Device and computer-implemented method for processing sensor data detected by a sensor element, in particular attached to a body

Abstract

The invention relates to a device for processing sensor data detected by a sensor element attached to a human or animal body.SOLUTION: For each sensor element 102, a time series of sensor data of the sensor element is provided, wherein for each time series, depending on the time series, a feature associated with the time series is determined, wherein depending on each part of the time series, a plurality of features associated with the time series is determined, wherein depending on the features associated with the time series, a value associated with the time series is determined, and wherein depending on the value associated with the time series, a time series value associated with the time series is determined; Depending on a part of the values associated with the time series, time series values associated with the time series are determined, and for at least one time series, depending on the time series values associated with the at least one time series, at least one predicted value of the values of the at least one time series associated with the at least one time series is determined.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] Conventional technology The present invention relates to an apparatus and computer-implemented method for processing sensor data, particularly detected by body-worn sensor elements. [Background technology]

[0002] Processing the sensor data is useful for a number of subsequent tasks. Summary of the Invention [Means for solving the problem]

[0003] Disclosure of the Invention A computer-implemented method for processing sensor data, particularly sensor data detected by sensor elements attached to a human or animal body, in particular acceleration sensors, angular velocity sensors, velocity sensors, magnetometers, pressure sensors, blood pressure sensors, VO2 sensors, and pulse sensors, envisages providing, for each sensor element, a time series of the sensor data of the sensor element; determining, for each time series, a feature associated with the time series depending on the time series; determining, for each time series, a plurality of features associated with the time series depending on respective portions of the time series; determining, for each time series, a value associated with the time series depending on the features associated with the time series; determining, for each time series, a time series value associated with the time series depending on the values ​​associated with the time series; determining, for each time series, a time series value associated with the time series depending on the portions of the values ​​associated with the time series; and determining, for at least one time series, at least one predicted value of the at least one time series value associated with the at least one time series depending on the time series value associated with the at least one time series. The sensor data included in the multivariate time series are processed for the predicted values. The predicted values ​​represent values ​​generated from the relationships of the sensor data contained in the multivariate time series, which are particularly suitable for post-processing, for example in the context of activity assessment.

[0004] For example, multiple time series are associated with a single sensor that includes multiple sensor elements for detecting the multiple time series, and for at least one time series, the time series value associated with the time series is determined depending on a portion of the time series associated with the sensor, thereby enabling the temporal characteristics of each time series to be evaluated with finer granularity.

[0005] For example, at least one sensor element is provided with meta-information at the time of detection of the time series by the sensor element, in particular the location or specifications of the sensor element or demographic information about the subject to which the sensor element is attached, the meta-information is mapped to a representation of the meta-information, and a signal for the time series is determined depending on the representation and at least one predicted value of the value of the time series, thereby enabling fusion with additional information.

[0006] For example, subsequent processing may envisage that the signal is classified, or that future values ​​of the time series are predicted depending on the signal, or that values ​​of the time series are reconstructed depending on the signal.

[0007] For example, sensor data values ​​provided during detection of the sensor data for a time series may be normalized by a transformation, the time series may include normalized values, future values ​​may be normalized and predicted, and the future values ​​may be determined by inversely transforming the normalized and predicted values, or reconstructed values ​​may be normalized and reconstructed, and the reconstructed values ​​may be determined by inversely transforming the normalized and reconstructed values. This allows for statistical differences in multivariate time series to be taken into account, and reduces the effects of sensor drift in each sensor element, particularly when the method is used to calculate predicted values ​​for multiple different multivariate time series multiple times in sequence.

[0008] To generate the training data, it is assumed that, for example, multiple sets of time series and meta-information are provided for each sensor element, the sets each comprising sensor data detected within a different time range, and a signal is determined for each set. In particular, the signals can be used for training a classifier for classifying the signals and / or for predicting or reconstructing at least one value from one or more of the time series.

