Device and computer-implemented method for monitoring an electric machine

Abstract

The invention relates to a device and computer-implemented method for monitoring an electric machine (102), wherein a plurality of distributions of values of different state variables of the electric machine (102), in particular state variables that each characterise a temperature of preferably a stator of the electric machine (102), a rotor of the electric machine (102), or a coolant for cooling the electric machine (102), are provided, wherein the distributions, in particular with a model for mapping distributions over values of different state variables to a multidimensional probability density function, in particular a multidimensional normal distribution, preferably with a normalising flow, are mapped to a multidimensional probability density function, in particular a multidimensional normal distribution. The invention also relates to a method and device (100, 112) for training the model.

Description

[0001] Description

[0002] title

[0003] Device and computer-implemented method for monitoring an electrical machine

[0004] State of the art

[0005] The invention is based on a device and a computer-implemented method for monitoring an electrical machine.

[0006] An electrical machine can be monitored by considering distributions of values ​​of state variables of the electrical machine.

[0007] Disclosure of the invention

[0008] A computer-implemented method for monitoring an electrical machine provides that a plurality of distributions are provided over values ​​of different state variables of the electrical machine, in particular state variables that each characterize a temperature, preferably of a stator of the electrical machine, a rotor of the electrical machine, or a coolant for cooling the electrical machine, wherein the distributions are mapped, in particular with a model for mapping distributions over values ​​of different state variables to a multidimensional probability density function, in particular a multidimensional normal distribution, preferably with a normalizing flow. The multidimensional probability density function, e.g.The multidimensional normal distribution represents a more manageable representation of the state variables for monitoring the machine compared to the multiple distributions, which do not have to be normally distributed in a real system. It can be provided that the multiple distributions are provided as a function of at least one operating variable of the electrical machine, in particular a time series of values ​​of the at least one operating variable, preferably a temperature of a stator of the electrical machine, a rotor of the electrical machine, or a coolant for cooling the electrical machine, wherein the at least one operating variable is measured and the multiple distributions are determined as a function of the at least one operating variable, in particular the time series.

[0009] In one example, it is provided that a physical model is provided which models a first part of a hybrid ordinary differential equation for determining the state variables as a function of the at least one operating variable, wherein a probabilistic data-driven model is provided which models a second part of the hybrid ordinary differential equation as a function of the at least one operating variable, wherein the time series of the at least one operating variable comprises values, wherein a respective value of the time series is mapped to a first part of a change in the state variables using the physical model, wherein the respective value of the time series is mapped to a second part of the change in the state variables using the probabilistic data-driven model,and wherein, for each value, the state variables are determined depending on the first part determined for the respective value and the second part determined for the respective value, in particular depending on a sum of the first part and the second part, and wherein the distributions over the values ​​of the state variables are determined depending on the values ​​of the state variables determined for the values ​​of the time series. The physical model and the probabilistic data-driven model represent a hybrid ordinary differential equation. This improves the quality of the state variables underlying the multidimensional normal distribution.

[0010] It can be provided that the data-driven model includes parameters, with the parameters being received, in particular, from a server. This allows the data-driven model to be parameterized with the server's parameters, for example, after training on a server.It can be provided that values ​​of a state variable of the state variables are determined as a function of the at least one operating variable, wherein for a distribution of the distributions at least one parameter which characterizes the distribution, in particular a mean value of the distribution, is determined as a function of the state variable, wherein the at least one parameter is mapped, in particular with the normalizing flow, to at least one parameter which characterizes the multi-dimensional probability density function, in particular the multi-dimensional normal distribution, in particular an expected value vector and a covariance matrix of the multi-dimensional probability density function, in particular the multi-dimensional normal distribution. In this way, the parameters which characterize the respective distribution orA multidimensional probability density function or normal distribution characterizes the distribution depending on one or more company variables. This makes the calculation manageable for large amounts of data.

[0011] For example, several operating variables of the electrical machine are measured, whereby the state variable is determined depending on the several operating variables.

[0012] It can be provided that the at least one parameter is determined for multiple distributions using values ​​of different state variables of the electric machine, wherein the parameters determined for the multiple distributions are mapped, in particular using the normalizing flow, to parameters that characterize the multidimensional probability density function, in particular the multidimensional normal distribution. This makes the calculation manageable for large data volumes for the multiple distributions.

[0013] It can be provided that the multidimensional probability density function, in particular the multidimensional normal distribution, is determined for different electrical machines, and depending on the multidimensional probability density functions, in particular multidimensional normal distributions, determined for the different electrical machines, an anomaly is detected or it is detected that no anomaly is present. Thus, the anomaly or the absence of an anomaly in an electrical machine can be identified from a comparison of the multidimensional probability density functions or multidimensional normal distributions of the different electrical machines.

