Method and device for providing a data-based model, in particular for realizing a virtual sensor

Abstract

The invention relates to a method, in particular a computer-implemented method, for providing a data-based sensor model for use as a virtual sensor in a technical system, comprising the following steps: - Providing training data sets resulting from a test bench measurement, wherein each training data set assigns one or more, in particular sensor-detected, state variables and / or one or more predetermined operating variables to several labels at a specific time, wherein the labels comprise one or more measured sensor variables to be modeled and one or more additional variables; - Training the data-based sensor model using the training data sets; - Implementing the sensor model in a control unit of the technical system so that during a model evaluation only the one or more sensor variables to be modeled are used as virtual sensors.

Description

Technical FieldThe invention relates to methods for providing data-based sensor models, in particular for realizing a virtual sensor. The invention further relates to a suitable training method for a data-based sensor model.Background ArtData-based models can be designed in the form of neural networks. These assign input quantities to one or more output quantities. Such neural networks may be used, for example, in applications for virtual sensors, anomaly detections, and the like.For the application as a virtual sensor in a technical system, a plurality of input variables, which indicate physical operating states of the technical system and which are correlated with the virtual sensor variable, can be used to ascertain the virtual sensor variable as an output variable. A virtual sensor indicates a sensor size representing a physical size, in particular at a specific detection position in the technical system. A virtual sensor is used when the provision of a real sensor is too complicated or is not possible due to the complexity or size of the technical system at the detection position.For this purpose, the data-based sensor model is trained with training data in a supervised manner. Training data generally comprise input variables in the form of operating and / or state variables of the technical system and a label of a physical variable which is recorded, for example, on a test stand and is intended to be mapped by the sensor model. In particular when used as a virtual sensor, the requirements for robustness and variability of the determined virtual sensor size are important quality criteria for use in technical systems.It is therefore an object of the present invention to provide an improved method of providing a data-based sensor model for use as a virtual sensor that provides greater robustness and less variability in model output.Disclosure of the InventionThis object is achieved by the method for providing a data-based sensor model, in particular for use as a virtual sensor, according to claim 1 and a corresponding device according to the subordinate claim.Further embodiments are given in the dependent claims.According to a first aspect, a method, in particular a computer-implemented method, for providing a data-based sensor model for use as a virtual sensor in a technical system is provided, having the following steps:providing training data sets which result from a test stand survey, wherein each training data set assigns one or more, in particular sensor-detected, state variables and / or one or more predefined operating variables to a plurality of labels at a specific point in time, wherein the labels comprise one or more measured sensor variables to be modeled and one or more additional variables;training the data-based sensor model using the training data sets;implementing the sensor model in a control unit of the technical system, so that during a model evaluation only the one or more sensor variables to be modeled are used as virtual sensors.Furthermore, the sensor model can be designed as an artificial deep neural network, wherein the additional variables relate to variables which are detected only during the test bench measurement and are not provided when using the sensor model in the technical system.Conventional deep neural networks assign one or more outputs to a series of inputs. For this purpose, the neural network is trained in a supervised manner with training data sets, which each assign an input variable vector to one or more output variables in a manner predetermined by the training data sets. Training is generally carried out with the aid of gradient-based learning methods, such as backpropagation, for example.The above method now proposes taking into account further features of the technical system in which the virtual sensor is to be used in the training datasets and during the training of the neural network in order to provide a data-based sensor model. For example, training data sets can be used for creating a data-based sensor model for a sensor variable, which training data sets comprise as input variables one or more state variables detected by sensors and / or one or more operating variables, e.g. variables with which the electric machine is actuated, and furthermore comprise as a label a sensor variable which is to be provided by the virtual sensor.The input variables correspond to variables which are also available during regular operation of the technical system. The labels of the one or more additional variables in the training datasets each correspond to a variable which is determined depending on one or more values acquired only on the test stand.Furthermore, the operating variables can comprise electrical control variables which are specified by a controller or regulator, and / or the state variables can comprise one or more variables, which are detected by sensors, of a state of the technical system, in particular of a rotational speed, a temperature, a torque, a force, a pressure, a movement and / or an acceleration.In the case of determining a component temperature of a component in an electric machine, the input variables of the training data sets can correspond to variables which comprise sensor-detected temperatures at other components of the electric machine, an ambient temperature, an electric variable (motor current) for actuating the electric machine, a rotational speed and / or the like. The label of the training data sets corresponds to a known component temperature, which is determined in each case by a value acquired only on the test stand.For training the sensor model, one or more additional variables are now provided, which can correlate with the sensor variable to be modeled, e.g. the component temperature, or influence it. The one or more additional variables can comprise, for example, an atmospheric humidity of the environment, strength of electromagnetic interference signals and the like. The additional variables are further variables which are detected by sensors and can be detected on a test stand, but do not have to be available in the technical system in the later application of the sensor model, since these do not represent input variables for the evaluation of the sensor model.Training data sets are thus made available for training the sensor model, these training data sets being assigned to the sensor size to be modeled and to one or more additional variables which are not detected or provided during the use of the sensor model during the later operation of the technical system, said input variables (state variables and / or operating variables) which can be detected and / or provided during the ongoing operation of the technical system.The sensor model may be augmented by one or more baseline layers for training for each of the one or more ancillary quantities, the one or more augmented baseline layers being removed from the sensor model after training.The training of the sensor model embodied as a neural network can be carried out in such a way that, given a predefined architecture of the artificial neural network, at least one output neuron is added for each of the additional variables, such that the output layer is extended by the corresponding number of additional variables and the corresponding number of neurons. In particular, one or more further hidden layers can be inserted for one or more of the additional variables.The training can now be carried out in a manner known per se, wherein the sensor size and the one or more additional variables in the training data sets influence the model parameters of the neural network as labels.Once the neural network formed in this way has been trained, the additional