[0009] For training, for example, an artificial neural network is provided, a first part of which is configured to determine a signal depending on the time series of sensor data of the sensor elements and depending on the meta-information, and a second part of which is configured to process, in particular classify, the signal or predict or reconstruct the value of one of the time series depending on the signal, the artificial neural network, in particular the first part of the artificial neural network, being trained based on the first part of the set to process, in particular classify, the signal or predict or reconstruct the value of one of the time series depending on the signal (312), and subsequently the second part of the artificial neural network is trained based on the second part of the set to process, in particular classify, the signal or predict or reconstruct the value of one of the time series depending on the signal (312).

[0010] For the purposes of regularization, it may be assumed that the predicted value is determined for at least one time series depending on the time series values ​​associated with the at least one time series as well as depending on the values ​​associated with the time series.

[0011] For the prediction over a certain period of time, it may be assumed that for each time series, a plurality of, in particular successive, time series values ​​associated with the time series are determined, and that at least one predicted value for the time series is determined depending on the plurality of time series values ​​determined for the respective time series.

[0012] For the purposes of regularization, it may be assumed that for at least one time series, the value associated with the time series is determined depending on the time series in addition to depending on the features.

[0013] An apparatus for processing sensor data detected in particular by body-worn sensor elements, in particular by acceleration sensors, angular velocity sensors, velocity sensors, magnetometers, pressure sensors, blood pressure sensors, VO2 sensors, pulse sensors, comprises at least one processor and at least one memory, wherein the at least one memory comprises instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the apparatus implements the method.

[0014] The computer program comprises instructions that are executable by a computer, which, when executed by the computer, perform the method.

[0015] Further exemplary embodiments can be seen from the following description and drawings. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a schematic diagram of an apparatus for processing sensor data detected in particular by body-worn sensor elements; [Figure 2] FIG. 2 is a schematic diagram of a block for processing sensor data. [Figure 3] FIG. 1 illustrates an exemplary architecture including blocks for processing sensor data. [Figure 4]FIG. 1 illustrates an exemplary processing of sensor data in blocks. [Figure 5] 1 is a flowchart with method steps for processing sensor data. DETAILED DESCRIPTION OF THE INVENTION

[0017] 1 shows a schematic representation of an apparatus 100 for processing sensor data, in particular sensor data detected by a sensor element 102 worn on the body.

[0018] Examples for the sensor element 102 are an acceleration sensor, an angular velocity sensor, a velocity sensor, a magnetometer, a pressure sensor, a blood pressure sensor, a VO2 sensor, and a pulse sensor.

[0019] It may be envisaged that the sensor element 102 acceleration sensor, the sensor element 102 angular velocity sensor, the sensor element 102 velocity sensor, the sensor element 102 magnetometer, or the sensor element 102 pressure gauge is mounted by the body of a machine, a robot, a vehicle, a home appliance, a personal assistant system, or a tool.

[0020] A single sensor may be provided that includes multiple sensor elements 102. For example, a three-axis acceleration sensor may be provided, with one sensor element 102 for each axis detecting acceleration as a sensor signal. For example, a three-axis angular velocity sensor may be provided, with one sensor element 102 for each axis detecting angular velocity as sensor data.

[0021] The device 100 includes at least one processor 106 and at least one memory 104 .

[0022] The at least one memory 104 includes instructions executable by the at least one processor 106 that, when executed by the at least one processor 106, cause the device 100 to perform a method for processing sensor data.

[0023] 2 shows a schematic diagram of a block 200 for processing sensor data. The block 200 is configured to process a multivariate time series 202. Processing of the multivariate time series 202 using the block 200 can be used to provide realistic signals of human activity, particularly for multivariate time series 202 including sensor data detected by sensor elements 102 attached to a human body. Correspondingly, the block 200 can also be used for sensor elements 102 attached to an animal body.

[0024] The block 200 is configured to process f time series 202, each of which has a length w. There are f time series 202 for the f sensor elements 102, and each time series 202 contains sensor data detected by one of the sensor elements 102.

