Computer-implemented method and system for data processing

Abstract

The invention relates to a computer-implemented method (100) for generating a prediction of a vehicle parameter by means of an AI model, wherein the AI model maps a functional relationship between a time series (Z) of a first vehicle parameter (P1) as an input variable and a second vehicle parameter (P2) as an output variable using a vector database, the method (100) comprising: - training (110) the AI model using at least one training time series (ZT) of the first vehicle parameter (P1), and - using (120) the trained AI model for generating at least one predicted value of the second vehicle parameter (P2) as a function of an input time series (ZE) of the first vehicle parameter (P1).

Description

[0001] Description

[0002] Computer-implemented method and system for data processing

[0003] The invention relates to a computer-implemented method for predicting at least one vehicle parameter by means of a computer model, a data processing system, a vehicle, a computer program product, a computer-readable storage medium and a data carrier signal.

[0004] In the operation of a vehicle, the time-resolved recording and evaluation of vehicle parameters is of great interest in the context of numerous applications.

[0005] Vehicle parameters in this context can include continuous data such as interior temperature, wheel speed, or similar information; categorical data such as the current gear, the status of a seatbelt buckle, the activation or deactivation of fault codes; or event-based data such as ESP interventions. Such vehicle parameters are acquired primarily via suitable sensors on the vehicle.

[0006] In the course of a vehicle parameter over time (time series of the vehicle parameter), recurring patterns typically occur, which are associated with specific driving situations or...

[0007] Vehicle states or driver actions correlate. The respective meaning of such patterns can be learned by AI models (AI, artificial intelligence), enabling them to generate predictions about a vehicle parameter or other vehicle parameters using a recorded time series of a vehicle parameter.

[0008] For example, such a computer simulation model can be used to predict a future value for a vehicle parameter based on its time series, thus making a prediction about the future course of the parameter's time series. Based on this, the vehicle's operation can be proactively adapted. A computer simulation model can also be used to predict the current value of another vehicle parameter based on the time series of at least one of its parameters. This allows the vehicle's operation to be adapted particularly efficiently to recognized patterns in vehicle parameters and the corresponding driving situations. Furthermore, this approach can be used, for example, to establish model-based redundancy for the vehicle's sensors, enabling monitoring of the relevant sensors.

[0009] Training AI models requires large amounts of training data, which must be processed as efficiently as possible. Furthermore, it is essential to make the patterns and relationships contained in the training data as easily accessible as possible for the AI ​​model in order to maximize the learning effect during training and the resulting accuracy of the AI ​​model's predictions.

[0010] Examples of data processing by AI models can be found in US 2020 / 0293527 A1, WO 2024 / 068781 A1 and KR 20220123845 A.

[0011] Known methods for training AI models have been found to suffer from the disadvantage that the preparation of both the training data and the input data used after training follows static patterns that do not take into account the nature and characteristics of the specific training data. This results in losses in data processing efficiency and the resulting accuracy of the AI ​​model.

[0012] It is therefore an object of the present invention to overcome at least one of the disadvantages described above, at least partially. In particular, it is an object of the invention to provide a computer-implemented method, a data processing system, a vehicle, a computer program product, a computer-readable storage medium, and a data carrier signal that enable efficient processing of training data or input data in general for AI models, as well as a high predictive quality of the AI ​​model based on the training performed.

[0013] The foregoing problem is solved by a computer-implemented method according to a first aspect of the present invention, by a system according to a

[0014] second aspect of the present invention, by a vehicle according to a third aspect of the present invention, by a computer program product according to a fourth aspect of the present invention, by a computer-readable storage medium according to a fifth aspect of the present invention, and by a data carrier signal according to a sixth aspect of the present invention.

[0015] Further features and details of the invention will become apparent from the dependent claims, the description, and the drawings. Features and details described in connection with the computer-implemented method according to the invention naturally also apply in connection with the system according to the invention and / or in connection with the vehicle according to the invention and / or in connection with the computer program product according to the invention and / or in connection with the computer-readable storage medium according to the invention and / or in connection with the data carrier signal according to the invention, and vice versa, so that with regard to the disclosure of the individual aspects of the invention, mutual reference is always made or can always be made.

[0016] According to a first aspect, the present invention relates to a computer-implemented method for generating a prediction of a vehicle parameter by means of a computer model, wherein the computer model maps a functional relationship between a time series of a first vehicle parameter as input and a second vehicle parameter as output using a vector database, the method comprising:

[0017] - Training the Kl model using at least one training time series of the first vehicle parameter and / or

[0018] - Using the trained AI model to generate at least one predicted value of the second vehicle parameter as a function of an input time series of the first vehicle parameter,

[0019] Training the AI ​​model includes, in particular, the following:

[0020] - Determining a discretized training time series from the training time series using a plurality of predefined main symbols, where each main symbol is characteristic of a value interval of the first vehicle parameter,

[0021] - Determining a compressed training time series from the discretized time series by replacing recurring symbol patterns with an additional symbol characteristic of the respective symbol pattern, particularly by iteratively determining the number of additional symbols used depending on the recognized symbol patterns, and

[0022] - Adapting the multidimensional data vectors encompassed by the vector database of the Kl model depending on the compressed time series to minimize a prediction error of the Kl model for the training time series, wherein in particular each data vector represents a specific symbol.

[0023] In particular, the input time series and / or the training time series represents a continuous or quasi-continuous progression of a vehicle parameter over time. Additionally or alternatively, the vector database can comprise a plurality of data vectors, especially multidimensional ones.

[0024] In other words, a computer-implemented method for predicting a vehicle parameter using a computer-aided model (CA) is proposed. The CA model characterizes a functional relationship between a time series of a first vehicle parameter as input and a second vehicle parameter as output. This functional relationship is represented, at least partially, within the CA model by a multidimensional vector database, which in particular comprises a plurality of multidimensional vectors. The use of multidimensional or high-dimensional vectors and corresponding vector databases is a well-established approach in CA models.

[0025] In particular, it is intended that the AI ​​model is trained using at least one training time series of the first vehicle parameter. Additionally or alternatively, the trained AI model can be used to generate at least one predicted value of the second vehicle parameter as a function of an input time series of the first vehicle parameter.

[0026] A time series of a vehicle parameter is understood here as a progression of the respective vehicle parameter over a given period. In particular, a time series can be characteristic of the progression of the vehicle parameter within a defined time window.