output neurons, including the model parameters assigned to them, can optionally be removed from the sensor model and the remaining neural network can be implemented as a sensor model into the technical system in order to provide the sensor size to a virtual sensor. This enables a significantly improved robustness of the sensor model formed in this way.Furthermore, the training of the sensor model can take place on the basis of an overall loss which results as an in particular weighted sum of partial loss values of the plurality of labels.The total loss for the gradient-based training of the sensor model may be determined based on the mean squared error (MSE), the log likelihood, the cross entropy, or the like as a partial loss for each of the labels, and the total loss determined as a weighted sum of the partial loss values.The weighting are hyperparameters that can be optimized based on approaches for hyperparameter optimization in order to ensure the convergence of the training of the sensor model in the case of a summation for ascertaining the total loss value.It can be provided that the one or more sensor variables comprise a differential sensor variable, wherein the sensor variable can be mapped with two output values of the sensor model, which represent the magnitude and the sign of the sensor variable.The sensor size may be an absolute indication or a differential indication. If the sensor variable is a differential variable which always indicates the change in the modeled sensor value per evaluation cycle, the sign of the sensor variable as a model output can be determined separately by a further model output as a classification output. Thus, model outputs for the sign and magnitude of the sensor variable can be provided in the output variables.Brief Description of the DrawingsEmbodiments are explained in more detail below with reference to the attached drawings. The following are shown: FIG. 1 shows a schematic illustration of a technical system which is operated via a control unit having a virtual sensor model implemented therein; FIG. 2 shows a schematic illustration of a neural network of the sensor model for use in training and use in the technical system; FIG. 3 shows a flow chart for illustrating a method for providing a sensor model for realizing a virtual one in the technical system of FIG. 1 ; andDESCRIPTION OF EMBODIMENTSFIG. 1 shows a schematic illustration of a technical system 1, such as an electric machine 2, for example, having a component, such as a stator 21 and a rotor 22, for example. The driver circuit 3 is controlled by a control unit 4.The electric machine 2 is provided with sensor elements 5, which can provide state variables of the electric machine 2 in the control unit 4.In the case of the electric machine 2, the motor current (phase currents), the motor rotational speed, temperatures of temperature sensors arranged in the electric machine, an ambient temperature or a coolant temperature in the case of a coolant circuit and the like can be detected as state variables, for example. As operation quantities, a supply voltage of the driver circuit 3, a torque specification M for the driver circuit, and the like can be provided. The state variables Z and the operating variables B can be provided as input variables for a data-based sensor model 41 implemented in the control unit 4. The sensor model 41 generates a temperature of a machine component, such as that of the rotor 22, for example, which cannot be measured without considerable effort in the technical system, as a sensor variable S.In the present exemplary embodiment, the control device 4 provides a sensor model 41 which represents a virtual sensor for determining the temperature of the determined machine component in the electric machine 2. The temperature can be processed in a temperature monitoring system. The temperature monitoring serves primarily to avoid overheating of the relevant machine component.The data-based sensor model 41 has a structure as described in more detail with reference to FIG. 2. The sensor model comprises a neural network 20 having an input layer 21 of a plurality of neurons 22, one or more hidden layers 23 (only one shown here) having a plurality of neurons 22 and an output layer 24 having a plurality of neurons 22, The output layer 24 firstly has one or more neurons for the sensor variable to be modeled, which neurons represent one or more virtual sensor variables to be modeled.Additionally, for training, the output layer 24 may include one or more neurons for one or more auxiliary quantities A that are used only as for training. The neural network 20 can be designed in a manner known per se as a deep neural network, as a recurrent neural network or the like. Alternatively, multiple layers of neurons may be added for the additional quantities A. These can preferably be connected to the neurons of the last hidden layer as a fully connected layer.The training of the sensor model 41 takes place according to a test bench survey, in which training data sets with operating variables, state variables, the sensor variable and the additional variables are provided. These state variables can be measured with the aid of sensors during the later operation of the technical system 1 and are then processed together with the one or more operating variables in the sensor model 41 implemented in the control unit 4 in order to obtain the corresponding sensor variable.The method for providing the sensor model 41 is explained in more detail with reference to the flow chart of FIG. 3.For training, training data sets are provided in step S 1, which result from the test bench survey. The test stand measurement in each case acquires the operating variables and state variables at a specific point in time and assigns the measured sensor variable to these as labels, which result from an additional sensor available only on the test stand.Furthermore, additional variables can be detected on the test stand as further labels, which additional variables are determined on the basis of additional sensors or the like at the determined point in time. The additional variable can also be a sign specification with respect to the sign of the sensor variable to be modeled.The training of step S 2 is carried out in a gradient-based forward and backward process. In the forward process, each of the neurons 22 of the output layer 24 yields a partial loss value which can be determined, for example, by MSE, log likelihood or cross entropy. The total loss for training the sensor model 41 is then obtained, for example, as a weighted sum (with predefined weights) of the individual partial loss values, so that the neural network can be trained with the aid of backpropagation or other gradient-based method.Once the sensor model 41 has been trained, the neurons 22 of the output layer 24 provided for the additional variables A can be removed or left unnoticed in step S 3 and the remaining sensor model 41 can be implemented in the control device 4 of the technical system in step S 4. If a plurality of layers of neurons 22 are added for the additional variables, these can be removed completely or partially analogously before implementing the sensor model 41.According to a further embodiment, the sensor model 41 can be provided such that the one or more sensor variables are provided differentially, such that, in the cyclical application of the sensor model 41, a difference from the previously determined value of the output variable of the sensor variable is determined. Thus, for example, positive and negative values of the output variable can occur when determining the sensor value, such as the temperature of the machine component.For improved robustness, a plurality of output values for representing the sensor variable S can be provided, one of which indicates the magnitude of the differential value of the sensor variable and the other output value correspondingly indicates the sign of the differential value of the sensor variable.The sign classification of the sensor variable offers the possibility of making an uncertainty estimate. That is, if the value of the model output for the sign is 0.5 (at 0:safe negative and 1:safe positive), the sensor model is maximally uncertain whether the sensor value to be modeled is positive or negative. Values between 0.5 and 0 or between 0.5 and 1 can be assigned a modeling uncertainty accordingly.The modeling uncertainty can be used in a superimposed control or regulating concept, for example, in the case of very uncertain predictions, a previous value is maintained. In this application, the associated neurons are of course retained.

