Prediction of converter fault behavior based on temperature estimation using machine learning algorithms

Abstract

A method for predicting the fault behavior of an electrical converter (12) is disclosed, the method comprising: receiving an operating point metric (I) of the electrical converter (12) indicative of the actual operating point of the electrical converter (12); receiving a measured device temperature (T d ) of a power semiconductor device (18) of the electrical converter (12) indicative of the actual temperature of the power semiconductor device (18); inputting the operating point metric (I) as input data into a machine learning algorithm (32) that is trained using historical data comprising the operating point metric and associated device temperatures; estimating an estimated device temperature #imgabs0# using the machine learning algorithm; and predicting a fault behavior (F) by comparing the estimated device temperature #imgabs1# with the measured device temperature (T d ).

Description

Technical Field

[0001] The present invention relates to the field of predictive maintenance of power semiconductor devices. In particular, the present invention relates to a method, a computer program, a computer readable medium and a controller for predicting the fault behavior of an electrical converter. Furthermore, the present invention relates to a converter having such a controller. Background Art

[0002] The case temperature or module temperature of a power semiconductor device such as an IGBT can be measured directly, for example, using a temperature sensor integrated into the semiconductor module housing the semiconductor device. However, little meaningful can be gained from the measured temperature other than that the temperature is too high. Often, because the operating point and conditions of the system using the power semiconductor device may change, it is difficult to see an overall temperature trend, such as an increase in temperature under the same conditions as before.

[0003] Temperature trending, on the other hand, is a simple way to monitor the static conditions of power semiconductor devices. A sudden increase could be seen as a component failure or aging (e.g. due to increased resistance), a cooling failure (which could be faulty or blocked), or an increase in ambient temperature (which could cause accelerated aging). However, this type of simple trending is often not applicable when temperatures vary greatly over a short period of time due to dynamic operating points and cyclic component usage, as it may be difficult to notice meaningful changes that occur slowly over time, such as higher maximum temperatures or slower cooldown times.

[0004] CN 109 101 738A describes a method for estimating the aging degree of an IBGT module. The method includes the step of training an artificial neural network using measured electrothermal characteristic data (such as current, voltage or temperature) of the IGBT module. In a further step, the actually measured electrothermal characteristic data is input into the trained artificial neural network, and an aging degree evaluation result is output, which can be the number of aging cycles or the aging degree.

[0005] EP 2 941 674 A2 describes a method for predicting the transformer oil temperature for an expected load based on a motion profile of the transformer. The motion profile is based on a machine learning algorithm that has been trained using historical data provided by transformer sensors.

[0006] EP 2 600 510 A1 describes a method for controlling an electrical converter having a power module which can be heated to a target temperature by generating a circulating current through the power module.The actual temperature of the power module can be estimated based on a thermal model of the power module and the ambient temperature.

[0007] AU 2013 260 082 A1 describes a method for reducing thermal cycling of a power switch in a power converter based on the junction temperature of the power switch. The junction temperature may be estimated based on the ambient temperature. Summary of the invention

[0008] It is an object of the present invention to improve predictive maintenance of power semiconductor devices.It is another object of the present invention to better predict temperature trends of power semiconductor devices.

[0009] These objects are achieved by the subject-matter of the independent claims. Further exemplary embodiments are evident from the dependent claims and the following description.

[0010] A first aspect of the invention relates to a method for predicting fault behavior of an electric converter. The method can be automatically performed by a controller of the electric converter, for example as a software module stored in the controller.

[0011] According to an embodiment of the invention, the method comprises: receiving an operating point indicator of the electrical converter indicative of an actual operating point of the electrical converter; and receiving a measured device temperature of the power semiconductor device indicative of an actual temperature of the power semiconductor device.