[0027] Within the scope of the present invention, a training time series of a vehicle parameter is understood to be a time series of the vehicle parameter used for training the AI ​​model. It may be provided that additional information is available for or included in a training time series, in particular regarding a second vehicle parameter. For example, the training time series may be characteristic of a value or state of the second vehicle parameter, especially at a specific time within the training time series. This can be used during the training of the

[0028] AI models, particularly when fitting data vectors, can be used to minimize prediction errors. Such additional information is especially important when the second vehicle parameter is different from the first. However, if the AI ​​model is used to predict a future value of the first vehicle parameter—that is, if the first and second vehicle parameters are the same—the required additional information is already present in the training time series. In this case, for training purposes, an AI model prediction for the second vehicle parameter can be generated at a point in time included in the training time series, and the prediction error can be minimized based on this prediction.

[0029] Within the scope of the present invention, an input time series of a vehicle parameter is understood to be a time series of the vehicle parameter which is used as an input variable of the trained AI model and on the basis of which a prediction of at least one predicted value of the second vehicle parameter is to be made by the AI ​​model. The focus of the present invention is in particular on an adaptive compression of the time series used as input variables for training as well as for generating predictions, with respect to a learning effect of the AI ​​model, based on a aggregation of time series values ​​adapted to the respective data or pattern-based.

[0030] Regarding the training of the AI ​​model, it is envisaged that the training process involves deriving a discretized training time series from the training time series using a plurality of predefined principal symbols, where each principal symbol is characteristic of a value interval of the first vehicle parameter. A principal symbol can be, for example, an alphabetic, numeric, or alphanumeric symbol. In other words, for discrete, particularly equidistant, time points along the training time series, the system checks in which value interval the vehicle parameter lies and assigns the corresponding principal symbol to each time point. In this way, a symbolized representation of the training time series is determined, which corresponds to a symbol-based discretization of the training time series.

[0031] The mechanisms and advantages described in the invention with regard to a training time series can, where appropriate, be transferred accordingly to an input time series.

[0032] Furthermore, regarding the training of the AI ​​model, it is planned that a compressed training time series is determined based on the discretized training time series. This is achieved by replacing recurring symbol patterns with a supplementary symbol characteristic of the respective symbol pattern, particularly through iterative processes. Supplementary symbols can also be alphabetic, numeric, or alphanumeric symbols. In particular, the supplementary symbols can continue a logic of the given primary symbols. For example, the digits 0 to 9 can be designated as primary symbols and the numbers from 10 onwards as supplementary symbols. Or the letters A to I can be designated as primary symbols and the remaining letters of the alphabet or other letter combinations as supplementary symbols.

[0033] In particular, the number of additional symbols used is dynamically determined based on the recognized symbol patterns. In other words, it can be implemented that whenever a new symbol pattern is recognized, a new additional symbol is defined, especially based on a predefined logic, and subsequently used for that symbol pattern. In this way, the symbol vocabulary is built up in a way that is adapted to the available data and not statically defined. The maximum number of additional symbols can be limited to a maximum number (additional symbol limit).

[0034] In particular, it can be provided that the symbol patterns have a predetermined length. This length can preferably be two symbols, i.e., two main symbols, or one main symbol and one additional symbol, or two additional symbols. This allows symbol patterns of all lengths to be reliably recognized in discretized time series. In an iterative replacement of symbol patterns, which will be described in more detail below, it is possible that in a first step a first additional symbol is inserted for a given symbol pattern. This additional symbol can then, in an iterative process, subsequently form symbol patterns with adjacent symbols or other additional symbols. The use of longer symbol patterns, i.e., symbol patterns with three or more symbols, is also conceivable within the scope of the invention.Replacing recognized symbol patterns with additional symbols offers the advantage of reducing the overall number of symbols and thus increasingly compressing the discretized time series. This enables more efficient data processing without making patterns within the discretized training time series invisible to the AI ​​model.

[0035] Furthermore, with regard to training the AI ​​model, it is planned that the data vectors encompassed by the AI ​​model's vector database, particularly the multidimensional ones, are adapted to the compressed training time series to minimize the AI ​​model's prediction error for that time series. This step thus maps the information contained in the training time series, or rather the relationship between the training time series and the second vehicle parameter, into the vector database and makes it subsequently available to the AI ​​model. Numerous approaches to this procedure are documented in the prior art. In addition to the vector database, the AI ​​model can, in principle, include further modules.

[0036] In particular, the compressed training time series is used as input to the AI ​​model, and the AI ​​model is used to generate at least one predicted value for the second vehicle parameter as a function of the compressed training time series. Preferably, the predicted value of the second vehicle parameter can then be compared with at least one target value (ground truth) known from the training time series to determine the prediction error of the AI ​​model. A higher deviation corresponds to a higher prediction error.

[0037] In particular, the adaptation of the data vectors encompassed by the vector database may include at least one of the following: - Generating at least one predicted value of the second vehicle parameter or an output symbol by the Kl model depending on the compressed training time series,

[0038] - Converting the output symbol into at least one predicted value of the second vehicle parameter.

[0039] - Comparing at least one vehicle parameter with at least one predetermined target value or a basic truth to determine a prediction error of the AI ​​model and / or

[0040] - Fitting at least one multidimensional data vector to minimize the prediction error.

[0041] With regard to the present invention, it may be provided that the conversion of the output symbol into the at least one predictive value of the second vehicle parameter comprises at least one of the following:

[0042] - Reducing the output symbol to a source symbol sequence if the output symbol is a supplementary symbol, where the source symbol sequence consists exclusively of main symbols,

[0043] - Reducing the output symbol sequence or output symbol to at least one predicted value of the second vehicle parameter using the value intervals of the first vehicle parameter corresponding to the main symbols of the output symbol sequence.

[0044] When converting an initial symbol, a correction factor or several correction factors can also be taken into account during training, as will be described below in relation to the use of the Kl model.

[0045] The method according to the invention offers the advantage that a high degree of compression can be achieved by replacing recurring symbol patterns in the time series used, thereby increasing the efficiency of data processing and / or reducing the required computing resources. At the same time, compressing the data based on recurring symbol patterns enables adaptive compression behavior that preserves recognizable patterns in the underlying time series. This ensures that such patterns remain visible to the AI ​​model, thus enabling improved training and ultimately higher predictive accuracy of the AI ​​model.

[0046] Within the scope of the invention, it can be provided that the number of data vectors corresponds to the number of main symbols and auxiliary symbols. In other words, it is conceivable that each symbol, i.e., each main symbol and each auxiliary symbol, is assigned exactly one data vector. Additionally or alternatively, it can be provided that at least one data vector is a multidimensional data vector, in particular a 512-dimensional data vector. Such data vectors allow the relationships between the respective symbols and the resulting meaning for the

[0047] Map the second vehicle parameter in a high-dimensional way in the vector database.

