Method for pattern recognition in drive cycle

Abstract

The invention relates to a method for pattern recognition in a drive cycle. The method herein includes (a) selecting an indefinite number of samples from a plurality of sensor datasets, wherein each sensor dataset represents a drive cycle. (b) performing unsupervised pattern recognition within the indefinite number of samples to identify a pattern is then performed. (c) performing supervised pattern recognition in the remaining sensor datasets of the plurality of sensor datasets to re-recognize a pattern and (d) determining, for each sensor dataset of the plurality of sensor datasets, a correspondence of the sensor dataset to the pattern. In this case, steps (a) to (d) are repeated until each correspondence exceeds a correspondence threshold.

Description

Technical Field

[0001] The present invention relates to a method for detecting patterns in driving cycles, in particular recurring patterns in sensor data sets. Background Art

[0002] Vehicle manufacturers already collect, or at least will in the near future, vast amounts of data from customer vehicles. The goal is to use this data, for example, to design and develop next-generation vehicles that are optimally tailored to customer needs while also being cost-effective to produce. This requires the ability to process large amounts of data on distributed and highly scalable computing systems (e.g., AWS, Azure).

[0003] A common approach for managing large amounts of data is unsupervised pattern recognition in multivariate time series. This approach can, for example, identify recurring traffic and driving situations in sensor datasets without requiring knowledge of their actual behavior.

[0004] This method allows the sensor data set to be broken down into recurring situations. Based on the frequency of these situations (or patterns), it is then possible to construct a driving cycle that is representative of the customer as a whole and serves as input for vehicle design and testing.

[0005] Basically, methods for pattern recognition are known.

[0006] A method for recognizing driving patterns is known from the scientific literature by M. SIAMI, M. NADERPOUR and J. LU “A Mobile Telematics Pattern Recognition Framework for Driving Behavior Extraction” IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 3, pp. 1459-1472, March 2021, doi: 10.1109 / TITS.2020.2971214.

[0007] US 2022 / 0242451 A1 relates to detecting malicious events in autonomous vehicles. To this end, recorded events are compared with normal pattern scenarios to identify deviations. If the deviation is large enough, a malicious event is identified. Normal patterns can be detected using supervised or unsupervised machine learning.

[0008] Search algorithms for unsupervised pattern recognition, for example based on the “Sequitur” algorithm or the “PatternEnumeration” algorithm, are known from the paper “Battery State Estimation of Electric Vehicles using Neural Networks”, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), 2022.

[0009] Document US 2002 / 0194159 A1 relates to a system for parallel object-oriented data mining.

[0010] Document DE 10 2008 013 002 A1 describes a method and apparatus for observing and analyzing real-time data sets from multiple internet cameras, in particular those publicly accessible via the internet, for identifying atypical situations and / or critical structures and reporting the real-time data sets identified thereby to a control instance. Metadata from the video images is evaluated, including, for example, the density of human mass, the distribution of human mass density, and the direction and speed of movement of human mass. Atypical situations and / or critical structures are identified primarily by comparing actual object mass characteristics with permitted object mass characteristics, for example, by querying a database for atypical situations and / or critical structures.

[0011] Disadvantages of previously known pattern recognition methods include the fact that cloud-based pattern recognition only works on known patterns, thus preventing unsupervised pattern recognition. Furthermore, previous unsupervised pattern recognition methods can be limited to serial application, meaning that the datasets to be examined can only be processed sequentially or on local applications. Consequently, only relatively small datasets can be examined. Finally, there are no cloud-based solutions for parallelizing unsupervised pattern recognition. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to provide a method that at least partially overcomes the above-mentioned disadvantages.

[0013] This object is achieved by the method according to the invention according to claim 1. Further advantageous embodiments of the invention are apparent from the dependent claims and the following description of preferred exemplary embodiments of the invention.

[0014] One aspect of the present invention relates to a method for pattern recognition in a driving cycle, comprising the following steps:

[0015] (a) Selecting an indefinite number of samples from multiple sensor datasets, where each sensor dataset represents a driving cycle;

[0016] (b) performing unsupervised pattern recognition to identify patterns within a variable number of samples;

[0017] (c) performing supervised pattern recognition on remaining sensor data sets in the plurality of sensor data sets to re-identify the pattern; and

[0018] (d) determining, for each sensor data set of the plurality of sensor data sets, a degree of correspondence between the sensor data set and the pattern;

[0019] Steps (a) to (d) are repeated until each correspondence exceeds a correspondence threshold.

[0020] The method can be executed on a distributed and highly scalable (cloud-based) computing system (AWS, Azure...).