[0012] The operating point of the electrical converter may be defined by the characteristics of the electrical converter at a particular point in time. For example, the operating point may be defined by the current and / or voltage handled by the electrical converter and / or by the output power of the electrical converter. The operating point indicator may be provided as a value or a set of values. The operating point indicator may include one or more values ​​and / or quantities related to the temperature of the power semiconductor device.

[0013] The operating point indicator may be determined from quantities measured in the electrical converter and / or quantities generated by a controller of the electrical converter controlling power semiconductor switches of the electrical converter.

[0014] The power semiconductor device may be a power semiconductor switch, such as an IGBT.

[0015] The measured temperature of the power semiconductor device may be provided as a value. The measured temperature may be measured with a temperature sensor attached to the power semiconductor device.

[0016] The operating point indicator and the measured device temperature may be determined over time, i.e., there may be an operating point indicator and a measured device temperature for a plurality of measurement steps. It may be necessary to sample one or more values ​​of the operating point indicator at a sufficiently high frequency, since temperature changes may occur within a few seconds. For example, the sampling frequency and / or the spacing of consecutive measurement steps may be in the range between 0.1 s and 5 s.

[0017] According to an embodiment of the present invention, the method includes: inputting an operating point indicator as input data into a machine learning algorithm trained using historical data (or training data), the historical data (or training data) including historical operating point indicators and related historical device temperatures; and estimating the estimated device temperature using the machine learning algorithm.

[0018] The estimated device temperature of the power semiconductor device can be determined using a machine learning algorithm that has been trained with historical data. The historical data can be data that has been recorded before the method is performed. The machine learning algorithm can be based on weights that are adjusted during training. The weights can define a function where the input is the input data and the output is the estimated device temperature.

[0019] Applicants have discovered that a quantity related to the operating point of an electrical converter, an operating point indicator, is suitable for training a machine learning algorithm that estimates the device temperature of the power semiconductor devices of the electrical converter. This may be because the operating point of the converter is related to the electrical stress on its power semiconductor devices. The machine learning algorithm can be used to determine the normal and faulty behavior of the converter and detect anomalies. Troubleshooting the root cause helps maintain the drive, extend the drive life, and prevent unexpected drive failures.

[0020] Since the machine learning algorithm has been trained using historical data recorded during normal operation of the power semiconductor device, the estimated device temperature represents the device temperature during normal operation, when the semiconductor device is in a good condition.

[0021] According to an embodiment of the invention, the method further comprises: predicting a fault behavior by comparing the estimated device temperature with the measured device temperature. The fault behavior may be provided as a value, for example, a "yes" / "no" value in one case.

[0022] If the machine learning algorithm is well trained and the drive is used consistently as described by the training data, there should be only a small error between the estimated and measured device temperatures. Comparing these two values ​​over time allows the temperature increase trend of the power semiconductor device to be seen, displayed as the difference between the estimated and measured values. In particular, when the measured device temperature is higher than the estimated temperature, assumptions can be made about the temperature increase trend.

[0023] The temperature increase trend is usually caused by faulty behavior of the power semiconductor device or other components of the electrical converter. For example, the power semiconductor device may be aged and may have an increased resistance. Other potential causes of temperature increase may be a clogged cooling filter or component aging of other components of the converter.

[0024] The fault behavior can be a fault behavior of the power semiconductor device and / or the electrical converter. By predicting the results, unexpected failures and downtime can be reduced, safety can be increased and other components of the converter can also be protected.

[0025] Using a machine learning algorithm for estimating temperature can provide a meaningful and automated way to notice temperature increases. The information gained can be used to understand which power semiconductor devices and / or which converter may need maintenance and / or require more detailed analysis to see if there is a problem in the power semiconductor module.

[0026] According to an embodiment of the present invention, the input data of the machine learning algorithm includes: the measured device temperature for the previous (measured) time step and the operating point indicator for the actual time step and / or the previous time step. In addition to the operating point indicator, the measured device temperature in the previous time step can also be input into the machine learning algorithm. This can increase the prediction accuracy of the estimated temperature. In this case, the historical data used for training must include additional historical measured device temperatures. The actual time step can be the last time step at which the measurement has been performed, and the previous time step can be the second to last time step.

