Computer-implemented method for generating a neural network to determine the state of a machine

Abstract

A computer-implemented method for generating a neural network is described. The network is configured to determine the state of a machine (2) based on physical measurement data. The method includes a reading step (S1) of physical measurement data. Furthermore, the method includes an evaluation step (S2) of the read physical measurement data using an evaluation function, where the evaluation function is configured to output an evaluation of the machine's (2) state based on the read physical measurement data. Finally, the method includes a training step (S3) of the neural network using the read physical measurement data as input data and the evaluation as output data, in order to learn the relationship between the input and output data.Furthermore, a method for training a neural network is described, which is configured to determine a state variable of a machine (2), wherein the neural network is trained (S7) depending on the state of the machine (2). For this purpose, the state is determined (S7.1) using the generated neural network. A computing device (4; 6) configured to execute such methods is also described, as well as a machine (2) with such a computing device (4).

Description

Technical FieldThe present invention relates to a computer-implemented method for generating a neural network which is configured to determine a state of a machine on the basis of physical measurement data. The present invention further relates to a method for training a neural network which is configured to determine a state variable of a machine. The training of the neural network takes place, inter alia, as a function of a state of the machine, wherein the state is determined using the previously generated neural network. The present invention further relates to a computing device which is configured to carry out one of the methods. The present invention also relates to a machine which has such a computing device designed as a control device.Prior ArtEvaluation functions are necessary for training many neural networks. Evaluation functions are used in supervised learning as well as in reinforcement learning. Executing evaluation functions to obtain an evaluation of a state is generally very computationally intensive, as evaluation functions generally use a rule-based model. Therefore, training using an evaluation function requires much computing power. Training is therefore often carried out in computing devices which can provide a great deal of computing power.SUMMARY OF THE INVENTIONThe present invention relates, in a first aspect, to a computer-implemented method for generating a neural network. Steps of the method can be carried out on a computing device. The neural network may be a fully connected neural network. To generate the neural network, at least one of a topology of the neural network and properties of individual neurons of the neural network can be determined. The topology of the neural network can be established, for example by a user. A transfer function, an activation function, or weights of one or more neurons may be defined, for example, by the user. After the neural network is generated, it can be provided for use.The neural network is configured to determine a state of a machine as a function of physical measurement data. The machine may be a motor vehicle, such as a passenger car or a truck. Alternatively, the machine may be a working machine such as a land or construction machine. Physical measurement data can be measured, for example, by means of one or more sensors of the machine. Physical measurement data can comprise, for example, pressure, torque, rotational speed, temperature or speed. For example, the physical measurement data can comprise a pressure in an element, for example in a transmission of the machine. The physical measurement data can comprise a torque at an input shaft of the transmission of the engine, as well as a rotational speed at the input shaft and, for example, a differential rotational speed between the input shaft and an output shaft of the transmission. The state of the machine may be, for example, an operating state, an error state or a danger state. For example, a fault state can be determined as physical measurement data as a function of a temperature.The method comprises reading in physical measurement data. In this case, physical measurement data relating to at least one variable can be read in, for example temperature or pressure. Alternatively, physical measurement data relating to a plurality of variables, such as temperature and pressure and torque, can be read in. Before the reading in, these physical measurement data, or at least a part of the physical measurement data, can be measured. Alternatively or additionally, the reading in can take place from a memory, whereupon the physical measurement data have been stored beforehand.Furthermore, the read-in physical measurement data is evaluated using an evaluation function. The evaluation function is configured to output an evaluation of the state of the machine as a function of the physical measurement data read in. The state of the machine can describe, for example, whether the machine is currently in a working state, in a rest state or in a maintenance state. Alternatively or additionally, the state of the engine may include whether an element, such as a transmission of the engine, is in a normal state or in a defective state. The defective state can be present, for example, if a particularly high temperature is present at an element of the machine. Thus, in the evaluation step, for example, as a function of a read-in temperature of the transmission, the evaluation can be carried out that the transmission is in a heated and thus in a faulty and defective state. In this case, the evaluation function can be configured to output the evaluation as a function of one or more physical measurement data read in. In this case, individual and simple, alternatively or additionally also a plurality of and complex behavior patterns with respect to the state of the machine can be implemented by the evaluation function. Thus, a comparatively simple