[0021] For example, multiple sensor datasets can be stored in a so-called "data lake." A data lake is a repository or system used in information technology for data storage. Sensor datasets can exist as measurement files in structured, semi-structured, or unstructured formats. The data lake can be stored on a server or in the cloud.

[0022] In this case, each sensor data set represents a driving cycle or a drive of the vehicle. The sensor data sets can come from different vehicles.

[0023] A sensor data set comprises a (if necessary time-resolved) sensor data stream, or in other words, the values ​​of one or more sensors detected and provided as a time series. In addition, a "sensor data set" should also include a data stream determined by a (vehicle-side) control device based on the provided sensor data. If the time series includes the values ​​of a single sensor, there is a univariate time series. If, on the other hand, there are values ​​of multiple sensors in the time series, there is a multivariate time series. Various vehicle measurement data (e.g. speed, acceleration, vehicle battery temperature, vehicle battery voltage and / or current, vehicle battery charge state, etc.) can be detected by the corresponding sensors. The sensors can be arranged on the vehicle side. Additionally or alternatively, sensors integrated in mobile terminal devices, such as smartphones, can also be used to detect vehicle measurement data.

[0024] In step (a), an indefinite number of samples are selected from a plurality of sensor data sets. Here, the indefinite number of samples correspond to samples used for unsupervised pattern recognition performed in step (b).

[0025] The initial selection of the variable number of samples can be performed randomly or based on a predetermined criterion. The predetermined criterion can be indicative of the relevance of the sensor dataset used for unsupervised pattern recognition, such as the duration / length of the drive cycle represented by the sensor dataset. For example, the sample can include the sensor dataset with the longest travel distance. Furthermore, other criteria for selecting samples that indicate the relevance of the sensor datasets can be used.

[0026] In step (b), unsupervised pattern recognition is performed within an indefinite number of samples (samples) to identify recurring patterns. A pattern comprises a temporal sequence of sensor signals / values ​​whose repetitions recur in the same or at least similar form within the sample. Repetition means that the sequence is present at least twice within the sample.

[0027] The pattern is indicative of the driving situation, traffic situation and / or operating situation of the vehicle. A recurring pattern can be understood as a similar recurring sequence in the sensor data set.

[0028] Unsupervised pattern recognition is performed by corresponding known unsupervised pattern recognition methods, which autonomously recognize such patterns. Such methods for unsupervised pattern recognition are based, for example, on the known Sequitur algorithm or “PatternEnumeration” algorithm.

[0029] An algorithm for unsupervised pattern recognition is applied to a sensor dataset of samples one at a time and serially. "Serially" means that the algorithm iterates through the sample from its start time point to its end time point. The pattern recognition algorithm identifies recurring patterns in the sample and creates a (global) dictionary from these recurring patterns and their definitions.

[0030] Algorithms used for unsupervised pattern recognition may require the preparation or adaptation of sample sensor datasets (sample datasets), i.e., the conversion of the sensor datasets into a form suitable for pattern recognition. For example, some algorithms require the sample datasets to exist as unique time series. This means that the time series of the sample datasets must be combined, in particular sequentially, into a unique time series in order to apply the pattern recognition algorithm to the sequential time series.

[0031] In some cases, a multivariate time series used as a sample dataset may need to be normalized, i.e., converted into a univariate time series. For example, a normalization algorithm can be used to determine a univariate time series from a multivariate time series. Common normalization algorithms are known to those skilled in the art.

[0032] In step (c), supervised pattern recognition is performed on the remaining sensor data sets in the plurality of sensor data sets to re-identify patterns.

[0033] Therefore, unsupervised pattern recognition is performed on all sensor datasets that are not part of the sample from multiple sensor datasets from the data lake.

[0034] Supervised pattern recognition aims to re-identify the patterns found in step (b). To do this, a previously created dictionary is used. In principle, the sensor dataset is compared with the dictionary.

[0035] Compared to unsupervised pattern recognition, re-identifying (discovered) patterns is significantly less complex in terms of computing time and memory and is therefore particularly suitable for parallelization. Pattern re-identification can therefore be performed in parallel. In other words, for each additional sensor data set, a corresponding worker can be set up for supervised pattern recognition. Due to the lower complexity of supervised pattern recognition, the workers used for this purpose can also have lower performance.

[0036] A worker is a program or part of a program that runs in its own thread or process. Multiple workers can run in parallel on multiple cores or processors. Multiple workers can be distributed, for example, across multiple servers with corresponding processors.

[0037] Known methods or algorithms can also be used for supervised pattern recognition.