[0025] Block 200 includes a device 204 for determining global features 206 associated with each one of the time series 202 .

[0026] The device 204 for determining global features 206 is configured to map each time series to a respective global feature associated with each time series.

[0027] The device 204 for determining global features 206 is configured, for example, to determine, for each time series, w global features associated with the time series. The device 204 for determining global features 206 may be configured, for each time series, to determine fewer than w or more than w global features associated with the time series.

[0028] The block 200 includes a device 208 for determining local features 210 associated with each one of the time series 202 .

[0029] The device 208 for determining local features 210 is configured to map each portion of each time series to a respective local feature associated with each time series.

[0030] The device 208 for determining local features 210 is, for example, configured to determine, for each time series, w local features associated with the time series. The device 208 for determining local features 210 may be configured to determine, for each time series, fewer than w or more than w local features associated with the time series.

[0031] For example, each time series may be subdivided into multiple portions, each of which may be, for example, of equal length.

[0032] The block 200 is configured to determine, inter alia, a future or reconstructed value 212 depending on the global features 206 and depending on the local features 210 .

[0033] Values ​​212 include, for each time series, a value associated with at least one of the time series. It may be assumed that values ​​212 include, for each time series, a plurality of, in particular consecutive, values ​​associated with the respective time series.

[0034] Block 200 is configured to determine, for example, when w′=1, for each time series a single value depending on the features associated with the time series in global features 206 and depending on the features associated with the time series in local features 210. Block 200 is configured to determine, for example, for each time series a single value depending on the sum of the features associated with the time series in global features 206 and the features associated with the time series in local features 210.

[0035] For example, block 200 is configured to determine, for each time series, if w'>1, an individual value of the plurality of values ​​as described above for a single value.

[0036] Optionally, when block 200 is configured to determine w′=w global features 206 for each time series and w′=w local features 210 for each time series, respectively, block 200 is configured to determine a value associated with each time series depending on the global features 206 and the local features 210, as well as depending on the respective time series, in particular depending on the respective time series and the sum of the global features 206 determined for the respective time series and the local features 210 determined for the respective time series.

[0037] The block 200 includes a device 214 for determining a first time series value 216 .

[0038] The device 214 for determining first time series values ​​216 is configured to determine w′ first time series values ​​216 respectively for the f time series depending on the values ​​212 .

[0039] The device 214 for determining first time series values ​​216 is configured to determine, for each time series, a single first time series value of the first time series values ​​216 depending on the value 212 when w′=1. For example, the device 214 for determining first time series values ​​216 is configured to determine, for each time series, a plurality of first time series values ​​of the first time series values ​​216 when w′>1, each of the first time series values ​​being determined as described above for a single first time series value.

[0040] The block 200 includes a device 218 for determining a second time series value 220 .

[0041] The device 218 for determining second time series values ​​220 is configured to determine w′ second time series values ​​220 for the f time series, each depending on a portion of the values ​​212 .

[0042] The device 218 for determining second time series values ​​220 is configured to determine, for each time series, a single second time series value of the second time series values ​​220 depending in part on the values ​​212, e.g., when w′=1. For example, the device 218 for determining second time series values ​​220 is configured to determine, for each time series, a plurality of second time series values ​​of the second time series values ​​220 when w′>1, where each second time series value is determined as described above for a single second time series value.

[0043] Block 200 is configured to determine at least one predicted value 222 depending on the time series 202, depending on the first time series values ​​216, and depending on the second time series values ​​220. The at least one predicted value 222 includes, for example, a future value or a reconstructed value of the respective time series.

[0044] Block 200 is configured to determine, for each time series, a single predicted value for each time series, depending on the respective time series, a single first time series value associated with the respective time series in first time series values ​​216, and a single second time series value associated with the respective time series in second time series values ​​220, for example, when w'=1. Block 200 is configured to determine, for each time series, multiple predicted values ​​of predicted values ​​222, depending on the respective time series, depending on a single first time series value associated with the respective time series in second time series values ​​220, for example, when w'>1, where each predicted value is determined as described above for a single predicted value.