behavior pattern can be reflected, for example, by the temperature and the heated state. A more complex behavior pattern can represent, for example, a relationship between a plurality of physical measurement data read in, such as pressure, torque and temperature of the transmission, and the associated evaluation. For example, as an evaluation, a load state of the transmission can be carried out as a function of these read-in measurement data. The evaluation function can represent a relationship of physical quantities and a state of the machine. The evaluation function can be defined, for example by a user. The evaluation function may have already been used previously, for example, for training another neural network, which is not described in further detail here. For example, the evaluation function may have been used in reinforcement learning of the further neural network. The evaluation function can be rule-based, for example.The method further comprises training the neural network with the read-in physical measurement data as input data and the evaluation as output data in order to learn a relationship of the input data and the output data. Training may additionally comprise validating, wherein the trained neural network may be validated with validation data. The evaluation function can output the evaluation, which can be used as a label during the training of the neural network. Monitored learning of the neural network can be applied in this case.With such a method, an evaluation function can be provided by a neural network. A rule-based evaluation can thus be dispensed with. Complex physical relationships can be reflected here by a relatively simple neural network, for example by means of a simple topology of the neural network. A comparable rule-based model for reflecting such an evaluation function is comparatively complicated to generate and computationally intensive during execution. Thus, an evaluation by the neural network can be compressed. The neural network, as it is executed, for example, on a computing device designed as a control device of the machine, can be efficiently stored on a memory of the computing device and executed with a processor or computing device. At the same time, this computing device may not be configured, for example, to store and execute a comparable rule-based model for such an evaluation function. The execution of the evaluation function depicted in the neural network can be carried out more quickly by using the neural network compared to a rule-based model. This can be realized due to the comparatively simple and less computation-intensive mode of operation and topology of the neural network. Thus, different previously defined evaluations, for example via the evaluation function, for different physical states and behavior patterns of the machine can be stored in the neural network. This stored score in the neural network may then be reflected in executing the neural network. This can also be carried out without performance or runtime losses on a comparatively low-computing device, such as a control device of an engine. The generated neural network can thus also be used on a control device with low computing power. This can be used, for example, for training a further neural network, for example when supervised learning or reinforcement learning is used and an assessment of the state of the machine is necessary for this purpose. Thus, a knowledge representation for evaluating physical behavior patterns with the neural network is provided. This can also be carried out at runtime on computing devices with low computing power, such as control devices of motor vehicles.According to a further embodiment, the method is characterized in that evaluation criteria for the evaluation function can be defined and the evaluation criteria can be transferred to the evaluation function. The defining can be performed based on a user input. Alternatively or additionally, the defining may be based on computer implemented logic. Thus, a user can define that a fault condition exists above a certain temperature. Alternatively or additionally, the logic may define that a load condition exists at elements of the transmission of the engine at a particular combination of temperature, differential speed, and torque. The conversion of the evaluation criteria into the evaluation function can be effected by a user and alternatively or additionally via a logic. Thus, a user can define that, for example, whenever an excessively high temperature is present, a fault state is detected, even if a load state of the transmission is simultaneously detected. Alternatively or additionally, the logic can be configured to transfer specific combinations of evaluation criteria into the evaluation function, such that the evaluation function can output a state of the machine. In this case, the logic itself can determine whether, for example, a faulty state of the machine or else a load state of the machine is output. Alternatively, both states can also be output simultaneously. This can be determined, for example, on the basis of information about previous evaluations, not described further here, and a statistical evaluation of these evaluations. The evaluation of the read-in physical measurement data can be carried out as a function of the defined evaluation criteria transferred into the evaluation function.Thus, by defining evaluation criteria, for example a plurality of different evaluation criteria, and converting these evaluation criteria into the evaluation function, this evaluation function can be provided for evaluation. Thus, adapted to a specific machine or specific elements which are installed in a specific machine, the evaluation function can be adapted to these elements. For example, a first machine may have a first transmission. A second machine may include a second and a third transmission. Different evaluation functions for the two machines can thus be defined and used for evaluating the state of the respective machine. The method can thus be used flexibly for determining states of different machines.According to a further embodiment, the method is characterized in that the evaluation function is further delayed over time, and