Method for determining a state variable of a machine by combining a rule-based model and a neural network
Patent Information
- Application Number
- PCT/EP2024/067580
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-07-24
- Filing Date
- 2024-06-24
- Publication Date
- 2025-03-20
AI Technical Summary
Existing methods for determining the state size of machines using either rule-based models or neural networks face limitations, with neural networks being restricted by a lack of training data and rule-based models simplifying physical reality, and both struggling to accurately represent complex machine states.
A procedure that combines a rule-based model and a neural network to determine a machine's state size by creating a training data set with threshold values to ensure sufficient data relevance, allowing the neural network to learn connections between state variables and using rule-based models for less data-rich areas, and strategically selecting between the two based on data availability and critical conditions.
This hybrid approach ensures accurate and precise determination of machine state sizes by leveraging the strengths of both models, improving usability and applicability across various machine conditions, especially in critical states where precise data is needed.
Smart Images
Figure EP2024067580_20032025_PF_FP_ABST
Abstract
Description
[0001] Method for determining a state variable of a machine by a combination of a rule-based model and a neural network
[0002] Technical area
[0003] The present invention relates to a method for providing a training data set for training a neural network, wherein the neural network is configured to determine a state variable of a machine. The present invention further relates to a method for training a neural network with a training data set provided in this way. The present invention further relates to a method for determining a rule-based model that is configured to determine a state variable of a machine. The present invention further relates to a method for determining a state variable of a machine by combining a rule-based model and a neural network. The present invention further relates to a computing unit configured to carry out steps of such methods.Furthermore, the present invention relates to a machine with such a computing unit.
[0004] State of the art
[0005] Methods and computing units are known from the prior art by means of which state variables of machines can be determined. On the one hand, methods and computing units are known in which the state variable of the machine is determined using a rule-based model. Alternatively, in the prior art, the state variable of the machine is determined using a neural network, wherein the neural network has been previously trained using training data so that it is configured to determine the state variable of the machine. However, the usability of such neural networks can be limited due to a lack of training data sets. On the other hand, the usability of rule-based models can be limited by the fact that they can often only represent the physical reality of the machine in a simplified manner. Description of the Invention
[0006] In a first aspect, the present invention relates to a method for providing a training data set for training a neural network. The training data set can comprise input data and output data, so that the neural network can be trained therewith in such a way that output data can be determined as a function of input data. The neural network can be a convolutional neural network or a fully connected neural network. The neural network is configured to determine a state variable of a machine. The machine can be a motor vehicle, for example a passenger car or a truck. The machine can be a work machine, such as an agricultural machine or a construction machine. The machine can have a device, for example a vehicle clutch. The neural network can, for example, be configured to determine a state variable of the vehicle clutch.The state variable can be, for example, a temperature, such as the clutch temperature of the vehicle clutch, or a sump temperature, for example of the transmission or engine; a speed, such as the engine speed; or a pressure, for example between two elements, such as two friction elements of the vehicle clutch. State variables can include operating parameters and, alternatively or additionally, physical measurement data of the machine.
[0007] The method comprises reading in operating data of the machine, wherein the operating data contain a relationship between at least two state variables of the machine, and wherein the operating data contain different value ranges of the state variables. A first state variable can be, for example, a rotational speed, and a second state variable can be, for example, a temperature. The different value ranges of the state variables can be, for example, equidistant value ranges of the state variables, or alternatively, value ranges of different sizes. For example, the temperature can be divided into value ranges in 10°C increments, while the rotational speed can be divided into value ranges in 100 revolutions per minute increments, for example.The operating data can, for example, be represented as a matrix, with value ranges of the first state variable plotted on a first axis and value ranges of the second state variable plotted on a second axis. Values can be entered in the individual fields of the matrix. These values can represent how often data is available for the respective value ranges of the state variables. The operating data can contain physical measurement data, and the values in the fields of the matrix can indicate how often measurement data is present in the operating data for certain value ranges. The operating data can therefore show a relationship between the two state variables of the machine. Alternatively or additionally, in addition to the operating data from the machine, further operating data from other machines can be read in.
