WO2022238082A1

WO2022238082A1 - Computer-implemented method and control device for controlling a drive train of a vehicle by means of a convolutional neural network

Info

Publication number: WO2022238082A1
Application number: PCT/EP2022/060026
Authority: WO
Inventors: Ivan SANCHES PROVASE
Original assignee: Vitesco Technologies GmbH
Priority date: 2021-05-10
Filing date: 2022-04-14
Publication date: 2022-11-17

Abstract

The invention relates to a computer-implemented method and a control device for controlling a drive train of a vehicle, the method comprising the following steps: operating the drive train of the vehicle; measuring a plurality of parameters of the drive train and/or determining a plurality of parameters of the drive train; providing a trained convolutional neural network designed to model a parameter of a virtual sensor of the drive train; providing the measured parameters of the drive train and/or the ascertained parameters of the drive train as input data for the trained convolutional neural network and calculating the parameters of the virtual sensor of the drive train as output data by means of the convolutional neural network; controlling the drive train of the vehicle using the calculated parameters of the virtual sensor.

Description

1

description

Computer-implemented method and control apparatus for controlling a powertrain of a vehicle using a convolutional neural network.

The present disclosure relates to a computer-implemented method and a controller configured to execute the computer-implemented method for controlling a powertrain of a vehicle, the computer-implemented method using a convolutional neural network.

The use of mathematical models based on physics is the standard in model development in the powertrain area. More specifically, surrogate models and diagnostic models (also known as virtual sensors) used to identify sensor failures are widely used in the powertrain field to meet on-board diagnostics (OBD) requirements. In the early years of calibration, the amount of data was limited. Today, due to the extensive and detailed validations required by the industry, vehicles are subjected to a greater number of tests before reaching the end customer.

A bad sensor can be identified by reading its value and classified according to its three electrical diagnoses: short to ground, short to positive, and open circuit. A more demanding diagnosis is the functional diagnosis: it consists in assessing whether the value provided by the sensor is appropriate for the given situation. This diagnostic is important to detect a drift in a sensor or a stuck sensor. The current functional OBD is based on the comparison of a specific model with the value of the corresponding sensor: if the difference between these two values is greater than a calibratable threshold, the sensor value is considered invalid. The difficulty here lies in the development of such a model: Most 2 of these models require a high level of knowledge about the physics underlying the system and the development of instruments to calibrate them. Also, these models are generally very specific, which means that very little can be extrapolated (generalized) to any other model. For example, the coolant temperature model of an internal combustion engine (ICE) cannot be used directly to model the charge air cooler system: much more than just an adjustment of the inputs is required, resulting in a new architecture of the physical model and new tools to calibrate it. In addition, both models require an equally large amount of data to be calibrated.

Physical models take as many inputs as necessary to create an image of a particular system, for example the value of the coolant temperature sensor. These physical models, also known as virtual sensors, are used as substitute values when the sensor is considered defective, or they are used to validate the sensor value when no electrical diagnostics are active, which is known as functional diagnostics. Functional diagnostics aim to detect sensor aging drift or sensor sticking, which is common for both negative and positive temperature coefficient (NTC/PTC) thermistors when operating in cold environments (e.g. below -10° C) are exposed. The current physical model used as a reference for emulating a coolant temperature sensor uses as inputs engine speed, intake air mass, vehicle speed, ambient temperature, engine run time, engine idle time, and/or cooler activation.

These quantities are relevant for the physical model. For example, vehicle speed is important because the faster the car drives, the higher the heat exchange between the radiator and the environment, so vehicle speed is relevant to the coolant temperature model. It is similar with the 3

Ambient temperature: The colder the ambient air, the colder the coolant temperature. Only the direction (positive or negative) of the correlations is described here. The thermal equations define the linearity or non-linearity of such correlations. The physical model is in most cases a Simulink model, where unknown factors (such as convection factors, fuel enthalpy, engine efficiency, etc.) are replaced with tables (also called "maps") and constants, and subjected to a regression analysis (using data from the vehicle) to be calibrated. Once the model is calibrated, the functional OBD can be calibrated.

