage in a 24V operating system, the current curve of the controlled solenoid valve will then have strong ripple, making the evaluation of the (second-order) time derivative very difficult. For example, complex filtering of the signal is required to avoid or minimize false detections.

[0011] It has now been demonstrated that by using artificial intelligence-based (machine) pattern recognition methods, such as artificial neural networks (often simply referred to as neural networks) or support vector machines (SVMs), regions characterizing opening or closing, such as bends, can be identified sufficiently accurately, even if the current curve signal may be noisy. Therefore, a method is proposed to determine or detect the current curve in the solenoid coil during solenoid valve operation, and to use an artificial intelligence-based pattern recognition method with the aid of at least one segment (but possibly the entire curve) of the curve to determine at least one variable characterizing the opening and / or closing of the solenoid valve's flow orifice.

[0012] In particular, for this purpose, the current curve is detected first, for example; this can be detected, for example, by reverse measurement of the control signal (current or voltage; the voltage signal here can represent the current curve) during control in a computational unit such as a controller or therein in a microcontroller. For solenoid valves, the control current is typically measured, for example, by a shunt (measuring resistor) located after the output stage. The voltage signal of the shunt is then converted, for example, into a measurement value that can be processed by software in the analog-to-digital converter of the microcontroller. Typically, these measurements are determined in an equidistant time grid, but any time pattern used for recording data can also be imagined, which can be redefined from measurement to measurement. These measurements (and thus a series of measurements) along with possible information about the measurement time points can then be provided as inputs or input values ​​(so-called input vectors) to a pattern recognition method, i.e., fed to the receptor of a neural network, for example. If the pattern recognition method has been trained accordingly, i.e. learned, it provides a corresponding output vector or one or more output values ​​based on the input vector, which in the present case are variables characterizing the opening and / or closing of the flow opening of the solenoid valve. The following statement also provides information on the training of the pattern recognition method or artificial neural network.

[0013] The representative variables are preferably selected from the opening time of the current-passing orifice, the closing time of the current-passing orifice, a value indicating whether the current-passing orifice has been opened or closed (in a categorical sense), and a value indicating the probability that the current-passing orifice has been opened or closed (e.g., values ​​can be given here in the form of percentages or ranges, such as "definitely open," "most likely to be open," etc.; this is also a form of categorization). Which of these variables is chosen as the output can be selected, for example, depending on the desired application; however, it goes without saying that the pattern recognition method is then accordingly selected for the desired representative variables. It should also be considered here that the segment of the current curve must be selected accordingly; that is, if the opening time point is to be determined, the segment fed to the pattern recognition method must also include the relevant time point. Similarly, if the pattern recognition method is selected and trained accordingly, several of these representative variables can also be output.

[0014] By using an AI-based pattern recognition method, information highly relevant to the operation of the solenoid valve can be obtained even when the current curve signal is (high) noisy. No additional filtering or even additional sensors are required. In particular, the characterizing variables can be determined directly from the current curve without the need for the (time) derivative; however, the derivative can be used instead of the curve itself.

[0015] Various types of pattern recognition methods based on artificial intelligence can be considered here, such as single-layer or multi-layer "feedforward" networks (also known as single-layer or multi-layer perceptrons), or recurrent networks. The appropriate computation can be selected based on the computational needs and desired accuracy and / or speed.

[0016] Of particular advantage is that the artificial intelligence-based pattern recognition method has been trained or learned before use. The method used to train the pattern recognition method is also the subject of this invention; this method is used to determine variables characterizing the flow opening of the solenoid valve—that is, the aforementioned purpose of use.

[0017] Therefore, for multiple curves or several variations of curves, such as the current curve in the solenoid coil during the operation of a solenoid valve under different boundary conditions (temperature, different valve samples, e.g., spring strength / center / edge position of the injection orifice plate, voltage curve, etc.), at least one segment is provided as an input value to the pattern recognition method. The segment used here should correspond to a segment that will also appear later when the pattern recognition method is used. For example, these segments or the current curves on which they are based can come from test measurements, etc.; in principle, simulated or analog current curves can also be envisioned here.

[0018] The machine pattern recognition method, such as the weights of an artificial neural network, is then adjusted using representative variables obtained as output values ​​from an AI-based pattern recognition method for these segments, along with their reference values. In other words, the pattern recognition method is thus provided with many different segments of the current curve that occur, for example, during the operation of a solenoid valve—where the relevant representative variables are known (these are the reference values ​​mentioned). The pattern recognition method then outputs a representative variable for each segment, but at least at the start of training, this representative variable typically does not, or rarely corresponds to, the actual representative variable, i.e., the reference value. Therefore, the AI-based pattern recognition method is adjusted so that the output representative variables are better or more accurate in subsequent calculations. This process can be repeated for many different, known segments of the current curve until the pattern recognition method has been trained accurately enough, i.e., until its output values ​​satisfactorily reflect the values ​​known and therefore expected during training.

