Method for processing a gearbox parameter signal, and method for training and using a neural network by means of the processed gearbox parameter signal
Patent Information
- Application Number
- EP2023722576
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-06-02
- Filing Date
- 2023-05-02
- Publication Date
- 2025-04-09
AI Technical Summary
Existing methods for training neural networks struggle with accurately processing time-continuous signals for transmission parameters, leading to undesirable transitions and incorrect training due to gaps and asymmetries, particularly around gear changes in transmissions.
A method that generates a synthetic signal based on physical measurement variables like clutch pressure and temperature, filters the continuous-time signal around gear changes, and applies a one-dimensional convolution using a kernel to focus on relevant time values, optimizing the training data for neural networks to emulate transmission sensors effectively.
This approach ensures that neural networks are trained optimally for states around gear changes by using only relevant time values, preventing jumps between cut-out areas and improving data quality, thereby enhancing the accuracy of transmission speed determination.
Smart Images
Figure 1.1
Abstract
Description
[0001] ZF Friedrichshafen AG
[0002] Method for processing a transmission parameter signal and method for training and using a neural network by means of the processed transmission parameter signal
[0003] Technical area
[0004] The present invention relates to a method for processing at least one continuous-time signal, wherein the continuous-time signal describes a temporal dependence of a transmission parameter. Furthermore, the invention relates to a method for training a neural network with a processed continuous-time signal, as well as to a method for using a neural network that has been trained using the method according to the present invention. Furthermore, the present invention relates to a computer for executing the method for training the neural network and to a vehicle having a control unit for executing the method for using the neural network.
[0005] State of the art
[0006] Data sets are provided for training neural networks. These are cleaned of undefined values or gaps before the neural network is trained by removing them. Furthermore, individual values, such as outliers, are identified and processed, and deleted if necessary. This contributes to improving data quality. Using probability distributions, such as normal distributions, the data is corrected for asymmetries before training. Furthermore, the data is normalized to scale the data to a uniform value range before training the neural network. When preparing a signal for training a neural network, it may be intended to use only certain value ranges of the signal for training the neural network in order to train the neural network precisely for the states that correspond to the specific value ranges.When selecting these value ranges, and especially when joining cut-out value ranges for training, undesirable transitions can occur at the boundaries of the value ranges, which then incorrectly train the neural network. Description of the invention.
[0007] A first aspect of the invention relates to a method for processing at least one continuous-time signal. The continuous-time signal describes a temporal dependence of a transmission parameter. The signal can be a transmission parameter signal. Transmission parameters can be, for example, clutch pressure, vehicle speed, rotational speed at one or more points within the transmission, or the temperature at one or more points within the transmission. The processed signal can be used for teaching, or training, a neural network.
[0008] The neural network can consist of one or more hidden layers.
[0009] The neural network can be a directed neural network, or alternatively or additionally a recurrent neural network. A transmission speed can be determined using the neural network trained with the processed, continuous-time signal. The trained neural network can represent a virtual transmission sensor at at least one location on the transmission and determine a transmission speed. This allows the neural network to emulate a transmission sensor.
[0010] The method comprises a step of generating a synthetic signal. The synthetic signal comprises information relating to a gear change. The synthetic signal can be a continuous-time signal, wherein information relating to the engaged gear of the transmission can be included over time. The method further comprises a step of filtering the continuous-time signal using the synthetic signal. The filtering of the continuous-time signal is performed as a function of information relating to a gear change. With the filtering step, values of the continuous-time signal which occur around the gear change can be filtered. These filtered values of the continuous-time signal can be used for the further method for processing the continuous-time signal. The method further comprises a step of determining a one-dimensional convolution of the filtered signal using a kernel.The kernel can be a one-dimensional kernel. The kernel can represent a convolution window. Using the kernel, only the filtered part of the continuous-time signal can be convolved in the determination step.
[0011] Thus, a method is provided in which only filtered values are used to determine the one-dimensional convolution by means of the kernel. In other words, the use of the kernel for one-dimensional convolution can be switched on and off using the synthetic signal. The one-dimensional convolution can be switched on and off depending on a gear change. Processing at least one continuous-time signal using such a method can result in only relevant time values of the continuous-time signal being used to train the neural network. As a result, by processing depending on information about a gear change, the neural network can be trained in such a way that a sensor can be optimally emulated for states surrounding a gear change by using only time values of the continuous-time signal surrounding a gear change for training.
