Method and device for predicting load time series for a component of a technical system

Abstract

A computer-implemented method for predicting load profiles of load variables of a component (2) to be monitored of a technical system (1, comprising the following steps: - Providing historical load time series for multiple loads, wherein the loads are quantities on which a measure of aging or degradation of the component (2) being monitored significantly depends; - Subdividing the historical load variable time series into successive load variable time windows (f1([;tsp]), f2([;tsp]), ..., fx([;tsp])) with a number of consecutive values; - Evaluating the successive load size time windows (f1([;tsp]), f2([;tsp]), ..., fx([;tsp])) using a data-based prediction model (10).

Description

Technical field

[0001] The invention relates to the lifetime prediction of components in a technical system based on the prediction of load profiles that indicate the wear or degradation of the components. In particular, the present invention relates to methods for predicting load time series based on historical load time series. Technical background

[0002] To ensure the reliable, trouble-free operation of a technical system, it is necessary to monitor the aging or degradation of wear-prone components within that system. For example, in a motor vehicle, components such as the battery, drive motor, bearings, seals, and the like can be monitored to detect any potential loss of performance or malfunction of the respective component at an early stage, so that the affected components can be replaced during scheduled, planned, or unscheduled maintenance.

[0003] The prediction of remaining service life is achieved by evaluating time series of load variables that significantly influence the aging measure, which indicates the degree of degradation of the component in question. For example, for a battery as a component to be monitored, load variables could include temperature, state of charge, and charging rate, as these significantly determine the battery's aging. For bearings as a component to be monitored, load variables could include rotational speed, transmitted torque, temperature, and similar parameters.

[0004] Using suitable aging models that evaluate historical time series, a current aging measure can be determined. To determine a remaining service life, the aging measure's progression must be predicted by the corresponding aging models, which requires predicting the load profiles.

[0005] German patent application DE 10 2016 202 306 A1 describes a method for operating an electrified motor vehicle. The motor vehicle has an electrical system with at least one electrical or electronic component that is subjected to stress during operation. The method determines the cumulative stress and the expected service life of the at least one component. If the expected service life deviates from the target service life, measures are taken to reduce the stress on the component and thereby increase its service life.

[0006] A method for determining the aging value of a rechargeable battery is known from German patent publication DE 10 2020 100 668 B4. In this method, measurement data for one or more state variables of the battery are obtained, one or more derived state variables of the battery are determined using a preceding algorithm, and the aging value is determined using at least one machine-learned algorithm. A key feature of this method is that the aging process is predicted. Disclosure of the invention

[0007] According to the invention, a method for predicting load profiles of a component of a technical system for determining an aging measure according to claim 1 and a corresponding device according to the dependent claim are provided.

[0008] Further details are specified in the dependent claims.

[0009] According to a first aspect, a method for predicting load characteristic profiles of load characteristics of a component of a technical system to be monitored is provided, in particular for determining a profile of an aging measure with the following steps: - Providing historical load time series for multiple loads, where the loads represent quantities on which a measure of aging or degradation of the component being monitored significantly depends; - Dividing the historical load variable time series into successive load variable time windows with a number of consecutive values; - Evaluating successive load time windows using a data-based prediction model up to a current time step to obtain predicted load time windows for at least one subsequent time step, wherein the data-based prediction model comprises a data-based encoder part, a data-based decoder part, and a data-based prediction model, wherein the encoder part is configured to compress the load time windows for a time step into a latent state vector, wherein the data-based prediction model is configured as a recurrent data-based model to model a predicted latent state vector for the subsequent time step from the latent state vector for the time step, and wherein the decoder part is configuredto model a predicted load-time window for the subsequent time step from the predicted latent state vector, with a number of successive values ​​of the multiple load variables; - Combining the historical load time windows and the predicted load time windows to obtain a complete load time series of the multiple loads; - Using the entire load size time series to determine the remaining service life of the component.

[0010] Furthermore, the course of an aging measure can be determined by a given aging model depending on the entire load time series, whereby the remaining service life is determined by a point in time at which the course of the aging measure reaches a given aging measure threshold.

[0011] To model the aging measure of a component in a technical system, it is necessary to consider the time-dependent profiles of load variables up to a point in time that corresponds to or lies after the end of the service life. The aging measure provides a numerical value that is representative of the component's aging. If the aging measure at a given time corresponds to a predefined aging measure threshold for a given end of service, then that corresponding time corresponds to the end of service.