[0048] Minimizing the prediction error of the AI ​​model for the training time series can be achieved primarily through optimization. This optimization can specifically involve optimizing a lookup table (or coefficient table) that characterizes the translation of a symbol into a data vector. Such a lookup table is also referred to as a vector table or embedding table. In other words, the lookup table, or rather the coefficients it contains, is optimized to minimize the AI ​​model's prediction error for the training time series.

[0049] If the AI ​​model is used to predict a future value of the first vehicle parameter, the prediction error can be determined by the deviation of a value of the first vehicle parameter at a given time within the training time series (ground truth) from the prediction of the first vehicle parameter at that time generated by the AI ​​model. The prediction can be generated, in particular, based on a time interval of the training time series prior to the given time. If the AI ​​model is used to predict a second vehicle parameter, different from the first, the prediction error can be determined using the value of the second vehicle parameter (ground truth), which is known from the training and characteristic of the training time series.

[0050] In particular, with regard to the present invention, it is conceivable that at least individual steps of the method are performed repeatedly and / or, at least partially, simultaneously. It is particularly preferred that the training of the AI ​​model and / or the use of the AI ​​model and the respective associated steps are carried out for a multitude of training time series. The advantages of the method according to the invention become particularly evident during extensive training of the AI ​​model, since efficient data handling is achievable due to the high compression, without obscuring patterns contained in the training data for the AI ​​model.

[0051] In particular, the procedure may include the following:

[0052] - Using at least one predicted value of the second vehicle parameter to adapt the vehicle's operation. In particular, adapting the vehicle's operation may include at least one of the following:

[0053] - Activating or deactivating at least one driver assistance system, in particular an electronic stability program (ESP) or anti-lock braking system (ABS) and / or an automatic emergency braking system and / or brake assist and / or lane keeping assist,

[0054] - Activating, deactivating, or adjusting a comfort system, in particular interior ventilation and / or seat heating or the like, and / or

[0055] - Adjusting at least one chassis parameter and / or lighting parameter of the vehicle, in particular performing a level control of the vehicle body and / or performing a headlight range control.

[0056] To adapt the operation of the vehicle, at least one predicted value of the second vehicle parameter can preferably be communicated to at least one control unit of the vehicle, which preferably performs a control or adaptation of corresponding vehicle systems or vehicle functions depending on the at least one predicted value of the second vehicle parameter.

[0057] With regard to the present invention, it is possible that the first vehicle parameter is the same as the second vehicle parameter. In this case, the Kl model is used to predict a future value of the first vehicle parameter or the further course of a given time series of the first vehicle parameter. It is also possible that the second vehicle parameter is a different vehicle parameter than the first. In this case, the Kl model is consequently used to predict the correlated state of a second vehicle parameter based on the time series of a first vehicle parameter.

[0058] Preferably, the method according to the invention can also be applied to multivariate time series as input. Accordingly, the training time series or the input time series can be a time series of a plurality of first vehicle parameters.

[0059] This allows the dependence of a second vehicle parameter as an output variable on a plurality of first vehicle parameters as input variables to be represented. In particular, it can be provided in this context that the recognition of symbol patterns takes place not only along the temporal dimension of the time series of the first vehicle parameters, but also between the respective first vehicle parameters for a given point in time.Within the scope of the invention, a vehicle parameter can be, in particular, a parameter characteristic of a vehicle's movement, a chassis parameter, or a vital parameter of a vehicle occupant, especially a wheel speed, an acceleration, particularly lateral or longitudinal acceleration, of the vehicle, and / or a speed, particularly longitudinal speed, of the vehicle, a steering angle, an engine speed, a currently engaged gear, the height of at least one wheel of the vehicle, and / or a chassis tilt, a vital function of the driver, or the like. A vehicle parameter can also be, for example, a categorical or event-based state parameter of at least one vehicle component or vehicle system. Such a state parameter can, for example, characterize the activity of an assistance system such as ABS or ESP (active or inactive) or the status (open or closed) of a vehicle's seat belt buckle.

[0060] The method according to the invention can be used in a variety of ways in the context of a vehicle. For example, the method according to the invention can be used to predict the current vehicle speed based on a time series of the engine speed and the currently engaged gear. In this way, redundancy can be established with an additional speed sensor provided in the vehicle, in particular a physical one, and the latter can be monitored.

[0061] Additionally or alternatively, the method according to the invention can be used to infer the distribution of vehicle occupants from the number and distribution of closed seat belt buckles in the vehicle, or to derive and implement necessary chassis or lighting adjustments within the framework of headlight range and / or level control. Additionally or alternatively, the method according to the invention can, for example, be used to infer a driver's fitness to drive from the time series of at least one vital function.

[0062] For example, the Kl model can depict a relationship between vehicle acceleration and the intervention of a vehicle assistance system such as ABS or ESP. The method according to the invention can thus be used to recognize, based on the temporal profile of the vehicle acceleration, the need to activate such an assistance system and to initiate appropriate actions.

[0063] Within the scope of the invention, it can be advantageous to select the number of predefined main symbols and / or the distribution of interval boundaries of the value intervals of the vehicle parameter characterized by the main symbols as a function of a discretization error resulting from the discretization. In particular, the number of main symbols and / or the distribution of the interval boundaries can be selected such that the discretization error is minimal. In other words, it can be provided that, for a given number of predefined main symbols and / or a given distribution of the respective interval boundaries, a discretization error between the discretized training or input time series and the actual training or input time series is determined, and the number of main symbols and / or the distribution of the interval boundaries is adjusted with the aim of minimizing the discretization error. This can, for example,This may be the subject of optimization. It may be stipulated that the number of main symbols is limited to a maximum number (main symbol limit).

[0064] Additionally or alternatively, the number of predefined main symbols and / or the distribution of interval boundaries for the vehicle parameter value intervals characterized by these main symbols can be selected such that a predefined minimum value range of the vehicle parameter is necessarily covered by the value intervals. This ensures that a relevant range of the vehicle parameter is always represented by the value intervals, with the respective granularity in individual ranges being selected to minimize the discretization error.

[0065] Within the scope of the invention, it is conceivable that a distribution of interval limits of the value intervals of the vehicle parameter characterized by the main symbols is possible.

[0066] - a uniform distribution is,

[0067] - is specified as a normal distribution around an expected value of the vehicle parameter determined from the training time series or

[0068] - corresponds or essentially corresponds to a distribution of the vehicle parameter in the training time series or

[0069] - is determined locally for a given time, depending on a given number of values ​​of the vehicle parameter from the training time series recorded chronologically before that time.