[0014] It can be provided that, depending on the multi-dimensional probability density function, in particular the multi-dimensional normal distribution, an anomaly is detected, or it is detected that no anomaly exists, or that the multi-dimensional probability density function, in particular the multi-dimensional normal distribution, is determined for the same electrical machine in different operating states of the electrical machine or at different operating times of the electrical machine, and wherein, depending on the multi-dimensional probability density functions, in particular multi-dimensional normal distributions, determined for the same electrical machine, an anomaly is detected, or it is detected that no anomaly exists.This makes it possible to identify the anomaly or absence of an anomaly of the electrical machine from a comparison of the multidimensional probability density functions or multidimensional normal distributions of the same electrical machine.

[0015] It can be provided that, depending on the multidimensional probability density functions, in particular multidimensional normal distributions, determined for the different electrical machines and for the same electrical machine, the anomaly is detected, or the absence of an anomaly is detected. Thus, the anomaly or absence of an anomaly in an electrical machine can be identified from a comparison of the multidimensional probability density functions or multidimensional normal distributions of the same and the different machines.

[0016] The method preferably provides that the model for mapping distributions over values ​​of different state variables to the multidimensional probability density function, in particular the multidimensional normal distribution, comprises model parameters, wherein the model parameters are received in particular from a server. This results in an update of the model parameters, which improves the quality of the prediction using the model.

[0017] It is preferably provided that the model for mapping distributions over values ​​of different state variables onto the multi-dimensional probability density function is designed to map a multi-dimensional probability density function onto distributions over values ​​of the different state variables, wherein a multi-dimensional probability density function which represents values ​​of different state variables of an electrical machine, in particular state variables which each characterise a temperature preferably of a stator of the electrical machine, a rotor of the electrical machine, or a coolant for cooling the electrical machine, is mapped onto distributions over values ​​of the different state variables using the model which is designed to map the multi-dimensional probability density function onto distributions over values ​​of the different state variables.This maps the multidimensional distribution to an explanation of the multidimensional distribution through the distributions of the different state variables.

[0018] A method can be provided for training a model for mapping distributions over values ​​of different state variables of an electrical machine, in particular state variables which each characterise a temperature, preferably of a stator of the electrical machine, of a rotor of the electrical machine, or of a coolant for cooling the electrical machine, onto a multi-dimensional probability density function, in particular a multi-dimensional normal distribution, wherein the model comprises model parameters, wherein distributions over values ​​of different state variables of the electrical machine or different electrical machines and a reference assigned to each of these for the multi-dimensional probability density function, in particular a multi-dimensional normal distribution, are provided, wherein a respective multi-dimensional probability density function,in particular a multidimensional normal distribution is determined depending on the respective distribution over the values ​​of the different state variables of the electrical machine or different electrical machines, and wherein the model parameters are determined depending on a difference between the respective multidimensional probability density function determined by the model, in particular multidimensional normal distribution, and the reference assigned to the respective multidimensional probability density function, in particular multidimensional normal distribution.

[0019] The training method may provide that the values ​​of the different state variables are determined depending on a data-driven model, wherein the data-driven model comprises parameters and models a part of a hybrid ordinary differential equation for determining a state of the electric machine depending on operating variables of the electric machine and the parameters, wherein the model parameters and / or the parameters are determined depending on operating variables of the electric machine or different electric machines. This provides updated values ​​of the parameters and / or model parameters.

[0020] A device for monitoring an electrical machine is configured to carry out the method. This means that the device, in particular the device with which the anomaly is detected depending on the multidimensional normal distribution, or it is detected that no anomaly is present, represents an early warning system that performs anomaly detection or anomaly prediction.

[0021] The device comprises, for example, at least one processor and at least one non-volatile memory, wherein the at least one processor is designed to execute instructions, upon execution of which by the at least one processor the device carries out the method, and wherein at least one non-volatile memory stores the instructions.

[0022] A device for training a model is designed to carry out the training method.

[0023] A computer program may be provided, whereby the

[0024] A computer program comprises instructions executable by a computer, the execution of which by the computer causes the respective method to run on the computer

[0025] Further advantageous embodiments can be found in the following description and the drawing. The drawing shows:

[0026] Fig. 1 is a schematic representation of a device for monitoring an electrical machine,

[0027] Fig. 2 is a flowchart with steps of a method for monitoring the electrical machine,

[0028] Fig. 3 multidimensional normal distributions as an exemplary example of a multidimensional probability density function.