As explained earlier, the goal of the functional diagnosis can be to detect a drifting or stuck sensor. This can be done in a number of ways, but the simplest method is to apply a positive and a negative offset around the modeled value (virtual sensor). As long as the temperature value provided by the real sensor is within this range, the sensor value is considered valid. A defective sensor is detected as soon as the value exceeds the extended range mentioned above. Typical values for the offset are around 20 °C for a temperature sensor (narrower windows are not usually used due to the performance of the current models used and the spread of the temperature sensors).

As explained above, physical models have the disadvantage that they require a high level of knowledge about the physics underlying the system and the development of tools for their calibration. Also, these models are generally very specific, which means that very little can be extrapolated (generalized) to any other model.

The aim of the present disclosure is therefore to provide a computer-implemented method and a control device which is designed to carry out the computer-implemented method which has a parameter for 4

Can calculate control of a drive train of a vehicle in a simple and reliable way.

This aim is achieved by a computer-implemented method having the features of the independent claim and by a control device used to execute the computer-implemented method according to the independent claim. Advantageous embodiments of the computer-implemented method and the control device are set out in the dependent claims.

A computer-implemented method for controlling a powertrain of a vehicle is specified, the method comprising the following steps:

operating the power train of the vehicle. In other words, the vehicle's powertrain is used to e.g. B. to heat the vehicle interior or to move the vehicle.

measuring a plurality of powertrain parameters using at least one powertrain sensor and/or determining a plurality of powertrain parameters using at least one powertrain model. According to one embodiment, the powertrain of the vehicle includes one or more sensors that are used to measure parameters of the powertrain such as temperatures or flow parameters. According to another specific embodiment, the powertrain includes a controller configured to control the powertrain models used to determine parameters of the powertrain. These models can use measured powertrain parameters as input data. A combination of measured and model-determined parameters is also conceivable.

Providing a trained convolutional neural network configured to model a parameter of a virtual powertrain sensor. A convolutional neural network is similar to an artificial neural network, except that the entries of a given vector or matrix are not evaluated flat, but evaluated together by the same "kernel". the kernel 5 consists of the weight matrices and lock values associated with the applied convolution filter. According to one embodiment, the neural convolutional network is trained during the development of the drive train with training data that originates from a test bench, for example.

providing the measured powertrain parameters and/or the determined powertrain parameters as input data to the trained convolutional neural network and computing the parameters of the virtual powertrain sensor as output data with the convolutional neural network.

Controlling the vehicle's powertrain based on the calculated parameters of the virtual sensor. The output of the convolutional neural network is used to control the vehicle's powertrain. For example, the output data of the convolutional neural network is used to control the internal combustion engine, a transmission or an electric motor of the powertrain.

The neuron in a neural network is a sum of linear combinations formed by multiplying the weights by the neuron's inputs and adding a gate value to Flin. Such a combination is then followed by a non-linear function.

Neural networks can be divided into three main types:

ANN - Artificial Neural Network CNN - Convolutional Neural Network RNN - Recurrent Neural Network

Convolutional neural networks are similar to an artificial neural network, except that the entries of a given vector or matrix are not evaluated flat, but evaluated together by the same "kernel". The kernel consists of the weight matrices and lock values associated with the applied convolution filter. The convolution process can be divided into three types:

1-dimensional: mostly used to handle time series data, it consists of a 2D filter applied unidirectionally to the data input 6

2-dimensional: is used for images, for example, and also consists of a 2D filter that moves in 2 dimensions (first horizontally and then vertically) over the input data (2D sweep)

3-dimensional: Filter: used for images that contain multiple layers (e.g. RGB layers). It consists of a 3D filter that moves in 2 dimensions (first horizontally and then vertically) over the entered data (2D sweep)

In other words, the one-dimensional convolution is actually a 2-D filter large enough that convolution filtering can only be applied in one axis (e.g. the time axis) instead of a 2-D sweep.

According to one embodiment, the convolutional neural network uses a max-pooling layer to reduce dimensionality, which has increased in proportion to the number of kernels used. Multiple convolutional layers and max-pooling layers can be used in different orders in a convolutional neural network. Then, a fully connected layer, following the classic structure of an artificial neural network, is used to deliver the neural network output.