[0070] In this context, a uniform distribution means that each symbol characterizes an equally sized interval of values ​​for the vehicle parameter. This represents a particularly simple and efficient distribution of the interval boundaries.

[0071] To distribute the interval boundaries according to a normal distribution (Gaussian distribution), the expected value, particularly as the time mean, as well as the standard deviation or variance of the vehicle parameter are determined from the training or input time series. A corresponding normal distribution can then be created based on these values. According to the structure of a normal distribution, the vehicle parameter exhibits narrower intervals of values ​​in the region of the expected value, which widen with increasing distance from the expected value. This results in a more nuanced distribution of the input or input time series.

[0072] The training time series is more highly resolved in the area around the expected value during discretization, and the discretization error can be reduced.

[0073] Alternatively, the distribution of the interval boundaries can be adapted to the distribution of the vehicle parameter in the training or input time series. This allows for consideration of a concentration of the vehicle parameter values ​​in specific ranges. Specifically, this approach involves first fitting a distribution function to the distribution of the vehicle parameter in the time series under consideration and then using this distribution function to determine the distribution of the interval boundaries.

[0074] It is also conceivable that the distribution of the interval limits for a given time point is determined locally along the training or input time series, depending on at least one chronologically prior value of the vehicle parameter, in particular depending on a distribution of a given number of chronologically prior values ​​of the vehicle parameter from the training or input time series.

[0075] Input time series. In other words, for a given point in time along the considered time series, a predetermined number of values ​​of the vehicle parameter prior to that point in time are examined. For this given number of values, a distribution of the vehicle parameter is determined, and this distribution is used at the given point in time to determine the distribution of the interval boundaries. The distribution of the interval boundaries thus varies over time and is locally adapted to the respective course of the vehicle parameter in the input or training time series. This approach has proven particularly effective in minimizing the discretization error. If an output symbol from the AI ​​model needs to be traced back to a time series value (e.g., a future value) using the value intervals characterized by the symbols, the local distributions of the interval boundaries must be taken into account accordingly.

[0076] Within the scope of the invention, it may be provided that determining a compressed training time series includes, in particular iteratively, the construction of an additional symbol vocabulary, wherein in particular the construction of the additional symbol vocabulary includes at least one of the following:

[0077] - Determining a most frequently occurring symbol pairing, in particular as a recurring symbol pattern, wherein the symbol pairing comprises two adjacent symbols, and / or - Defining an additional symbol characteristic of the most frequently occurring symbol pairing and / or

[0078] In particular, it may be provided that the replacement of recurring symbol patterns by a, in particular iterative, replacement of recurring symbol patterns includes at least one of the following:

[0079] - Identifying a most frequently occurring symbol pairing, in particular as a recurring symbol pattern, wherein the symbol pairing comprises two adjacent symbols, and / or

[0080] - Replacing the symbol pairing with an additional symbol characteristic of the symbol pairing.

[0081] In other words, it may be planned that, for the construction of the additional symbol vocabulary, i.e., the set of additional symbols used, a most frequently occurring symbol pairing is first determined, particularly based on the discretized training time series. The symbol pairing comprises two adjacent symbols, i.e., either

[0082] Two main symbols, or one main symbol and one additional symbol, or two additional symbols. Within the scope of the invention, adjacent symbols are understood to be symbols that are positioned directly next to each other. Thus, no further symbol is located between two adjacent symbols. For the most frequently occurring symbol pairing, an additional symbol is then defined, in particular according to a predetermined logic, the meaning of which subsequently characterizes the corresponding sequence of symbols.

[0083] In the process of replacing recurring symbol patterns, the most frequently occurring symbol sequence, particularly in a given step, is preferentially replaced by the defined additional symbol, thereby compressing the time series without any loss of information. This procedure is preferably performed iteratively, so that successively recurring symbol patterns are replaced by additional symbols, thus increasing the degree of compression.

[0084] Furthermore, it is conceivable that the development of the additional symbol vocabulary is completed when the number of defined additional symbols reaches or exceeds a predefined limit. In other words, it may be stipulated that the maximum number of additional symbols to be assigned is limited. This allows for the definition of a termination criterion for the development of the additional symbol vocabulary.

[0085] Additionally or alternatively, it can be stipulated that a supplementary symbol is only assigned to a most frequently occurring symbol pair if the frequency of the symbol pair reaches or exceeds a predefined frequency threshold. It is understood that in this case, the symbol pair is not replaced with a supplementary symbol. Depending on the length of the time series under consideration, the frequency threshold can, for example, be...

[0086] The frequency of a symbol pair can be 1000, 500, or 100. In other words, it can be stipulated that a symbol pair is only replaced with an additional symbol if its frequency reaches or exceeds a predefined threshold (frequency threshold). Otherwise, no additional symbol is assigned to that symbol pair. This approach allows a threshold to be defined at which frequency a recurring symbol pair is recognized as a relevant pattern and treated accordingly. In this way, compression can be focused on particularly prominent patterns in terms of frequency, and a corresponding termination criterion for building the additional symbol vocabulary can be defined.

[0087] It may be planned that the development of the additional symbol vocabulary is carried out for a predetermined number of training time series, representing a subset of all training time series used. For example, the predetermined number of training time series could be a maximum of 10%, 5%, or exactly 1% of all training time series used. The training time series used to develop the additional symbol vocabulary can be randomly selected from the total number of training time series used. This approach is based on the understanding that the relevant patterns repeat themselves in the training time series, and that the patterns to be considered can therefore be identified in a subset of the training data.For the remaining training time series and also for the processing of input time series, the already established additional symbol vocabulary can then be used, thus enabling more efficient data processing.

[0088] It is also conceivable that adapting the multidimensional data vector encompassed by the vector database of the Kl model includes the following:

[0089] - Adjusting the data vectors such that a distance between two data vectors, each representing an additional symbol, is proportional to a distance between the time series subsequences underlying the additional symbols, in particular determining the distance between the data vectors and / or the distance between the time series subsequences as a Euclidean distance.

[0090] This approach is based on the assumption that the Kl model can make qualitatively better predictions if substantively similar data, or their representation by data vectors of corresponding proximity to one another, are stored in the vector database. Each additional symbol characterizes a sequence of at least two main symbols. This sequence of main symbols can be translated back into a portion of the discretized time series, which is to be understood as a time series subsequence, according to the steps described within the scope of the present invention. The distance between two data vectors, or the distance between the respective time series subsequences, can preferably be determined as a Euclidean distance (L2 loss).