Claims

Method, in particular a computer-implemented method, for providing a data-based sensor model (41) for use as a virtual sensor in a technical system (1), having the following steps: - providing (S1) training data sets which result from a test bench survey, wherein each training data set assigns one or more, in particular sensorically detected, state variables (Z) and / or one or more predefined operating variables (B) to a plurality of labels at a specific point in time, wherein the labels comprise one or more sensor variables (S) to be modeled, in particular measured, and one or more additional variables (A); - training (S2) the data-based sensor model (41) with the aid of the training data sets; implementing (S 4) the sensor model (41) in a control unit (4) of the technical system (1), so that during a model evaluation only the one or more sensor variables (S) to be modeled are used as a virtual sensor.Method according to claim 1, wherein the sensor model (41) is designed as an artificial deep neural network, wherein the sensor model (41) is extended by one or more output layers for the training for each of the one or more additional variables (A), wherein the one or more extended output layers are removed from the sensor model (41) after the training.Method according to Claim 1 or 2, wherein the one or more additional variables (A) relate to variables which are detected only during the test bench measurement or are provided as a function of detected values and are not provided when using the sensor model (41) in the technical system (1), wherein the one or more additional variables (A) furthermore comprise feature variables, in particular a sign, which are derived from the sensor variable (S) to be modeled.Method according to one of Claims 1 to 3, wherein the operating variables (B) comprise control variables which are predefined by a controller or regulator, and / or wherein the state variables (Z) comprise one or more variables, which are detected by sensors, of a state of the technical system (1), in particular of a rotational speed, a temperature, a torque, a force, a pressure, a movement and / or an acceleration.The method according to any one of claims 1 to 4, wherein one or more sensor quantities (S) comprises a differential sensor quantity.Method according to one of Claims 1 to 6, wherein the sensor model (41) is designed as an artificial deep neural network, wherein the training of the sensor model (41) takes place on the basis of a total loss which results as an in particular weighted sum of partial loss values for the plurality of labels.Method according to claim 6, wherein the weighting of the partial lot values is optimized with the aid of an optimization method in order to ensure convergence of the training of the sensor model (41) in the case of summing for ascertaining the total lot.An apparatus for carrying out any of the methods according to any of claims 1 to 7.A computer program product comprising instructions which, when the program is executed by at least one data processing device, cause the program to carry out the steps of the method according to any one of claims 1 to 7.A machine readable storage medium comprising instructions which, when executed by at least one data processing device, cause the at least one data processing device to carry out the steps of the method according to any one of claims 1 to 7.

Images