Recurrent neural networks have the advantage that, as the name suggests, they are recursive, meaning that the output of any neuron can be connected to the input of the other neurons. In addition, true bi-directional forward propagation can be implemented, reversing the inputs at the same time. The easiest way to illustrate this is when translating languages: in German, for example, the action verb is usually at the end of the sentence. Therefore, the translator (in this case from German to English) must first read the entire sentence before starting the translation (in English, the action verb is usually in the second position of the sentence), because the meaning of the sentence changes significantly as it is not known until the last word is considered. Such an analogy can also be applied to other examples of time series. Due to their architecture, however, recurrent neural networks are usually much larger and cannot be used in engine control units 7

(ECU) can be deployed without significantly increasing the cost (it even requires more memory and CPU power compared to other neural network architectures).

To compensate for this lack of recursion, a hybrid convolutional neural network (HCNN) can be created: in this architecture, the output of the HCNN is written as input for the next interaction. Therefore, not only does the input matrix contain the 9 system inputs (explained in the next section), but the model itself is added as the 10th input. This approach is not only inspired in the recurrent neural networks, but also in the physical model used as a reference: when solving the heat transfer equation, the heat flow is recalculated in each iteration considering the current state of the system, taking the temperature into account .

A neural network requires a large amount of training data to be able to correctly predict its results. Such an amount of training data is usually only available at the end of the calibration process. However, this is not a problem as the OBD tasks are also performed at the end of the calibration process due to their dependency on the other parameters of the calibration. This allows the use of a neural network instead of physical equations to model the OBD reference models (virtual sensors). The advantages here are:

1 ) The model calibrates itself: no additional tools required

2) No specific physics knowledge required: the convolutional neural network is robust enough to ignore irrelevant entries and still provide good accuracy

3) High portability: Once the main architecture is complete, the model can be applied to other systems with just a few changes to the inputs. 8th

The use of the convolutional neural network therefore has the advantage that controlling the powertrain based on the calculated parameters is simple and robust.

According to one embodiment, controlling the powertrain of the vehicle includes using the calculated parameter of the virtual sensor for on-board diagnostics (OBD) of the powertrain of the vehicle, wherein the calculated parameter of the virtual sensor is compared with a measured parameter of a corresponding sensor of the Drive train is compared and wherein the measured parameter of the corresponding sensor is classified as valid if the comparison between the calculated parameter of the virtual sensor and the measured parameter of the corresponding sensor is within a predefined threshold. According to one embodiment, the predefined threshold is stored in a control device of the powertrain.

According to one embodiment, controlling the vehicle's powertrain includes using the calculated virtual sensor parameter as an input parameter for a control function used to control the vehicle's powertrain. In this case, the calculated parameter is used, for example, for a control function for controlling the internal combustion engine or an electric motor of the drive train.

According to one embodiment, the convolutional neural network is a hybrid convolutional neural network, ie a convolutional neural network with at least one additional feedback loop that also uses the output as input. A hybrid convolutional neural network (FICNN) is a convolutional neural network with an additional feedback loop that exhibits behavior similar to a recurrent neural network, where the output (modeled value/calculated parameter) is also used as input. The convolutional aspect of the neural network is important because the correlations between the entries are not just additive or linear correlations, so multiple filters must be used to get the physical correlations correct 9 reconstruct. According to one embodiment, the modeled temperature value is also used as an input, mimicking the widely used approach to solving ordinary differential equations (ODE). This feedback loop is the reason for the term "hybrid" to be added to the classical convolutional neural network.

According to one embodiment, the hybrid convolutional neural network comprises an input layer, at least two one-dimensional convolutional layers, a maximum pooling layer, a flattening layer and an output layer.