[0027] According to an embodiment of the present invention, the estimated temperature for the actual time step is estimated by a machine learning algorithm. As mentioned above, measurements can be made regularly on the measurement or sampling time step. The input data may include: the value of the operating point indicator and the optional measured temperature of at least one previous (measurement) time step. This may take into account the development of these quantities in a timely manner.

[0028] According to an embodiment of the present invention, the input data includes measured device temperatures for a number of previous time steps. The input data may also include operating point indicators for a number of previous time steps. In this case, the machine learning algorithm directly takes into account the timely behavior of these quantities.

[0029] In general, the input data may include: measured device temperatures for only the previous time step and operating point metrics for only one previous time step. The input data may include: measured device temperatures for only one previous time step and operating point metrics for multiple previous time steps. The input data may include: measured device temperatures for only multiple previous time steps and operating point metrics for only one previous time step. The input data may include: measured device temperatures for only multiple previous time steps and operating point metrics for only multiple previous time steps.

[0030] According to an embodiment of the present invention, the input data includes: one or more differences between measured device temperatures for a plurality of previous time steps. The input to the machine learning algorithm may be one or more temperature differences between two consecutive temperature measurements. When using the absolute values ​​of the temperature measurements as input, the machine learning algorithm may learn to predict that the temperature at the next time step is the same as the temperature at the previous time step. When using differences instead of absolute values, the actual temperature values ​​may be hidden and the input and output parameters may be decoupled.

[0031] According to an embodiment of the invention, the method further comprises: receiving an ambient temperature of the converter indicating an ambient temperature of the power semiconductor device and / or the converter. The electrical converter and / or the controller may comprise an ambient temperature sensor. The ambient temperature may be a temperature inside the converter, for example a temperature inside a converter housing.

[0032] According to an embodiment of the present invention, the input data of the machine learning algorithm also includes the ambient temperature. The ambient temperature can be used as further input data of the machine learning algorithm. This can improve the prediction accuracy of the estimated temperature because the temperature of the power semiconductor device depends on the ambient temperature and the heat generated inside the power semiconductor device.

[0033] According to an embodiment of the present invention, the measured device temperature in the input data is provided relative to the ambient temperature. In other words, the ambient temperature can be subtracted from the measured device temperature before the result is input into the machine learning algorithm.

[0034] According to an embodiment of the present invention, the estimated device temperature output by the machine learning algorithm is provided relative to the ambient temperature. It is also possible that the output of the machine learning algorithm, i.e., the estimated temperature, is a relative value relative to the ambient temperature. The relative value can be added to the ambient temperature to determine the absolute temperature value.

[0035] However, it is also possible that one or more absolute values ​​of the measured temperature are input to the machine learning algorithm, and / or the estimated temperature output by the machine learning algorithm is provided in the form of an absolute value. According to an embodiment of the present invention, the method further comprises: determining a temperature error based on the difference between the measured temperature and the estimated temperature; and comparing the temperature error with a threshold value for predicting fault behavior of the power semiconductor device. The temperature error may be an indicator of fault behavior because when the measured temperature is much higher than the estimated temperature, the operation of the power semiconductor device and / or the converter is different from that shown in the historical data reflecting normal operation.

[0036] The temperature errors may be averaged or the average of the temperature errors may be determined. For example, a moving average error and / or a median error may be tracked to see when the measured temperature values ​​begin to deviate from the estimated temperature values. When the temperature error has reached a threshold and optionally remains above the threshold for a predefined period of time, then fault behavior may be predicted. It may be assumed that something has changed in the converter system and a warning may be given along with instructions on how to proceed.

[0037] According to an embodiment of the invention, a machine learning algorithm has been trained using historical data of the same converter. The historical data may have been recorded during a commissioning phase of the converter, where normal operation can be assumed. This may be beneficial because data can be recorded for the same power semiconductor device and subsequently supervised using the method.