[0012] According to a further embodiment, the step of generating the synthetic signal can be carried out as a function of at least one physical measurement variable, wherein the physical measurement variable can comprise information about a gear change. The physical measurement variable can comprise a clutch pressure within the transmission, information about rapid filling within the transmission, and alternatively or additionally information about the state of synchronously rotating parts within the transmission. The step of generating the synthetic signal can be carried out as a function of more than one physical measurement variable. For example, the synthetic signal can be generated as a function of the clutch pressure and temperature at several points in the transmission. Thus, a synthetic signal can be generated which can represent a gear change as realistically as possible as a synthetic signal.
[0013] Thus, a method can be demonstrated in which the processing of the continuous-time signal can depend on physical measured variables. The filtering, and thus the subsequent training, can be performed depending on information about a physically occurring gear change in a transmission. This allows the neural network to be trained with such a processed signal in the best possible way depending on a gear change.
[0014] According to a further embodiment, the step of filtering the continuous-time signal can filter those values of the continuous-time signal that lie within a time range around a gear change time. The time range can be symmetrical around the gear change time. In the step of determining the convolution, only the signal values filtered within a time range around a gear change time can be used. Thus, using the synthetic signal, the determination of the one-dimensional convolution can be enabled within the time ranges around gear change times and disabled in the complementary time ranges.
[0015] Thus, a method can be demonstrated that can process a continuous-time signal in such a way that a neural network can be trained only within time ranges surrounding a gear shift. These time ranges can represent the relevant ranges for training the neural network, since a sensor is to be emulated for these time ranges using the trained neural network. By defining time ranges, it is possible to avoid cutting out sections of the continuous-time signal and then combining such cut-out sections of the continuous-time signal to train the neural network. This prevents the neural network from learning jumps between cut-out sections of the continuous-time signal.
[0016] According to a further embodiment, the method may further comprise a step of determining the kernel. The size and, alternatively or additionally, the structure of the kernel may be determined. For example, the kernel may comprise 50 or 100 points in time of the continuous-time signal. Alternatively or additionally, the kernel may have a specific structure by means of which specific gear changes, for example, gear changes in which one or more gears are skipped, are filtered. The method may further comprise a step of determining the time range around a time of a gear change depending on the step of determining the kernel. The step of determining the time range may be carried out depending on the step of generating the synthetic signal. Thus, the time range around the time of the gear change can be determined. The time range may be a multiple of the kernel, or window.For example, the time domain can be twice or three times as large as the kernel. For example, for a kernel comprising 50 time points of the continuous-time signal, the time domain can comprise 100 or 150 time points of the continuous-time signal. The kernel can be shifted within the time domain to determine the one-dimensional convolution of the filtered signal. For a kernel with 50 time points and a time domain with 100 time points, the kernel can be shifted over a range of 51 time points. The filtering step can be performed independently of the step of determining the time domain. For example, the kernel can be determined first, followed by the time domain. The determined time ranges can then be used for filtering.
[0017] Thus, a method can be demonstrated whereby a specific kernel and a specific time range enable a better response to the continuous-time signal, thus optimally preparing the continuous-time signal for training the neural network. A shorter time range can be selected if dependencies between continuous-time signals are only relevant for training the neural network within a relatively short time range around a gear change. A larger time range can be used if these dependencies exist over a longer period around a gear change. By selecting the optimal time range, the amount of data used to train the neural network can be optimized, i.e., reduced to the relevant data of the continuous-time signal.
[0018] According to a further embodiment, the step of determining the time range around a time of a gear change can be performed depending on a time of a subsequent gear change. Thus, the time range around a time of a gear change can be determined such that it ends at the latest when half the time has elapsed between the time of the currently considered gear change and the subsequent gear change. Thus, a method can be demonstrated whereby jumps during training of the neural network within the continuous-time training signal are avoided. By shifting the kernel, or the window, from one time range to another, and precisely not across the jump, the neural network can be prevented from being trained with such jumps.
[0019] According to a further embodiment, the method can be implemented for processing at least two continuous-time signals. The two continuous-time signals can include, for example, clutch pressures, temperatures at various points in the transmission, or the driving speed of the vehicle. Furthermore, the at least two processed signals can be used to train a neural network. Exactly one neural network can be trained using the at least two processed signals. Thus, the method can first be implemented for the first continuous-time signal, and then the method can be implemented for the second continuous-time signal. Alternatively or additionally, the method can be implemented in parallel for two continuous-time signals. The two processed continuous-time signals can then be used to train the same neural network.