[0091] Within the scope of the invention, it is optionally possible that the use of the trained

[0092] The AI ​​model includes at least one of the following:

[0093] - Determining a discretized input time series from the input time series using the given main symbols,

[0094] - Determining a compressed input time series from the discretized input time series by replacing recurring symbol patterns with the additional symbol characteristic of the respective symbol pattern, particularly iteratively,

[0095] - Generating at least one predicted value of the second vehicle parameter or an output symbol by the Kl model as a function of the compressed input time series, in particular using the vector database, and / or - converting the output symbol into the at least one predicted value of the second vehicle parameter.

[0096] The mechanisms described in relation to the use of the AI ​​model can, where appropriate, be applied to training the AI ​​model and vice versa.

[0097] In other words, when using the Kl model, a discretized input time series is also derived from the input time series. The main symbols used or defined during training are employed for this purpose. Furthermore, a compressed input time series is derived from the discretized input time series. This is achieved by replacing recurring symbol patterns, particularly iteratively, with the additional symbol characteristic of each symbol pattern. The recognizable symbol patterns and their corresponding additional symbols are known from the training of the Kl model or the additional symbol vocabulary developed during training.

[0098] Using the compressed input time series, a predicted value of the second vehicle parameter can be generated as an output variable by the Kl model.

[0099] In particular, using the compressed input time series, an output symbol can first be generated or provided as an output variable by the Kl model, where the output symbol is characteristic of at least one predicted value of the second vehicle parameter. This occurs particularly when the first vehicle parameter is equal to the second vehicle parameter, i.e., when the Kl model is used to predict a future value of the first vehicle parameter. The output symbol can be a primary symbol or an additional symbol and characterizes the further temporal progression of the input time series predicted by the Kl model. If the Kl model generates an output symbol as an output variable, it can preferably be provided that this output symbol is converted into at least one value of the second vehicle parameter.To convert the output symbol, the mechanisms described in the invention for processing a training or input time series are applied in reverse.

[0100] Furthermore, it may be provided within the scope of the invention that the conversion of the output symbol into the at least one predictive value of the second vehicle parameter comprises at least one of the following:

[0101] - Reducing the output symbol to a source symbol sequence if the output symbol is a supplementary symbol, where the source symbol sequence consists exclusively of main symbols, and / or

[0102] - Reducing the output symbol sequence or output symbol to at least one predicted value of the second vehicle parameter using the value intervals of the first vehicle parameter corresponding to the main symbols of the output symbol sequence.

[0103] In other words, it may be possible to convert the output symbol by first reducing it to an output symbol sequence if the output symbol is an auxiliary symbol. In other words, the auxiliary symbol is replaced by the symbol pattern for which it is characteristic. If the symbol pattern contains another auxiliary symbol, this is also replaced until the resulting output symbol sequence consists exclusively of primary symbols. For each primary symbol in the output symbol sequence, or the output symbol generated by the AI ​​model, a specific value of the second vehicle parameter can then be determined using the value intervals of the first vehicle parameter assigned to the respective primary symbols. In the case of an output symbol sequence, the AI ​​model thus provides a number of predicted values ​​for the second vehicle parameter corresponding to the number of primary symbols in the output symbol sequence.In particular, preprocessing of the input time series must be taken into account when converting the output symbol sequence or an output symbol back to at least one predicted value of the second vehicle parameter. Specifically, it may be provided that using the trained AI model includes creating at least one modified input time series based on the input time series (sampling the input time series). Modifying the input time series may, for example, include adding noise. Generating at least one predicted value of the second vehicle parameter can be performed, in particular, for the input time series and each modified input time series. Preferably, the respective predictions of the AI ​​model for the input time series and the at least one modified input time series can be averaged to further improve the prediction quality of the AI ​​model.

[0104] In particular, a probabilistic forecast can also be generated from the various prediction values ​​that were additionally determined on the basis of the modified input time series(s).

[0105] Preferably, with regard to the present invention, it may be provided that the reduction of an output symbol sequence to at least one predictive value of the second vehicle parameter comprises the following:

[0106] - Taking into account a correction factor when reducing at least one main symbol of the original symbol sequence to a predicted value of the

[0107] second vehicle parameter, wherein in particular the correction factor is determined depending on the main symbol returned and / or at least one main symbol preceding the main symbol returned in the output symbol sequence.

[0108] At least one correction factor can be additive or multiplicative. Correction factors can be stored in a lookup table. In particular, one correction factor and / or multiple correction factors can be stored for each main symbol. For example, a correction factor can be stored for each combination of one main symbol with another. With 20 main symbols, this would result in...

[0109] 20 2 or 400 correction factors. In this way, the correction factor for a main symbol to be derived can be determined depending on the main symbol preceding it in the original symbol sequence.

[0110] In particular, correction factors can be determined as part of an optimization process to minimize the prediction error of the AI ​​model. This can be done, for example, when new training data is available and a complete retraining of the AI ​​model is to be avoided. Such use of correction factors thus allows the AI ​​model's predictive behavior to be adjusted without having to perform entirely new training. Furthermore, it is conceivable that preprocessing to determine a preprocessed time series from the training time series and / or the input time series is also included, wherein the preprocessing comprises at least one of the following:

[0111] - Filtering the training time series and / or the input time series, in particular with a low-pass filter, and / or smoothing the training time series and / or the input time series using hysteresis,

[0112] - Normalizing the training time series and / or the input time series, in particular to a time mean and / or

[0113] - Determining the first and / or second and / or higher derivative of the training time series and / or the input time series,

[0114] - Converting the training time series and / or input time series into an equidistantly sampled (pre-processed) time series, in particular by interpolation,

[0115] where the pre-processed time series is used as the basis for determining the discretized training time series and / or the discretized input time series.

[0116] In other words, it may be possible to preprocess a training time series and / or an input time series before discretization. The goal of this preprocessing is, in particular, to reveal similarities between seemingly dissimilar time series and thus to treat patterns that are present in all the time series equally.

[0117] For example, two time series of a vehicle parameter may differ essentially only by an offset and / or noise. Further differences may lie in a linear or amplitude scaling, a linear drift, or a shift along the time axis. Nevertheless, both time series may contain similar patterns that the AI ​​model should recognize as such. For example, angle sensors in vehicles may have an offset due to their installation position. This is irrelevant to the sensor's function, as changes in angle can be easily detected with the desired accuracy. However, a time series from such an angle sensor will include this offset compared to one from an angle sensor without an offset.