[0029] Figure 1 schematically illustrates a device 100 for monitoring an electrical machine 102. The device 100 is, for example, an embedded system, ie, an embedded device, in particular an Internet of Things device, ie, an Internet of Things device, IoT device.

[0030] The device 100 is designed to carry out a method for monitoring the electrical machine described below.

[0031] The device 100 comprises, for example, at least one processor 104 and at least one memory 106. The at least one memory 106 comprises, for example, volatile and non-volatile memory.

[0032] The at least one processor 104 is configured to execute instructions, which, when executed by the at least one processor 104, cause the device 100 to execute the method. The at least one memory 106 is configured to store the instructions.

[0033] In the example, the device 100 comprises an input 108 for detecting at least one operating variable / 1; ...,I n the electrical machine 102.

[0034] The device 100 optionally includes an interface 110 for communication with a server 112. Figure 2 shows a flowchart with steps of a method for monitoring the electrical machine 102.

[0035] In the method, several distributions x over values ​​of state variables of the electrical machine 102 are determined depending on at least one operating variable , ...,I n the electric machine 102.

[0036] The method includes, for example, a step 202.

[0037] In step 202, the at least one operating variable / 1; ...,I n For example, time series of at least one operating variable / 1; ...,I n measured.

[0038] It can be provided that a plausibility of at least one operating size / 1; ...,I n It may be provided that the procedure is continued if the plausibility of at least one operating variable / 1; ...,I n is determined, and otherwise a new measurement is carried out or the procedure is aborted. The plausibility is checked, for example, with threshold values ​​for measured values ​​of at least one operating variable , ...,I n checked. The threshold values ​​are, for example, predefined for the at least one operating variable. Plausibility can be checked using a trained model for plausibility detection, which, for example, classifies the time series of the at least one measured operating variable as plausible or implausible. The model can be trained based on a knowledge graph that identifies plausible values ​​of the at least one operating variable as plausible.

[0039] At least one company size / 1; ...,I n is, for example, a temperature preferably of a stator of the electrical machine 102, a rotor of the electrical machine 102, or a coolant for cooling the electrical machine 102, or a voltage applied to the electrical machine 102 or a part of the electrical machine, or a current flowing through the electrical machine 102 or a part of the electrical machine 102, or a rotational speed of the electrical machine.

[0040] The method comprises a step 204. In step 204, the plurality of distributions x are determined depending on the at least one operating variable certainly.

[0041] For example, the multiple distributions x are determined depending on the time series of at least one operating variable / 1; ...,I n certainly.

[0042] In the example, the multiple distributions x are provided via values ​​of different state variables of the electric machine 102.

[0043] In the example, the state variables each characterize voltage, current or a non-measured temperature, preferably of the stator of the electrical machine 102, the rotor of the electrical machine 102, or the coolant for cooling the electrical machine 102.

[0044] It can be provided that the state variables are described by a hybrid ordinary differential equation, which is determined by the at least one operating variable , ...,I n For example, a first model and / or a second model are used to determine characteristics F which, in the example, depend on the at least one company size / 1; ...,I n are.

[0045] The first model is used, for example, to determine a hotspot temperature T Hof the stator depending on a temperature T measured with a temperature sensor s of the stator. The first model can be an empirical, physical, or data-driven model. The first model is used, for example, to extract features F, such as aggregated data signals or model states, and to provide them.

[0046] The second model is used, for example, to determine a temperature T SR of the stator at a region of the stator facing the rotor depending on a measured temperature T sof the stator and the current through the stator and / or the voltage applied to the stator. The second model can be an empirical, a physical, or a data-driven model. The second model is used, for example, to extract features F, such as model states, and to provide them. The first model can have parameters that depend on pairs of temperatures T measured by the temperature sensor. s and one of the measured temperatures T s associated measured hotspot temperature of the stator, are or will be learned or validated.

[0047] The second model can have parameters that depend on pairs of measured temperatures of the T s and one of the measured temperatures T s associated measured temperature of the stator at the area of ​​the stator facing the rotor, are or will be learned or validated.

[0048] The features F can be states of the first model or the second model or control variables of the first model or the second model.

[0049] For example, the first model comprises a first artificial neural network, wherein the features F are provided to a layer of the first artificial neural network at a time t.

[0050] For example, the second model comprises a second artificial neural network, wherein the features F are provided to a layer of the second artificial neural network at a time t.

[0051] The hybrid ordinary differential equation has, for example, the structure x T = (a, x T , F, t) + g(ß, x T , F, t) with a physical part f and a data-driven part g, where x T is a state vector that includes the state variables.