The applied network structure is structured as follows:

Input layer: 5x10

First one-dimensional convolution layer

Second one-dimensional convolution layer

Maximum pooling tier

flatten layer

Output layer: 1x1

The input layer consists of the relevant parameters (e.g. ten parameters: vehicle speed, ambient temperature, etc.) stacked vertically. The height of such an array is five rows for the last five values measured at a 1 Hz grid. The first one-dimensional convolutional layer consists of 45 filters. Each filter has a size of 3x10 (ie 30 parameters per filter), which gives 1395 parameters of the first layer (1350 multiplicative parameters and 45 preamps). The second one-dimensional convolution layer consists of 20 filters. Each filter has a size of 2x45 (so 90 parameters per filter), which gives 1820 parameters for the first layer (1800 multiplicative parameters and 20 preset values). The maximum pooling tier is a maximum selection applied to the previous tier with a 2x2 window. The Flatten layer converts the folded structure into a flat structure. The output layer includes 21 parameters. The total number of parameters is 3236 with rectified linear activation function (ReLu) in all neurons. 10

According to one embodiment, no padding is used. Once the first 5x10 inputs are filled (which takes 5 seconds due to the recursion of the convolutional neural network), the convolutional neural network can output its first value. The first line is then deleted. And the new inputs are used in the lower part of the input window in addition to the first output used as the new value for the tenth column.

The process is repeated by moving the input window down one line with each interaction.

With such a reading scheme, the system can be used online: only the last 5 repetitions of each of the 10 input variables need to be stored. Such a small input window is important not only to keep the convolutional neural network small in relation to memory (RAM), but also to reduce the load on the ECU in general by reducing the amount of variables to be stored (larger input windows could result in an impractical implementation - discussions of memory and deployment are covered later in this work). The final architecture is the result of a grid search that included four main sweeps:

Number of filters in the first layer Number of filters in the second layer Drop layers (optional)

activation function

The number of filters in the first and second layer was tested in a range of 5 to 45 and 5 to 20 (with a step of 5), respectively. According to one embodiment, a drop layer with a drop ratio of 10% or 20% can be added after the second convolution layer. The drop layer is useful to avoid overfitting (regularization). According to one embodiment, no overfitting was found since the testing accuracy was similar to the training accuracy.

According to one embodiment, the powertrain parameter calculated using the convolutional neural network is a temperature parameter of the 11

powertrain. According to one embodiment, the measured powertrain parameters and/or the determined powertrain parameters used as input parameters to the convolutional neural network are a vehicle speed, an intake air mass flow rate, a fuel cut-off Boolean value, an engine idling time, an engine running time, a pedal value, a Boolean value indicating that the engine is running and/or a PWM signal from the coolant fan. According to one embodiment, the convolutional neural network is initialized with a measured ambient temperature after a predefined exposure time. According to one embodiment, a setpoint of the neural network is a measured coolant temperature value.

The modeled coolant temperature is the output of the hybrid convolution neural network. Since the input layer is 5x10, this means that at least the first five rows of this column must be filled. The modeled temperature is initialized with the ambient temperature. Although intake air mass flow is in many cases a reflection of driver demand (pedal value), this is not true for conditions such as idle speed, so pedal value can be used together with intake air mass flow to complete the engine load picture.

According to one embodiment, the powertrain parameter calculated using the convolution neural network is a coolant temperature at a coolant flow input of a belt starter generator of the powertrain. According to one embodiment, the measured powertrain parameters and/or the determined powertrain parameters used as input parameters are a belt starter generator current, a vehicle speed, an intake air mass flow rate, a fuel cutoff boolean value, an ambient temperature, an engine speed, a cooling fan activation boolean value , an engine coolant temperature, an intake air temperature before the charge air cooler, an intake air temperature after the charge air cooler, an intake air temperature at the manifold, a pump activation signal, and/or a bypass valve activation signal. 12

According to this embodiment, a virtual sensor for the temperature at the input of a belt starter generator (BSG) electric motor cooled by the water/air intercooler system is reconstructed using the convolutional neural network. The aim of this system is to reduce the cost of implementing a parallel hybrid powertrain by not using dedicated radiators to cool the electric motor, but using a control logic: with a pump and a bypass valve that allows both the intercooler and the BSG to work under the desired conditions. The system provides data, in particular temperature data, which can be used as input data for the convolutional neural network. The inputs to the system are: belt starter generator current, vehicle speed, intake air mass flow, fuel cutoff (boolean), ambient temperature, engine speed, radiator fan activation, engine coolant temperature, intake air temperature before the charge air cooler, intake air temperature after the charge air cooler, intake air temperature at the manifold, a pump activation (PWM signal) and/or a bypass valve activation (PWM signal).