[0012] Particularly when multiple load variables are primarily responsible for degradation, these variables, individually and / or in combination, can have a non-linear influence on the aging behavior. Therefore, the load variable profiles must be considered as a time series, starting from a commissioning point, in order to determine the component's degree of aging at a specific point in time. The underlying aging models depend on the type of component and are often formulated as linear or non-linear systems of differential equations.

[0013] To determine the aging level of an operational component, time series of several load variables are evaluated starting from the commissioning point (BOL state, BOL: Begin of Life) to determine the aging level over time. When the aging level reaches a predefined threshold, the end of the component's service life is indicated. If the determined point in time lies in the past, the need for immediate component replacement can be identified. If the point in time lies in the future, a remaining service life can be determined, which dictates the type of measures to be taken and enables the planning of a component replacement. The above method involves predicting future load profiles based on historical load profiles.This results in a time series of load variables, from which the remaining service life can be determined using the procedure described above. Alternatively, the end of service life can also be determined using a rule-based approach, e.g., by comparing the threshold values ​​of one or more load variables, so that the remaining service life can be determined depending on when one or more load variables reach a predefined threshold value.

[0014] Specifically, the above method employs a prediction model that combines an autoencoder with a predictive model. The autoencoder comprises an encoder section and a decoder section and is trained to reconstruct load time windows. These load time windows represent time segments for a given time step, each containing a predefined number of consecutive values ​​or samples of each of the considered load variables. Each time window thus comprises a load time series of the relevant load variables for the component being monitored.

[0015] Prediction is based on the latent state vector. A trained, data-driven prediction model is used for this purpose. The latent state vector at the output of the encoder section of the autoencoder is used to generate a predicted latent state vector for a subsequent time step using the prediction model. From this predicted latent state vector, a predicted load time window for the following time step is then modeled using the decoder section.

[0016] It may be provided that the predicted load time windows are predicted for a large number of future time steps, e.g. by recursively evaluating the prediction model by evaluating the most recently determined predicted load time windows as inputs for the next time step using the prediction model.

[0017] The prediction model can thus be used to predict a load time window for the next time step, based on the current load time windows of a given time step. By recursively applying the prediction model, a number of predicted load time windows can be determined starting from the current time.

[0018] By recursively applying the prediction model with the load time windows of the most recently predicted time step, the series of load time windows can be continued by determining the predicted time windows sequentially. By combining the load time windows, the complete time series of the load variables required to determine the aging curve and the remaining service life can be determined.

[0019] The encoder and decoder parts of the autoencoder could each be constructed with several simple layers of neurons (fully connected layers). Alternatively, they could also include one or more convolutional layers, which can be implemented one-dimensionally across the time axis or two-dimensionally together with the dimension of the load variables.

[0020] The latent state vector, which is modeled by the encoder part from the load size time windows of a specific time step, is fed into the data-based prediction model, which outputs a predicted state vector for the following time step.

[0021] Furthermore, the evaluation of the prediction model can include feeding one or more aggregated quantities from the load parameter time windows into the prediction model in addition to the latent state vector of the encoder part, so that the determination of the predicted latent state vector depends on the one or more aggregated quantities.

[0022] Additionally, the latent state vector can be augmented with load characteristic features as aggregated quantities to achieve improved prediction by the forecasting model. An aggregated quantity can comprise a mean value or an average gradient of at least one respective load characteristic within the corresponding load characteristic time window.

[0023] Furthermore, the values ​​of at least one load variable time window can be corrected by the amount of the initial value of the time series of the load variable time window before being fed to the encoder part, wherein the initial value of the load variable in question is fed to the prediction model in addition to the latent state vector of the encoder part, so that the determination of the predicted latent state vector depends on the initial value of the load variable in question, wherein the prediction model is designed to further output a predicted initial value with which the predicted load variable time window of the load variable in question is applied, in particular additively.

[0024] The load variable's behavior within a load variable time window can have an initial value of 0 if the load variable time window is corrected for this initial value. This is useful if the load variable is an integrating load variable, such as the state of charge of a battery.