[0052] It can be provided that a physical model is provided which models a first part of the hybrid ordinary differential equation, in the example the physical part f, depending on the at least one operating variable.

[0053] It may be provided that a probabilistic data-driven model is provided that models a second part of the hybrid ordinary differential equation, in the example the data-driven part g, depending on the farm size.

[0054] The physical part f is modeled, for example, by the physical model with parameters a of the physical part f, where the physical part f has the state x T at a time t depending on the parameters a of the physical part f to an output of the physical part f.

[0055] The output of the physical part f in the example depends on the temperature of the coolant, ie the cooling temperature T K from the input variables / 1; ...,I n For example, a heat flow from a coolant with the cooling temperature T K to the rotor depending on a difference between cooling temperature T K and the last calculated temperature of the rotor. In the example where the state x T the temperature of the rotor, the difference T K — x T calculated. In the example where the state x T characterizes the temperature of the rotor, the difference T K — T R determined, where T R the temperature of the rotor, which depends on the state x T The heat flow in the example is determined by multiplying the difference between the cooling temperature T Kand the last calculated rotor temperature with a specific thermal conductivity of the coolant. The specific thermal conductivity of the coolant is calculated in the example within the physical part f using the parameter a and the characteristics F.

[0056] Characteristics F can be, for example, the flow rate Q K of the coolant and the temperature of the coolant T K according to the calculation rule

[0057] G K = («1 + «2 QK 3 + a A) 1 serve.

[0058] It can be provided that the state x T is described probabilistically by taking uncertainties of the temperature T R of the rotor using a Laplace approximation of the parameter a, where the state x T depending on the Laplace approximation of the parameter a. This means that the state x Tand the uncertainties for state x T Alternatively or additionally, Monte Carlo methods can be used to determine labels for a desired value of the state x T or a value of the desired change x T of state x T are sampled and several parameterizations of the respective model for calculating or predicting the state x T It can be provided that uncertainties or tolerances of a label are incorporated, provided that the labels are described with uncertainties or tolerances. Probabilistic modeling can also be carried out by using a plurality of differently parameterized models with respect to the state x. T evaluated and then the state x T is probabilistically quantified by re-fitting a probability density function.

[0059] For example, a heat flow from the stator to the rotor is determined depending on a difference between the last calculated temperature T s of the stator and the last calculated temperature of the rotor. In the example where the state x T the temperature of the rotor, the difference T s — x T calculated. In the example where the state x T characterizes the temperature of the rotor, the difference T s — T R determined, where T R the temperature of the rotor, which depends on the state x T The heat flow in the example is determined by multiplying the difference between the last calculated temperature T sof the stator and the last calculated temperature of the rotor 104 with a specific thermal conductivity of the stator. The specific thermal conductivity of the stator is determined in the example for the stator using the physical parameters a and the characteristics F by the rule

[0060] G s = (a5n“ 6 ) -1 calculated. Here, n is the speed of the electric machine.

[0061] Furthermore, in the example, in the context of the physical part f, the power loss P L of the rotor 104 is calculated. This is done based on the characteristics F and the physical parameters a. For example, the power loss is calculated using a characteristic map whose parameters are contained in the physical parameters a. In the example, the input variables for the characteristic map are the phase currents I d and I q of the electrical machine and the speed n of the electrical machine.

[0062] The temperature differential of the rotor temperature calculated by the physical part f in the example is therefore f(a,x T ,F, t) = P L (a,n + G s (a,F')(T s - T R ) + G K (a, F)(T K - T R ) where G s the transfer function for the heat flow from the stator to the rotor and G K the transfer function for the heat flow from the coolant to the rotor.

[0063] The described transfer functions G s and G K represent an illustrative example of equations in the physical part f. The concept is also applicable to other equations for these quantities.

[0064] The data-driven part g is modeled, for example, by a data-driven model with parameters ß of the data-driven part g, where the data-driven part g has the state x Tat a time t, depending on the parameters ß of the data-driven part g, to an output of the data-driven part g. This output serves to correct the predictions of the physical part f by data-drivenly mapping effects not represented in the physical part f.

[0065] In the example, the expenditures are estimated to change x T added. Another way of combining the outputs, e.g., a weighted addition, is also possible.

[0066] Predictions of the hybrid model for the course of state x T are determined, for example, using numerical methods for solving ordinary differential equations. These are methods for the temporal discretization of ordinary differential equations, such as the Euler method.

[0067] The hybrid model is designed, the size x Twhich solves the differential equation at time t, depending on the measured temperature T s , which for this particular temperature T SR , the characteristics F, and the input variables ß, to be provided at time t.