[0038] In step (d), for each of the plurality of sensor data sets, a degree of correspondence between the sensor data set and the pattern is determined. The degree of correspondence indicates how well the pattern discovered in step (b) covers the sensor data set. In other words, the degree of correspondence indicates how much of the sensor data set can be explained by the discovered pattern. Therefore, the degree of correspondence represents a quality criterion for unsupervised pattern recognition. The degree of correspondence can be expressed, for example, as a percentage. The degree of correspondence can also be understood as coverage.

[0039] Steps (a) to (d) are repeatedly (iteratively) performed until the correspondence of each sensor data set exceeds a correspondence threshold, which may be, for example, a minimum of 85% or a minimum of 90%. Therefore, exceeding the correspondence threshold is a termination condition.

[0040] Alternatively, other termination conditions may be provided. For example, other termination conditions may include traversing steps (a) to (d) according to a predetermined number of loop iterations. Other termination conditions may include terminating the method if no new pattern is found during the traversal of steps (a) to (d).

[0041] The method according to the present invention aims to describe or cover all journeys (and therefore sensor data sets) as completely as possible (or at a corresponding threshold level) by the identified patterns. The method ends when this goal is achieved. However, since the presence of patterns depends on the driver and the driving environment, not every pattern is present in every sensor data set. If the unsupervised pattern recognition in step (b) has not yet identified a specific pattern (or specific situation), the supervised pattern recognition in step (c) will no longer be able to identify this pattern.

[0042] Steps (a) to (d) are thus repeated until each of the plurality of sensor data sets is sufficiently covered or described by the discoverable pattern. Thus, a particularly efficient pattern recognition method can be provided.

[0043] In another embodiment, an indefinite number of samples (samples) may be changed during the repetition of step (a). This means that the samples are changed during the second, third, fourth, etc. iterations of step (a). By changing the samples, an attempt is made to identify patterns in the current iteration that were not identified in the previous iterations of steps (a) to (d) so that the newly determined correspondence of the sensor data set meets the correspondence threshold.

[0044] Changing the sample can be performed in various ways. A first method may include supplementing the sample with an additional sensor dataset that was not previously included in the sample. As a further condition, the additional sensor dataset may have a correspondence less than a correspondence threshold and optionally have a significant travel length / distance. A second method may include selecting a completely new sample from the plurality of sensor datasets, wherein the sensor dataset of the new sample differs from the sensor dataset of the old sample.

[0045] Pattern recognition methods can be constructed particularly efficiently by combining an iterative process with changing samples.

[0046] In another embodiment, as described above, during repeated execution of step (a), an indefinite number of samples (samples) can be supplemented with one or more sensor datasets from the plurality of sensor datasets whose correspondence is less than a correspondence threshold. The sample is modified by supplementing the sensor datasets with a lower correspondence after the previous iteration of steps (a) to (d), i.e., below the correspondence threshold, and optionally with a specific travel distance. For example, a predetermined number of sensor datasets with the lowest correspondence can be selected. This attempts to generate samples for repeated execution of step (b) that fill in gaps in patterns not identified in the previous samples.

[0047] In another design, when step (a) is initially executed, an indefinite number of samples may be randomly selected from multiple sensor data sets.

[0048] In another embodiment, the supervised pattern recognition can be performed in parallel on the remaining sensor data sets in step (c). This means that multiple remaining sensor data sets can be examined in parallel during the supervised pattern recognition. To this end, each available worker is assigned one of the remaining sensor data sets, so that each worker performs supervised pattern recognition independently.

[0049] Therefore, relatively less complex supervised pattern recognition can be performed faster through parallelization, while the available computing power (i.e., workers) is also efficiently utilized.

[0050] Supervised pattern recognition can be parallelized by applying the corresponding pattern recognition algorithms to various data sets in parallel on cloud computing platforms (e.g., AWS, Azure, etc.). Cloud-based parallelization allows the method to scale particularly well with increasing data volumes.

[0051] In another design, step (a) may also include concatenating sensor data sets of an indefinite number of samples (sample data sets) into a unique data set. This means that multiple time series in the sensor data set are concatenated into a unique time series. The algorithm for unsupervised pattern recognition is applied once and serially to the consecutive time series of samples. By concatenating the sample data sets and thereby forming a relatively long time series, repeated patterns can be better identified compared to relatively short data sets. Therefore, unsupervised pattern recognition is improved by concatenating the sample data sets into a unique time series. In addition, compared to methods in which multiple data sets are examined simultaneously, the serial application of the algorithm up to a certain size of the consecutive time series is fast because there is less coordination work because only one unique algorithm is applied to the data set. In addition, the serial application of the algorithm is particularly simple compared to parallel application to multiple data sets. In addition, the subsequent merging of the results of the unsupervised pattern recognition performed in multiple data sets is eliminated.