[0025] The prediction model can be implemented in various ways, especially as a recurrent neural network, which can be designed in the form of an LSTM (Long Short-Term Memory), GRU or simple recurrent neural network.

[0026] According to another aspect, a procedure for providing a prediction model for predicting load profiles of load variables of a component to be monitored in a technical system is provided, with the following steps: - Providing historical load time series for multiple loads, where the loads represent quantities on which a measure of aging or degradation of the component being monitored significantly depends; - Subdividing the historical load variable time series into load variable time windows with a number of values ​​for successive time steps; - In a first training step, training an autoencoder with a data-based encoder part and a data-based decoder part using the load size time windows as training data; - in a second training step; training a data-based prediction model with latent state vectors as training data, which are determined using the encoder part from the load variable time windows for successive time steps, so that a predicted latent state vector is determined for each load variable time window of the last evaluated time step; - Combining the autoencoder with the prediction model, so that the output of the encoder part serves as the input for the data-based prediction model and the output of the data-based prediction model serves as the input for the decoder part.

[0027] The training of the prediction model can be performed in two stages. In the first training stage, only the autoencoder is trained based on historical load time windows. The encoder and decoder components are trained in a known manner based on a reconstruction loss.

[0028] Additionally, the autoencoder can also be trained as a variational autoencoder to achieve a predefined distribution of the elements of the latent state vector on the output side of the encoder part.

[0029] The prediction model can be trained independently of the autoencoder for more efficient training. For example, the prediction model can be trained based on a time series of latent state vectors derived from sequences of historical load time windows across successive time steps. The training loss for the prediction model then arises, in a known manner, from the difference between the prediction model output of a latent state vector from a specific time step and the decoder output of the load time window from the corresponding subsequent time step. The second training step is performed while keeping the model parameters of both the encoder and decoder sections constant.

[0030] Multi-stage training reduces training time because the individual functions are easier to learn. A key advantage is that the number of values ​​per time window can be reduced, resulting in a significant reduction in the amount of data without compromising the predictive accuracy of the largest time series.

[0031] The above method, combining a data-driven autoencoder with a data-driven prediction model, provides a way to predict load time series. Prediction is based on compressed information from the latent state vector, encompassing the load time windows of one time step to the next, followed by a reconstruction of the predicted load time windows. Compressing the load time windows significantly reduces the amount of data in the prediction model by filtering out redundant information from multiple load time series. Furthermore, the encoder and decoder components of the autoencoder enable better generalizability of the overall prediction model, thereby avoiding the risk of overfitting.

[0032] By using the autoencoder, the characteristics can be decoupled with respect to a correlation between the load variables by compressing the load variable time windows for a specific time step onto a latent state vector, so that the state variables of the state vector can be predicted independently of each other by the prediction.

[0033] Furthermore, debiasing can be achieved using two-stage training, i.e., training the autoencoder followed by training the prediction model. Debiasing can be performed by using only the relevant time steps of the load windows during training; that is, load windows in which there is no load on the components in question are ignored. Alternatively, the time steps of the load windows can be weighted accordingly, so that during training, load windows in which there is no significant load on the component have only a very minor impact on the prediction model. Brief description of the drawings

[0034] The embodiments are explained in more detail below with reference to the accompanying drawings. These show: Fig. 1 a vehicle as an example of a technical system with a battery as a component for which a remaining service life is to be determined; Fig. 2 a structure of the prediction model for predicting load variable time series; Fig. 3. A flowchart illustrating a procedure for determining the remaining service life of a component; and Fig. 4. A flowchart illustrating a procedure for creating a prediction model for predicting load time series for multiple loads. Description of embodiments

[0035] Fig. Figure 1 schematically shows a vehicle 1 with a large number of wear-prone and operationally degradable components 2. These components 2 can include, for example, a battery 3, a bearing, a drive motor 4, a seal, or the like.

[0036] The degradation of the corresponding components 2 occurs based on characteristic load parameters and their time profiles. Using the battery 3 as an example, these load parameters can correspond to battery temperature, state of charge, and charging throughput. In the case of a bearing, torque, rotational speed, and bearing temperature can be assumed as load parameters.

[0037] To determine the degradation of a specific component 2 of vehicle 1, historical time series of the load variables assigned to the component must be evaluated in order to determine an aging measure. The load variable profiles must be evaluated at a high temporal resolution (0.1 Hz to 100 Hz) to obtain a reliable indication of the aging measure.