[0020] A method can thus be demonstrated by which several continuous-time signals can be processed. This allows a neural network to be trained, whereby the relationship between different continuous-time signals can be trained. All continuous-time signals with which the neural network is to be trained can be processed using this method. Thus, the neural network can be trained to determine transmission states during a gear change, alternatively or additionally, temporally around a gear change.
[0021] A second aspect of the invention relates to a method for training a neural network. The method comprises a step of conditioning at least one continuous-time signal according to a method according to an embodiment of the first aspect of the invention. In addition, the method for training a neural network can comprise a step of conditioning at least a second continuous-time signal according to an embodiment of the first aspect of the present invention. The method for training the neural network further comprises a step of training the neural network with the conditioned continuous-time signal. Alternatively, the step of training the neural network is carried out with two or more conditioned continuous-time signals. Such training of the neural network can be supervised training of the neural network.
[0022] A third aspect of the invention relates to a method for using a neural network that has been trained using the method according to an embodiment of the second aspect of the invention. The method for using the neural network comprises a step of reading in a first continuous-time signal. The first continuous-time signal can be a clutch pressure. The method for using the neural network further comprises a step of reading in a second continuous-time signal. The second continuous-time signal can be a temperature signal at a location in the transmission. The method for using the neural network further comprises a step of determining a transmission speed using the neural network based on the first and second read-in signals.The method for using the neural network further comprises a step of outputting the determined transmission speed. The steps of reading in the first and second continuous-time signals can occur sequentially or simultaneously. The determining step can occur after the reading steps. The outputting step can occur after the determining step.
[0023] Using such a method, a neural network can be used as a virtual transmission sensor.
[0024] A fourth aspect of the invention relates to a computer which is configured to carry out the method for training a neural network according to an embodiment according to the second aspect of the invention.
[0025] A fifth aspect of the invention relates to a vehicle comprising a control unit, wherein the control unit is configured to carry out the method for using a neural network according to an embodiment according to the third aspect of the invention.
[0026] Short description of the characters
[0027] Figure 1 shows schematically steps of a method for preparing a
[0028] Signals according to one embodiment.
[0029] Figure 2 schematically shows steps of a method for training a neural network according to an embodiment.
[0030] Figure 3 schematically shows steps of a method for using a neural network according to an embodiment.
[0031] Figure 4 shows a schematic diagram of a computer for training a neural
[0032] Network and a vehicle for using the neural network according to one embodiment.
[0033] Detailed description of embodiments
[0034] Fig. 1 schematically shows steps of a method for processing S11 a signal according to one embodiment. The method comprises a step of generating S1 a synthetic signal. The synthetic signal comprises information about a gear change. The generation S1 of the synthetic signal takes place as a function of several physical measured variables, including clutch pressure within a transmission, temperatures at locations within the transmission, and states of synchronously rotating parts within the transmission, such as states of a dog clutch. The method further comprises a step of determining S4 a kernel. The kernel represents a window for a one-dimensional convolution. The kernel, or window, is determined such that a specific number of time values of the time-continuous signal are included in the window. The method further comprises a step of determining S5 a time range around a time of a gear change.The step of determining S5 the time range around a gear shift time depends on the step of determining S4 the kernel. The step of determining S5 the time range depends on the step of generating S1 the synthetic signal. In this way, time ranges around gear shift times are determined. The time ranges are determined depending on the kernel, or window. Thus, symmetrically located time ranges are determined as multiples of the kernel, or window, around the gear shift time.
[0035] The method further comprises a step S2 of filtering the continuous-time signal. In the filtering step S2, those values of the continuous-time signal that lie within a time range around a time of a gear shift are filtered. The time ranges were determined in the determining step S5 such that they do not overlap.
[0036] The method further comprises a step S3 of determining a one-dimensional convolution of a filtered signal using the kernel. In the step S3 of determining the one-dimensional convolution, the kernel, or window, slides within a time domain from one end of the time domain to the other end of the time domain and performs the one-dimensional convolution of the filtered signal. Upon reaching an end of the time domain, the kernel, or window, is shifted to the beginning of a time domain of a temporally subsequent gear change. Such a shift does not result in any convolution beyond a time domain boundary. No one-dimensional convolution is determined across jumps between the time domains. The convolution is only determined in individual time domains around the times of a gear change.
[0037] Fig. 2 schematically shows steps of a method for training a neural network. The method comprises the step S11 of processing at least one continuous-time signal. In the embodiment shown, the processing steps S11 are performed for exactly two continuous-time signals. According to the embodiment shown, a processing step S11 is performed for a clutch pressure signal and a processing step S11 is performed for a temperature signal at a location of the transmission. Furthermore, the method for training the neural network comprises a step S12 of training the neural network with the two processed continuous-time signals.