[0038] According to an embodiment of the invention, the machine learning algorithm has been trained with historical data of at least one different converter, which may have the same and / or similar topology. In this case, a generic temperature model can be provided for all similar converters. The first historical data collection from a new drive can be omitted, but the temperature behavior can be monitored directly after installation.

[0039] However, it is also possible that a machine learning algorithm trained with historical data from different converters is additionally trained with historical data from the same converter. This can further improve the prediction results.

[0040] According to an embodiment of the present invention, the machine learning algorithm is an artificial neural network. An artificial neural network may typically have three layers, such as an input layer, a hidden layer, and an output layer. There may also be multiple hidden layers (e.g., for deep learning). For this application, it has been shown that only one hidden layer is required to obtain acceptable performance.

[0041] According to an embodiment of the invention, the operating point indicator comprises at least one of the following: converter current, converter voltage, switching frequency, DC link voltage. All these quantities represent the heat generated by the power semiconductor devices of the converter.

[0042] According to an embodiment of the present invention, an electric converter is connected to a rotating electric machine for driving the rotating electric machine, and an operating point indicator includes at least one of the following: a torque of the rotating electric machine, and a rotational speed of the rotating electric machine. Drive parameters such as torque and rotational speed can be used as input data to train a machine learning algorithm and estimate device temperature.

[0043] If these parameters are positively correlated with the device temperature and are optionally sampled at a high enough frequency (e.g. more than one sample per second), another quantity can be considered along with the operating point indicator. The sampling frequency can range between 0.1s and 5s.

[0044] Another aspect of the invention relates to a computer program which, when executed by a processor, is suitable for performing a method as described above and below, and to a computer-readable medium having such a computer program stored thereon. The computer-readable medium may be a floppy disk, a hard disk, a USB (Universal Serial Bus) storage device, a RAM (Random Access Memory), a ROM (Read Only Memory), an EPROM (Erasable Programmable Read Only Memory) or a flash memory. The computer-readable medium may also be a data communication network (e.g. the Internet) which allows downloading program code. In general, the computer-readable medium may be a non-transitory or a transient medium.

[0045] Another aspect of the invention relates to a controller of an electric converter, the controller being adapted to perform the method as described above and below. The controller may comprise a processor and a memory, and the method may be implemented as a computer program in the controller. The computer program may be stored in the controller as a software module. The controller may also be adapted to control a power semiconductor switch of the electric converter. This task may be performed by another software module stored in the controller.

[0046] Another aspect of the invention relates to an electrical converter comprising a plurality of power semiconductor devices and at least one temperature sensor arranged to measure the device temperature of at least one power semiconductor device. A plurality of power semiconductor devices, which may include semiconductor switches (e.g., IGBTs) and optionally diodes, may be housed in one or more semiconductor modules. The semiconductor module may include a housing and / or wiring for mechanically supporting and / or electrically interconnecting one or more power semiconductor devices. Each module may also include a temperature sensor for measuring the device temperature of one or more semiconductor devices.

[0047] The power semiconductor devices may be connected in one or more half-bridges which are interconnected to form a specific converter topology.An electrical converter may be interconnected between an electrical grid and a rotating electrical machine, such as a motor or a generator.

[0048] The electrical converter may further comprise a controller, as described above and below, adapted to estimate the temperature of at least one of the power semiconductor devices.

[0049] The method may be performed for one, some or all power semiconductor devices. For example, the method may be performed for at least one power semiconductor device of each semiconductor module. When a fault behavior of at least one power semiconductor device in a module is predicted, the semiconductor module may be shut down and / or replaced.

[0050] It shall be understood that features of the method described above and below may also be features of the electrical converter, the computer program, the computer readable medium and the controller described above and below, and vice versa.

[0051] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The subject matter of the invention will be explained in more detail hereinafter with reference to exemplary embodiments shown in the drawings.