the evaluation is carried out at least with a temporally delayed evaluation function. The time delay can be effected, for example, with a filter, such as a bandpass filter or a lowpass filter. A temporally delayed evaluation can be determined with the temporally delayed evaluation function.With such a method, a time-delayed evaluation can be determined and the neural network can thus be generated and trained with a time-delayed evaluation. It is thus possible to dispense with the fact that a time-delayed evaluation has to be reflected in the neural network itself, for example in the topology of the neural network. Thus, the memory requirement, the generation effort and the requirement for computing power when executing the neural network can be minimized and optimized. A recurrence can thus be reflected via the time delay of the evaluation function in the neural network. Alternatively to the temporal delay of the evaluation function, the recurrence can also be generated by evaluating with a recurrent evaluation function. If both a recurrent evaluation function and a time delay of the evaluation function for evaluation are present, the training and generation of the neural network can take place more stably and thus more efficiently and quickly.According to a further embodiment, the method is characterized in that the time delay can be carried out using different parameters in order to obtain evaluation functions with different time delays. For example, the evaluation function can be delayed using differently parameterized filter functions, such as, for example, using differently parameterized low-pass or bandpass filters. A time delay can also be carried out with a low-pass filter and simultaneously with a band-pass filter. The assessment of the physical measurement data read in can be carried out using assessment functions with different time delays. Thus, the temporal relationship of the physical measurement data and the temporal relationship of the evaluation with evaluation functions with different time delays can be used for training and for generating the neural network.The state can thus be determined at any desired points in time and operating points with the neural network thus generated, since the temporal profile of physical measurement data and the associated evaluation can be reflected in the neural network by means of time-delayed evaluation functions. Thus, a recurrence of the assessment, even for different time delays by assessment functions with different time delays, can be reflected in the neural network. In this case, a recurrent neural network can be dispensed with, since the recurrence can be swapped out from the topology of the neural network into the input data and output data for training. The neural network can thus also be executed on low-computing-power control devices. Alternatively or additionally, the evaluation function can thus likewise be used for feature engineering, in which it may be necessary to be able to determine temporal dependencies of the state of temporally dependent read-in physical measurement data. Different time dependencies of the physical measurement data and the associated evaluations can be covered by evaluation functions with different time delays and reflected in the neural network. The time delay of the evaluation can thus be trained by the neural network without having to simultaneously reproduce the time profile itself in the topology, for example via a recurrent neural network.According to a further embodiment, the method is characterized in that the evaluation functions with different time delays are also evaluated, and the physical measurement data read in is evaluated as a function of the evaluation of the evaluation functions with different time delays. The evaluation of the read-in physical measurement data can be carried out as a function of the evaluated evaluation functions. The evaluation of the evaluation functions delayed in different times can be effected, for example, via a user or via a logic. The evaluation of the evaluation functions with different time delays can be carried out, for example, using a further secondary evaluation function, which is not described in more detail. It can thus be evaluated, for example, that evaluation functions that are more delayed in time are used less strongly in evaluating the read-in physical measurement data using the evaluation function. Thus, evaluations concerning physical measurement data from the near past can be evaluated more strongly than physical measurement data and the evaluation thereof originating from a more past. Thus, the state can be less affected by the physical measurement data read in for a longer time and can be increasingly determined by the current physical measurement data read in. Training can take place as a function of differently evaluated time-delayed evaluation functions.With such a method, complex behavior patterns, for example represented by time-dependent physical measurement data and their associated states of the machine, can also be trained by the neural network. The neural network can be configured to output these time-dependent complex behavior patterns later in the form of specific states.According to a further embodiment, the method is characterized in that a synthetic signal is generated based on at least one of the read-in physical measurement data. The synthetic signal can be generated, for example, via a time delay, for example, via a filter such as a low-pass filter or a band-pass filter, as a function of one or more physical measurement data read in. For example, a synthetic signal may be determined as an engine performance, as a function of torque, speed, and speed difference of elements of the engine transmission. In training, the generated synthetic signal may be used. For example, for training the neural network, the synthetic signal, for example the specific power of the machine, can be used as input data. Alternatively, the synthetic signal may be used as output data.Thus, further complex physical behavior patterns may be trained by the neural network. The neural network can be configured to be able to determine a state of the machine for such complex physical behavior patterns