[0045] For example, the device described above in block 200 may be configured as a layer of an artificial neural network.

[0046] The layers of the neural network are configured to map a multivariate time series 202 at the input of the block 200 to at least one predicted value 222 at the output of the block 200 .

[0047] These devices are configured, for example, as simple layers and particularly as dense feed forward layers, respectively, of an artificial neural network.

[0048] The dense feedforward layer is one example of a type of layer that can be used. Other types of layers may be provided. The device may be implemented using multiple layers.

[0049] FIG. 3 illustrates an exemplary architecture 300 including block 200 for processing sensor data.

[0050] The architecture 300 assumes that block 200 is used to determine predicted values ​​222 depending on the time series 202. In this example, w predicted values ​​are determined for f time series, respectively.

[0051] The architecture 300 assumes that the locations 302 of the sensor elements 102 at the time of detection of the f time series 202 are provided. The locations 302 are names of locations, for example, the names of the parts of the body where the respective sensor elements 102 are located. This means that multiple sensor elements 102 are located in the same location on the body in this example.

[0052] The wearing location in this example relates to the body of a living being, for example a human or animal body, for example the wearing locations in the case of a human being are the hand joints, chest, waist, ankle, torso, pocket, head.

[0053] The architecture 300 assumes that mounting locations 302 are mapped to representations 306 using an encoder 304. In this example, w identical representations 306 are determined for a time series 202 of length w.

[0054] The architecture 300 assumes that an extended representation 310 is determined using a replicator 308 depending on the representation 306, and the extended representation 310 includes, for each predicted value of the predicted values ​​222, a representation 306 of the mounting location 302 associated with the respective predicted value.

[0055] The architecture 300 assumes that a signal 312 is determined depending on the predicted value 222 and the extended representation 310. The signal 312 is, for example, a realistic signal of human activity.

[0056] The signal 312 is determined, for example, depending on the sum of the associated individual predicted values ​​222 and the individual values ​​of the extended representation 310 .

[0057] For example, the architecture 300 is configured as layers of an artificial neural network that is configured to map the multivariate time series 202 and meta-information at the input of the architecture 300, in this example, the wearing location 302, to a signal 312 at the output of the architecture 300.

[0058] The architecture 300 assumes that the output quantity 318 is determined using the head 316 in dependence on the signal 312. This means that the output quantity 318 is provided in dependence on the multivariate time series 202 without the need for further costly data collection.

[0059] For example, the head 316 is configured as one or more layers of an artificial neural network.

[0060] The head 316 is, for example, a classifier configured to classify the signal 312 .

[0061] The classifier can be trained, for example, in dependence on the signal 312, to identify or track variations in the activities of a human wearing the sensor element 102 for detecting the multivariate time series 202, so as to operate without the need for further laborious data collection.

[0062] The architecture 300 assumes, for example, that the signal 312 is classified using a classifier.

[0063] The head 316 is configured, for example, similarly to the block 200, to predict future values ​​of at least one of the time series 202 depending on the signal 312 and / or to reconstruct values ​​of at least one of the time series 202 depending on the signal 312.

[0064] The part of the architecture 300 located upstream of the head 316 forms a base model. The base model is configured to identify characteristic patterns in the multivariate time series 202, particularly depending on meta-information, in this example, depending on the mounting location 302. The mounting location 302 represents an example of meta-information regarding the placement of the sensor elements 102. The meta-information may include specifications of each sensor element 102 or demographic information regarding the recipient of the sensor elements 102. The specifications may be, for example, the sampling rate, sensitivity, or measurement range of each sensor element 102. The demographic information may be, for example, the age, gender, height, and weight of the recipient of the sensor elements 102. The head 316 is interchangeable depending on the task. Multiple different heads 316, each for a different task, can be simultaneously implemented. The multiple heads 316 can be simultaneously trained depending on the signal 312.