[0008] The method further comprises determining, based on a threshold, for which value ranges sufficient operating data is available. The determination can be made based on one or more thresholds. The threshold can be an absolute value. For example, all value ranges for which the corresponding value is above the threshold can be determined. For example, for each field in the matrix, it can be determined whether sufficient data is available for the value ranges of the two state variables, based on the threshold.
[0009] The method further comprises adding the operating data of the specific value ranges to the training data set. For example, only the operating data that was determined in the previous step based on the threshold value can be added to the training data set. For example, the operating data for which more data is available than defined by the threshold value can be added to the training data set. For example, if the threshold value defines that 100 values must be available for a specific speed and a specific temperature, i.e., for example, 100 measured physical measurement data, these specific value ranges of speed and temperature are added to the training data set. This operating data can then be added to the training data set as training data. The training data set can therefore contain the combination of value ranges, for example, a specific temperature and a specific speed.
[0010] Such a method can ensure that only the operating data for which sufficient data, e.g. physical measurement data, is added to the training dataset. By determining based on the threshold value, it can be ensured that sufficient data is available for certain value ranges of the state variables. This can ensure that the relationship between the two state variables can be extracted from the operating data based on statistical relevance and transferred to the training dataset. If a neural network is trained with such a training dataset, it can be ensured that the neural network has learned the relationship between the two state variables sufficiently well to be able to determine one state variable as a function of the other with sufficient accuracy after training, for example.
[0011] According to a further embodiment, when determining for which value ranges sufficient operating data is available, different threshold values can be used for different value ranges. For example, a different threshold value can be used for a first value range of the first state variable than for a second value range of the first state variable. Alternatively or additionally, a different threshold value can be used for a first value range of the second state variable than for a second value range of the second state variable. Alternatively or additionally, there can be clusters of value ranges that have the same threshold values, whereby different clusters can have different threshold values.
[0012] For example, for the speed as a state variable, it can be determined that value ranges corresponding to the idle speed of a machine's drive motor, where the speed corresponds to the speed of an output shaft of the motor, have a different threshold than speeds that do not correspond to idle speed. For example, the threshold for the value ranges corresponding to idle speed can be infinite, so that no operating data is determined for these value ranges and added to the training data set.
[0013] This can, for example, prevent the neural network from learning a correlation with idle time from the training data. Different thresholds in different value ranges can be used to functionally define the thresholds and thus the training data set. This allows for a functional selection of specific value ranges of the operating data for addition to the training data set.
[0014] According to a further embodiment, a user specification for the threshold value can also be read in. The user specification can, for example, include information that idling occurs at a certain speed. A user specification can be read in at one or more threshold values. Furthermore, the threshold value can be determined depending on the read-in user specification. The threshold value can be determined depending on one or more, for example, even all, read-in user specifications. The user specification can, for example, define that idling occurs at 1000 revolutions per minute. Consequently, the threshold value for this speed can be set to infinity.
[0015] This allows the user to define which thresholds should be used for which value ranges. A functional definition of which operating data can be included in the training dataset can thus be made by the user.
[0016] According to a further embodiment, the threshold value can also be determined as a function of the read-in operating data. The threshold value can be determined as a function of the read-in operating data as an alternative to or in addition to the threshold value being determined as a function of the read-in user specification. For example, it can be detected that the rotational speed is used as the first state variable. Based on information stored in a memory of the machine, the method can be used to detect that idling occurs at a specific rotational speed, for example 1000 revolutions per minute. A suitable threshold value can be determined as a function of this. The threshold value can be determined as a function of at least one state variable of the operating data.
[0017] This makes it possible to automatically define different thresholds for different value ranges of the operating data. User input is not necessary. For example, the different thresholds can ensure that only operating data from specific value ranges is added to the training dataset. This ensures that only certain relationships between the state variables, reflected in the operating data, are also found in the training dataset and thus learned by the neural network.
