DE102020205688A1 - Method and device for evaluating an operating variable of an injection valve - Google Patents

Abstract

A method for evaluating an operating variable of an injection valve, in particular a solenoid valve for an internal combustion engine, in particular a motor vehicle, having the following steps: providing a signal characterizing the operating variable, evaluating the signal by means of an, in particular artificial, neural convolution network.

Description

State of the art

The disclosure relates to a method for evaluating an operating variable of an injection valve, in particular a solenoid valve for an internal combustion engine, in particular a motor vehicle.

The disclosure also relates to a device for evaluating an operating variable of an injection valve, in particular a solenoid valve for an internal combustion engine, in particular a motor vehicle.

Disclosure of the invention

Exemplary embodiments relate to a method for evaluating an operating variable of an injection valve, in particular a solenoid valve for an internal combustion engine, in particular a motor vehicle, having the following steps: providing a signal characterizing the operating variable, evaluating the signal by means of an, in particular artificial, neural convolution network : convolutional neural network, CNN). This enables an efficient evaluation of the operating size with a comparatively low requirement for computing time and / or memory resources.

In further exemplary embodiments, it is provided that the evaluation includes carrying out at least one one-dimensional convolution operation based on the signal and / or data that can be derived therefrom.

In further exemplary embodiments, the signal is time- and value-discrete and is present, for example, in the form of a large number, in particular series or time series, of digital data values, as in other exemplary embodiments, for example by means of an analog / digital converter based on an output voltage, for example a sensor of the injector can be obtained. In further exemplary embodiments, the provision of the signal can include, for example, determining and / or at least temporarily storing the signal. The sensor can, for example, output an electrical voltage that characterizes a position of a valve needle of the injection valve.

In further exemplary embodiments, the signal can be fed to the convolution network, for example in the form of an input vector, the number of elements of the input vector corresponding, for example, to the number of data values.

In further exemplary embodiments, it is provided that the convolution network has at least one input layer, an output layer, and at least one, in particular hidden, layer arranged between the input layer and the output layer ). For example, the signal, for example in the form of the above-mentioned input vector, can be fed to the at least one input layer.

In further exemplary embodiments it is provided that the convolution network has a first network part which carries out at least one or the at least one (one-dimensional) convolution operation.

In further exemplary embodiments, it is provided that the convolution network has a second network part, which in particular is different from the first network part, and which processes an output signal of the first network part, the second network part in particular at least one further signal, which is at least one further operating variable of the injection valve characterized, and / or data that can be derived therefrom, processed.

In further exemplary embodiments it is provided that one or more further signals are fed to the second network part, e.g. a signal characterizing a fluid pressure (e.g. fuel pressure) in the region of the injection valve. For example, in fuel injection systems with a high-pressure fuel accumulator (“rail” or “common rail”) the further signal can characterize a rail pressure.

In further exemplary embodiments it is provided that the second network part is supplied with a signal that characterizes an activation duration for an actuator (e.g. electromagnet) of the injection valve.

In further exemplary embodiments it is provided that the second network part has at least one layer with fully connected neurons, the second network part in particular determining an output signal, the output signal in particular indicating an operating state and / or a transition between operating states of the Injector characterized.

In further exemplary embodiments it is provided that at least some neurons of at least one layer of the convolution network use at least one of the following elements as an activation function: a) ReLU, b) linear, c) SoftMax.

Further exemplary embodiments relate to a method for training a, in particular artificial, neural convolution network, wherein the convolution network can be used in particular for a method according to the embodiments, the method for training having the following steps: providing training data, dividing the training data into learning data and test data, with the splitting in particular taking place on a random or pseudo-random basis, where in particular an association of the training data with individual injection valves is not taken into account, training of the convolution network based on the learning data and / or the test data.

In further exemplary embodiments, disregarding the association of the training data with individual injection valves increases the spread of the data, so that in a subsequent pattern recognition the property of an individual injection valve (“injector”) is not learned, but the physically based signal curve. This can help the convolution network in further exemplary embodiments to achieve increased generalization.

Further exemplary embodiments relate to a device for carrying out the method according to the embodiments.

Further exemplary embodiments relate to a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the method according to the embodiments.

Further exemplary embodiments relate to a computer program, comprising instructions which, when the program is executed by a computer, cause the computer to execute the method according to the embodiments.