on the basis of the read-in physical measurement data. This can be made possible by also reading synthetic signals into the neural network in addition to the physical measurement data read in in order to train the neural network. Thus, a complex physical behavior pattern partially reflected in the synthetic signal may be presented to the neural network for training. This makes it possible to dispense with a complex topology during the generation of the neural network, since the complexity can be predefined to the neural network via the training data and the synthetic signal. This can reduce the effort involved in determining the state using the then trained and generated neural network, in order thus to be able to execute the method, for example, also on computing units with low computing power losses.In a second aspect, the invention relates to a computer-implemented method for training a neural network. This neural network can be a second neural network to the previously described neural network, a first neural network, for example. This second or further neural network is to be distinguished from the previously described or first neural network. The second neural network may be configured to determine a state variable of a machine. The state variable may be, for example, a temperature, such as, for example, a clutch temperature of a transmission of the engine. Alternatively, the state variable may be one of a pressure, a rotational speed, a rotational speed difference, a current intensity or a speed of the machine or of an element of the machine. Training of the second neural network can be carried out with training data and as a function of a state of the machine. The training data can have information about the state variable of the machine, for example as output data. Furthermore, the training data can comprise information about operating parameters of the machine, for example about physical measurement data of the machine, such as a pressure or a temperature. These operating parameters or physical measurement data can be used, for example, as input data during the training of the neural network. The state can be determined using a further neural network which has been generated using a method according to one embodiment of the first aspect of the present invention. Thus, the state can be determined with the generated first neural network. For example, the state of the machine can be determined on the basis of physical measurement data, which can also be used as input data in the form of the training data for training the second neural network.Thus, with the aid of the method described under the first aspect of the invention, a second neural network can be trained with the method described under the second aspect of the invention. On the basis of the generated first neural network for determining the state of the machine, the training can also be carried out on a computing device with low computing power. For example, the second neural network can thus be trained on a transmission control unit of the machine designed as a motor vehicle, it being possible to determine the state of the machine on the transmission control unit using the first neural network.According to a further embodiment, the method is characterized in that the training of the neural network for determining the state variable can be a monitored training and can comprise an optimization function which can be provided by means of the further neural network. During the monitored training, for example, the state variable can be used as output data and physical measurement data and alternatively or additionally synthetic signals dependent on physical measurement data can be used as input data. The optimization function can be configured to be used within the scope of an error function during the training of the second neural network. In this case, for example, the difference between the output data and an output of the second neural network to be trained can be minimized and thus optimized.The method for training a neural network in the context of the monitored training can thus be used on a computing device with low computing power. The first neural network can provide the optimization function for the monitored training and for the error function within the scope of the monitored training. Thus, even on a low-computing-power transmission control device, the state of the machine with the first neural network, generated by a method according to one embodiment of the first aspect, can be determined and used for training the second neural network.According to a further embodiment, the method is characterized in that the training of the neural network for determining the state variable can be an intensifying training and can comprise an evaluation function which can be provided by means of the further neural network. The evaluation function can be configured to evaluate the output of the second neural network and to return the evaluation to the second neural network during training of the second neural network. In this case, it can be defined beforehand which states are good and which states are poor states during the training of the second neural network, in order to define the evaluation function for the reinforcement training.Thus, a neural network, here the second neural network, can also be trained on a computing device with low computing power, such as a transmission control unit of a machine designed as a motor vehicle, in the course of the reinforcement training. The state necessary for training the second neural network can be determined on the transmission control unit at runtime using the generated first neural network.In a third aspect, the invention relates to a computing device which is configured to execute a method according to one embodiment of the first two aspects. For example, the computing device can be configured to carry out the method according to one embodiment of the first aspect of the invention for generating the neural network, in this case the first neural network. Such a computing device can be, for example, a computing device located externally to the machine, such as, for example, a computer or a server. According to a further