[0018] In a second aspect, the present invention relates to a method for training a neural network. The method for training the neural network can be configured to train the neural network such that the neural network learns a relationship between a first state variable and a second state variable of a machine. The neural network is configured to determine a state variable of a machine, for example, to determine a second state variable of the machine as a function of a first state variable of the machine. The method for training the neural network can be carried out using a training data set that was provided according to an embodiment of the first aspect of the present invention.This allows the neural network training method to precisely train the neural network to the value ranges for which sufficient operating data is available, based on the threshold value. This ensures that the neural network can be trained to detect a relationship between the two state variables that is reflected in the operating data. The training method can thus automatically define the value ranges of the operating data for which the neural network can be used after training. In addition to training, the neural network can be validated after training.
[0019] In a third aspect, the present invention relates to a method for determining a rule-based model. The rule-based model can, for example, deterministically describe a relationship between a first state variable and a second state variable of the machine using physical relationships. For example, the rule-based model can describe the relationship between the two state variables using a function and, in doing so, determine a second state variable depending on a first state variable. The method for determining the rule-based model comprises reading in operating data of the machine. With regard to the definition of the terms state variable, operating data, and machine, reference is made to the definition relating to the first aspect of the invention. The operating data comprise a relationship between at least two state variables of the machine, wherein the operating data comprise different value ranges of the state variables.With regard to the definition of value ranges and the relationship, reference is also made to the definition in the description of the first aspect of the present invention. Furthermore, the method comprises determining, based on a threshold value, for which value ranges less operating data is available than the threshold value specifies. The threshold value can, for example, be the same threshold value that was also used in the first aspect of the present invention; alternatively, the determination can be based on a different threshold value. The determination can be based on one or more threshold values. During the determination, as already defined with regard to the first aspect, a value of the value ranges can be compared with the threshold value. The value can indicate how much physical measurement data is available for the value ranges in the operating data.Furthermore, at least one rule-based model is determined for the specific value ranges below the threshold value. One or more rule-based models can be determined for specific value ranges. The determination of at least one rule-based model can only take place for the specific value ranges below the threshold value. Specific rule-based models can be determined for specific clusters of value ranges of the operating data, wherein different rule-based models can be determined for different clusters. For example, if only a few values are available for a specific speed and a specific temperature as the first and second state variables of the operating data, and these values are below the threshold value, it can be determined that less operating data is available for these value ranges than defined by the threshold value. A rule-based model can then be determined for these value ranges.This rule-based model can represent the relationship in the value range between the first and the second state variable.
[0020] This demonstrates a method by which a relationship between the first and second state variables of the machine can be described using a rule-based model. The rule-based model can be determined only for those value ranges for which fewer values are present in the operating data than defined by the threshold. It is not necessary to create a rule-based model for those value ranges of the operating data for which more values are present than defined by the threshold. This allows the determination of the rule-based model to be specifically tailored to certain value ranges, which can make determining the rule-based model easier. This also allows a more precise representation of the physical relationship to be implemented using the rule-based model.
[0021] According to a further embodiment, a determination can also be made, based on a threshold, as to which value ranges have sufficient operating data. This threshold can be the same threshold as the threshold for determining which value ranges have less operating data than the threshold; alternatively, it can be a different threshold. Furthermore, a rule-based model can be determined for the determined value ranges above the threshold. One or more rule-based models can be determined for the determined value ranges above the threshold.
[0022] Thus, a rule-based model can be determined for some value ranges for which training data has also been provided using the method described in the first aspect. Using such a method, a rule-based model can also be determined for value ranges that have more values than the threshold defined.
[0023] In a fourth aspect, the present invention relates to a method for determining a state variable of a machine using a combination of a rule-based model and a neural network. With regard to the definition of state variable, machine, rule-based model, and neural network, reference is made to the definition in the two previous aspects. The method comprises reading in a value of a first state variable. Furthermore, based on the read-in value of the first state variable, it is determined whether the rule-based model or the neural network should be used to determine a second state variable of the machine. The determination can be made, for example, via a filter or a software function. If the rule-based model is to be used, a value of the second state variable is determined using the rule-based model depending on the read-in value of the first state variable.If the neural network is to be used, the value of the second state variable is determined using the neural network based on the input value of the first state variable. The value of the second state variable can be determined using either the rule-based model or the neural network.