Further exemplary embodiments relate to a data carrier signal that the computer program transmits and / or characterizes according to the embodiments.

Further exemplary embodiments relate to a use of the method according to the embodiments and / or the device according to the embodiments and / or the computer-readable storage medium according to the embodiments and / or the computer program according to the embodiments and / or the data carrier signal according to the embodiments for at least one of the the following elements: a) determining an operating state of the injection valve, b) determining a transition between operating states of the injection valve, c) determining a point in time or the transition between (the) operating states of the injection valve, d) determining a needle closing time of a valve needle of the injection valve, e ) Controlling and / or regulating an operation of the injection valve.

Further features, possible applications and advantages of the invention emerge from the following description of exemplary embodiments of the invention, which are shown in the figures of the drawing. All of the features described or illustrated form the subject matter of the invention individually or in any combination, regardless of how they are summarized in the claims or their reference and regardless of their formulation or representation in the description or in the drawing.

• 1 schematisch ein vereinfachtes Blockdiagramm eines Einspritzventils gemäß beispielhaften Ausführungsformen in einem Zielsystem,
• 2A schematisch ein vereinfachtes Flussdiagramm gemäß weiteren beispielhaften Ausführungsformen,
• 2B schematisch ein vereinfachtes Flussdiagramm gemäß weiteren beispielhaften Ausführungsformen,
• 3 schematisch ein beispielhaftes Faltungsnetzwerk gemäß weiteren beispielhaften Ausführungsformen,
• 4 schematisch ein Signal gemäß weiteren beispielhaften Ausführungsformen,
• 5 schematisch ein vereinfachtes Blockdiagramm einer Vorrichtung gemäß weiteren beispielhaften Ausführungsformen,
• 6 schematisch Signale gemäß weiteren beispielhaften Ausführungsformen, und
• 7, 8 jeweils schematisch ein Histogramm gemäß weiteren beispielhaften Ausführungsformen.

In the drawing shows:

• 1 schematically a simplified block diagram of an injection valve according to exemplary embodiments in a target system,
• 2A schematically a simplified flowchart according to further exemplary embodiments,
• 2 B schematically a simplified flowchart according to further exemplary embodiments,
• 3 schematically an exemplary convolution network according to further exemplary embodiments,
• 4th schematically a signal according to further exemplary embodiments,
• 5 schematically a simplified block diagram of a device according to further exemplary embodiments,
• 6th schematically signals according to further exemplary embodiments, and
• 7th , 8th each schematically a histogram according to further exemplary embodiments.

1 schematically shows a simplified block diagram of an injection valve 10 according to exemplary embodiments in a target system, which in the present case is, for example, an internal combustion engine 20th of a motor vehicle 1 acts. For example, is the injector 10 designed as a solenoid valve and intended to deliver fuel into a combustion chamber (not shown) of the internal combustion engine 20th inject.

Further exemplary embodiments, cf. 2 , relate to a method for evaluating an operating variable of an or the injection valve 10 comprising the following steps: Deploy 100 one the company size characterizing signal S1 , Evaluate 110 of the signal S1 by means of an, in particular artificial, neural network (English: convolutional neural network, CNN) 30th , see e.g. 3 . This enables an efficient evaluation of the operating size with a comparatively low need for computing time and / or memory resources for the evaluation 110 executing device 200 , see below 5 . For example, it can be the operating size of the injection valve 10 be a position or a time course of the position of a valve needle.

In further exemplary embodiments it is provided that the evaluation 110 performing at least one one-dimensional convolution operation 112 based on the signal S1 and / or has data which can be derived therefrom.

In further exemplary embodiments, the signal is S1 Discrete time and values and is available, for example, in the form of a large number of digital data values, as in further exemplary embodiments, for example by means of an analog / digital converter (not shown) based on an output voltage, for example from a sensor (not shown) of the injection valve 10 ( 1 ) can be obtained.

Providing 100 ( 2A) of the signal S1 may, in further exemplary embodiments, have, for example, ascertaining and / or at least temporarily storing the signal.

In further exemplary embodiments, the signal is S1 the convolution network 30th ( 3 ) can be supplied, for example, in the form of an input vector, the number of elements of the input vector corresponding, for example, to the number of data values.