[0053] Figure 1 A drive system with an electrical converter according to an embodiment of the invention is schematically shown.

[0054] Figure 2 A block diagram illustrating a method and a controller according to an embodiment of the present invention is shown.

[0055] Figure 3 A block diagram illustrating a method and a controller according to another embodiment of the present invention is shown.

[0056] Figure 4A and Figure 4B Graph showing estimated and measured temperatures of a power semiconductor device.

[0057] Figure 5A , Figure 5B , Fig. 6A , Figure 6B A graph showing temperature errors determined using a method according to an embodiment of the present invention is shown.

[0058] Figure 7 The structure of a neural network trained to perform temperature estimation in a method according to an embodiment of the present invention is schematically shown.

[0059] The reference symbols used in the drawings and their meanings are listed in summary form in the reference symbol list. In principle, identical components are provided with the same reference symbols in the figures. DETAILED DESCRIPTION

[0060] Figure 1 A drive system 10 is shown, including an electrical converter 12 and a rotating electrical machine 14 (eg, a motor or a generator). The electrical converter 12 is connected to a power grid 16 and converts the grid voltage into an output voltage that is provided to the rotating electrical machine 14.

[0061] like Figure 1 As shown schematically, the converter 12 includes a power semiconductor device 18, which may be a semiconductor switch, in particular an IGBT. A freewheeling diode connected in parallel with the semiconductor switch may also be regarded as a power semiconductor device.

[0062] As shown, the power semiconductor devices 18 may be connected in series to form a half bridge 20 , which may be connected to a DC link 22 and / or may provide a phase output therebetween. The power semiconductor devices 18 of one half bridge may be assembled into a power semiconductor module 24 .

[0063] For each power semiconductor device 18, there may be a temperature sensor 26, which may also be integrated into the respective power semiconductor module 24. However, it is also possible that there is only one temperature sensor 26 per power semiconductor module 24. With the temperature sensor 26, the actual device temperature of the relevant power semiconductor device 18 may be measured.

[0064] The drive system 10 further comprises a controller 28, which is adapted to control the converter 12 and, as indicated, the power semiconductor switches. To this end, gate signals for the power semiconductor switches may be generated by the controller 28. Figure 1 As shown, the controller 28 may therefore receive measurement signals of currents and / or voltages from the converter 12 , such as input voltage, DC link voltage, output voltage, output current, etc.

[0065] In addition to controlling the power and / or torque of the drive system 10 , the controller 28 is also adapted to predict fault behavior of the converter 12 and / or its power semiconductor devices 18 based on temperature measurements and estimates.

[0066] To this end, the controller 28 is further adapted to receive a measurement signal from the temperature sensor 26. The converter 12 may further comprise an ambient temperature sensor 30, which may be arranged in the housing of the converter 12. The controller 28 may further be adapted to receive a measurement signal from the ambient temperature sensor 30.

[0067] Figure 2 and Figure 3 A diagram of the components of the controller 28 that may perform prediction of fault behavior is shown. Figure 2 and Figure 3 , a method for predicting the fault behavior of an electrical converter 12 is also described.

[0068] The method and controller 28 are based on a machine learning algorithm that can Figure 2 and 3 The module 32 is shown as being executed.

[0069] like Figure 2As shown, the operating point index I of the electric converter 12 is input as input data into the machine learning algorithm 32, and the operating point index I represents the actual operating point of the electric converter 12. The operating point index I may include at least one of the converter current, the converter voltage, the switching frequency, the DC link voltage, the torque of the rotating electric machine 14, the speed of the rotating electric machine 14, etc.

[0070] Generally speaking, the operating point index I can be any quantity that is positively correlated with the device temperature.

[0071] The operating point indicator I may be provided and / or determined by other control functions of the controller 28 and / or may be based on current and / or voltage measurements in the converter 12 .