[0019] Various types of training or learning can be considered. In the case of supervised learning, an AI-based pattern recognition method is given, for example, an input pattern (here, a known segment of a current curve), and the output or output value produced by the pattern recognition method in its current state is compared with the value it is actually intended to output. By comparing the target output and the actual output, a conclusion can be drawn about which configuration needs to be changed. For neural networks with single-layer perceptrons, the so-called Delta rule (also known as the perceptron learning rule) can be used. Multilayer perceptrons are typically trained using backpropagation, which is a generalization of the Delta rule.

[0020] Unsupervised learning involves learning a pattern from input data. Artificial intelligence-based pattern recognition methods adapt to the input pattern. There is also reinforcement learning, which is used when not every input set has a suitable output set (comparison value) available for training.

[0021] Particularly preferably, multiple current curves are selected for different values ​​of the solenoid valve parameters (selecting segments from these; of course, a complete current curve can also be used). Parameters include, for example, the control voltage of the solenoid coil, the elastic strength of the spring pressing against the armature, the size of the flow opening (e.g., diameter), the temperature of the fluid and the solenoid valve, the operating time and wear of the solenoid valve, the number of open / close cycles that have occurred, and the function of the solenoid valve (e.g., the current curve of a solenoid valve that may be intentionally blocked). By training with current curves for different parameter values, AI-based pattern recognition methods can, in later use, particularly identify different training conditions, i.e., whether the current curve indicates that the solenoid valve is blocked and therefore not open. Similarly, however, for different operating conditions, and thus for different current curves obtained, it is always possible to find the actual physical time point of opening / closing, or to identify valve blockage. Multiple current curves can be selected, particularly for different values ​​of the solenoid valve parameters, from input vectors provided via remote computing and / or storage systems (e.g., the so-called cloud). The computing unit according to the invention, such as a controller of a motor vehicle or machine, or a PC (for training), is designed, in particular, to execute the method according to the invention using programming techniques.

[0022] It is advantageous to implement the method according to the invention in the form of a computer program or a computer program product having program code for performing all method steps, as this results in particularly low costs, especially if the controller performing the action is also used for other tasks and therefore already exists, such as for controlling valves. Finally, a machine-readable storage medium is specified on which the computer program as described above is stored. Suitable storage media or data carriers for providing the computer program are, in particular, magnetic, optical, and electrical memories, such as hard disks, flash memory, EEPROM, DVDs, etc. The program can also be downloaded via a computer network (Internet, intranet, etc.). Such downloading can be performed wired or wirelessly (e.g., via WLAN networks, 3G, 4G, 5G, or 6G connections, etc.).

[0023] Furthermore, an implementation can be envisioned where the detection of the input vector (the value and timestamp of the current curve) is performed on-site in a mobile controller (motor vehicle, machine) and then transmitted to a neural network in the cloud to determine the output vector (upload). In this case, the controller receives the output vector of the neural network as a response to a request (download). This upload / download can be performed wired or wirelessly (e.g., via a WLAN network, 3G, 4G, 5G, or 6G connection, etc.). The neural network used to evaluate the output vector in the cloud is particularly advantageous because, here, even during continuous operation, supervised learning can be more easily continued by the observer (operator / supervisor) by further adjusting the neural network, and permanent optimization of the parameters can be achieved through a large amount of data from different samples operating under actual operating conditions.

[0024] Other advantages and designs of the present invention will become apparent from the description and drawings. Attached Figure Description

[0025] The present invention is schematically illustrated in the accompanying drawings with the aid of embodiments, and is described below with reference to the drawings.

[0026] Figure 1 A solenoid valve is schematically shown in which the method according to the invention can be performed;

[0027] Figure 2 A schematic curve of the current in the solenoid coil of the solenoid valve is shown to illustrate the method according to the invention in a preferred embodiment;

[0028] Figure 3 The current curves and associated time derivatives in the solenoid coils of the operating and blocked solenoid valves during the pull-in phase are shown to explain the method according to the invention in a preferred embodiment.

[0029] Figure 4The flowchart of the method according to the invention in a preferred embodiment using the circuit is shown;

[0030] Figure 5 The flow of the method according to the invention in another preferred embodiment is shown. Detailed Implementation

[0031] exist Figure 1 A solenoid valve 100 is schematically shown, which is exemplaryly applied in a fuel injector 101 and in which the method according to the invention can be performed. The solenoid valve 100 has an electromagnet 110 with an electromagnetic coil 111, which may be designed as a ring, for example. When a (control) voltage U is applied, for example by a computing unit 180 designed as a controller, a current I flows in the electromagnetic coil 111.