[0038] To determine the remaining service life of a component, the future load profiles must also be predicted. By combining historical and predicted load profiles, an aging trajectory, or a time course of the aging measure, can be determined for the component in question based on an aging model. This allows the remaining service life to be determined depending on when an aging measure threshold is reached.

[0039] Predicting load time series is necessary to determine the aging measure for the component in question based on a degradation or aging model. Such aging models are often based on linear or non-linear differential equations, so the availability of the complete load time series, preferably starting from a commissioning point (BOL point: Begin of Life), is required.

[0040] Suitable sensors 6 can be arranged on the components 2 to detect the load quantities.

[0041] Using a control unit 5, in which the aging model for the component to be monitored or aging models for the components 2 are implemented, the degree of aging in the vehicle 1 can be determined. Alternatively, the degree of aging can also be determined remotely in a central unit (cloud, not shown) if the corresponding load time series are transmitted to it.

[0042] The control unit 5 records and stores the load time series from the time of commissioning and uses them to determine the remaining service life of the relevant component 2 based on predicted load time series. For this purpose, a prediction model and an aging model are implemented in the control unit 5 or in the remote central unit. The predictive model is used to determine the predicted load time series. Based on the combined historical and predicted load time series, the aging model determines the progression of an aging measure, from which the remaining service life can be calculated. The prediction model and the aging model can be implemented in hardware, in software, or a combination of both.

[0043] As in Fig. As shown in Figure 2, the prediction model 10 comprises an encoder section 11 and a decoder section 12 of an autoencoder 13 and a prediction model 14, which is arranged in the signal flow direction between the encoder section 11 and the decoder section 12. On the input side, load variable time windows f1([;tsp}), f2([;tsp]), ..., fx([;tsp]) are fed to the encoder section 11 to provide a latent state vector z on the output side of the encoder section 11. tsp corresponds to a considered time point, so that the time step ending at that time is considered. This is processed in the prediction model 14 into a predicted state vector z'. The predicted state vector z' is modeled by the decoder part 12 into corresponding predicted load quantity time series f*1([tsp;]), f*2([tsp;), ..., f*x([tsp;]) of a time step closest to the considered time tsp.

[0044] If the load variable is an integrating load variable, such as the state of charge of a battery, the load variable in question can be corrected by the magnitude of the initial value of the time series of this load variable within the load variable time window, so that the corresponding time series of the time window starts at 0, as illustrated in the example of the load variable time window for load variable f2. The encoder section 11 is then fed f2([tn;tn+1])-f2(tn). The magnitude of the initial value f2(tn) can then be provided to the prediction model 14 as an additional input alongside the latent state vector z. This model then models a predicted initial value f2*(tp) for the load variable in question as an output, for applying to the time series of the corresponding load variable f*2 of the predicted load variable time window f2*( [tp;tp+1])-f2*[tp].

[0045] While the encoder part 11 generates the latent state vector z based on the load variable time windows f1([;tsp]), f2([;tsp]), ..., fx([;tsp]) of the relevant time step, which is fed to the prediction model as input, load variable characteristics fmerk can additionally be fed as aggregated quantities from the load variables of the load variable time windows as input for the prediction model 14 in order to improve the modeling of the predicted state vector z'.

[0046] The encoder part 11 and the decoder part 12 of the autoencoder 13 can each be constructed with several simple layers of neurons (fully connected layer). Alternatively, they can also include one or more convolutional layers, which can be implemented one-dimensionally across the time axis of each of the load variable time series, or two-dimensionally across the time axis of each of the load variable time series together with the dimension across the load variables. The encoder part 11 serves for compression and dimensionality reduction, thereby reducing the correlations between the load variables. The decoder part 12 serves to reconstruct the load variable time series from the predicted state vector.

[0047] The data-based prediction model 14 can be implemented in various ways, in particular as a recurrent neural network, which can be designed in the form of an LSTM (Long Short-Term Memory), GRU or simple recurrent neural network.

[0048] In Fig. Section 3 describes a procedure for determining the remaining service life of component 2 using a flowchart.

[0049] In step S1, load variables are continuously recorded and stored as time series of samples. This provides the control unit with 5 load variable time series for the several considered load variables, starting from the commissioning time.