[0068] The parameters a,ß are or will be learned or validated depending on tuples, which are measured input variables ß, a measured temperature T R of the rotor, a measured temperature T s of the stator, a measured hotspot temperature of the stator, a temperature of the stator at a region of the stator facing the rotor, which are associated with each other in particular at time t.

[0069] For example, the parameters a are determined using a least squares method. For example, the parameters ß are determined using a gradient descent method. In the example where the state x Tthe temperature of the rotor, for example, the parameters a,ß are determined for which a deviation of the state x T of the measured temperature of the rotor for the tuples is as small as possible.

[0070] In the example where the state x T characterizes the temperature of the rotor, the procedure is accordingly, whereby the parameters a,ß are determined for which a deviation of the value for the respective state x T certain temperature T R of the measured temperature of the rotor for the tuples is as small as possible.

[0071] The time series of at least one operating variable comprises values.

[0072] A respective value of the time series is mapped, for example, to a first part of a change in the state variables using the physical model.

[0073] The respective value of the time series is mapped to a second part of the change in the state variables, for example, using the probabilistic data-driven model.

[0074] For each value, for example, the state variables are determined depending on the first part determined for the respective value and the second part determined for the respective value.

[0075] For example, the state variables are determined depending on a sum of the first part and the second part.

[0076] The distributions x over the values ​​of the state variables will be determined, for example, depending on the values ​​of the state variables determined for the time series. The models can be designed to determine the values ​​of a state variable depending on one operating variable or depending on several operating variables. This means that the values ​​of a state variable can be determined depending on one operating variable or depending on several operating variables.

[0077] In this example, at least one parameter characterizing a distribution over values ​​of a state variable is determined. In this example, the parameter "mean value of the distribution" is determined as a function of the state variable.

[0078] The method includes a step 206.

[0079] In step 206, the distributions x are mapped to a multidimensional normal distribution.

[0080] The distributions x are mapped to the multidimensional normal distribution using a normalizing flow.

[0081] In the example, for each distribution, at least one parameter is mapped, in particular with the normalizing flow, to at least one parameter that characterizes the multidimensional normal distribution.

[0082] In the example, for each distribution, the mean of the distribution is mapped to an expected value vector of the multidimensional normal distribution and the covariance matrix of the multidimensional normal distribution.

[0083] The distributions x are calculated in the example with the normal flow, ie with a chain of k bijective functions z = fg(x') = fg mapped to the multidimensional normal distribution z. 9 denotes the function parameters of the chain of bijective functions f d (x) is denoted.

[0084] The distributions x are mapped, for example, on the device 100 to the multidimensional normal distribution z.

[0085] The distributions x are mapped, for example, to the multidimensional normal distribution z on the server 112. For example, it is provided that the distributions x are transmitted from the device 100 to the server 112 via the interface 110.

[0086] For example, it is provided that the distributions x on the server 112 are mapped to the multidimensional normal distribution z.

[0087] For example, it is provided that the multidimensional normal distribution z, in particular the expected value vector and the covariance matrix of the multidimensional normal distribution z, is transmitted via the interface 110 to the device 100.

[0088] The method optionally includes a step 208.

[0089] In step 208, it may be provided that, depending on the multidimensional normal distribution, an anomaly is detected, or it is detected that no anomaly exists.

[0090] For example, it is provided that the anomaly is detected on the device 100, or that no anomaly is detected.

[0091] For example, it is provided that the distributions x on the server 112 are mapped to the multidimensional normal distribution z and in a check on the server 112 the anomaly is detected or it is detected that no anomaly exists.

[0092] For example, it is provided that a result of the check, ie that the anomaly is detected or that it is detected that no anomaly exists, is transmitted from the server 112 to the device 100.

[0093] It can be provided that the multidimensional normal distribution is determined for different electrical machines 102.

[0094] For example, the distributions of different devices 100 monitoring different electrical machines 102 are transmitted from the respective device 100 to the server 112 via a respective interface 110. For example, the multidimensional normal distribution for the distributions received from the different devices 100 is determined on the server 112.

[0095] It can be provided that the multidimensional normal distribution is determined for the same electrical machine 102 in different operating states of the electrical machine 102 or at different operating times of the electrical machine 102.

[0096] For example, on the device 100, a respective multidimensional normal distribution is determined for the same electrical machine 102 in different operating states of the electrical machine 102 or at different operating times of the electrical machine 102.

[0097] For example, a respective multidimensional normal distribution is determined on the server 112 for the same electrical machine 102 in different operating states of the electrical machine 102 or at different operating times of the electrical machine 102.