[0072] Based on the input data, the machine learning algorithm 32 estimates the device temperature Generally speaking, the machine learning algorithm 32 can be an estimated device temperature that can be calculated based on the input data. The function may include parameters or weights that have been adjusted during the training phase of the machine learning algorithm 32.

[0073] In particular, the machine learning algorithm 32 has utilized the operating point indicator I and the associated device temperature T d The machine learning algorithm 32 is trained on historical data that has been recorded for the same converter 12 and / or from different converters that may be of the same type and / or topology. With the help of historical data, the weight parameters of the machine learning algorithm 32 can be adjusted so that the function outputs similar estimated device temperatures for similar input data.

[0074] For example, historical data may be recorded at the beginning of the useful life of the power semiconductor devices 18 and may be used until the end of the useful life of these devices 18 .

[0075] For example, the machine learning algorithm 32 may be an artificial neural network. It has been shown that a simple artificial neural network with an input layer, only one hidden layer and one output layer is sufficient to estimate the temperature with high accuracy.

[0076] Figure 4A shows the measured device temperature T of an IGBT as part of the converter 12 d , and the estimated device temperature generated by this artificial neural network that has been trained accordingly Temperature is shown in °C on the vertical axis. Figure 4A is generated for a set of sample time steps at the beginning of the IGBT's active life, as shown in the figure on the right.

[0077] Figure 4B is with Figure 4ASimilar diagram, but with the device temperature T measured at the end of the IGBT's lifetime d and estimated device temperature Since the ANN was trained with normal data, it could not correctly predict the high temperature at the end of the shelf life. Figure 4B The higher measured device temperature T d If the difference between the estimated and measured values ​​is clearly visible, this indicates that the temperature behavior has changed compared to the training data accumulated at the beginning of the validity period. This error can be monitored, for example, using the moving error index.

[0078] Back to Figure 2 , the input data of the machine learning algorithm 32 may include: the measured device temperature T for the previous sampling and / or measurement time step t-1 d (t-1) and the operating point indicator I(t-1) for the previous time step t-1. The operating point indicator I(t) for the actual time step t may also be included in the input data in addition or alternatively. Then, the estimated device temperature for the actual time step t may be obtained by the machine learning algorithm 32. Make an estimate.

[0079] By estimating the device temperature from the actual The actual measured device temperature T at the actual time step t is subtracted from d (t), the temperature error E(t) at the actual time step t can be determined, which is input into the averaging and / or comparator module 34.

[0080] The averaging and / or comparator module 34 may average the temperature error E over a time range of a plurality of time steps. The result may be compared with a threshold, and if the average error is above the threshold, the fault behavior signal F may become 1. Otherwise, the fault behavior signal F may be 0.

[0081] It should be noted that, alternatively, the median error E(t) may be calculated by the averaging and / or comparator module 34 based on the last measured device temperature T d and estimated device temperature to be sure.

[0082] exist Figure 3 , it is shown that the input data of the machine learning algorithm 32 may include: operating point indicators I(t-1), I(t-2), ... for multiple previous time steps t-1, t-2, etc. By using more than one previous operating point indicator I, the accuracy of the machine learning algorithm 32 can be improved. It should be noted that in Figure 3 In , the operating point indicator I(t) of the actual time step t can also be included in the input data.

[0083] Figure 3 It also shows that the measured device temperature T at the previous time step d (t-1) may be part of the input data. It is also possible that the input data of the machine learning algorithm 32 includes the measured device temperature T for a plurality of previous time steps t-1, t-2, etc. d (t-1), T d (t-2), ... Again, this can improve forecast accuracy.

[0084] When measuring the ambient temperature T a When the ambient temperature T of the previous time step is a The ambient temperature at (t-1) and optionally for a number of previous time steps may also be included in the input data.