[0038] Fig. 3 schematically shows steps of a method for using a neural network according to one embodiment. The neural network has been trained using the method shown in Fig. 2. The method for using the neural network comprises a step of reading in S21 a first continuous-time signal. The first continuous-time signal is a clutch pressure. The method for using the neural network further comprises a step of reading in S22 a second continuous-time signal. The second continuous-time signal is a temperature signal at a location in the transmission. The method for using the neural network further comprises a step of determining S23 a transmission speed using the neural network. The determining step S23 is carried out based on the first and second read-in signals, i.e., based on the clutch pressure and the temperature signal.By training the neural network using the method schematically shown in Fig. 2, the neural network is trained to train areas around a gear change. Areas that are not directly around a gear change can be calculated using simple physical relationships. Thus, training the neural network on these areas is not advantageous, as this leads to neglecting the training of the neural network around areas of the gear change. Therefore, in the step S23 of determining the transmission speed, a transmission speed around the time range of a gear change is determined as accurately as possible. The method for using the neural network further comprises a step S24 of outputting the determined transmission speed. The determined transmission speed, determined by the neural network as a virtual transmission sensor, is thus sent to other control units of a vehicle.These other control units can use the signal for calculations.
[0039] Fig. 4 schematically shows a computer 2 for training a neural network. The computer 2 is configured to carry out the method for training a neural network according to the embodiments shown here. Fig. 4 further schematically shows a vehicle 4, comprising a control unit 6, wherein the control unit 6 is configured to carry out the method for using the neural network according to the embodiments shown here. The computer 2 has a transmitter 8, and the vehicle 4 has a receiver 10. The computer 2 is configured to train the neural network. The computer 2 is further configured to send information about the trained neural network to the vehicle 4 via the transmitter 8 and the receiver 10. The control unit 6 of the vehicle 4 is configured to use the neural network sent by the computer 2.
[0040] Reference numerals Computer Vehicle Control unit Transmitter Receiver Generating a synthetic signal Filtering a time-continuous signal Determining a one-dimensional convolution Determining a kernel Determining a time domain Preparing a time-continuous signal Training a neural network Reading in a first signal Reading in a second signal Determining a transmission speed Outputting the transmission speed
Claims
Patent claims 1. Method for processing (S11) at least one continuous-time signal, wherein the continuous-time signal describes a time dependence of a transmission parameter, and wherein the processed signal is usable for training a neural network and wherein a transmission speed can be determined by means of the trained neural network, comprising the steps: generating (S1) a synthetic signal which includes information about a gear change, filtering (S2) the continuous-time signal using the synthetic signal, wherein the filtering (S2) is performed depending on information about a gear change, and determining (S3) a one-dimensional convolution of the filtered signal using a convolution.
2. Method according to claim 1, wherein the step of generating (S1) the synthetic signal is performed depending on at least one physical measurement variable, the physical measurement variable comprising information about a gear change.
3. Method according to one of claims 1 or 2, wherein in the step of filtering (S2) the continuous-time signal those values of the continuous-time signal are filtered which lie within a time range around a time of a gear change and wherein in the step of determining (S3) the convolution only the values of the signal filtered within a time range around a time of a gear change are used.
4. Method according to one of the preceding claims, further comprising a step of determining (S4) the kernel, and a step of determining (S5) the time domain around a time of a gear change depending on the step of determining (S4) the kernel, wherein the step of filtering (S2) is performed depending on the step of determining (S5) the time domain.
5. Method according to claim 4, wherein the step of determining (S5) the time range around a time point of a gear change is performed as a function of a time point of a temporally subsequent gear change.
6. A method according to any of the preceding claims, wherein the method is capable of processing at least two continuous-time signals, and the at least two processed signals are capable of being used to train a neural network.
7. Method for training a neural network, comprising the steps of: preparing (S11 ) at least one continuous-time signal according to a method according to one of claims 1 to 6, and training (S12) the neural network with the prepared continuous-time signal.
8. Method for using a neural network trained using the method according to claim 7, comprising the steps: reading (S21) a first continuous-time signal, reading (S22) a second continuous-time signal, determining (S23) a gear speed using the neural network based on the first and second read-in signals, and outputting (S24) the determined gear speed.
9. Computer (2) which is configured to perform the method for training a neural network according to claim 7.
10. Vehicle (4) comprising a control unit (6) wherein the control unit (6) is configured to execute the method for using a neural network according to claim 8.