[0118] Filtering a time series with a low-pass filter removes high-frequency noise. Smoothing using hysteresis has a similar effect, whereby changes in a vehicle parameter are only recognized or accepted as such when they reach or exceed a threshold value relative to a reference value. Normalizing a time series can correct offsets or amplitude scaling. The same applies if the first derivative of the time series is used instead of the actual time series. When the first derivative is calculated, it is understood that the pre-processed time series provided is the first derivative of the vehicle parameter. The same applies to higher-order derivatives. If a time series exhibits linear drift, this effect can be compensated for by calculating the second or a higher-order derivative.

[0119] Additionally or alternatively, the input time series can be converted into an equidistantly sampled time series. This can be achieved, in particular, by interpolating corresponding sample points if the input time series was originally a non-equidistantly sampled time series. An equidistantly sampled time series is defined as a time series whose sample points are distributed at equidistant time intervals. In this way, after interpolation, the methodology presented within the scope of the invention can also be applied to time series that were originally not equidistantly sampled.

[0120] If an output symbol from the AI ​​model needs to be traced back to a time series value (e.g., a future value) using the value intervals characterized by the symbols, the type of preprocessing used must be taken into account accordingly. For example, if the first derivative of the vehicle parameter over time is provided as a preprocessed time series, then conversely, integration is required to derive a value of the vehicle parameter from the AI ​​model's output.

[0121] Within the scope of the invention, it can be advantageous for preprocessing to be represented by at least one preprocessing symbol in the preprocessed time series, wherein the preprocessing symbol is characteristic of the type of preprocessing. A preprocessing symbol thus represents a kind of additional symbol specifically reserved for identifying preprocessing. By inserting a preprocessing symbol into the preprocessed time series, information about whether and what type of preprocessing has taken place can be made accessible to the AI ​​model. Accordingly, offsets, noise, scaling, and the like can be taken into account during the training of the AI ​​model and also in its subsequent application.

[0122] Preprocessing symbols can be inserted into the time series sequence as independent additional symbols. Alternatively, a main symbol or an additional symbol can simultaneously be characteristic of a completed preprocessing. This ensures that the degree of compression of the input or training time series is not affected by preprocessing symbols. Within the scope of the invention, the Kl model can be implemented as a transformer Kl model, a state-space model, or a recurrent neural network, and known model architectures can be used.

[0123] The application of the present invention in connection with transformer-Kl models has proven particularly advantageous. A transformer model can, in particular, comprise an embedding layer and at least one encoder module and / or a decoder module. Several encoder modules and / or decoder modules, especially those connected in series, can be provided. In particular, the encoder modules can have the same or substantially the same architecture. The same applies to decoder modules. The embedding layer serves, in particular, to convert input data into a vector representation. The present invention thus relates specifically to the embedding layer. For the architecture of the respective encoder and / or decoder modules or the embedding layer, known architectures can generally be used.

[0124] Training the AI ​​model can include training additional modules of the AI ​​model. For a transformer AI model, this can include, for example, training at least one encoder module and / or decoder module. Training can be particularly useful for minimizing the AI ​​model's prediction error over a training time series. This can involve, for example, assigning weights to neural networks of at least one encoder or decoder module.

[0125] The decoder module will be adapted as part of the training.

[0126] According to a second aspect, the present invention further relates to a data processing system comprising means for carrying out a computer-implemented method according to the first aspect of the present invention. The means may, in particular, comprise at least one processor and / or at least one main memory and / or at least one non-volatile main memory. Thus, with respect to a system according to the invention, the same advantages arise as those already described with respect to a computer-implemented method according to the invention.

[0127] In particular, it can be provided that the training of the AI ​​model in the manner described within the scope of the invention and the use of the trained AI model in the manner described within the scope of the invention are carried out on different data processing systems. For example, the training of the AI ​​model can be carried out independently of a vehicle and the use of the AI ​​model can be carried out within a vehicle.

[0128] According to a third aspect, the present invention further relates to a vehicle. Preferably, the vehicle is configured to be operated according to a computer-implemented method according to the first aspect of the present invention and / or the vehicle includes a data processing system according to the second aspect of the present invention. This provides the same advantages with respect to a vehicle according to the invention as have already been described with respect to a computer-implemented method and / or a system according to the invention.

[0129] According to a fourth aspect, the present invention further relates to a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to execute a computer-implemented method according to the first aspect of the present invention. This provides the same advantages with respect to a computer program product according to the invention as have already been described with respect to a computer-implemented method and / or a system and / or a vehicle according to the invention.

[0130] According to a fifth aspect, the present invention relates to a computer-readable storage medium on which a computer program product according to the fourth aspect of the present invention is stored. This provides the same advantages with respect to a computer-readable storage medium according to the invention as have already been described with respect to a computer-implemented method and / or a system and / or a vehicle and / or a computer program product according to the invention.

[0131] According to a sixth aspect, the present invention further relates to a data carrier signal that transmits a computer program product according to the fourth aspect of the present invention. This provides the same advantages with respect to a data carrier signal according to the invention as have already been described with respect to a computer-implemented method and / or a system and / or a vehicle and / or a computer program product and / or a computer-readable storage medium according to the invention.

[0132] Further advantages, features, and details of the invention will become apparent from the following description, in which several embodiments of the invention are described in detail with reference to the drawings. The features mentioned in the claims and in the description can each be essential to the invention individually or in any combination. Figure 1 shows a schematic view of a method.

[0133] Fig. 2 shows a schematic view of a process,

[0134] Fig. 3 shows a schematic view of a time series,

[0135] Fig. 4 shows a schematic view of a discretized time series,

[0136] Fig. 5 shows a schematic view of a distribution of interval boundaries,

[0137] Fig. 6 shows a schematic view of a distribution of interval boundaries,

[0138] Fig. 7 shows a schematic view of a distribution of interval boundaries,

[0139] Fig. 8 shows a schematic view of a compressed time series,

[0140] Fig. 9 shows a schematic view of a process,

[0141] Fig. 10 shows a schematic view of a time series and

[0142] Fig. 11 is a schematic view of a vehicle.

[0143] The figures use identical reference numerals for the same technical features, even for different embodiments.

[0144] Fig. 1 shows a method 100 for generating a prediction of a vehicle parameter by a Kl model, wherein the Kl model has a functional relationship between a time series Z of a first vehicle parameter Pi as input and a

[0145] The second vehicle parameter P2 is mapped as an output variable using a vector database, comprising the procedure 100:

[0146] - Training 110 of the Kl model using at least one training time series ZT of the first vehicle parameter Pi and

[0147] - Use 120 of the trained Kl model to generate at least one prediction value of the second vehicle parameter P2 as a function of an input time series ZE of the first vehicle parameter Pi.

