FIELD OF THE INVENTION

[0001] The invention relates to a method for operating an electric drive system for electricals driving a vehicle.

BACKGROUND INFORMATION

[0002] CN 112350555 A shows a multi-phase 2-level inverted space vector pulse width modulation method for suppressing common-mode voltage. Furthermore,

KR 102091589 B1 shows a device for reducing a common-mode voltage in a multiphase motor control system. Furthermore, CN 109194208 A shows a speed sensor less control method.

SUMMARY OF THE INVENTION

[0003] It is an object of the present invention to provide a method for operating an electric drive system for electrically driving a vehicle such that common-mode voltage may be kept particularly low.

[0004] This object is solved by a method having the features of patent claim 1. Advantageous embodiments with expedient developments of the invention are indicated in the other patent claim.

[0005] The invention relates to a method for operating an electric drive system for electrically driving a vehicle such as, for example, a passenger vehicle. The method comprises a first step of selecting, from a plurality of pulse width modulation (PWM) patterns for operating the electric vehicle system and stored in a memory, one of the patterns on the basis of at least one machine learning (ML) process estimating respective common-mode voltages (CMV) resulting from the respective patterns. The method comprises a second step of using the selected pattern for operating, in particular controlling, the electric drive system. Thus, for example, the invention may use machine learning technology to online predict a maximum vote current of a three-phase motor drive inverter. In particular, by the invention, common-mode noise resulting from common-mode voltage may be reduced, in particular by reducing a modal phase current and keeping a high efficiency of the inverter. Preferably, the invention may reduce the common-mode (CM) current thereby minimizing the usage of the CM choke. This may result in less EMI while keeping inverter loss low.

[0006] Further advantages, features, and details of the invention derive from the following description of a preferred embodiment as well as from the drawings. The features and feature combinations previously mentioned in the description as well as the features and feature combinations mentioned in the following description of the figures and/or shown in the figures alone can be employed not only in the respectively indicated combination but also in any other combination or taken alone without leaving the scope of the invention.

BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The novel features and characteristic of the disclosure are set forth in the appended claims. The accompanying drawings, which are incorporated in and constitute a part of this disclosure, illustrate exemplary embodiments and together with the description, serve to explain the disclosed principles. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. The same numbers are used throughout the figures to reference like features and components. Some embodiments of system and/or methods in accordance with embodiments of the present subject matter are now described below, by way of example only, and with reference to the accompanying figures.

[0008] Fig 1 shows a schematic view of a 3-phase voltage source inverter for an electric drive system. [0009] Fig. 2 shows diagrams for illustrating a method for operating the electric drive system.

[0010] Fig 3 shows 3 equations for illustrating the method.

[0011] Fig. 4 shows a further diagram for further illustrating the method.

[0012] Fig 5 shows a table for further illustrating the method.

[0013] Fig. 6 shows a schematic view of a neural network for further illustrating the method.

[0014] Fig. 7 shows an algorithm for an implementation of a lookup table.

[0015] Fig. 8 shows an algorithm for an implementation of a neural network.

[0016] Fig 9 shows a flow diagram for further illustrating the method.

[0017] In the figures the same elements or elements having the same function are indicated by the same reference signs.

DETAILED DESCRIPTION

[0018] In the present document, the word "exemplary" is used herein to mean "serving as an example, instance, or illustration". Any embodiment or implementation of the present subject matter described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

[0019] While the disclosure is susceptible to various modifications and alternative forms, specific embodiments thereof have been shown by way of example in the drawing and will be described in detail below. It should be understood, however, that it is not intended to limit the disclosure to the particular forms disclosed, but on the contrary, the disclosure is to cover all modifications, equivalents, and alternatives falling within the scope of the disclosure.

[0020] The terms “comprises”, “comprising”, or any other variations thereof, are intended to cover a non-exclusive inclusion so that a setup, device or method that comprises a list of components or steps does not include only those components or steps but may include other components or steps not expressly listed or inherent to such setup or device or method. In other words, one or more elements in a system or apparatus preceded by “comprises” or “comprise” does not or do not, without more constraints, preclude the existence of other elements or additional elements in the system or method.

[0021] In the following detailed description of the embodiment of the disclosure, reference is made to the accompanying drawing that forms part hereof, and in which is shown by way of illustration a specific embodiment in which the disclosure may be practiced. This embodiment is described in sufficient detail to enable those skilled in the art to practice the disclosure, and it is to be understood that other embodiments may be utilized and that changes may be made without departing from the scope of the present disclosure. The following description is, therefore, not to be taken in a limiting sense.