[0098] For example, the distributions x determined in the different operating states are transmitted from the device 100 to the server 112 via the interface 110. For example, the respective multidimensional normal distribution for the distributions determined in the different operating states is determined on the server 112.

[0099] In step 208, it can be provided that, depending on the multidimensional normal distributions determined for the different electrical machines 102, an anomaly is detected, or it is detected that no anomaly is present.

[0100] For example, the distributions of different devices 100 monitoring different electrical machines 102 are transmitted from the respective device 100 to the server 112 via a respective interface 110. For example, a respective multidimensional normal distribution is determined for the distributions received from the different devices 100 on the server 112. For example, depending on the respective multidimensional normal distributions determined for the respective devices 100, it is recognized that an anomaly exists or that no anomaly exists.

[0101] This means that in step 208 it can be provided that, depending on the multidimensional normal distributions determined for the same electrical machine 102, the anomaly is detected, or it is detected that no anomaly is present 208.

[0102] This means that in step 208 it can be provided that, depending on the multidimensional normal distributions determined for the different electrical machines 102 and those determined for the same electrical machine 102, the anomaly is detected, or it is detected that no anomaly is present 208.

[0103] In an optional step 210, a multidimensional normal distribution z with an inverse x = f e (z) to the distributions x.

[0104] In the example, a trained model is provided that maps the distributions to the multidimensional normal distribution. For the optional step 210, for example, the trained model is designed to map a respective multidimensional normal distribution to the distributions underlying it.

[0105] In the example, the trained model maps the parameters of the distributions x to the parameters of the multidimensional normal distributions.

[0106] For example, in a procedure for training the model with a maximum likelihood learning method, the model is trained to approximate the bijective mapping of the chain of bijective functions.

[0107] The model includes, for example, an encoder q(z|x) and a decoder p(z|x) for which: p(x|z) = N(f e (z), he 2 / ) where θ is the Dirac delta function, f0(x) is a tractable normal flow, and I is the identity matrix of appropriate dimension. The model includes model parameters, e.g., model parameters that define the encoder and decoder. The model parameters are determined in the process of training the model. For example, the model is an artificial neural network with weights, where the model parameters include the weights. The model can include a residual flow instead of an architecture with an encoder and decoder.

[0108] The training is carried out, for example, by evaluating with the maximization of

[0109] .

[0110] 0* e argmaxE^zi logq(e) + log det logp(x|z) + logp(z) executed.

[0111] Furthermore, it can be planned that the training is carried out as part of an optimization of a negative likelihood. In this example, the likelihood is determined based on the Kullback-Leibler divergence between the actual multidimensional probability distribution and the multidimensional probability distribution output by the model to be trained. For example, a negative log likelihood is minimized in order to use the then trained model to predict a probability distribution that is as similar as possible to the actual multidimensional probability distribution.

[0112] In the example, the training is performed on server 112. Server 112 receives, for example, measured values ​​of at least one operating variable of electrical machine 102 or various electrical machines 102.

[0113] In one example, device 100 includes the trained model. In one example, server 112 includes the trained model. The training method includes, for example, sending the measured values ​​from device 100 to server 112 and receiving the model parameters from server 112 by device 100.

[0114] Figure 3 illustrates multidimensional normal distributions. Multidimensional normal distributions are an example of a multidimensional probability density function. The method is generally applied to a multidimensional probability density function as described for the multidimensional normal distribution.

[0115] The multidimensional normal distributions can be determined on the basis of operating variables of the same electrical machine 101 or of several electrical machines.

[0116] In the example, two-dimensional normal distributions are shown, each with a first state 302 and a second state 304. In Figure 3, the multidimensional normal distributions are each represented by a center point 306 and a variance 308. The variance 308 of a respective normal distribution is represented as a circle around the center point 306 of the respective normal distribution.

[0117] A limit 310 forms a boundary between a first range 312, in which the normal distribution lies when no anomaly is present, and a second range 314, in which the normal distribution lies when an anomaly is present. As shown in Figure 3, a third range 316 may be provided, defined by a tolerance band around the limit 310, in which the normal distribution lies when an anomaly may be present or when no anomaly may be present.

[0118] In the example, limit 310 is specified. Limit 310 is advantageously designed using simulation or a model. This allows limit 310 to be interpreted very well in terms of physical parameters.

[0119] In the example, the presence of the anomaly is detected if a normal distribution is determined for the electrical machine 101 that lies in the second range 314. In the example, the presence of no anomaly is detected if a normal distribution is determined for the electrical machine 101 that lies in the first range 312. In the example, no statement is made about the anomaly if a normal distribution is determined for the electrical machine 101 that lies in the third range 316.