[0032] In addition, an armature 120 is provided, which also serves as a valve needle, through which the flow opening 150 can be closed or opened. Furthermore, a spring 130 is provided, which acts on the armature 120 and, without energizing the solenoid coil 111 and therefore without magnetic force, presses the armature 120 into or against the flow opening 150, thus closing the flow opening. The spring 130 can rest against a suitable component of the solenoid valve 100 on its side opposite to the armature.

[0033] When the electromagnetic coil 111 is energized, a magnetic force is generated, and the armature 120 is lifted against the elastic force of the spring 130 and pulled toward the electromagnetic coil 111 or electromagnet 110. The current-passing opening 150 is opened in this situation. When the electromagnetic coil is energized accordingly, the armature 120 can be lifted to the stop 115.

[0034] Figure 2 A schematic signal curve V of the current I in the solenoid coil of the solenoid valve versus time t is shown to illustrate the method according to the invention in a preferred embodiment.

[0035] At time t=0, activation begins by applying a control voltage to the electromagnetic coil, and the current increases (closing phase). At time t... O The bend in the current curve V can be observed; according to the "BIP" (Blow-In-Pulse) method described above, this is the point at which the solenoid valve opens. This bend is caused by the movement of the armature (or valve needle), which lifts up via the solenoid coil when the magnetic force is strong enough.

[0036] Then the current increases further, reaching its maximum value; from then on, it usually transitions from the pull-in phase to a holding phase with a lower current, until the activation period Δt has elapsed. AThe energizing process ends afterward. This can be accomplished by removing the voltage or by applying a clearing voltage. The armature then falls back, and at time t... S Close the flow opening again. According to the mentioned "Ejection Pulse End" (EIP) method, this is the solenoid valve closing time. Here, the total opening time Δt O The difference between the closing and opening times can be used, for example, to determine the amount of fluid injected at that time.

[0037] Figure 3 Various curves of the current I in the solenoid coil of the solenoid valve and the associated second-order time derivative I'' (i.e., d) are shown. 2 I / dt 2 The method according to the invention is described in a preferred embodiment; in particular, only the portion of the curve containing the opening time point is shown here.

[0038] In this case, curve V1 corresponds to the current curve that may appear in a normally operating solenoid valve; specifically, this curve is related to... Figure 2 The curve V or its corresponding segment is equivalent. Because the control voltage used here is relatively high, the opening time point t... O The bend at the point is not like Figure 2 It's not as obvious as the diagram in the image.

[0039] Curve V''1 corresponds to the second time derivative of curve V1, and therefore represents the change in curvature. At the opening time point t... O The maximum value is clearly visible. Traditional detection methods, for example, use a significant maximum value following the minimum value as a characteristic of the valve opening time. However, the curve V''1 shown here is filtered; otherwise, the maximum value would be more difficult to determine or less accurately determined.

[0040] Curve V2 corresponds to the current curve that might occur when the solenoid valve is blocked or malfunctioning; therefore, there is no bend at the turn-on time point. Curve V''2 corresponds to the second time derivative of curve V2, thus representing the change in curvature. Here, for the turn-on time point, the corresponding sharp maximum value following the minimum cannot be determined; however, in some cases, conventional evaluation methods may misinterpret the curve preceding the expected turn-on time point as a less obvious minimum-maximum value and incorrectly report the turn-on time point.

[0041] However, as stated above, within the scope of this invention, in particular, current curves, i.e., without derivatives, i.e., curves V1 or V2, are used directly to determine the variables characterizing the flow opening of the solenoid valve when it is opened and / or closed.

[0042] Figure 4A flowchart of the method according to the invention in a preferred embodiment using the circuit is shown. The current I flowing during the energization of the magnet coil is detected by detecting a measured voltage u decreasing on the measuring shunt 400 and amplifying it, for example, by an amplifier 410. This measured voltage, or its curve, therefore corresponds to the current curve in the magnet coil, for example in… Figure 2 In the middle of V or Figure 3 The voltage is represented by V1 or V2. In the execution computing unit 180, the measured voltage can then be converted into a measurement value that can be processed by software, for example, via an analog-to-digital converter 420. These measurement values, along with possible information about the measurement time points, are then fed as input values ​​425 to the receptor of the (previously trained) artificial neural network, as the pattern recognition method 430 used here.

[0043] Then, the artificial neural network 430 will consider at least one feature variable, such as the opening time point t. O Or close at time t S Alternatively, both can be specified as the output value 435. Similarly, characteristic variables can include values ​​indicating whether the solenoid valve is open (yes / no); in Figure 3 In the case of curve V2, the solenoid valve will not open and can output a value such as 0 (a number indicating "not open").