[0050] In step S2, the load time series are divided into consecutive load time windows. Each load time window contains a predefined number of consecutive values ​​of the respective load variables. Additionally, an initial value of an integrating load variable and load variable characteristics, as described above, can be provided separately.

[0051] In step S3, the sequence of load time windows is used to evaluate prediction model 10, which has a recurrent model component, in order to predict a load time window following the current time. This predicted load time window is then fed into prediction model 10 to supplement the sequence of load time windows for evaluation and to successively predict load time windows for future time steps. This results in a sequence of historical load time windows and predicted load time windows, from which a complete load time series can be constructed.

[0052] In step S4, the entire determined load quantity time series is evaluated using the aging model to obtain a trend of an aging measure over time.

[0053] In step S5, a future point in time is determined, based on an aging threshold value that indicates the end of the component's service life. This point in time is then used to calculate the remaining service life of the component.

[0054] In step S6, depending on the determined remaining service life, instructions can be issued to a user, or maintenance or replacement can be scheduled. Furthermore, the remaining service life can be signaled visually, audibly, or in another way.

[0055] The procedure for training the prediction model 10 to determine the predicted load time series is illustrated using the flowchart of the Fig. 4 described in more detail.

[0056] The training process begins with the provision of historical load variable time series in step S11.

[0057] Subsequently, in step S12, the historical load time series are divided into time windows with a predetermined number of values ​​of the selected load variable in order to obtain load variable time windows for each of the load variables under consideration.

[0058] In step S13, the load parameter time windows are preprocessed, filtered, outlier-cleaned, or similarly modified. For an integrative load parameter, the load parameter time window can be corrected by an initial value so that the first value of the corresponding load parameter time window is always zero.

[0059] In step S14, in a first training step, the autoencoder 13 is trained based on the historical load size time windows, whereby the latent state vector z is immediately provided to the decoder part 12 for evaluation.

[0060] The training is performed in a known manner based on minimizing the reconstruction error by adjusting the model parameters of the encoder part 11 and decoder part 12. This training can be performed as a normal autoencoder or as a variational autoencoder, in which the distribution of the load quantities of the latent state vector z is taken into account.

[0061] After training the autoencoder 13, the prediction model 14 can be trained in step S15 based on the sequence of historical load time windows. The load time windows for two consecutive time steps are used to train a batch of data for the prediction model, preferably while fixing or maintaining the model parameters of the encoder part 11 and decoder part 12. The batch is determined based on the difference between an output vector of the prediction model 14 when the load time windows of the earlier time step are created and the latent state vector of the load time windows output of the encoder part 11 for the later time step. This allows for training of the prediction model exclusively.

[0062] Alternatively, the training of the prediction model 14 can be carried out in two steps: in a first step, in which only the prediction model 14 is carried out with fixed model parameters of the encoder part 11, and in a second step, in which the prediction model 14 including the encoder part 11 is trained.

[0063] In addition to the latent state vector, the prediction model 14 can also receive further input variables that comprise aggregated load characteristics for one or more load time windows, such as a mean value, a gradient value of a load time window, or the like. This improves the prediction of the predicted state vector.

[0064] If a load variable time window is corrected by an initial value, this initial value can also be fed into the prediction model 14 as an input. The prediction model 14 is then trained to model the predicted latent state vector on a predicted initial value. After generating the predicted load variable time windows, the load variable time window assigned to the integrating load variable can then be populated with the predicted initial value.

[0065] In this way, a prediction model 10 can be trained and deployed, which can then model predicted load time windows. The modeling of load time series can be achieved through recursive application of the prediction model. This involves feeding a predicted load time window for a given time step into the prediction model to predict the load time window for the next time step. In this way, a predicted load time series can be determined from historical load time series. Using a suitable aging model, this predicted load time series can then be used to determine the temporal progression of an aging measure. This results in a remaining service life, from which appropriate instructions for action or interventions in the operation of the technical system or vehicle can be derived. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2016 202 306 A1

[0005] DE 10 2020 100 668 B4

[0006]