The training data set for training the convolutional neural network for this embodiment was 5600 seconds. The training data base was obtained through measurements on a test bench. However, this relatively small data set was sufficient, mainly because the system is rich in temperature information. An 85%/15% split between training and testing was used.

According to one embodiment, an activation function of the convolutional neural network is a rectified linear unit function (ReLu) that returns 0 for negative inputs and passes the input as an output for positive inputs, or a hyperbolic tangent function (tanh) that satisfies equation

13 follows, where s(z) is the output and where z is any number representing any parameter. According to this embodiment, two different functions are used: Rectified linear unit (ReLu): returns 0 for negative entries and passes the input as an output for positive entries. Hyperbolic tangent (tanh): "squeezes" the inputs into an output from -1 to +1 according to the equation above. These two activation functions provide the required accuracy. However, other activation functions can also be used.

However, according to one embodiment, the sweeps for optimizing the hybrid convolutional neural network are performed randomly and not incrementally as in a "grid search" approach. Therefore, not all combinations were carried out. Once a particular combination gave the expected result, only similar combinations were evaluated.

According to one embodiment, the number of hidden layers is chosen based on the lowest number required to achieve a relatively good result: A first test was performed with only one convolutional layer, but even with several combinations of hyperparameters the convolutional neural network was not able to predict the correct parameter (temperature). Since the results with two layers of convolution were satisfactory, no further increment was made to prevent possible overfitting.

As an initial value, 8700 seconds of measurements divided into lots of 300 seconds (a reasonable average warm-up time for the combustion engine at different ambient temperatures) were collected and used as the main training data for the hybrid convolutional neural network. However, the system exhibited a combination of overfitting and poor test accuracy. The data set was then increased by feature engineering in two steps: increasing the ambient temperature by 3 °C and the measured coolant temperature by 3 °C, with all other parameters remaining unchanged. Reducing the ambient temperature by 3°C and the 14 measured coolant temperature by 3 °C, with all other parameters remaining unchanged. After these steps, the final training dataset was 26100 seconds. According to one embodiment, a training strain split of 85%/15% was used. According to another embodiment, K-fold validation may be applied.

The main 29 measurements for the training data examined a wide variety of conditions and driving cycles of the vehicle, for example: a. Cold start at -30°C followed by 5 minutes idle b. Cold start at -15 °C followed by a city-like driving style c. Start at +30 °C, followed by highly dynamic driving at high loads d. Cold start at -20°C followed by a high load, highly dynamic drive cycle (possible as the test was conducted on a frozen lake in a controlled environment) e. Start at +40 °C, followed by 5 minutes idle

However, all tests had one thing in common: the engine idle time was more than 8 hours, which indicates that the engine could be idle for a long time, and the coolant temperature could be approximated by the ambient temperature.

The reason for this was that the system did not correctly distinguish the initial exposure time from the stop/start strategy, as the engine idle time and the engine running time are reset both on the first start and on each subsequent start. Since the stop/start strategy is mainly active when the engine is hot, this caused the convolutional neural network to first calculate a temperature drop (attempting to regain ambient temperature once the engine dwell time was set, just like when the model was initialized), which was not plausible. This problem could be solved according to one embodiment in the ECU software by setting a flag and inhibiting the model update during the stop/start phase, but no strategy could be easily applied when developing the model, so that was the easiest way , the exposure time and 15 to differentiate the stop/start duration with a higher margin (which facilitates the learning of the convolutional neural network during the training process).

According to one embodiment, three hyperparameters were examined to optimize the model: the compiler, the loss function, and the integration step. According to one embodiment, three different compilers can be used: a. Stochastic gradient descent b. RMSprop c. Adam

Adam can be interpreted as a mixture of RMSprop and Adagrad. Adam implements the exponential moving average of the gradient to scale the learning rate instead of a simple average as in Adagrad. Also, it includes a pulse that accelerates convergence and combines it with the vibration reduction performed by RMSprop.