[0052] In another embodiment, in step (b), unsupervised pattern recognition can be performed by a single worker. Here, the algorithm for unsupervised pattern recognition is executed by a single worker. In some examples, the worker with the highest performance can be selected from multiple workers for this purpose. This allows for relatively rapid execution of supervised pattern recognition.

[0053] In another design, step (b) may further include determining a global pattern dictionary based on unsupervised pattern recognition, wherein the global pattern dictionary defines the recognized patterns and is used for supervised pattern recognition.

[0054] In another embodiment, in step (b), unsupervised pattern recognition can be performed in parallel for sensor datasets from an indefinite number of samples. In other words, unsupervised pattern recognition is performed separately for each sensor dataset from the samples (sample datasets). This allows the sample datasets to be checked for patterns at least partially simultaneously.

[0055] This approach differs from designs where unsupervised pattern recognition is performed by only one worker. This means that all other workers are not needed at this stage, even though they are enabled and available for use. Therefore, all available workers are used here to perform unsupervised pattern recognition on a single sample dataset.

[0056] Similar to the parallelization of supervised pattern recognition, unsupervised pattern recognition can be parallelized by applying the corresponding pattern recognition algorithms (for unsupervised pattern recognition) to various data sets in parallel on cloud computing platforms (e.g., AWS, Azure, etc.). Cloud-based parallelization allows the method to scale particularly well with increasing data volumes.

[0057] Furthermore, in a design approach in which an algorithm for unsupervised pattern recognition is applied serially to successive sample data sets, it may happen that the time series generated by the successive sample data sets are too long for the algorithm. Therefore, in this case, unsupervised pattern recognition is parallelized.

[0058] Therefore, in another embodiment, step (b) can be performed by multiple workers. To this end, each available worker from the plurality of workers is provided with a sample data set, so that each worker performs supervised pattern recognition independently. Parallelization of unsupervised pattern recognition allows for efficient utilization of available computing power (i.e., workers).

[0059] In another embodiment, step (b) may further include determining a specialized pattern dictionary for each sensor data set of the variable number of samples based on unsupervised pattern recognition, wherein the specialized pattern dictionary defines the recognized patterns. Furthermore, the method may further include combining the specialized pattern dictionaries into a global pattern dictionary, wherein the global pattern dictionary is used to perform the supervised pattern recognition according to step (c).

[0060] The design approach is based on the fact that unsupervised pattern recognition is parallelized and executed by different workers. As a result, the individual unsupervised pattern recognition of the sample data set has a corresponding specialized pattern dictionary as a result. In order to use these results in the supervised pattern recognition in step (c), the specialized pattern dictionary (of the sample data set) is unified into or constitutes the above-mentioned global dictionary. For example, the specialized dictionary can contain consistent patterns. It is only included once in the global dictionary. Therefore, the global dictionary includes all patterns present in the specialized dictionary at one time. Therefore, the redundancy of the global dictionary in the specialized dictionary is eliminated.

[0061] In another embodiment, the above method may further include preparing a sensor dataset for unsupervised pattern recognition. This means preparing sensor signal values / data in the sensor dataset. For example, discretization of the values ​​detected by the sensor can be performed in a multivariate time series. In the case of a multivariate time series, discretization is performed for each dimension of the time series. In other words, each sensor signal of the different sensor signals of the (multivariate) time series is discretized. Alternatively or optionally, frequency filtering can also be applied to smooth the sensor signals.

[0062] Various methods known to those skilled in the art can be used to discretize a time series. For example, a multivariate (real) time series of a sensor dataset can be converted into a symbolic time series (i.e., a sequence of strings). This conversion (or discretization) can be performed, for example, by applying the Symbolic Aggregate Approximation (SAX) algorithm.

[0063] As mentioned above, it may also be necessary to normalize the multivariate time series, that is, convert the multivariate time series into a univariate time series.

[0064] In another embodiment, the preparation of the sensor data sets can be parallelized. To this end, multiple workers can be provided so that the data preparation can be performed at least partially in parallel. The number of workers present determines how many sensor data sets can be prepared in parallel. Parallelization can be performed on a cloud-based platform, as described above for pattern recognition.

[0065] Another aspect of the present disclosure relates to a method for vehicle design. The design method uses the results of a pattern recognition method, i.e., the identified patterns, for vehicle design. For example, the design method includes one of the above-described methods for identifying patterns in a driving cycle and further includes the following steps:

[0066] - determining a design data set based on the identified patterns; and

[0067] -Use design datasets for vehicle design.