[0024] This shows a method in which the value of the second state variable is determined using either the rule-based model or the neural network. If the rule-based model is used, the second state variable can be determined deterministically as a function of the first state variable. It is possible to understand how the step of determining the second state variable is carried out. Furthermore, the rule-based model can be used for different values of the first state variable. The use of the rule-based model does not necessarily have to be restricted to previously known values. If the neural network is used to determine the second state variable, the second state variable can be determined more precisely as a function of the first state variable. This allows for a more precise determination than if the model were used with a rule-based model.However, determining the value with the neural network can be restricted to a predefined value range of the first state variable. For example, the predefined first value range can depend on training data with which the neural network was trained. This allows the advantage of the rule-based model to be combined with the advantage of the neural network. A strategic combination of the rule-based model and the neural network can be used, whereby the neural network can be used whenever, for example, the use of the neural network is possible and sensible. This can be the case, for example, if the corresponding value range for a specific value is also part of the training data set. If the use of the rule-based model is necessary, it can be used, for example, to determine the second state variable.The method can also be used to determine the second state variable for values of the first state variable for which no neural network has been trained. This makes it possible to provide a hybrid model for determining the state variable. Such a method can therefore be used in different situations to determine the state variable. According to a further embodiment, the rule-based model can be used if a critical state of the machine exists. A critical state of the machine can exist, for example, if there is a particularly high mechanical load on the machine, for example due to a high speed, or a high thermal load on the machine, for example due to a high temperature. In a critical state, it may be necessary, for example, for the state variable to be determined deterministically and to be traceable.The rule-based model can determine the state variable deterministically.
[0025] Thus, the method can also be used to determine state variables of the machine in a critical state and, alternatively or additionally, for system-critical functions. At the same time, the method can use the neural network for non-critical states to determine the state variables. This expands the applicability of the method.
[0026] According to a further embodiment, the critical state of the machine can be determined depending on the read-in value of the first state variable. For example, the first state variable can be a rotational speed. The critical state can be determined if, for example, the rotational speed is above a certain threshold, for example, above 5000 revolutions per minute. The critical state can be determined alternatively or additionally depending on read-in values of other state variables.
[0027] This allows the critical state to be determined automatically. The method can then automatically determine whether the rule-based model or the neural network should be used. This allows system-critical functions to be automatically operated using the determined state variable.
[0028] According to a further embodiment, the neural network may have been trained using a method according to an embodiment according to the second aspect of the present invention. Alternatively or additionally, the rule-based model may have been determined using a method according to an embodiment according to the third aspect of the present invention.
[0029] According to a further embodiment, the neural network can be used if the read-in value of the first state variable lies within a first specific value range of the first state variable. The first specific value range can include the values that lie above the threshold value, which is also used to provide the training data set. Alternatively or additionally, the neural network can be used if the read-in value of the first state variable is present in the training data set.
[0030] Thus, the method can always use the neural network when it has been trained for the first specific value range. This allows the value of the second state variable to be determined most accurately, and always when the neural network has been trained for a suitable value of the first state variable.
[0031] According to a further embodiment, the rule-based model can be used if the read-in value of the first state variable lies within a second specific value range of the first state variable. The second specific value range can include the values that lie below the threshold value that is also used to provide the training data set. This second specific value range can be the inverse of the first specific value range of the previous embodiment with regard to all operating data. Thus, for example, the rule-based model can always be used when the neural network is not being used, and vice versa. Alternatively or additionally, the rule-based model can be used if, for a value of the operating data, this value is not present in the training data. Alternatively, the first specific value range and the second specific value range can at least partially overlap.
[0032] This ensures that the rule-based model is always used when the neural network has not been trained for a specific value of the first state variable. This ensures that the value of the second state variable is determined using the rule-based model whenever the neural network has not been trained to determine the value of the second state variable.