embodiment, the computing device can be a control device, such as a transmission control unit, of the machine. The computing device can be configured to execute a method according to one embodiment of the second aspect of the invention for training the neural network, here the second neural network.A fourth aspect of the present invention relates to a machine which has a computing device designed as a control device according to one embodiment of the third aspect. The machine may be a motor vehicle, such as a passenger car or a truck. Alternatively, the machine may be a working machine such as a land mesh or a construction machine. The machine can have, for example, a computing device designed as a transmission control unit. This computing device can be configured to carry out a method for training a neural network, here the second neural network, according to an embodiment of the second aspect of the present invention. In addition, the machine can be communicatively connected to a further computing device. This further computing device can be configured to execute a method for generating a neural network, here the first neural network, according to an embodiment of the first aspect of the present invention. This further computing device can be a server or a computer. Thus, the first neural network can first be generated on the further computing device. The information about the neural network generated can then be transmitted to the computing device comprised by the machine. Using this information, the computing device of the machine can determine the state of the machine by executing the first neural network in order to be able to train the second neural network.Brief Description of the FiguresFIG. 1 schematically shows a machine with a computing device and a further computing device according to embodiments of the invention. FIG. 2 schematically illustrates steps of a method for generating a neural network for determining a state of the machine schematically shown in FIG. 1, according to an embodiment of the present invention. FIG. 3 schematically shows steps of a method for training a further neural network depending on a state which is determined using a neural network generated after the steps of a method shown schematically in FIG. 2.Detailed Description of EmbodimentsFIG. 1 schematically shows a machine 2 which has a computing device 4 designed as a control device 4. Furthermore, a computing device 6 is shown which is external to the machine 2. The computing device 6 is designed as a server. The computing device 4 is designed as a control device of the machine 2. The computing devices 4, 6 are communicatively connected to one another.FIG. 2 schematically shows steps of a computer-implemented method for generating a neural network. A so-called first neural network is generated. The first neural network is configured to determine a state of the machine 2 from FIG. 1 as a function of physical measurement data.The method comprises reading in S 1 physical measurement data. In this case, for example, pressure, temperature and rotational speed, measured using sensors of machine 2 that are not shown in greater detail, are read in. Furthermore, the read-in physical measurement data is evaluated S 2 using an evaluation function. The evaluation function is configured to output an evaluation of the state of the machine 2 as a function of the physical measurement data read in. The evaluation function is defined by a user and represents a physical relationship between physical measurement data and the state of the machine 2.Evaluation criteria for the evaluation function are defined S 4. This definition S 4 is carried out by a user via a user input. In this case, the user defines that a critical state is present, for example, at a specific temperature of an element of the machine 2. Furthermore, the evaluation criteria are transferred S 5 into the evaluation function. The transfer S 5 takes place, for example, as a function of a user, which specifies via a user input that different evaluation criteria influence the evaluation function to different extents. Thus, the evaluation criterion that a faulty state exists at a specific temperature flows more strongly into the evaluation function than that a load state of the engine 2 exists at specific rotational speeds and torques of elements of a transmission of the engine 2.Furthermore, a temporal delay S2.1 of the evaluation function takes place. The time delay takes place via a low-pass filter. The assessment S 2 is carried out at least with a temporally delayed assessment function. For example, the evaluation function with the evaluation criterion that a critical state is present above a certain temperature is delayed in time. Thus, with the generated first neural network, a time-dependent value for the evaluation can also be output with different temporal physical measurement data relating to the temperature.Furthermore, the delaying S2.1 takes place with different parameters. For example, the time delay of the evaluation criterion to the temperature is effected with differently parameterized low-pass filters. Thus, a different temporal behavior of the physical measurement data and the associated evaluation is used in evaluating S 2 the read-in physical measurement data for determining the evaluation of the state of the machine 2. The delay S2.1 takes place with different parameters in order to obtain evaluation functions with different time delays. The evaluation S 2 of the physical measurement data read in is carried out using evaluation functions with different time delays.Furthermore, training S 3 of the neural network is carried out with the physical measurement data read in as input data and the evaluation as output data in order to learn a relationship of the input data and the output data. Training S 3 is a supervised training, wherein an optimization function, not described in more detail, is used for the error function.Furthermore, the evaluation functions delayed in different times are evaluated S2.2. In this case, the evaluation S2.2 takes place as a function of the deceleration S2.1. When evaluating S2.2 the evaluation functions with different time delays, the evaluation functions with