[0068] By means of the above-described method for pattern recognition, the sensor data set can be broken down into recurring patterns (and thus recurring driving situations), which can then be used to determine the design data set.

[0069] In one example, a design dataset can be determined (identified) from a plurality of sensor datasets, wherein the design dataset is a representative sensor dataset for the plurality of sensor datasets. To this end, a characterizing parameter can be determined for each sensor dataset, which indicates the frequency of a pattern in the sensor dataset. Based on the characterizing parameter, i.e., the frequency of the occurring (and re-identified) pattern, each sensor dataset can be described.

[0070] The characteristic parameters may include, for example, an absolute pattern frequency (APF), which indicates the absolute frequency of a pattern in the sensor data set. The absolute pattern frequency spectrum includes the absolute pattern frequency of each pattern in the sensor data set.

[0071] Characteristic parameters may also include, for example, relative pattern coverage (RPC), which indicates how many time points are covered by a pattern (including its repetitions in the sensor dataset) relative to the length of the sensor dataset. Whether or not some of these time points are also covered by other patterns is irrelevant. Accordingly, the relative coverage spectrum includes the relative coverage of each pattern present in the sensor dataset.

[0072] Identifying the design data set from a plurality of sensor data sets can be performed, for example, by first performing an average calculation on the characteristic parameters of each sensor data set, and then selecting the sensor data set closest to the average as the design data set.

[0073] In another example, the design dataset can be determined in other ways. To this end, a plurality of representative patterns can be selected from the patterns identified in the global dictionary based on characteristic parameters. For example, a predetermined number of the most frequently occurring patterns and / or the patterns with the greatest relative coverage can be used to construct the design dataset. However, other methods for selecting representative patterns and their frequencies based on characteristic parameters are also feasible.

[0074] The thus determined (i.e., identified or constructed) design dataset is used for vehicle design. For example, the vehicle's component design, operating strength, etc. can be designed and tested based on the design dataset. Because the design dataset is particularly representative of vehicle operation, vehicle design can be improved by using the design dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Now, an embodiment of the present invention will be described by way of example and with reference to the accompanying drawings.

[0076] Figure 1 A flow chart illustrating a first method for identifying recurring patterns in driving cycles;

[0077] Figure 2 A flow chart illustrating a second method for identifying recurring patterns in driving cycles;

[0078] Figure 3 a flow chart illustrating a method for identifying and using a design dataset for vehicle design; and

[0079] Figure 4 Flowchart showing a method for constructing and using a design dataset for vehicle design DETAILED DESCRIPTION

[0080] Reference Figure 1 A method for identifying recurring patterns in driving cycles is described.

[0081] In block 101, multiple sensor data sets are prepared. The sensor data sets are stored in a data lake and each represent a driving cycle or a journey through the vehicle. In other words, the sensor signals in the sensor data sets are prepared. Optionally, block 101 can be omitted if the sensor data sets have already been prepared for pattern recognition.

[0082] For preparing the sensor data set, a plurality of workers may be provided so that the data preparation can be performed at least partially in parallel. In this case, the number of workers present determines how many sensor data sets can be prepared in parallel.

[0083] The preparation of sensor datasets can include discretization of the time series. In the case of multivariate time series, normalization can also be performed. Furthermore, the preparation of sensor datasets can also include the application of frequency filtering to smooth the sensor signals.

[0084] Optionally, in block 101, a preselection may be performed from the plurality of sensor datasets based on predetermined criteria, so that preparation is performed only for the preselection. The predetermined criteria may include, for example, a plausibility check that identifies unreasonable transitions in the sensor dataset. Unreasonable sensor datasets are then excluded from the plurality of sensor datasets for pattern recognition.

[0085] For preparation, the sensor dataset can be retrieved from the data lake and prepared by the computing unit. After preparation, the sensor dataset can be stored in the data lake again.

[0086] Hereinafter, “sensor dataset” also refers to a prepared sensor dataset.

[0087] In block 102, one or more samples are selected from a plurality of sensor datasets from a data lake. Unsupervised pattern recognition is performed on the samples to identify patterns within the samples. The selection of samples can be random or based on predetermined criteria.

[0088] In block 102a, samples are adapted for a pattern recognition algorithm (for unsupervised pattern recognition). In this case, the sample data sets are combined, in particular concatenated, into a single time series, so that the pattern recognition algorithm is applied only to the concatenated time series.