[0085] The input data may also include the difference (T d (t)-T d (t-1), T d (t-1)-T d (t-2), ...). Figure 3 As shown, the estimation performed by the machine learning algorithm 32 can be relative to the ambient temperature T a Execute. When the ambient temperature T a Before inputting into the machine learning algorithm 32, the device temperature T d Subtract the ambient temperature T a Then, the machine learning algorithm is used to calculate the ambient temperature T d Estimate the device temperature, that is, output It must be noted that in order to determine the temperature error E, the relative measured temperature T must also be subtracted d -T a .

[0086] Figure 5A and Figure 5B The moving average value of the temperature error E at the beginning and end of the IGBT life is shown. Note that the error plotted on the vertical axis is Figure 5A The value shown as 2 in Figure 5B The value shown in is 12. The sampling time step is shown on the horizontal axis.

[0087] Fig. 6A and Figure 6B is a similar plot, but with median error

[0088] from Figure 5A and Figure 5B or from Fig. 6A and Figure 6B It can be seen that the threshold value can be used for the temperature error E, To trigger a fault behavior signal F and generate a warning about the change in temperature behavior.

[0089] It can be seen that the temperature error E, At the beginning of the validity period it is less than 2. At the end of the validity period it is always above 4 and may briefly reach values ​​above 10. After a certain time when the threshold value has been exceeded, a fault behavior signal can be triggered and / or a warning can be generated about the temperature increase in the power semiconductor module 24. This measure can be taken even if an overtemperature fault has not been triggered so far. Once a warning is triggered, troubleshooting can begin to find the cause of the fault behavior, thus possibly preventing future sudden damage or failures.

[0090] The following table includes a sample of test results output by a neural network used to implement the machine learning algorithm 32. Testing was done using Azure Machine Learning Studio because it provides easy access to stored drive data as well as suitable machine learning libraries. In this example, a data set of torque values ​​"Torque 15" to "Torque 20" indicating the torque of the motor 14 is used as input data for training a neural network to estimate the IGBT case temperature. The "Score" column contains the corresponding output data of the neural network, i.e., the estimated value given by the neural network. The "TC" column includes the measured temperature value for each estimated value. It can be seen that there is only a small difference between the estimated and measured values. As an alternative or in addition to the torque, the rotational speed or ambient temperature of the motor 14 can be used as input data for the neural network.

[0091]

[0092] Figure 7 The structure of the neural network is schematically shown. The neural network has an input layer 36 with a plurality of torque input neurons T for inputting torque values ​​and a plurality of speed input neurons n for inputting speed values; and an output layer 37 with an output neuron for outputting the corresponding temperature estimate. In this example, the input layer 36 and the output layer 37 are interconnected by a hidden layer 38, and ten previous samples from each signal have been used to estimate the temperature.

[0093] Although the invention has been described and illustrated in detail in the drawings and the foregoing description, such description and illustration should be considered illustrative or exemplary rather than restrictive; the invention is not limited to the disclosed embodiments. Other variations of the disclosed embodiments may be understood and effected by a person skilled in the art in practicing the claimed invention by studying the drawings, the disclosure and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single processor or controller or other unit may perform the functions of several items listed in a claim. The fact that certain measures are listed in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be interpreted as limiting the scope.