[0065] In this example, the architecture 300 assumes one block 200 located upstream of the fusion with meta information. The architecture 300 can also assume multiple blocks 200 located upstream of the fusion. The architecture 300 assumes, for example, one block 200 as a head 316 located downstream of the fusion with meta information. The architecture 300 can also assume multiple blocks 200 as heads 316 located downstream of the fusion.

[0066] Figure 4 shows m time series with length of w=n values: x 1,1 ,···,x 1,n x 2,1 ,···,x 2,n x m-1,1 ,···,x m-1,n x m,1 ,···,x m,n , and for each time series, a single predicted value x out of the predicted values ​​222 1,n+1 ,···,x m,n+12. An exemplary processing of sensor data in block 200 is shown schematically for

[0067] In FIG. 4, the global features 206 include:

number

number

number

[0068] In FIG. 4, the first time series value 216 includes:

number

number

[0069] In FIG. 4, the second time series value 220 includes:

number

[0070] In the example shown in Figure 4, the first two time series are associated with the same sensor, i.e., the second time series value of the first time series

number

number

number

number

number

number

number

number

number

number

number

number

[0071] For example, in an application using a three-axis acceleration sensor, it is assumed that three axis sensor elements 102, i.e., three time series containing sensor data from the three sensor elements 102 of the three-axis acceleration sensor, are associated with the same sensor, and each second time series value is determined depending on the three time series values ​​associated with these three time series.

[0072] For example, in an application using a three-axis angular velocity sensor, it is assumed that three axis sensor elements 102, i.e., three time series containing sensor data from the three sensor elements 102 of the three-axis angular velocity sensor, are associated with the same sensor, and each second time series value is determined depending on the three time series values ​​associated with these three time series.

[0073] Individual predicted values ​​x i,n+1 is the value determined for each time series i

number

number

number

[0074] FIG. 5 shows a flow chart with the steps of a method for processing sensor data.

[0075] The method includes step 502 .

[0076] In step 502, for each sensor element 102, a time series of sensor data for the sensor element 102 is provided. For example, m time series with n values ​​each: x 1,1 x 1,n ;···;x m,1 ,x m,n is provided.

[0077] The method includes step 504 .

[0078] In step 504, for each time series, a feature associated with the time series is calculated in a time series dependent manner.

number

[0079] The method includes step 506 .

[0080] In step 506, for each time series, a plurality of features associated with the time series are calculated, each feature depending on a respective portion of the time series.

number

[0081] The method includes step 508 .

[0082] In step 508, for each time series, the features associated with the time series are

number

number

[0083] For example, for at least one time series i, the value associated with the time series

number

number

[0084] The method includes step 510 .

[0085] In step 510, for each time series, the value associated with the time series is

number

number

[0086] The method includes step 512 .

[0087] In step 512, for each time series, the value associated with the time series is

number

number

[0088] For example, multiple time series are associated with a single sensor that includes multiple sensor elements 102 for detecting the multiple time series. For example, for at least one time series i associated with the sensor, the time series value

number

[0089] The method includes step 514 .

[0090] In step 514, for time series i, at least one time series value associated with time series i is calculated.

number

[0091] It may be assumed that forecasts for values ​​of multiple time series are to be determined.

[0092] It may be assumed that for each time series, multiple predictions are determined.

[0093] For example, for multiple time series, the predicted value x of each time series is 1,n+1 ,···,x m,n+1 is determined.

[0094] Predicted value x i,n+1 For example, for at least one time series i, the time series value associated with at least one time series is

number

number

[0095] It may be assumed that the method for considering the wearing location includes step 516 .

[0096] In step 516, at least one sensor element 102 is provided with the mounting location 302 of the sensor element 102 at the time of detection of the time series by the sensor element 102.

[0097] It may be assumed that the method for considering the wearing location includes step 518 .