[0089] Depending on the pattern recognition algorithm used, conditioning the sample may include additional steps, such as normalization for multivariate time series, which may be required for some algorithms.

[0090] In other examples (not shown), the adjustments according to block 102 a may also be performed in block 101 on all sensor data sets from the data lake.

[0091] In summary, in blocks 102 and 102a, samples are selected and / or adjusted to apply a pattern recognition algorithm for unsupervised pattern recognition.

[0092] In block 103, unsupervised pattern recognition is performed to identify patterns in the sample (or in a time series from the sample). For unsupervised pattern recognition, only a pattern recognition algorithm is used. Corresponding pattern recognition algorithms are known to those skilled in the art. For example, a pattern recognition algorithm based on the Sequitur algorithm can be used. Alternatively, a pattern recognition algorithm based on the "Pattern-Enumeration algorithm" can also be used. In some examples, additional parameterization of the pattern recognition algorithm for the unsupervised model can also be performed to optimize the algorithm for the sample.

[0093] The pattern recognition algorithm is applied to the adjusted samples one at a time and serially. The pattern recognition algorithm identifies repeated patterns in the samples and creates a (global) dictionary from the repeated patterns and their definitions.

[0094] In other words, a plurality of recurring patterns in the sample are determined in block 103 , based on a pattern recognition algorithm for unsupervised pattern recognition for identifying (recurring) patterns.

[0095] In block 104, supervised pattern recognition is performed on the remaining (or other) prepared sensor datasets that are not part of the sample. That is, based on the previously created global dictionary, repeated patterns are identified in the remaining sensor datasets in the plurality of sensor datasets that are not part of the sample. Appropriate pattern recognition algorithms known to those skilled in the art are used to perform supervised pattern recognition.

[0096] In block 104a, a degree of correspondence is determined for each of the plurality of sensor data sets. The degree of correspondence represents a quality criterion for the pattern recognition method.

[0097] In block 105, the quality criterion is compared with respect to a termination condition. That is, it is checked whether each of the plurality of sensor data sets has a correspondence equal to or greater than a predetermined correspondence threshold.

[0098] The predetermined correspondence threshold may be, for example, 80%. In other examples, the predetermined correspondence threshold may also be 85% or 90%.

[0099] In one example, pieces 102a, 103, 104, 104a, 105 and piece 106 described later can be repeated as a loop. In this example, another termination condition can be specified, and in piece 105, this other termination condition is checked. For example, another termination condition can be selected so that the loop consisting of pieces 102a, 103, 104, 104a, 105 and 106 should be traversed through a predetermined number of cycles. In other words, another termination condition can include a maximum number of traversals. In this case, another termination condition is compared with respect to another quality standard that is indicative of the number of traversals. Additionally or alternatively, another termination condition can be selected so that at least one new repetitive pattern must be found in each traversal. If no new pattern is found, this another termination condition is satisfied.

[0100] The additional termination condition may also include that no anomaly in pattern recognition occurs during the unsupervised pattern recognition (block 103). For example, an anomaly may occur if too few patterns are recognized during the unsupervised pattern recognition or if the pattern recognition algorithm for the unsupervised pattern recognition terminates itself. In this case, the method 100 is repeated as described above.

[0101] If the termination condition (or conditions) is not met in block 105, the method proceeds to block 106. If the termination condition is met, the method 100 ends.

[0102] In block 106, the sample is reselected or changed. In principle, block 102 is repeated in block 106. The sample can be changed in a variety of ways. For example, the sample can be supplemented with sensor data sets from multiple sensor data sets whose correspondence, determined in block 104a, is less than a predetermined correspondence threshold and optionally has a certain travel length. Alternatively, a completely new sample that was not previously part of the sample can be selected from the multiple sensor data sets. Method 100 then proceeds to block 102a. If block 102a is omitted, as described above, method 100 proceeds to block 103.

[0103] In a subsequent (repeated) block 103 , a new global dictionary is then created based on the changed samples, ie the previously created global dictionary is replaced.

[0104] After the method is finished, ie if the termination condition is met, the result contains a globally valid dictionary comprising a plurality of recurring patterns. "Globally valid" here means that all data sets can be overwritten with respect to predetermined values ​​by the patterns stored in the dictionary.

[0105] In method 100, in block 103 (unsupervised pattern recognition), only the unique time series that is grouped together by the sample data set is processed by a unique worker, which means that other workers are not used even if they are available.