different time delays are evaluated differently. The evaluation S2 takes place as a function of the evaluation S2.2 of the evaluation functions with different time delays. Thus, less time-delayed evaluation functions of the temperature are evaluated more strongly than more time-delayed evaluation functions of the temperature. Training S 3 thus takes place more strongly with evaluations that lie in the closer past than with those that lie in the further past.Furthermore, a synthetic signal is generated S 6 on the basis of at least one of the read-in physical measurement data. For example, a power of the machine 2 is determined as a function of pressure, temperature and rotational speed. This power is then used as a synthetic signal in training S 3. Thus, the power is used as part of the input data in training S 3 the first neural network.FIG. 3 schematically shows training S 7 of a second neural network. This second neural network is configured to determine a state variable of machine 2. This second neural network is configured to determine, for example, a clutch temperature of machine 2. The training of the second neural network is carried out with training data and as a function of a state of the machine 2. a determination S 7.1 of the state of the machine 2 is carried out with the first neural network, which has previously been generated with a method which is schematically illustrated in FIG. 2.According to an embodiment not shown in further detail, the training of the second neural network for determining the state variable takes place with a monitored training. In this case, training the second neural network comprises an optimization function which is provided with the first neural network.In a further, alternative embodiment, which is not shown in further detail, the training of the second neural network for determining the state variable is carried out with an intensifying training. The training comprises an evaluation function which is provided with the first neural network.The computing device 6 is configured to carry out steps of the method for generating the first neural network. If the first neural network has been generated, information about the generated first neural network is sent to the computing device 4 of the machine 2. This information is stored on the computing device 4. Furthermore, the computing device 4 is configured to carry out steps of the method for training the second neural network. In this case, the computing device 4 is configured to determine a state of the machine 2 with the generated first neural network.Reference numerals denote reference numerals2 Machine 4, 6 Computing device S 1 Reading in physical measurement data S 2 Evaluating read-in physical measurement data with an evaluation function S 2.1 Delaying the evaluation function S 2.2 in time Evaluating evaluation functions S 3 delayed in different times Training a (first) neural network S 4 Defining evaluation criteria for the evaluation function S 5 Converting the evaluation criteria into the evaluation function S 6 Generating a synthetic signal S 7 Training a (second) neural network S 7.1 Determining a state of the machine

Claims

Computer-implemented method for generating a neural network which is set up to determine a state of a machine (2) on the basis of physical measurement data, having the steps: - reading in (S1) physical measurement data; - evaluating (S2) the read-in physical measurement data using an evaluation function, wherein the evaluation function is set up to output an evaluation of the state of the machine (2) on the basis of the read-in physical measurement data; - training (S3) the neural network using the read-in physical measurement data as input data and the evaluation as output data, in order to learn a relationship between the input data and the output data.Method according to Claim 1, characterized in that evaluation criteria for the evaluation function are also defined (S4) and the evaluation criteria are converted (S5) into the evaluation function.Method according to one of the preceding claims, characterized in that the evaluation function is further delayed over time (S2.1), and the evaluation (S2) is carried out at least using a temporally delayed evaluation function.Method according to Claim 3, characterized in that the time delay (S2.1) is carried out using different parameters in order to obtain evaluation functions with different time delays, and the evaluation (S2) of the physical measurement data read in is carried out using evaluation functions with different time delays.Method according to Claim 4, characterized in that the evaluation functions delayed in different times are also evaluated (S2.2), and the physical measurement data read in is evaluated (S2) as a function of the evaluation (S2.2) of the evaluation functions delayed in different times.Method according to one of the preceding claims, characterized in that a synthetic signal is generated (S6) on the basis of at least one of the physical measurement data read in, and the generated synthetic signal is used during training (S3).Computer-implemented method for training a neural network, which is configured to determine a state variable of a machine (2), wherein training (S7) of the neural network takes place with training data and as a function of a state of the machine (2), wherein determination (S7.1) of the state takes place with a further neural network, which has been generated with a method according to one of the preceding claims.Method according to Claim 7, characterized in that the training (S7) of the neural network for determining the state variable is a monitored training and comprises an optimization function which is provided by means of the further neural network.Method according to Claim 7, characterized in that the training (S7) of the neural network for determining the state variable is reinforcement training and comprises an evaluation function which is provided by means of the further neural network.Computing device (4; 6) which is configured to carry out a method according to one of the preceding claims.Machine (2) which has a computing device (4) designed as a control device (4) according to the preceding claim.

Images