[0033] According to a further embodiment, for a specific read-in value of the first state variable, both the value of the second state variable can be determined using the rule-based model and the value of the second state variable can be determined using the neural network. The simultaneous determination using the rule-based model and the neural network can be performed for several specific read-in values of the first state variable, or alternatively, only for one read-in value of the first state variable. Two values of the second state variable can be determined using the method. A first value of the second state variable can be determined by the rule-based model, and a second value by the neural network. These two values can be identical or different. If they are identical, the plausibility of determining the second state variable using the neural network is verified by the determination using the rule-based model.Two values that lie within a predetermined error tolerance can also be considered identical. If they differ, the value determined using the rule-based model can be used as the value of the machine's state variable determined by the method.
[0034] This allows validation or plausibility checks of the neural network with the rule-based model. This makes it possible to check whether the neural network has been trained correctly. This can provide information about how well the method for determining the machine's state variable functions and operates. This can be advantageous, for example, for system-critical functions or for critical value ranges. The neural network can determine the value of the second state variable with particular precision. At the same time, by determining it using the rule-based model, this value of the second state variable can be verified and validated.
[0035] According to a further embodiment, the rule-based model can comprise different physical sub-models for different value ranges of the first state variable. Different physical sub-models can have different functions and, alternatively or additionally, can be parameterized differently. For example, a first physical sub-model can be provided for a first value range of the first state variable, and a second physical sub-model can be provided for a second value range that is different from the first value range, wherein the first and second physical sub-models are different. Furthermore, a sub-model can be selected depending on the read-in value of the first state variable. If the read-in value is part of the first value range, the first sub-model is selected. If the read-in value is part of the second value range, the second sub-model is selected.Furthermore, the value of the second state variable can be determined using the selected submodel. Thus, if the first submodel has been selected, the value of the second state variable can be determined using the first submodel. Alternatively, the value of the second state variable can be determined using the second submodel.
[0036] Thus, different relationships or different physical submodels can be defined for different value ranges of the first state variable. These can be selected depending on the value of the first state variable to determine the value of the second state variable. This can increase accuracy by using specific submodels for specific value ranges, and the submodels can be different and independent of each other. Using different submodels can also simplify the determination of the rule-based model, since physical relationships only need to apply to specific value ranges.
[0037] In a fifth aspect, the present invention relates to a computing unit configured to execute steps of a method of an embodiment of the preceding aspects. All of the methods described above can be computer-implementable methods. The computing unit can be a control unit of a machine embodied as a motor vehicle. Alternatively or additionally, the computing unit can be a server or a computer, which can be external to the machine.
[0038] In a sixth aspect, the present invention relates to a machine with a computing unit according to an embodiment of the fifth aspect of the invention. The machine can be a passenger car or a truck, alternatively a work machine, such as an agricultural machine or a construction machine. The machine can have a first computing unit, for example a control unit, which can be configured to carry out a method according to an embodiment of the fourth aspect of the invention. The machine can be connected to a second computing unit via communication interfaces. The second computing unit can be configured to carry out a method according to embodiments of the first, second and third aspects of the invention. The second computing unit can thus be configured to provide a training data set and to train the neural network with such a data set.Furthermore, the second computing unit can be configured to determine a rule-based model. The second computing unit can be configured to read operating data from multiple machines located in the field and thereby execute the method. Alternatively, only the machine's operating data is read. The second computing unit can be configured to send both the determined neural network and the determined rule-based model to the machine and the first computing unit via the communication interface. Thus, the first computing unit can execute the method for determining the state variable of the machine using the rule-based model and the neural network.
[0039] Short description of the characters
[0040] Figure 1 schematically shows steps of a method for providing a training data set for training a neural network.
[0041] Figure 2 shows schematically steps of a method for determining a rule-based model.
[0042] Figure 3 schematically shows steps of a method for determining a state variable of a machine by a combination of a rule-based model and a neural network.
[0043] Figure 4 shows schematically a machine and computing units for carrying out the methods schematically shown in Figures 1 to 3.