[0106] Therefore, the following reference Figure 2 The second method 200 is described. The second method 200 takes advantage of additional available workers so that unsupervised pattern recognition can be performed on multiple univariate time series in parallel. Figure 2 A second method 200 for identifying recurring patterns in driving cycles is shown as a flow chart. Second method 200 differs from first method 100 in blocks 202a, 203, and 203a. Therefore, to avoid repetition, only the different blocks 203 and 203a are described.

[0107] In block 202a, as in block 102a, the sample sensor data set is adjusted for the pattern recognition algorithm used for unsupervised pattern recognition. To this end, the sample data set may be unified, if necessary. Unlike block 102a, in block 202a, the sample data sets are not combined or concatenated into a single time series. Instead, the sample data sets are individually subjected to unsupervised pattern recognition in subsequent block 203. If adjustment of the sample data sets is not required, block 202a can be omitted.

[0108] In other examples (not shown), the adjustments according to block 202 a may also be performed in block 101 on all sensor data sets from the data lake.

[0109] In block 203, unsupervised pattern recognition is performed separately for each sample data set. The pattern recognition algorithm for unsupervised pattern recognition is applied separately to each sample data set in order to determine a special (pattern) dictionary for each sample data set. Since the pattern recognition algorithm runs on the sample data set separately and is therefore independent of other sample data sets, the sample data sets can be checked in parallel in block 203. The pattern recognition of the sample data sets is performed by workers. As described in relation to block 103, algorithms for unsupervised pattern recognition are known to those skilled in the art. The result of performing unsupervised pattern recognition on multiple sample data sets in parallel is the creation of a special dictionary for each sample data set.

[0110] In block 203a, the results of block 203, i.e., the specialized dictionaries, are combined into a global dictionary. For example, the specialized dictionaries may contain consistent patterns. These patterns are only included once in the global dictionary. The second method 200 then proceeds to block 104.

[0111] Similar to Figure 1 Method 100, Figure 2 The method 200 proceeds to block 202a after changing the sample in block 106.

[0112] Based on the recurring patterns identified in one of the methods 100, 200 for identifying recurring patterns and stored in the global dictionary, a design dataset can be determined. The design dataset can then be used for vehicle design. Two methods 300, 400 for determining and using a design dataset are described below.

[0113] Based on the recurring patterns identified in one of the methods 100, 200 for identifying recurring patterns, a design dataset can be determined. The design dataset can then be used for vehicle design. Two methods 300, 400 for determining and using a design dataset are described below.

[0114] Figure 3 A method 300 is shown as a flow chart for identifying a design dataset from a plurality of sensor datasets for use in vehicle design.

[0115] In block 301 , characterizing parameters are determined for each sensor data set.

[0116] The characteristic parameter indicates the frequency of a recurring pattern in the sensor data set. Based on the characteristic parameter, ie, by the frequency of the occurring (and re-identified) pattern, each sensor data set can thus be described.

[0117] In block 302 , a design data set is determined (or identified) from a plurality of sensor data sets based on characterizing parameters.

[0118] In block 303, the design dataset is used for vehicle design.

[0119] Figure 4 A method 400 is shown as a flow chart for constructing a design dataset based on re-identified patterns and for use in vehicle design.

[0120] In block 401 , characterizing parameters are determined for each sensor data set as in block 301 .

[0121] In block 402 , a plurality of representative patterns are determined from the identified patterns based on characterizing parameters.

[0122] In block 403 , a design data set is constructed based on a plurality of representative patterns.

[0123] In block 404 , a design dataset for vehicle design is used.

[0124] This disclosure relates to the following:

[0125] (1) A method for pattern recognition in a driving cycle, comprising:

[0126] (a) Selecting an indefinite number of samples from multiple sensor datasets, where each sensor dataset represents a driving cycle;

[0127] (b) performing unsupervised pattern recognition to identify patterns within a variable number of samples;

[0128] (c) performing supervised pattern recognition in remaining sensor data sets of the plurality of sensor data sets to re-identify patterns; and

[0129] (d) determining, for each sensor data set of the plurality of sensor data sets, a degree of correspondence between the sensor data set and the pattern;

[0130] Steps (a) to (d) are repeated until each correspondence exceeds a correspondence threshold.

[0131] (2) The method according to (2), wherein the indefinite number of samples is changed when step (a) is repeatedly performed.

[0132] (3) A method according to (1) or (2), wherein, when step (a) is repeatedly performed, the indefinite number of samples are supplemented with one or more sensor data sets from multiple sensor data sets whose correspondence is less than a correspondence threshold.

[0133] (4) A method according to any one of (1) to (3), wherein when step (a) is first performed, an indefinite number of samples are randomly selected from multiple sensor data sets.