[0148] Fig. 2 shows a more detailed representation of the training 110 of the Kl model, wherein the training 110 of the Kl model comprises the following: - Determining 110.1 a discretized training time series ZT,D from the training time series ZT using a plurality of predefined principal symbols Si , wherein each principal symbol Si is characteristic of a value interval I of the first vehicle parameter Pi,

[0149] - Determining 110.2 a compressed training time series ZT,C from the discretized training time series ZT,D by replacing, in particular iteratively, recurring symbol patterns M with an additional symbol S2 characteristic of the respective symbol pattern M, and

[0150] - Adjusting 110.3 of multidimensional data vectors encompassed by the vector database of the Kl model as a function of the compressed training time series ZT,C to minimize a prediction error of the Kl model for the training time series ZT.

[0151] Fig. 3 shows a schematic view of a time series Z of a first vehicle parameter Pi. It can be seen that the time series Z characterizes the course of the first vehicle parameter Pi over time t. In this case, the time series Z is a training time series ZT, from which a discretized training time series ZT,D is determined according to the procedure shown in Fig. 4.

[0152] To determine the discretized training time series ZT.D, a plurality of predefined principal symbols Si are used, where each principal symbol Si is characteristic of a value interval I of the first vehicle parameter Pi. The distribution of the interval limits IG shown in Fig. 4 is a uniform distribution.

[0153] The digits from 0 to 9 are defined as the primary symbols Si. To determine the discretized training time series ZT,D, the system checks, for equidistantly distributed time points T along the training time series ZT, the interval I in which the first vehicle parameter Pi lies, and assigns the primary symbol Si corresponding to the determined interval I to each time point T. This procedure is shown purely schematically. In particular, the number of primary symbols Si and / or the distribution of the interval boundaries IG can be adjusted to minimize the discretization error resulting from the described procedure.

[0154] Fig. 5 schematically shows a distribution of the interval limits IG that differs from Fig. 3, where the distribution of the interval limits IG is defined as a normal distribution around an expected value of the first vehicle parameter Pi determined from the training time series ZT. Furthermore, Fig. 6 shows another distribution of the interval limits IG, where the distribution of the interval limits IG corresponds to, or essentially corresponds to, the distribution of the first vehicle parameter Pi in the training time series ZT.

[0155] Fig. 7 shows a further distribution of the interval limits IG, where the distribution for a given time T is determined locally along the training time series ZT, depending on a predetermined number of chronologically recorded values ​​of the first vehicle parameter Pi from the training time series ZT before time T.

[0156] Fig. 8 shows a schematic representation of a discretized training time series ZT,D, which is subsequently converted into a compressed training time series ZT,C. Here, too, the digits 0 to 9 are used as the main symbols Si. To determine the compressed training time series ZT,C, recurring symbol patterns M are replaced by characteristic additional symbols S2. In an iterative process, the most frequently occurring symbol pairing is determined in each step. In the first step, shown in Fig. 7, this is the symbol pairing of symbols 3 and 6, which is replaced by the additional symbol 10. In the next step, the most frequently occurring symbol pairing is again determined as a recurring symbol pattern M, which then consists of symbols 10 and 4. This symbol pattern M is replaced accordingly with the additional symbol 11.

[0157] The compression is then complete, as there are no further symbol pairs that occur multiple times.

[0158] The determined compressed training time series ZT.c can subsequently be used to fit at least one multidimensional data vector encompassed by the vector database of the Kl model in order to minimize a prediction error of the Kl model with respect to the training time series ZT ZU underlying the compressed training time series ZT,C.

[0159] Fig. 9 further shows a more detailed representation of the use 120 of the Kl model to generate at least one predictive value of the second vehicle parameter P2, wherein the use 120 of the Kl model comprises the following:

[0160] - Determine 120.1 a discretized input time series ZE,D from the input time series ZE using the given main symbols S1,

[0161] - Determining 120.2 a compressed input time series ZE, c from the discretized input time series ZE,D by replacing, in particular iteratively, recurring symbol patterns M with the one for the respective symbol pattern M

[0162] characteristic additional symbol S2 and

[0163] - Generate 120.3 an output symbol S1 , S2 by the Kl model as a function of the compressed input time series ZE,C,- Convert 120.4 the output symbol Si , S2 into at least one prediction value of the second vehicle parameter P2.

[0164] In this way, the method 100 presented within the scope of the invention can be used such that the second vehicle parameter P2 corresponds to the first vehicle parameter Pi, and thus at least one future value of the first vehicle parameter Pi, i.e., a further progression of the input time series ZE, is determined by the Kl model. This further progression of the input time series ZE is shown schematically in Fig. 10 by the dashed section.

[0165] The determination of the discretized input time series 120.1 is carried out analogously to the previously described determination of the discretized training time series ZT,D 110.1. Similarly, the determination of the compressed input time series 120.2 is carried out analogously to the previously described determination of the compressed training time series ZT,C 110.2. To convert the output symbol S1, S2 into a prediction value, the procedure described with reference to Fig. 8 and / or Fig. 4 is applied in reverse, taking into account any preprocessing of the input time series ZE.

[0166] Fig. 11 further shows a schematic view of a vehicle 200. The vehicle 200 comprises a data processing system 50, which includes means for carrying out a computer-implemented method 100 according to the first aspect of the present invention.

[0167] The preceding explanation of the embodiments describes the present invention solely by way of examples. Naturally, individual features of the embodiments can be freely combined with one another, provided this is technically feasible, without departing from the scope of the present invention. List of reference numerals