In further exemplary embodiments it is provided that the convolution network 30th at least one input layer 31 has an output layer 39 , and at least one between the input layer 31 and the output layer 39 arranged, in particular hidden, layer (English: hidden layer) 32 , 33 , 34 , 37 , 38 . For example, the at least one input layer 31 the signal S1 , for example in the form of the above-mentioned input vector.

For example, in further exemplary embodiments, the input vector may have 50 elements.

In further exemplary embodiments it is provided that the convolution network 30th a first network part 30a which performs at least one or the at least one convolution operation.

In further exemplary embodiments, the first network part 30a by way of example the following layers: The input layer 31 , a folding layer 32 showing the one-dimensional convolution 112 ( 2A) based on that of the input layer 31 entered data or the signal S1 executes (and the vector from the layer 31 transformed into a 50 x 4 matrix), one layer 33 which is designed to do so, at least temporarily, in particular during training of the convolution network 30th to carry out a regularization method, in particular a predeterminable number of neurons, in particular of the preceding convolution layer 32 to switch off ("dropout"), and a shift 34 that is designed to be used by the dropout layer 33 to transform (“flatten”) the multidimensional data obtained (for example in the form of a 50 x 4 matrix) into one-dimensional data (vector) S30a, the vector having, for example, 200 elements.

In further exemplary embodiments it is provided that the convolution network 30th a second network part 30b has, in particular from the first network part 30a is different, and the one output signal S30a of the first network part 30a processed, in particular the second network part 30b at least one further signal, the at least one further operating variable of the injection valve 10 ( 1 ) characterized, and / or data that can be derived therefrom, processed, in the present case two further signals by way of example S2 , S3 ( 3 ).

In further exemplary embodiments it is provided that the second network part 30b one or more additional signals S1 , S2 are supplied, for example a fluid pressure (for example fuel pressure) in the area of the injection valve 10 characterizing signal S2 . For example, in the case of fuel injection systems with a high-pressure fuel accumulator (“rail” or “common rail”), the further signal can be used S2 characterize a rail pressure.

In further exemplary embodiments it is provided that the second network part 30b a signal S3 is supplied, which is a control duration for an actuator (eg electromagnet) of the injection valve 10 characterized.

In further exemplary embodiments, the second network part 30b an input layer 35 to input the signal S2 and / or an input layer 36 to input the signal S3 on.

In further exemplary embodiments, the signal is S2 and / or the signal S3 each characterized by a scalar quantity.

In further exemplary embodiments it is provided that the second network part 30b a layer 37 for concatenation, i.e. for joining or appending, the input data of the layers 34 , 35 , 36 having. The layer 37 can accordingly have a vector with 202 elements as output data, where 200 elements correspond to the signal S30a, and where the further two elements are the (further) signals S2 , S3 characterize.

In further exemplary embodiments it is provided that the second network part 30b at least one shift 38 , 39 with fully connected neurons (English: fully connected layer), in particular the second network part 30b an output signal S4 determined, in particular the output signal S4 an operating state and / or a transition between operating states of the injection valve 10 characterized.

In further exemplary embodiments, the two layers are 38 , 39 trained as fully connected layers. For example, the layer 38 202 Neurons on, and the layer 39 has 50 neurons, for example. In particular, the layer 39 also the output layer of the convolution network 30th form.

In further exemplary embodiments it is provided that at least some neurons of at least one layer of the convolution network 30th Use at least one of the following elements as an activation function: a) ReLU (rectified linear unit), b) linear, c) SoftMax. In further exemplary embodiments, other activation functions that characterize a non-linearity can also be used instead of ReLU.

In further exemplary embodiments, the convolution network 30th eg to be used for the signal S1 of the injector 10 ( 1 ) to evaluate. 4th shows an example of a signal over time S1 where in one area B1 Operating state transitions occur that can be characterized, for example, by the fact that a valve needle of the injection valve 10 assumes one or more predeterminable positions, in particular an open position or a closed position. In further exemplary embodiments, the assumption of the closed position of the valve needle is synonymous with an injection end of an injection effected by the injection valve. This operating state transition or the point in time of its occurrence can advantageously be used in further exemplary embodiments for monitoring and / or diagnosis and / or control and / or regulation of an operation of the injection valve. In further exemplary embodiments, this operating state transition or the time of its occurrence can be performed using the method according to, for example 2A can be determined efficiently and in a way that conserves resources.