According to one embodiment, two different loss functions can be used: a. Mean absolute error b. Mean squared error

Both methods had similar final accuracy, with mean square errors allowing for faster convergence (10 loops instead of 20 loops).

According to one embodiment, three different integration steps can be used: a. 0.001 b. 0.005c. 0.01

The value of 0.001 was used because when using RMSprop, the final convergence was severely compromised for larger integration steps. Because even with the smallest step integration, the processing time was less than 20 minutes. It's common in deep learning, along with the three above 16 to carry out a fourth analysis: a sensitivity test with a random initialization value. The reason for this is that, as with any other optimization problem, depending on the initialization (seed) of the weights (“w”), the compiler can get stuck in a local minimum and not find the way to the global minimum; the robustness of the optimization algorithm can be checked by varying the type of initialization. However, a global minimum point is not a mandatory requirement when it comes to optimizing multiple variables. Since satisfactory convergence was achieved when using all of the compilers mentioned in the various runs of the neural convolutional networks (albeit with different performance), initialization sensitivity is not absolutely necessary.

Another characteristic of convolutional networks of any kind is their size and CPU requirements: their precision and high generalization come at a price when it comes to memory and CPU.

Memory requirements are divided into two areas: RAM and Flash. The additional RAM requirement compared to current physical models results from the fact that larger inputs are typically required. In the case of the study, a matrix of 5x10 rather than a similar 1x7 input in the reference model, so memory needs to be provided for these additional 43 variables. Another problem is the calculation itself: the physical model used as a reference has about 800 variables divided into maps and constants, and at each iteration an average of 80 variables must be stored in RAM so that the final parameter (e.g. a coolant temperature) can be calculated.

Even taking into account that each map requires a two-dimensional linear interpolation to calculate its output (so that four memory locations are actually used for one output), the total amount of working memory is still much smaller than the 17 simultaneously processing the larger input and all 3236 parameters of the convolutional neural network.

For the same reason that RAM requirements increase during processing, so does CPU load: the convolutional (hybrid) neural network has only addition and multiplication and a simple ReLu activation (which, compared to a hyperbolic tangent, requires very little required by the CPU), but the number of parameters (about 40 times higher) leads to an average 12-fold CPU load or 8-fold calculation time with the same CPU performance. This value is still reasonable and requires minor trade-offs in the functions in the ECU. The memory requirement in the flash is also high. Basically, the direct relationship between the variables is applied for the memory requirement, which according to one embodiment is 3236 in the (hybrid) convolutional neural network versus 250 for the physical reference model, which corresponds to a 13-fold increase in flash.

The use of neural networks instead of classic physical-mathematical modeling can be done not only when using accuracy as a metric, but also considering the implementation of such a method in currently available flardware (ECU). Such an approach not only reduces development and calibration time, but also has high transferability potential. In this way, resources (time, number of vehicles for the project, team size, etc.) can be saved.

Especially when developing new flybrid powertrains and completely new hardware (DC/DC converters, batteries, BSG), the cost savings in development by using a (hybrid) convolutional neural network for modeling a virtual sensor can become an important source of cost reduction. 18

According to a further aspect of the present disclosure, a control device is specified, the control device comprising a control unit configured to execute a computer-implemented method for controlling a powertrain of a vehicle as described above. The control device is, for example, the control device dedicated to the control of the vehicle engine. According to another embodiment, it is also conceivable that the control device is part of another control device that controls other parts of the vehicle.

Further advantageous embodiments of the present disclosure emerge from the detailed description of exemplary embodiments in connection with the figures.

In the figures shows:

1 shows an example of a cold start at -30° C. with a stuck coolant temperature sensor,

2 shows an exemplary representation of a neuron from a neural network,

3 shows a schematic representation of a convolutional neural network (CNN),

4 shows an exemplary graph of a ReLu activation function,

5 shows an exemplary graph of a hyperbolic tangent activation function (tanh).