[0168] 50 System

[0169] 100 procedures

[0170] 110 Training

[0171] 110.1 Determine

[0172] 110.2 Determine

[0173] 110.3 Adjust

[0174] 120 Use

[0175] 120.1 Determine

[0176] 120.2 Determine

[0177] 120.3 Generate

[0178] 120.4 Convert

[0179] 200 vehicles

[0180] I Value interval

[0181] IG interval limit

[0182] M Symbol pattern

[0183] Pi is the first vehicle parameter

[0184] P2 second vehicle parameter

[0185] 51 Main symbol

[0186] 52 Additional symbol

[0187] T Time

[0188] t time

[0189] Z time series

[0190] ZE Input time series

[0191] ZT Training Time Series

[0192] ZT,D discretized training time series

[0193] ZT,C compressed training time series

Claims

Patent claims 1. Computer-implemented method (100) for generating a prediction of a vehicle parameter by a Kl model, wherein the Kl model maps a functional relationship between a time series (Z) of a first vehicle parameter (Pi) as input and a second vehicle parameter (P2) as output using a vector database, the method (100) comprising: - Training (110) of the KL model using at least one training time series (T) of the first vehicle parameter (Pi) and - Using (120) the trained Kl model to generate at least one prediction value of the second vehicle parameter (P2) as a function of an input time series (ZE) of the first vehicle parameter (Pi), where training (110) the Kl model includes the following: - Determining (110.1) a discretized training time series (ZT,D) from the training time series (ZT) using a plurality of predefined principal symbols (Si), where each principal symbol (Si) is characteristic of a value interval (I) of the first vehicle parameter (Pi), - Determining (110.2) a compressed training time series (ZT,C) from the discretized training time series (ZT,D) by replacing, in particular iteratively, recurring symbol patterns (M) with an additional symbol (S2) characteristic of the respective symbol pattern (M) and - Fitting (110.3) multidimensional data vectors encompassed by the vector database of the Kl model as a function of the compressed training time series (ZT,C) to minimize a prediction error of the Kl model for the training time series (ZT).

2. Computer-implemented method (100) according to claim 1, characterized by that the number of specified main symbols (Si) and / or a distribution of interval limits (IG) of the value intervals (I) of the vehicle parameter (P) characterized by the main symbols (Si) is chosen depending on a discretization error resulting from the discretization.

3. Computer-implemented method (100) according to one of the preceding claims, characterized in that a distribution of interval limits (IG) of the value intervals (I) of the vehicle parameter (P) characterized by the main symbols (Si) - is specified as a normal distribution around an expected value of the first vehicle parameter (Pi) determined from the training time series (ZT), or - corresponds to or essentially corresponds to a distribution of the first vehicle parameter (Pi) in the training time series (ZT), or - is determined locally for a given time (T) along the T training time series (ZT), depending on a distribution of a given number of chronologically recorded values ​​of the first vehicle parameter (Pi) from the training time series (ZT) prior to time (T).

4. Computer-implemented method (100) according to one of the preceding claims, characterized in that that the determination (110.2) of a compressed training time series includes, in particular iteratively, the construction of an additional symbol vocabulary, encompassing: - Determining a most frequently occurring symbol pairing, where the symbol pairing (SP) comprises two adjacent symbols (Si , S2), and - Defining an additional symbol (S2) characteristic of the most frequently occurring symbol pairing.

5. Computer-implemented method (100) according to claim 4, characterized by that the development of the additional symbol vocabulary is completed when the number of defined additional symbols (S2) reaches or exceeds a predetermined additional symbol limit and / or that for a most frequently occurring symbol pairing (SP) an additional symbol (S2) is only defined if the frequency of the symbol pairing (SP) reaches or exceeds a predetermined frequency limit.

6. Computer-implemented method (100) according to one of the preceding claims, characterized in that that the fitting (110.3) of the multidimensional data vector encompassed by the vector database of the Kl model includes the following: - Adjusting the data vectors so that a distance between two data vectors, each representing an additional symbol (S2), is proportional to a distance between the time series subsequences underlying the additional symbols (S2), wherein in particular the distance between the data vectors and / or the distance between the time series subsequences is determined as a Euclidean distance.

7. Computer-implemented method (100) according to one of the preceding claims, characterized in that that using (120) the trained AI model includes at least one of the following: - Determining (120.1) a discretized input time series from the input time series (ZE) using the specified main symbols (Si), - Determining (120.2) a compressed input time series from the discretized input time series by replacing, in particular iteratively, recurring symbol patterns (M) with the additional symbol (S2) characteristic of the respective symbol pattern (M) and - Generating (120.3) at least one predictive value of the second vehicle parameter (P2) or an output symbol (S1 , S2) by the Kl model depending on the compressed input time series and / or - converting (120.4) the output symbol (S1 , S2) into the at least one predicted value of the second vehicle parameter (P2).

8. Computer-implemented method (100) according to claim 7, characterized by that the conversion of the output symbol (S1 , S2) into the at least one predicted value of the second vehicle parameter (P2) includes at least one of the following: - Reducing the output symbol (S1 , S2) to an output symbol sequence if the output symbol (S1 , S2) is an additional symbol (S2), where the output symbol sequence consists exclusively of main symbols (S1), - Reducing the output symbol sequence or output symbol (S1, S2) to at least one predicted value of the second vehicle parameter (P2) using the value intervals (I) of the first vehicle parameter (Pi) corresponding to the main symbols (S1) of the output symbol sequence.

9. Computer-implemented method (100) according to claim 8, characterized by that reducing an output symbol sequence to at least one predicted value of the second vehicle parameter (P2) includes the following: - Taking into account a correction factor when converting at least one main symbol of the output symbol sequence to a predicted value of the second vehicle parameter, wherein in particular the correction factor is determined as a function of the converted main symbol (Si) and / or at least one main symbol (Si) preceding the converted main symbol (Si) in the output symbol sequence.

10. Computer-implemented method (100) according to one of the preceding claims, characterized in that that furthermore, preprocessing to determine a preprocessed time series (Zv) from the training time series (ZT) and / or the input time series (ZE) is included, wherein the preprocessing includes at least one of the following: - Filtering the training time series (T) and / or the input time series (I), in particular with a low-pass filter and / or smoothing the training time series (T) and / or the input time series (I) using hysteresis, - Normalizing the training time series (TS) and / or the input time series (IT), in particular to a time mean and / or - Determining the first and / or second derivative of the training time series (TS) and / or the input time series (IT), - Converting the training time series (ZT) and / or the input time series (ZE) into an equidistantly sampled time series, in particular by interpolation, whereby the preprocessed time series (Zv) is used as the basis for determining the discretized training time series (ZT,D) and / or the discretized input time series.

11. Computer-implemented method (100) according to claim 10, characterized by that preprocessing has taken place is represented by at least one preprocessing symbol (S3) in the preprocessed time series (Zv), where the preprocessing symbol (S3) is characteristic of the type of preprocessing.

12. System (50) for data processing, comprising means for carrying out a computer-implemented method (100) according to any one of the preceding claims.

13. Computer program product comprising instructions which, when the program is executed by a computer, cause the computer to execute a computer-implemented method (100) according to any one of claims 1 to 11.

14. Computer-readable storage medium on which a computer program product according to the preceding claim is stored.

15. Data carrier signal that transmits a computer program product according to claim 13.

Images