[0022] In the following, a method for operating an electric drive system for electrically driving a vehicle will be explained on the basis of the figures. As will be described in greater detail below, the method comprises a first step of selecting, from a plurality of pulse width modulation patterns for operating the electric drive system installed in a memory, one of the patterns on the basis of at least one machine learning (ML) process estimating respective common-mode (CM) voltages resulting from the respective pattern. The respective common-mode voltage is also referred to as CMV. The method further comprises a second step of using the selected pattern for operating the electric drive system. As shown in Fig. 9, a control system 10 may comprising a computing device 12, wherein the control system 10 is configured to perform the method (i.e. the steps of the method). Preferably, the vehicle comprises the control system 10 such that the vehicle is configured to perform the method.

[0023] To reduce the common-voltage (CMV) in a pulse width modulation (PWM) based motor drive system, many CMV reduction method have been proposed. However, the performance of such methods has limitations, such as only implemented on particular operating conditions with fixed switching frequency or PWM patterns, relying on a simulation or experimental data. The method may allow to build a smart controller that may actively evaluate CM performance through artificial intelligence. Machine learning methods are employed to actively analyze 3 popular PWMs (SVPWM, AZSPWM, and DPWMMin) on chip. In this way one way based on the talk and speed command to online determine the best PWM patterns and switching frequency with minimum requirements of computation resources.

[0024] A three-phase voltage source inverter (VSI), as shown in Fig. 1, is commonly used in AC motor drive systems. As a result of common-mode voltage (CMV), commonmode current (CM I) flows through the motor case to the capacitor middle point. This may cause insulation failure, greatly shorten the motor lifespan and may cause significant EMI problems. To reduce the CMV, and ultimately reduce the CMI, some PWM methods or pattern have been proposed, such as active zero-state PWM (AZSPWM) and discontinuous PWM clammed to negative bus (DPWMMin). As shown in Fig. 2, the common idea for these PWMs is to reduce or avoid the usage zero-state vectors. Compared with space vector (PWM, SVPWM) the peak value of the CMV has been shaped by AZSPWM and DPWMMin.

[0025] To quantify the actual CMI reduction performance, traditional approaches count on simulation and experiments to collect the data with respect to specific operating points, which, however, lack of analytic models and tools. A double-Fourier integral (DFI) based analytic model may be used to characterize the CM performance of various PWMs. However, such DFI model is significantly time consuming, taking more than 15 minutes to finish the computation on an ARM cortex-A9 processor on Xilinx Zynq-7000 Zedboard. This makes it infeasible to do online calculation of CMI. Although using a lookup table (LUT) is a good alternative, it may require a lot of memory space, especially considering the table could be multi-dimensional. Another method is to use a machine learning (ML) algorithm to learn a mapping between input parameters and a target, thus only a trained model needs to be stored in the memory. The space caused than depends on the complexity of the model rather than the data itself with acceptable run time, which makes the online estimation of the CMI thereby further selecting PWM patterns possible.

[0026] Lookup tables may estimate the data at the non-existing points inevitable by using some interpolation methods. For a 3-D lookup table, the double linear interpolation (DLI) is the simplest method and the easiest one to be implemented in the embedded systems/microprocessors. Assuming there are 4 data points (fi, Mi, In), (fi, M2, I12), (f2, Mi, I21) and (f2, M2, I22) where f the inverter switching frequency, M is the modulation index and I is the peak CMI (fi < f ₂ , Mi < M2). Any data on the surface formed by these 4 data points may be estimated using DLI. Assuming the unknown data is (f, M, I) where f and M are given, and I needs to be estimated. The estimation result may be obtained by equation (1) shown in Fig. 3.

[0027] Estimating the maximum CM I is a regression problem. Various ML algorithms to solve such problems have already been proposed and applied. Linear regression is the simplest regression algorithm where each feature has linear relationship with the target, as shown in equations (2) and (3) shown in Fig. 3.

[0028] therein < , x and b are vectors, and the loss function is used to update the set (<o,b) that gives the minimum error given the dataset D={(xi,yi),(x2,y2),... ,(x _m ,y _m )}, xi eR ^d , eR ¹ .