[0094] Reference numerals list

[0095] 10. Drive system

[0096] 12 Electric converter

[0097] 14 Rotating motor

[0098] 16 Power Grid

[0099] 18 Power semiconductor devices

[0100] 20 Half Bridge

[0101] 22 DC Link

[0102] 24 Power semiconductor modules

[0103] 26 Device Temperature Sensor

[0104] 28 Controller

[0105] 30 Ambient temperature sensor

[0106] 32 Machine Learning Algorithms / Modules

[0107] 34 Averaging module and / or comparator module

[0108] 36 Input Layer

[0109] 37 Output layer

[0110] 38 hidden layers

[0111] n Speed ​​input neuron

[0112] t actual time step

[0113] t-1,t-2 previous time step

[0114] I Operating Point Indicator

[0115] T Torque input neuron

[0116] T d Measured device temperature

[0117] Estimated device temperature

[0118] T a Ambient temperature

[0119] E Error

[0120] F Fault signal

[0121] Moving average error

[0122] Median Error

Claims

1. A method for predicting the fault behavior of an electrical converter (12), the method include: receiving an operating point indicator of the electric converter (12) indicating an actual operating point of the electric converter (12), wherein the electric converter (12) is connected to a rotating electric machine (14) for driving the rotating electric machine (14), and the operating point indicator comprises at least one of the following: a torque of the rotating electric machine (14), a rotation speed of the rotating electric machine (14); receiving a measured device temperature of the power semiconductor device (18) of the electrical converter (12) indicative of an actual temperature of the power semiconductor device (18); inputting the operating point indicator as input data into a machine learning algorithm (32) trained using historical data including the operating point indicator and an associated device temperature, wherein the machine learning algorithm (32) has been trained using historical data recorded during normal operation of the power semiconductor device (18); estimating an estimated device temperature using the machine learning algorithm, wherein the estimated device temperature represents a device temperature during normal operation; The fault behavior is predicted by comparing the estimated device temperature to the measured device temperature.

2. The method according to claim 1, The input data of the machine learning algorithm include: the measured device temperature at a previous time step and the operating point indicator at an actual time step; wherein the machine learning algorithm (32) estimates an estimated device temperature for the actual time step.

3. The method according to claim 1, wherein the input data comprises measured device temperature for a plurality of previous time steps; and / or wherein the input data comprises operating point indicators for a plurality of previous time steps; and / or Wherein the input data comprises one or more differences in measured device temperature for a plurality of previous time steps.

4. The method according to claim 2, wherein the input data comprises measured device temperature for a plurality of previous time steps; and / or wherein the input data comprises operating point indicators for a plurality of previous time steps; and / or Wherein the input data comprises one or more differences in measured device temperature for a plurality of previous time steps.

5. The method according to any one of the preceding claims, further comprising: include: receiving an ambient temperature of the electrical converter (12), the ambient temperature being indicative of an ambient temperature of the power semiconductor device (18) and / or the electrical converter (12); and / or The input data of the machine learning algorithm (32) also includes the ambient temperature.

6. The method according to claim 5, wherein the measured device temperature in the input data is provided relative to the ambient temperature; Wherein the estimated device temperature output by the machine learning algorithm is provided relative to the ambient temperature.

7. The method according to any one of claims 1 to 4, further comprising: include: determining a temperature error based on a difference between the measured device temperature and the estimated device temperature; The temperature error is compared to a threshold value for predicting the fault behavior.

8. The method according to any one of the preceding claims 1 to 4, wherein the machine learning algorithm (32) has been trained using historical data of the same electrical converter (12); and / or Wherein the machine learning algorithm (32) has been trained using historical data of at least one different electrical converter.

9. The method according to any one of the preceding claims 1 to 4, Wherein the machine learning algorithm (32) is an artificial neural network.

10. The method according to any one of the preceding claims 1 to 4, The operating point indicator includes at least one of the following: Converter current, Converter voltage, Switching frequency, DC link voltage.

11. The method according to any one of the preceding claims 1 to 4, The power semiconductor device (18) is a power semiconductor switch; and / or The power semiconductor device (18) is an IGBT.

12. A computer program adapted, when executed by a processor, to perform the method according to any one of the preceding claims.

13. A computer readable medium having stored therein a computer program according to claim 12.

14. A controller (28) of an electric converter, adapted to perform the method according to any one of claims 1 to 11.

15. An electrical converter (12), include: A plurality of power semiconductor devices (18); at least one temperature sensor (26) arranged to measure a device temperature of at least one of the power semiconductor devices (18); The controller (28) of claim 14, adapted to estimate an estimated device temperature of at least one of the power semiconductor devices (18) and for predicting the fault behavior.

Images