[0044] Detailed description of embodiments
[0045] Figure 1 schematically shows steps of a method for providing a training data set for training a neural network. Operating data from a machine 6, which is schematically shown as a motor vehicle in Fig. 4, is read in S1. In an alternative embodiment, operating data from one or more machines (not shown) are read in. The operating data show a relationship between at least two state variables of the machine 6. The operating data have a plurality of values. Furthermore, the operating data have different value ranges of the state variables, wherein the values of the operating data are arranged in the different value ranges of the state variables. Furthermore, a determination S2 is made, based on a threshold value, for which value ranges sufficient operating data is available. During determination S2, it is determined whether or not more values than defined by the threshold value are present for a value range.This determines which value ranges have sufficient values, i.e., more than the threshold value defined. Furthermore, the operating data for the specific value ranges is added to the training data set (S3). If there are more values for a specific value range than the threshold value defined, these values from these specific value ranges are added to the training data set.
[0046] Different thresholds are used for different value ranges when determining S2. The first state variable is a speed, and the second state variable is a temperature. Different thresholds exist for specific value ranges of speed and temperature. For low speed and low temperature, there is a first threshold, and for high speed and high temperature, there is a second threshold that is different from the first threshold. Accordingly, it can happen that the operating data contains the same number of values for different value ranges. However, in the step of determining S2, only one of the two values is determined to be higher than the threshold and is therefore added to the training dataset when adding S3.
[0047] Furthermore, a user specification for the threshold value is read in S4. This is done via a user interface, e.g., a GUI. Furthermore, the threshold value is determined S5.1 depending on the read in user specification. The user specifies that above a speed of 1000 revolutions per minute, a lower threshold value should be used than at a speed below 1000 revolutions per minute. This ensures that idling, which occurs at speeds below 1000 revolutions per minute, is not reflected in the training data set. Furthermore, the threshold value is determined S5.2 depending on the read in operating data. In this case, S5.2 is determined depending on the value ranges of the first state variable. For example, at lower speeds, idling is specified, and thus the threshold value is defined higher than at other speeds.
[0048] Steps of such a method for providing the training data set are performed to provide a training data set and to train a neural network such that the neural network is configured to determine the second state variable as a function of the first state variable of a machine 6. The machine 6 is shown schematically in Figure 4.
[0049] Figure 2 schematically shows steps of a method for determining a rule-based model. The rule-based model represents a physical relationship between a function input value and a function output value via a function. The rule-based model is configured to determine a state variable of the machine 6 schematically shown in Fig. 4. The rule-based model is configured to determine a second state variable of the machine 6 depending on a first state variable. Operating data of the machine 6 is read in S6. In an embodiment not shown, operating data of one or more other machines not shown are read in. The operating data exhibit a relationship between two state variables of the machine 6. The operating data exhibit a relationship between a first state variable and a second state variable of the machine 6.The operating data exhibits various value ranges of the state variables. For example, a speed as the first state variable exhibits various speed ranges, and a temperature as the second state variable exhibits various temperature ranges.
[0050] Based on a threshold, a determination is made (S7.1) for which value ranges less operating data is available than the threshold specifies. In doing so, it is determined whether more or fewer values are available for a specific value range than defined by the threshold. Furthermore, a determination (S8.1) of at least one rule-based model is made for the specific value ranges below the threshold. Thus, a rule-based model is determined for all value ranges in which fewer values are available than defined by the threshold. In doing so, the determination (S8.1) of the rule-based model is made based on user input.
[0051] Furthermore, S7.2 determines, based on a threshold, the value ranges for which sufficient operating data is available. In this step of S7.2, all value ranges for which more values than defined by the threshold are available are determined. Furthermore, S8.2 determines a rule-based model for the determined value ranges above the threshold. A rule-based model is determined for all value ranges for which more values than defined by the threshold are available.
[0052] Figure 3 schematically shows steps of a method for determining a state variable of a machine 6. The method is formed by a combination of a rule-based model and a neural network. The method thus uses a hybrid model consisting of a rule-based model and a neural network. A value of a first state variable is read in S9. In the embodiment shown, the first state variable is a rotational speed of an engine of the machine 6, designed as a motor vehicle, from Figure 4. Furthermore, a determination S10, based on the read-in value of the first state variable, takes place as to whether the rule-based model or the neural network should be used to determine a second state variable of the machine 6. The second state variable is a temperature of the machine 6. Furthermore, a determination S13 of a critical state of the machine 6 takes place depending on the read-in value of the first state variable.For example, if the speed is low, no critical condition is determined. If the speed is high, a critical condition of machine 6 is determined. Both high and low speeds are determined using predefined thresholds. If the speed of machine 6 is in a critical condition, the rule-based model is used.