[0029] There are other regression models such as Ridge Regression and Lasso Regression which aim to address overfitting issue and learn a sparse model. However, these are linear regression models which give the best performance in the linear regression problems, where the features have high linearity with the target. For the nonlinear regression problems, without any feature engineering, those models should not be considered as the first choice. Thus, other models need to be raised to solve non-linear regression problems. Neural networks (NN) consist of series of hidden layers with multiple neurons. Each neuron processes the equation (2) and passes the result to an activation function which is non-linear. Because of activation functions in the neural networks (NN), such models yield better performance and are able to learn any functions in theory.

[0030] As mentioned above, the DFI takes a long time to execute on an ARM processor, making it impossible to online estimate the CMI. Here either lookup table or NN can replace DFI and directly obtain the maximum CMI at any operating point. Such maximum CMI then may be compared with vehicle standards such as CISPR 25 to online examine if any EMI requirement has been violated.

[0031] To form the lookup table or train the NN, the dataset needs to be collected. For example, it may be assumed that the motor inverter has no CM choke. The DC bus voltage and fundamental frequency are constant which are 400 V and 100 Hz, respectively. The modulation index (Ml) and switching frequency (fs) are the input while the peak CMI are the output of the lookup table. For NN, modulation index and switching frequency are the features and the peak common mode current is the target.

[0032] The modulation index in the training dataset varies from 0.1 to 0.9 with a step size of 0.05, while the switching frequency is from 10 kHz to 40 kHz with the step size of 250 Hz. Another dataset used as the testset is collected with less data where the modulation index is from 14 kHz to 38 kHz with the step size of 5 kHz, and the switching frequency is from 0.13 to 0.87 with the step size of 0.1. Three PWMs are used, therefore the peak CM I values under each PWM are also collected for the training dataset and test dataset. In summary, there are totally 2057 samples in the training dataset and 40 samples in the test dataset, where each sample contains F _s , Ml and three peak CMI values for SVM, AZPWM and DPWM. Fig. 4 shows the distribution of the data in the training dataset and test dataset in 2D point of view (fs and Ml). Furthermore, surface plots of the dataset may be created. From such surface plots, the non-linearity between the features and the target may be obviously gathered.

[0033] To implement and test the lookup table, the whole training dataset is stored in the memory and the test dataset is used for evaluation only. For the NN, the training dataset is used to train the model in a PC and only the trained parameters (weights and biases) are stored in the memory. The structure of the NN is shown in table 1 in Fig. 5 and in Fig. 6.

[0034] In the training process of the NN, the training dataset is split into training set and validation set based on the ratio of 4:1. The training process will stop when the validation performance has increased more than the maximum failure times, which is 10 in this training algorithm. Levenberg-Marquardt backpropagation is used as the training function. The pseudo code of the implementation of lookup table and NN are given in algorithm 1 and algorithm 2, respectively, as shown in Figs. 7 and 8.

[0035] The prediction performance of the lookup table and the NN is evaluated using the test dataset (containing 40 samples), and the results may be shown in corresponding figures or plots. In such plots, for example, the symbol "." may denote the real peak CMI value, the symbol "+" may denote the estimated peak CMI value using lookup table and the symbol "o" may denote the estimated value using NN. For example, a table may show the summary of the models and time/space expense inside the ARM Cortex-A9 processor on Xilinx Zynq-7000 Zedboard. The program runtime is measured using a global timer on the board and the occupied space is calculated based on the data type and number of the data. Based on the results, the lookup table has shorter runtime (at microsecond level), but it requires much larger space to store data. On the other hand, the NN sacrifices time to save space, which has millisecond-level runtime but only needs 30 times less the memory. The prediction results of the lookup table on SVM and AZSPWM are slightly better than NN while on DPWM, NN is better than the lookup table. It should be addressed that the neural network may still be utilized to get prediction results such as current ripple prediction results. In real applications, the runtime of neural network is already acceptable for the on-chip real time operation. Once such model is embedded in the ARM, one may then online decide the best switching fs and PWM choices to comply with the CM I standard.

[0036] The method may be used to solve the computation issue and the CM I estimation through machine learning perspective compared with lookup table, aiming to save the computation resource consumed by the DFI. The lookup table and neural network give the similar performance. However, the current stage, DC bus voltage and fundamental frequency are only constant. In electric vehicles, such two parameters are variable, yielding the lookup table is not an ideal choice because it will take too much memory space. After that, one may embed such ML based CM I estimation module inside ARM to online decide PWM patterns. Furthermore, Fig. 9 shows a flow diagram further illustrating the method. The method may be used when the initialization frequency values are different from the switching frequency values, 10 kHz to 40 kHz, as shown in Figure 9.

List of Reference Signs

10 control system

12 computing device