6 shows an exemplary accuracy in a training strain starting at 23° C. followed by a city drive.

7 shows an exemplary accuracy in a training strain with a cold start at -13° C., followed by idling and parking maneuvers. 19

8 shows an exemplary test accuracy in a WLTC emissions test cycle,

9 shows an exemplary environment of a belt starter generator mounted together with a charge air cooling water cooling system.

10 shows an exemplary training accuracy for an inlet temperature model of a belt starter generator (training example number 1),

11 shows an exemplary training accuracy for an inlet temperature model of a belt starter generator (training example number 3),

12 shows an exemplary test accuracy for the inlet temperature of a belt starter generator.

FIG. 1 shows an example of a stuck sensor after a cold start at an ambient temperature of minus 32°C. As can be seen at the beginning of the measurement, the sensor value remains stationary and does not move. The modeled coolant temperature continues to increase, as could be physically expected from the coolant temperature itself. As soon as the difference between these two values reaches the calibratable threshold (in this case 20 °C, around the 80-second mark), the diagnosis is made: the system switches to the modeled coolant temperature model and uses its values as for all others Functions of the ECU. Later, as can be seen in figure 1, the sensor starts working again: it is the healing phase. Once the reading enters the extended range, the sensor is considered valid again and the system temperature switches back to the sensor reading from the surrogate model.

FIG. 2 shows a brief summary of the functioning of the basic neuron of a neural network. The neuron is nothing more than a sum of linear combinations (z) formed by multiplying the weights (W) by the neuron's inputs and adding a preload value (b). This combination is then followed by a non-linear function (g). 20

FIG. 3 shows the schematic representation of a convolutional network with a global maximum pooling intermediate layer. The input has n rows as a time entry with k columns as characteristics. In Figure 3, the kernel is a 2x6 matrix. The output after the one-dimensional kernel convolution filtering is always a column vector. Several filters are used to form the convolution layer, which is a horizontal concatenation of the outputs from the kernels.

Figure 4 shows an exemplary graph of a rectified linear unit (ReLu) activation function. The ReLu activation function returns 0 for negative entries and passes the input as an output for positive entries.

Figure 5 shows an exemplary graph of a hyperbolic tangent activation function (tanh). The hyperbolic tangent activation function "squeezes" the inputs into an output from -1 to +1 according to the following equation:

follows, where s(z) is the output and where z is any number representing any parameter.

Figures 6 to 8 show the test results where the convolutional neural network was used to determine the coolant temperature of an internal combustion engine. The convolutional neural network is therefore used as a virtual sensor. The standard metric used in neural networks when using the MSE loss function is the Root Means Squared (RMS) accuracy, ie the square root of the sum of all deviations. The temperature difference between the sensor value and the modeled temperature value is used as a default metric to check the accuracy of the (hybrid) convolutional network. As can be seen in Figures 6 to 8, in 100% of the cases (training and test) a deviation within a range of +-12°C was achieved, with 85% of the cases remaining within a range of +-6°C . In FIGS. 6 to 8, the darker curve represents the measured coolant temperature and the lighter curve represents that through the hybrid convolution network 21 computed modeled coolant temperature. Figure 6 shows exemplary accuracy in a training strain starting at 23°C followed by city driving. Figure 7 shows an exemplary accuracy in a training strain with a cold start at -13 °C, followed by idling and parking manoeuvres. FIG. 8 shows an exemplary test accuracy in a WLTC emissions test cycle.

Figure 9 shows an exemplary environment of a belt starter generator mounted in conjunction with the charge air cooling water cooling system of a vehicle powertrain. The hybrid convolutional neural network architecture was applied to a belt starter generator model. The aim is to reconstruct a virtual sensor for the temperature at the input of an electric motor - belt starter generator (BSG) cooled by the water/air intercooler system. This system aims to reduce the cost of implementing a parallel hybrid powertrain by not using dedicated radiators to cool the electric motor, but using a control logic: with a pump and a bypass valve that allows both the intercooler and the enable the BSG to work under the desired conditions. Such an environment is shown in FIG. Figures 10, 11 and 12 show the verification accuracy of the hybrid convolution neural network applied to this system.