Further exemplary embodiments, cf. 2 B , relate to a method for training an, in particular artificial, neural convolution network 30th ( 3 ), where the convolution network 30th can be used in particular for a method according to the embodiments (cf. eg 2A) wherein the method of training comprises the steps of: providing 150 of training data TrD, splitting 160 the training data TrD into learning data LD and test data TD, in particular the splitting 160 takes place on a random or pseudo-random basis, whereby in particular an association of the training data TrD with individual injection valves is not taken into account, training 170 of the convolution network 30th based on the learning data LD and / or the test data TD. For exercising 170 In further exemplary embodiments, training methods known to those skilled in the art for neural networks, in particular convolution networks, can be used.

Further exemplary embodiments, cf. 5 , refer to a device 200 to carry out the procedure (see e.g. 2A and or 2 B) according to the embodiments.

The device 200 has, for example: one at least one computation kernel 202a having computing device (“computer”) 202, one of the computing device 202 associated storage device 204 for at least temporary storage of at least one of the following elements: a) data DAT, b) computer program PRG, in particular for executing a method according to the embodiments.

In further exemplary embodiments, the computer program PRG can also characterize at least one application (hereinafter also referred to as “application”) APP1, APP2, APP3, which are at least temporarily on the computing device 202 is carried out, for example to control the injection valve 10 or of components (eg magnetic actuator) thereof.

In further preferred embodiments, the memory device 204 volatile memory 204a (e.g. main memory (RAM)) and / or a non-volatile memory 204b (e.g. Flash EEPROM).

In further exemplary embodiments, the computing device 202 at least one of the following elements has or is as at least one of these elements is formed: microprocessor (µP), microcontroller (µC), application-specific integrated circuit (ASIC), system on chip (SoC), programmable logic module (e.g. FPGA, field programmable gate array), hardware circuit, graphics processor (GPU), accelerator circuit for evaluating / executing the convolution network 30th , or any combination of these.

Further exemplary embodiments relate to a computer-readable storage medium SM, comprising instructions PRG which, when executed by a computer 202 cause them to carry out the method according to the embodiments.

Further exemplary embodiments relate to a computer program PRG, comprising instructions that are used when the program is executed by a computer 202 cause them to carry out the method according to the embodiments.

Further exemplary embodiments relate to a data carrier signal DCS that characterizes and / or transmits the computer program PRG according to the embodiments. The data carrier signal DCS is, for example, via an optional data interface 206 the device 200 receivable.

In further exemplary embodiments, the data interface 206 also to receive at least one of the signals S1 , S2 , S3 and / or for outputting the signal S4 be used.

Further exemplary embodiments relate to a use of the method according to the embodiments and / or the device according to the embodiments and / or the computer-readable storage medium according to the embodiments and / or the computer program according to the embodiments and / or the data carrier signal according to the embodiments for at least one of the the following elements: a) Determining an operating state of the injection valve 10 , b) determining a transition between operating states of the injection valve 10 , c) determining a point in time or the transition between (the) operating states of the injection valve 10 , d) determining a needle closing time of a valve needle of the injection valve 10 , e) controlling and / or regulating an operation of the injection valve 10 .

Further exemplary aspects and embodiments are described below, each of which can be combined individually or in any combination with at least one of the exemplary embodiments described above.

In further exemplary embodiments, a target position of the needle closing of the valve needle of the injection valve can be used during training of the convolution network 30th ( 3 ) can be used as a scalar value, with the output layer 39 for example can have a (single) neuron. In further exemplary embodiments, this can also be interpreted to the effect that the convolution network 30th solves a regression problem, which is referred to below as "variant 1" as an example.

In further exemplary embodiments, hereinafter referred to as “variant 2, variant 3”, the target position of the needle closing can be OneHot-coded (ie a vector is formed that has a “one” at the target position during learning and otherwise “zero” at all other positions), whereby the number of possibilities or elements of the vector corresponds to the desired accuracy, e.g. 50 possible positions.