[0044] The characteristic variables (or several of them) can then be transmitted, for example, to the correction, substitution, emergency, and / or diagnostic function 440, which, for example, corrects the control timing or control voltage used for the electromagnetic coil in the next control cycle or additional injection. These control times or control voltages can then be implemented in the control software 450 to control the electromagnetic coil, for example, by applying a specific control voltage; this is illustrated, for example, by switch 460.

[0045] exist Figure 5 In another preferred embodiment, a flowchart of the method according to the invention is schematically shown, specifically for training a pattern recognition method, such as an artificial neural network 430. For this purpose, several different current curves are supplied as input values ​​525 to the neural network 430, which generates one or more output values ​​535 respectively. The neural network or its weights can then be adjusted, for example, in step 545 by comparison with a comparison value 540.

[0046] As mentioned earlier, it is beneficial to select several current curves for different parameter values ​​of the solenoid valve; for example, three parameters P1, P2 and P3 are shown, which may represent, for example, the control voltage of the solenoid coil, the elastic strength of the spring pressing on the armature and the size of the flow opening.

[0047] In the case of neural networks in the cloud, the input vectors used can be used as parameters P1, P2, and P3 or other parameters for permanent training.

[0048] Then pattern recognition methods trained or learned in this way can be used, such as references. Figure 4 This is explained in order to, for example, determine the opening and closing times of the solenoid valve.

Claims

1. A method for determining variables characterizing the opening and / or closing of a flow opening of a solenoid valve (100), wherein a solenoid coil (111) is energized for the solenoid valve (100) to raise an armature (120) for opening a flow opening (150) for fluid. in, During the operation of the solenoid valve (100), the curve (V, V1, V2) of the current (I) in the solenoid coil (111) is determined. Specifically, an artificial intelligence-based pattern recognition method (430) is used to determine the representative variable by means of at least one segment of the curve or a curve derived therefrom. The characteristic variable is selected from: the opening time of the flow opening (t). O The closing time point of the flow passage opening (t) S The values ​​indicate whether the flow opening is open or closed, and the probabilities of the flow opening being open or closed.

2. The method according to claim 1, wherein, The operation of the solenoid valve (100) and / or diagnostics are performed based on the determined characteristic variables.

3. The method according to claim 1 or 2, wherein, The solenoid valve (100) is used to introduce fuel into the cylinder of the internal combustion engine, or to introduce a reducing agent or reducing agent solution into the exhaust system of the internal combustion engine.

4. The method according to claim 1 or 2, wherein, The solenoid valve (100) is used to meter or dispense the fluid into the cavity.

5. A method for training an artificial intelligence-based pattern recognition method (430) for determining variables characterizing the opening (150) of a flow passage (100) of a solenoid valve (100), wherein an electromagnetic coil (111) is energized for the solenoid valve (100) to raise an armature (120) for opening the flow passage (150) of a fluid. in, For multiple current curves (V, V1, V2) that appear in the electromagnetic coil (111) during the operation of the electromagnetic valve, or curves derived therefrom, at least one segment is provided as input value (525) to the pattern recognition method (430) based on artificial intelligence. The artificial intelligence-based pattern recognition method (430) is adjusted by using the representative variables and their reference values ​​(540) obtained by the artificial intelligence-based pattern recognition method (430) as output values ​​(535) for the aforementioned segment. The multiple current curves are selected for different values ​​of the parameters (P1, P2, P3) of the solenoid valve. The different parameters are selected from: the operating temperature of the fluid and the solenoid valve, the operating time and wear of the solenoid valve, the number of open / close cycles that have occurred, the control voltage (U) of the solenoid coil, the elastic strength of the spring (130) pressing on the armature (120), the size of the flow opening (150), and the operability of the solenoid valve (100).

6. The method according to claim 5, wherein, The multiple current curves are selected from input vectors provided by a remote computing and / or storage system for different values ​​of the parameters (P1, P2, P3) of the solenoid valve.

7. The method according to claim 5 or 6, wherein, The pattern recognition method (430) based on artificial intelligence uses artificial neural networks or "support vector machines".

8. A computing unit (180) configured to perform all method steps of the method according to any one of claims 1 to 7.

9. A computer program product having program code, which, when executed on a computing unit (180), causes the computing unit (180) to perform all method steps of the method according to any one of claims 1 to 7.

10. A machine-readable storage medium having a computer program stored thereon, which, when executed on a computing unit (180), causes the computing unit (180) to perform all the method steps of the method according to any one of claims 1 to 7.

Images