In Figures 10 to 12, the darker curve represents the measured coolant temperature and the lighter curve represents the modeled coolant temperature through the hybrid convolution network. Figure 10 shows an exemplary training accuracy for a BSG inlet temperature model (training example number 1). FIG. 11 shows an exemplary training accuracy for a BSG inlet temperature model (training example number 3). FIG. 12 shows an exemplary test accuracy for a BSG inlet temperature.

A number of implementations have been described. Nonetheless, it should be understood that various modifications can be made without departing from the spirit and scope of the disclosure. Accordingly, other implementations are within the scope of the following claims.

Claims

22 patent claims

A computer-implemented method for controlling a powertrain of a vehicle, comprising the steps of:

operating the powertrain of the vehicle;

measuring a plurality of powertrain parameters using at least one powertrain sensor and/or determining a plurality of powertrain parameters using at least one powertrain model;

providing a trained neural network, the neural network being a convolutional neural network configured to model a parameter of a virtual sensor of the powertrain;

providing the measured powertrain parameters and/or the determined powertrain parameters as input data to the trained convolutional neural network and using the convolutional neural network to calculate the parameters of the virtual powertrain sensor as output data of the convolutional neural network;

Controlling the vehicle's powertrain based on the calculated parameters of the virtual sensor.

2. The computer-implemented method of claim 1, wherein controlling the vehicle's powertrain includes using the calculated virtual sensor parameter for on-board diagnostics of the vehicle's powertrain, wherein the calculated virtual sensor parameter is compared with a measured parameter of a corresponding sensor of the powertrain is compared and wherein the measured parameter of the corresponding sensor is classified as valid if the comparison between the calculated parameter of the virtual sensor and the measured parameter of the corresponding sensor is within a predefined threshold.

3. The computer-implemented method of claim 1, wherein controlling the powertrain of the vehicle includes that the calculated 23

Virtual sensor parameters are used as input parameters for a control function used to control the vehicle's powertrain.

4. A computer-implemented method according to any one of the preceding claims, wherein the convolutional neural network is a hybrid convolutional neural network, i. H. a convolutional neural network with at least one additional feedback loop that also uses the output as input.

5. The computer-implemented method of claim 4, wherein the hybrid convolutional neural network comprises an input layer, at least two one-dimensional convolutional layers, a maximum pooling layer, a flattening layer, and an output layer.

6. The computer-implemented method of claim 1, wherein the powertrain parameter calculated using the convolutional neural network is a powertrain temperature parameter.

7. The computer-implemented method according to claim 6, wherein the measured powertrain parameters and/or the determined powertrain parameters used as input parameters for the convolutional neural network are a vehicle speed, an intake air mass flow rate, a Boolean value for the fuel cut-off, an engine idle time, a engine run time, a pedal value, a Boolean value indicating that the engine is running, and/or a coolant fan PWM signal.

8. A computer-implemented method according to claim 6 or 7, wherein the convolutional neural network is initialized with a measured ambient temperature. 24

9. The computer-implemented method of claim 1, wherein the powertrain parameter calculated using the convolutional neural network is a coolant temperature at a coolant flow input of a belt starter generator of the powertrain.

10. The computer-implemented method of claim 9, wherein the measured powertrain parameters and/or the determined powertrain parameters used as input parameters include a belt starter generator current, a vehicle speed, an intake air mass flow, a boolean value of a fuel cut, an ambient temperature, an engine speed, a boolean value for radiator fan activation, an engine coolant temperature, an intake air temperature before the charge air cooler, an intake air temperature after the charge air cooler, an intake air temperature at the manifold, a pump activation signal, and/or a bypass valve activation signal.

11. Computer-implemented method according to any one of the preceding claims, wherein an activation function of the neural convolutional network is a rectified linear unit function (ReLu) which returns 0 for negative inputs and passes the input as an output for positive inputs, or a hyperbolic tangent (tanh) function , which the equation

12. Control device, comprising a control unit, which is designed to execute the computer-implemented method according to one of claims 1 to 11.