In further exemplary embodiments, in particular also in variant 2, in particular the last layer (output layer) is 39 of the convolution network 30th as "fully connected" (all connections to the previous layer 38 are occupied). The activation function is, for example, either linear (variant 2) or "SoftMax" (variant 3). The neuron with the largest output corresponds, for example, to the predicted position, see fig. 6th .

6th shows schematically signals according to further exemplary embodiments, the signal S4 'a setpoint behavior of the injection valve 10 regarding the through the signal S1 characterized company size S1 and the signal S4 ″ a behavior of the injection valve predicted using exemplary embodiments (in particular variant 3 mentioned above) 10 reproduces.

7th and 8th each schematically show a histogram according to further exemplary embodiments, based on variant 3 described above. 7th indicates a frequency of deviation of the prediction from the target value, and 8th gives this frequency in contrast to 7th with logarithmic scaling of the vertical axis.

In further exemplary embodiments, the principle according to the embodiments enables efficient detection of, for example, a predeterminable pattern in a signal S1 or in one the signal S1 characterizing time course. In particular, the principle according to the embodiments can also be applied to other signals or other systems and is not limited to the evaluation of signals from an injection valve.

Claims (13)

Method for evaluating an operating variable of an injection valve (10), in particular a solenoid valve for an internal combustion engine (20), in particular a motor vehicle (1), comprising the following steps: providing (100) a signal (S1) characterizing the operating variable, evaluating (110) of the signal (S1) by means of an, in particular artificial, neural convolution network (30). Procedure according to Claim 1 wherein the evaluation (110) comprises the execution of at least one one-dimensional convolution operation (112) based on the signal (S1) and / or data which can be derived therefrom. Method according to at least one of the preceding claims, wherein the convolution network (30) has at least one input layer (31, 35, 36), one output layer (39), and at least one, in particular hidden, layer (32, 32, 33, 34, 37, 38). Method according to at least one of the preceding claims, wherein the convolution network (30) has a first network part (30a) which carries out at least one or the at least one convolution operation (112). Procedure according to Claim 4 , wherein the convolution network (30) has a second network part (30b) which processes an output signal (S30a) of the first network part (30a), in particular the second network part (30b) at least one further signal (S2, S3), the at least one further operating variable of the injection valve (10) characterized and / or data that can be derived therefrom, processed. Procedure according to Claim 5 , the second network part (30b) having at least one layer (38, 39) with completely connected neurons, the second network part (30b) in particular determining an output signal (S4), the output signal (S4) in particular having an operating state and / or a Characterized transition between operating states of the injection valve (10). Method according to at least one of the preceding claims, wherein at least some neurons of at least one layer of the convolution network (30) use at least one of the following elements as an activation function: a) ReLU, b) linear, c) SoftMax. Method for training a, in particular artificial, neural convolution network (30), wherein the convolution network (30) can in particular be used for a method according to at least one of the preceding claims, wherein the method for training has the following steps: providing (150) training data ( TrD), splitting (160) the training data (TD) into learning data (LD) and test data (TD), with the splitting (160) taking place in a random or pseudo-random manner, whereby in particular an association of the training data (TrD) with individual injection valves is not taken into account , Training (170) the convolution network (30) based on the learning data (LD) and / or the test data (TD). Device (200) for carrying out the method according to at least one of the preceding claims. Computer-readable storage medium (SM), comprising instructions (PRG) which, when executed by a computer (202), cause the computer (202) to perform the method according to at least one of Claims 1 until 8th to execute. Computer program (PRG), comprising instructions which, when the program (PRG) is executed by a computer (202), cause the computer (202) to execute the method according to at least one of the Claims 1 until 8th to execute. Data carrier signal (DCS), which the computer program (PRG) according to Claim 11 transmits and / or characterizes. Use of the method according to at least one of the Claims 1 until 8th and / or the device (200) Claim 9 and / or the computer-readable storage medium (SM) according to Claim 10 and / or the computer program (PRG) Claim 11 and / or the data carrier signal (DCS) Claim 12 for at least one of the following elements: a) determining an operating state of the injection valve (10), b) determining a transition between operating states of the injection valve (10), c) determining a point in time or the transition between operating states of the injection valve (10), d ) Determining a needle closing time of a valve needle of the injection valve (10), e) controlling and / or regulating (120) an operation of the injection valve (10).

Images