[0053] If the rule-based model is to be used, a value of the second state variable is determined S11 using the rule-based model depending on the read-in value of the first state variable. If the neural network is to be used, a value of the second state variable is determined S12 using the neural network depending on the read-in value of the first state variable.
[0054] The neural network is used if the read-in value of the first state variable lies within a first specific value range of the first state variable. The first specific value range is identical to the operating data of the specific value ranges added to the training data set according to the method for providing a training data set shown in Figure 1. The rule-based model is used if the read-in value of the first state variable lies within a second specific value range of the first state variable. The second specific value range is the inverse of the first specific value range with respect to the operating data. Thus, the second specific value range includes all the value ranges of the operating data that were not added to the training data set as part of the operating data in the method for providing the training data set, i.e., which have values less than the defined threshold.In an alternative embodiment (not shown), for at least one specific read-in value of the first state variable, both the determination S11 of the value of the second state variable using the rule-based model and the determination S12 of the value of the second state variable using the neural network are performed. Thus, the plausibility of the value of the second state variable determined using the neural network is verified by determining S11 of the second state variable using the rule-based model.
[0055] The rule-based model comprises different physical sub-models for different value ranges of the first state variable. A first physical sub-model is included in a first value range of the first state variable, and a second, different physical sub-model is included in a second, different value range of the first state variable. The two physical sub-models describe physical relationships differently, for example using different functions or function parameters. Furthermore, a sub-model is selected S11.1 depending on the read-in value of the first state variable. If, for example, a value from the first value range of the first state variable is read in, the first physical sub-model is selected; otherwise, the second physical sub-model is selected. Furthermore, a determination S11.2 of the value of the second state variable using the selected sub-model takes place.Thus, the second state variable is determined solely via the selected submodel. This ensures that the submodels, as part of the rule-based model, are adapted to different value ranges and represent the physical relationship between the first and second state variables as accurately as possible.
[0056] The method, shown schematically in Figure 3, uses a neural network that has been trained with a training data set provided by the method shown schematically in Figure 1. Furthermore, the method shown schematically in Figure 3 uses a rule-based model that has been determined using a method shown schematically in Figure 2.
[0057] Figure 4 schematically shows computing units 2, 4 and the machine 6. The machine 6 has a first computing unit 2. The machine 6 is communicatively connected to a second computing unit 4. The machine 6, designed as a power machine, has the computing unit 2, as a control unit of a transmission of the
[0058] Machine 6. The computing unit 4 is designed as a server. The computing unit 4 is configured to carry out steps of the method shown schematically in Figure 1 for providing a training data set. Furthermore, the computing unit 4 is configured to train a neural network using this specific training data set. Furthermore, the computing unit 4 is configured to carry out the method, shown schematically in Figure 2, for determining a rule-based model. The computing unit 4 is further configured to transmit both the trained neural network and the determined rule-based model to the machine 6 and the computing unit 2. The computing unit 2 is configured to store both the neural network and the rule-based model in a memory (not shown). Furthermore, the computing unit 2 is configured to carry out steps of the method shown schematically in Figure 3 for determining a state variable of the machine 6.Reference symbol ,4 Computing unit machine 1 (step) Reading in operating data from the machine 2 (step) Determining, based on a threshold, for which value ranges sufficient operating data is available 3 (step) Adding the operating data of the determined value ranges to the training data set 4 (step) Reading in a user specification 5.1 (step) Determining the threshold as a function of the user specification 5.2 (step) Determining the threshold as a function of the read-in operating data 6 (step) Reading in operating data from the machine 7.1 (step) Determining for which value ranges less operating data is available than the threshold 7.2 (step) Determining, based on the threshold, for which value ranges sufficient operating data is available 8.1 (step) Determining a rule-based model for value ranges below the threshold 8.2 (step) Determining a rule-based model for value ranges above the threshold 9 (step) Reading in a value of the first state variable 10 (step) Determining whether the rule-based model or the neural network should be used 11 (step) Determining a value of the second state variable using the rule-based model 11.1 (step) Selecting a sub-model 11.2 (step) Determining the value of the second state variable using the selected sub-model 12 (step) Determining a value of the second state variable using the neural network 13 (step) Determining a critical state of the machine.