[0120] For example, a distance measure is evaluated that indicates a distance of the expected value vector of the respective multidimensional distribution from a mean value of the expected value vectors of the multidimensional distributions.

[0121] Limit 310, for example, specifies a distance of the expected value vector of a multidimensional distribution to be evaluated from the mean value, at which an anomaly is detected, and below which no anomaly is detected. It can be provided that, in addition to the distance, the 99.99%, 99%, or 95% confidence of the respective multidimensional distribution, or a specific quantile, is considered.

[0122] For example, no statement is made about the anomaly if the 95% confidence intersects with the limit 310, and otherwise a statement is made about the anomaly, ie either the anomaly is detected or not.

[0123] It may be provided that the described models are provided in a trained form before executing the method. It may be provided that the models are retrained during the method using newly measured values.

[0124] For example, it is provided that the parameters ß of the data-driven part g are replaced by parameters ß that are determined during training of the data-driven part g on the server 112. The parameters ß of the data-driven part g are trained, for example, depending on the measured values ​​of the at least one operating variable of the electrical machine 102 or various electrical machines 102 that are used for training the model, ie, for determining the model parameters.

[0125] For example, the parameters ß determined during the training of the data-driven part g on the server 112 are transmitted from the server 112 to the device 102 via the interface 110. For example, the parameters ß are learned or validated depending on tuples that each contain input variables ß measured for different machines 102, measured temperature T R of the rotor, measured temperature T sof the stator, measured hotspot temperature of the stator, and / or temperature of the stator at a region of the stator facing the rotor, which are associated with each other in particular at a specific time t.

[0126] For example, the parameters ß are determined on the server 112 using a gradient descent method. In the example where the state x T the temperature of the rotor, for example, the parameters ß are determined for which a deviation of the state x T of the measured temperature of the rotor for the tuples is as small as possible.

[0127] In the example where the state x T characterizes the temperature of the rotor, the server 112 proceeds accordingly, whereby the parameters ß are determined for which a deviation of the value for the respective state x T certain temperature T R of the measured temperature of the rotor for the tuples is as small as possible.

Claims

Claims 1 . Computer-implemented method for monitoring an electrical machine (102), characterized in that a plurality of distributions over values of different state variables of the electrical machine (102), in particular state variables which each characterize a temperature preferably of a stator of the electrical machine (102), a rotor of the electrical machine (102), or a coolant for cooling the electrical machine (102), are provided (202, 204), wherein the distributions are mapped (206), in particular with a model for mapping distributions over values of different state variables onto a multidimensional probability density function, in particular a multidimensional normal distribution, preferably with a normalizing flow.

2. The method according to claim 1, characterized in that the plurality of distributions are provided as a function of at least one operating variable of the electrical machine (102), in particular a time series of values of the at least one operating variable, preferably a temperature of a stator of the electrical machine (102), a rotor of the electrical machine (102), or a coolant for cooling the electrical machine (102), wherein the at least one operating variable is measured (202) and the plurality of distributions are determined as a function of the at least one operating variable, in particular the time series (204).

3. Method according to claim 2, characterized in that a physical model is provided which models a first part of a hybrid ordinary differential equation for determining the state variables as a function of the at least one operating variable, wherein a probabilistic data-driven model is provided which models a second part of the hybrid ordinary differential equation as a function of the at least one operating variable, wherein the time series of the at least one operating variable comprises values, wherein a respective value of the time series is mapped to a first part of a change in the state variables using the physical model, wherein the respective value of the time series is mapped to a second part of the change in the state variables using the probabilistic data-driven model, and wherein for each value the state variables are determined as a function of the first part determined for the respective value and of the second part determined for the respective value, in particular as a function of a sum of the first part and the second part, and wherein the distributions over the values of the state variables are determined as a function of the values of the state variables determined for the values of the time series (204).

4. The method according to claim 3, characterized in that the data-driven model comprises parameters, wherein the parameters are received in particular from a server (112).

5. Method according to one of claims 2 to 4, characterized in that values of a state variable of the state variables are determined as a function of the at least one operating variable (204), wherein for a distribution of the distributions at least one parameter which characterizes the distribution, in particular a mean value of the distribution, is determined as a function of the state variable (204), wherein the at least one parameter, in particular with the normalizing flow, is mapped to at least one parameter (206) which characterizes the multi-dimensional probability density function, in particular the multi-dimensional normal distribution, in particular an expected value vector and a covariance matrix of the multi-dimensional probability density function, in particular the multi-dimensional normal distribution.