Claims

[1] Computer-implemented method for predicting load magnitude curves of load variables of a component (2) to be monitored of a technical system (1), in particular for determining a curve of an ageing measure, comprising the following steps: - providing historical load size time series for several load sizes, where the load sizes indicate sizes on which a degree of ageing or degradation of the component to be monitored (2) depends significantly; - Dividing the historical load size time series into consecutive load size time windows (f1([;tsp]), f2([;tsp]), ..., fx([;tsp])) with a number of consecutive values; - Evaluating the consecutive load size time windows (f1([;tsp}), f2([;tsp]), ..., fx([;tsp])) using a data-based prediction model (10) up to a current time step in order to generate predicted load size time windows (f*1([tsp;]), f*2([tsp;]), ..., f*x([tsp;])) for a subsequent time step, wherein the data-based prediction model (10) has a data-based encoder part (11), a data-based decoder part (12) and a data-based prediction model (14), wherein the encoder part (11) is designed to compress the load size time windows for a time step into a latent state vector (z), wherein the data-based prediction model (14) is designed as a recurrent data-based model in order to model a predicted latent state vector (z') for the subsequent time step from the latent state vector (z) for the time step, and wherein the decoder part (12) is designed to model predicted load size time windows for the subsequent time step from the predicted latent state vector (z'). - Merging the historical load size time windows and the predicted load size time windows (f*1([tsp;]), f*2([tsp;]), ..., f*x([tsp;])) to obtain an overall load size time series; - Using the entire load magnitude time series to determine a remaining lifetime of the component. [2] Method according to claim 1, wherein a course of an aging measure is determined by a predetermined aging model depending on the entire load size time series, wherein the remaining service life is determined by a point in time at which the course of the aging measure reaches a predetermined aging measure threshold value. [3] Method according to claim 1 or 2, wherein the predicted load size time windows are predicted for a plurality of future time steps by recursively evaluating the prediction model by evaluating the last determined predicted load size time windows on the input side for a next time step by the prediction model (10). [4] Method according to one of claims 1 to 3, wherein the evaluation of the prediction model (10) comprises that one or more aggregated variables from the load variable time windows are fed to the input side of the prediction model (14) in addition to the latent state vector of the encoder part (11), so that the determination of the predicted latent state vector (z') is carried out depending on the one or more aggregated variables. [5] Method according to one of claims 1 to 4, wherein the values ​​of at least one load size time window are corrected by the amount of the initial value of the time series of the load size time window before being fed to the encoder part (11), wherein the initial value of the load size in question is fed to the input side of the prediction model (14) in addition to the latent state vector (z) of the encoder part (11), so that the determination of the predicted latent state vector (z') is carried out as a function of the initial value of the load size in question, wherein the prediction model (14) is designed to further output a predicted initial value with which the predicted load size time window of the load size in question is applied, in particular additively. [6] Computer-implemented method for providing a prediction model (10) for predicting load size curves of load sizes of a component (2) to be monitored of a technical system (1), comprising the following steps: - providing historical load size time series for several load sizes, where the load sizes indicate sizes on which a degree of ageing or degradation of the component to be monitored (2) depends significantly; - Dividing the historical load size time series into load size time windows (f1([;tsp]), f2([;tsp]), ..., fx([;tsp])) with a number of consecutive values ​​for consecutive time steps; - In a first training step, training an autoencoder (13) with a data-based encoder part (11) and a data-based decoder part (12) with the load size time windows (f1([;tsp]), f2([;tsp]), ..., fx([;tsp])) as training data; - in a second training step; training a data-based prediction model (14) with latent state vectors (z) as training data, which are determined with the aid of the encoder part (11) from the load size time windows (f1([;tsp]), f2([;tsp]), ..., fx([;tsp]))) for successive time steps, so that in each case a predicted latent state vector (z') is determined for a load size time window of the last evaluated time step; - Combining the autoencoder (13) with the prediction model (14) such that the output of the encoder part (11) serves as input for the data-based prediction model (14) and the output of the data-based prediction model serves as input for the decoder part (12). [7] Apparatus for carrying out one of the methods according to one of claims 1 to 6. [8] Computer program product comprising instructions which, when the program is executed by at least one data processing device, cause the device to carry out the steps of the method according to one of claims 1 to 6. [9] Machine-readable storage medium comprising instructions which, when executed by at least one data processing device, cause it to carry out the steps of the method according to one of claims 1 to 6.

Images