Claims
Patent claims 1. A method for providing a training data set for training a neural network which is configured to determine a state variable of a machine (6), the method comprising the steps of: reading in (S1) operating data of the machine (6), the operating data having a relationship between at least two state variables of the machine (6) and the operating data having different value ranges of the state variables; Determining (S2), based on a threshold, for which value ranges sufficient operating data is available; and adding (S3) the operating data of the determined value ranges to the training data set.
2. Method according to claim 1, characterized in that when determining (S2) for which value ranges sufficient operating data are available, different threshold values are used for different value ranges.
3. Method according to one of the preceding claims, characterized in that furthermore a reading (S4) of a user specification for the threshold value and a determination (S5.1) of the threshold value in dependence on the read-in user specification takes place.
4. Method according to one of the preceding claims, characterized in that the threshold value is further determined (S5.2) as a function of the read-in operating data.
5. A method for training a neural network which is configured to determine a state variable of a machine (6), with a training data set which was provided according to a method of the preceding claims.
6. A method for determining a rule-based model which is configured to determine a state variable of a machine (6), the method comprising the steps of: reading in (S6) operating data of the machine (6), the operating data containing a relationship between at least two state variables of the machine (6) and wherein the operating data have different value ranges of the state variables; determining (S7.1), based on a threshold value, for which value ranges less operating data is available than the threshold value specifies; and determining (S8.1) at least one rule-based model for the determined value ranges below the threshold value.
7. The method according to claim 6, characterized in that a determination (S7.2) is further carried out, based on a threshold value, as to which value ranges sufficient operating data are available, and a determination (S8.2) of a rule-based model is carried out for the determined value ranges above the threshold value.
8. A method for determining a state variable of a machine (6) by a combination of a rule-based model and a neural network, the method comprising the following steps: reading in (S9) a value of a first state variable; determining (S10), based on the read-in value of the first state variable, whether the rule-based model or the neural network should be used to determine a second state variable of the machine (6); if the rule-based model is to be used, determining (S11) a value of the second state variable with the rule-based model depending on the read-in value of the first state variable; if the neural network is to be used, determining (S12) a value of the second state variable with the neural network depending on the read-in value of the first state variable.
9. The method according to claim 8, characterized in that the rule-based model is used if a critical state of the machine (6) exists.
10. The method according to claim 9, characterized in that a determination (S13) of the critical state of the machine (6) takes place as a function of the read-in value of the first state variable. 11 . Method according to one of claims 8 to 10, characterized in that the neural network was trained with a method according to claim 5 and the rule-based model was determined using a method according to one of claims 6 or 7.
12. Method according to one of claims 8 to 11, characterized in that the neural network is used if the read-in value of the first state variable lies in a first specific value range of the first state variable.
13. Method according to one of claims 8 to 12, characterized in that the rule-based model is used if the read-in value of the first state variable lies in a second specific value range of the first state variable.
14. Method according to one of claims 8 to 13, characterized in that for a specific read-in value of the first state variable, both the determination (S11) of the value of the second state variable is carried out with the rule-based model and the determination (S12) of the value of the second state variable is carried out with the neural network.
15. Method according to one of claims 8 to 14, characterized in that the rule-based model comprises different physical sub-models for different value ranges of the first state variable, and further comprising a Selecting (S11 .1 ) a submodel depending on the read-in value of the first state variable, and determining (S11 .2) the value of the second state variable with the selected submodel.
16. Computing unit (2; 4) which is configured to carry out steps of a method according to one of the preceding claims.
17. Machine (6) with a computing unit (2) according to the preceding claim.