6. The method according to claim 5, characterized in that a plurality of operating variables of the electrical machine (102) are measured (202), wherein the state variable is determined as a function of the plurality of operating variables (204).

7. The method according to claim 5 or 6, characterized in that the at least one parameter is determined (204) for a plurality of distributions via values of different state variables of the electrical machine (102), wherein the parameters determined for the plurality of distributions are mapped (206), in particular with the normalizing flow, to parameters which characterize the multidimensional probability density function, in particular the multidimensional normal distribution.

8. Method according to one of the preceding claims, characterized in that the multi-dimensional probability density function, in particular the multi-dimensional normal distribution, is determined for different electrical machines (102), and wherein, depending on the multi-dimensional probability density functions, in particular multi-dimensional normal distributions, determined for the different electrical machines (102), an anomaly is detected, or it is detected that no anomaly is present (208).

9. Method according to one of the preceding claims, characterized in that depending on the multi-dimensional probability density function, in particular the multi-dimensional normal distribution, an anomaly is detected, or it is detected that no anomaly is present (208), or that the multi-dimensional probability density function, in particular the multi-dimensional normal distribution, is determined for the same electrical machine (102) in different operating states of the electrical machine (102) or at different operating times of the electrical machine (102), and wherein depending on the multi-dimensional probability density functions, in particular multi-dimensional normal distributions, determined for the same electrical machine (102), an anomaly is detected, or it is detected that no anomaly is present (208).

10. Method according to claim 8 or 9, characterized in that depending on the multi-dimensional probability density functions determined for the different electrical machines (102) and those determined for the same electrical machine (102), in particular multi-dimensional Normal distributions, the anomaly is detected, or it is detected that no anomaly is present (208).

11. Method according to one of the preceding claims, characterized in that the model for mapping distributions over values of different state variables to the multidimensional probability density function, in particular the multidimensional normal distribution, comprises model parameters, wherein the model parameters are received in particular from a server (112).

12. Method according to one of the preceding claims, characterized in that the model for mapping distributions over values of different state variables onto the multidimensional probability density function is designed to map a multidimensional probability density function onto distributions over values of the different state variables, wherein a multidimensional probability density function representing values of different state variables of an electrical machine (102), in particular state variables that each characterize a temperature, preferably of a stator of the electrical machine (102), of a rotor of the electrical machine (102), or of a coolant for cooling the electrical machine (102), with the model that is designed to map the multidimensional probability density function onto distributions over values of the different state variables,is mapped to distributions over values of the different state variables (210)., 13. Method for training a model for mapping distributions over values of different state variables of an electrical machine (102), in particular state variables which each characterize a temperature, preferably of a stator of the electrical machine (102), of a rotor of the electrical machine (102), or of a coolant for cooling the electrical machine (102), onto a multi-dimensional probability density function, in particular a multi-dimensional normal distribution, wherein the model comprises model parameters, wherein distributions over values of different state variables of the electrical Machine (102) or different electrical machines (102) and a reference for the multidimensional probability density function, in particular a multidimensional normal distribution, assigned to each of them, wherein the model is used to determine a respective multidimensional probability density function, in particular a multidimensional normal distribution, depending on the respective distribution across the values of the different state variables of the electrical machine (102) or different electrical machines (102), and wherein the model parameters are determined depending on a difference between the respective multidimensional probability density function, in particular multidimensional normal distribution, determined by the model, and the reference assigned to the respective multidimensional probability density function, in particular multidimensional normal distribution.

14. The method according to claim 13, characterized in that the values of the different state variables are determined as a function of a data-driven model, wherein the data-driven model comprises parameters, and models a part of a hybrid ordinary differential equation for determining a state of the electrical machine (102) as a function of operating variables of the electrical machine (102) and the parameters, wherein the model parameters and / or the parameters are determined as a function of operating variables of the electrical machine (102) or of different electrical machines (102).

15. Device (100) for monitoring an electrical machine (102), characterized in that the device (100) is designed to carry out the method according to one of claims 1 to 12.

16. Device (100) according to claim 15, characterized in that the device (100) comprises at least one processor (104) and at least one non-volatile memory (106), wherein the at least one processor (104) is designed to execute instructions, upon execution of which by the at least one processor (104), the device (100) carries out the method according to one of claims 1 to 12, and wherein at least one non-volatile memory (106) stores the instructions.

17. Device (100, 112) for training a model, characterized in that the device (100, 112) is designed to carry out the method according to one of claims 13 or 14.

18. Computer program, characterized in that the A computer program comprising computer-executable instructions, the execution of which by the computer causes the method according to any one of claims 1 to 14 to run on the computer.

Images