[0134] (5) The method according to any one of (1) to (4), wherein in step (c) supervised pattern recognition is performed in parallel on the remaining sensor data sets.

[0135] (6) The method according to any one of (1) to (5), wherein step (a) further comprises concatenating sensor data sets of an indefinite number of samples into a single data set.

[0136] (7) The method according to any one of (1) to (5), wherein in step (b), unsupervised pattern recognition of a sensor data set from an indefinite number of samples is performed by a single worker.

[0137] (8) The method of (7), wherein step (b) further comprises determining a global pattern dictionary based on unsupervised pattern recognition, wherein the global pattern dictionary defines the recognized patterns and is used for supervised pattern recognition.

[0138] (9) The method according to any one of (1) to (6), wherein in step (b), unsupervised pattern recognition of a sensor data set from an indefinite number of samples is performed in parallel.

[0139] (10) The method according to (6), wherein step (b) is performed by multiple workers.

[0140] (11) A method according to (6) or (7), wherein step (b) further comprises determining a specialized pattern dictionary for each sensor data set of an indefinite number of samples based on unsupervised pattern recognition, wherein the specialized pattern dictionary defines the recognized pattern, and the method further comprises combining the specialized pattern dictionaries into a global pattern dictionary, which is used to perform supervised pattern recognition according to step (c).

[0141] (12) A method according to any one of (1) to (11), further comprising preparing a sensor dataset for unsupervised pattern recognition.

[0142] (13) The method of (12), wherein the preparation of the sensor data set is performed in parallel.

[0143] Reference Signs List

[0144] 100 Pattern Recognition Methods

[0145] Preparation of 101 sensor dataset

[0146] 102 Select Sample

[0147] 102a Adjustment Sample

[0148] 103Unsupervised pattern recognition in samples

[0149] 104 Perform supervised pattern recognition on the remaining sensor dataset

[0150] 104a Determine the correspondence

[0151] 105 Check abort conditions

[0152] 106 Change Sample

[0153] 200 Pattern Recognition Methods

[0154] 203 Perform unsupervised pattern recognition on each dataset from the sample separately

[0155] 203a Combined Specialized Dictionary

[0156] 300 First Design Method

[0157] 301 Determine the representative parameters of the sensor data set

[0158] 302 identification design dataset

[0159] 303Using Design Datasets for Transportation Design

[0160] 400 Second Design Method

[0161] 401 Determine the representative parameters of the sensor data set

[0162] 402 Determine multiple representative patterns from the identified patterns

[0163] 403 Design dataset based on multiple representative patterns

[0164] 404 Using design dataset for transportation design

Claims

1. A method for pattern recognition in a driving cycle, comprising: a) selecting an indefinite number of samples (102) from a plurality of sensor data sets, wherein each sensor data set represents a driving cycle; b) performing unsupervised pattern recognition (103, 203) on an indefinite number of samples to identify patterns; c) performing supervised pattern recognition (104, 204) in remaining sensor data sets of the plurality of sensor data sets to re-identify the pattern; and d) for each sensor data set of the plurality of sensor data sets, determining a correspondence between the sensor data set and the pattern (105); Wherein, steps a) to d) are repeatedly performed until each correspondence exceeds a correspondence threshold.

2. The method according to claim 1, wherein: When step a)(106) is repeatedly executed, the indefinite number of samples is changed.

3. The method according to claim 1 or 2, wherein: When step a) (106) is repeatedly executed, the indefinite number of samples are supplemented with one or more sensor data sets from the plurality of sensor data sets whose correspondence is less than a correspondence threshold.

4. The method according to any one of claims 1 to 3, wherein: When step a)(102) is first performed, an indefinite number of samples are randomly selected from a plurality of sensor data sets.

5. The method according to any one of claims 1 to 4, wherein: In step c) (104) supervised pattern recognition is performed in parallel on the remaining sensor data sets.

6. The method according to any one of claims 1 to 5, wherein: Step a) (102a) also includes concatenating the sensor data sets of the indefinite number of samples into a unique data set.

7. The method according to any one of claims 1 to 5, wherein: In step b) (103), unsupervised pattern recognition on a sensor data set from an indefinite number of samples is performed by a single worker.

8. The method according to claim 7, wherein: Step b) further comprises determining a global pattern dictionary (103) based on the unsupervised pattern recognition, wherein the global pattern dictionary defines the recognized patterns and is used for the supervised pattern recognition. 9 . The method according to claim 1 , further comprising the preparation of a sensor data set for unsupervised pattern recognition ( 102 a ).

10. The method according to claim 9, wherein: The preparation of the sensor data set is performed in parallel (102a).

Images