[0098] In step 518 , the mounting location 302 is mapped to a representation 306 of the mounting location 302 .

[0099] It may be assumed that the method for considering the wearing location includes step 520 .

[0100] In step 520, a signal 312 for the time series is determined depending on the representation 306 of the mounting location 302 and at least one predictor 222 for the time series.

[0101] The method includes step 522 .

[0102] In step 522, for example, the predicted value 222 is classified or the signal 312 is classified to take into account the wearing location 302.

[0103] For example, the head 316 is trained in dependence on the signal 312 to identify or track a variety of activities of a person wearing the sensor element 102 to detect the multivariate time series 202. For example, the head 316 is trained based on the time series data 202 to classify the time series data 202 according to an activity performed by a person wearing the sensor element 102 at the time of detecting the sensor data. For example, the head 316 is trained based on the time series data 202 to classify the time series data 202 according to a body state of a person wearing the sensor element 102 at the time of detecting the sensor data.

[0104] In step 522, for example, future values ​​of the time series are predicted depending on the predicted values ​​222, or the values ​​of the time series are reconstructed.

[0105] To take into account the wearing location 302, for example, in step 522, future values ​​of the time series are predicted or the values ​​of the time series are reconstructed depending on the signal 312.

[0106] For example, the heads 316 may be trained to predict future values ​​of the time series or to reconstruct values ​​of the time series depending on the signals 312. For example, the heads 316 may be trained to reconstruct or predict at least one aspect of the multivariate time series 202.

[0107] For prediction, the head 316 is trained based on the multivariate time series 202, for example, to predict the nth next value of at least one multivariate time series of the multivariate time series 202.

[0108] For the reconstruction, the head 316 is trained based on the multivariate time series 202, for example, to predict n masked values ​​of at least one multivariate time series of the multivariate time series 202.

[0109] It may be assumed that the head 316 for the three-axis acceleration sensor or three-axis angular velocity sensor is trained based on the multivariate time series 202 to determine the time series of sensor data for the third axis depending on the time series containing the sensor signals for two of the three axes.

[0110] It may be assumed that the signal 312 is subsequently processed using multiple heads 316. For example, the signal 312 is classified, values ​​are reconstructed, and / or values ​​are predicted using multiple heads depending on the signal 312. It may be assumed that the multiple heads 316 are trained simultaneously with each other depending on the signal 312.

[0111] Training may involve, for example, basic training of layers of an artificial neural network depending on the multivariate time series 202 and depending on meta-information including, for example, the position of each sensor element 102.

[0112] Training may envisage, for example, specifically training the head 316, i.e., one or more output layers, of a neural network for a given task. It may be envisaged that the specific training may be performed without using meta-information.

[0113] This means that it is possible to determine the prediction 222 and the signal 312 without using meta-information.

[0114] It may be assumed that step 502 is then performed.

[0115] In step 502, it can be assumed that the values ​​of the sensor data provided during detection of the sensor data for the time series are normalized by a transformation.

[0116] The sensor data is normalized, for example, using a Reverse Instance Normalization layer from RevIN, which removes the mean value of the sensor data for each sensor element and standardizes the variance in the values ​​of the sensor data for each sensor element.

[0117] This means that the time series contains normalized values.

[0118] For example, future values ​​are normalized and predicted, and future values ​​are determined by inverse transforming the normalized predicted values.

[0119] The sensor data is normalized, for example, using a Reverse Instance Denormalization layer from RevIN, e.g., the respective pre-removed mean values ​​are added to the sensor data.

[0120] For example, the reconstructed values ​​may be normalized and reconstructed, and the reconstructed values ​​may be determined by inverse transforming the normalized and reconstructed values.

[0121] It can be assumed that for each time series, a plurality of, in particular successive, time series values ​​associated with the time series are determined in step 512. For example, in step 514, at least one predicted value of a value of the time series is determined depending on the plurality of time series values ​​determined for each time series.

Claims

1. 1. A computer-implemented method for processing sensor data detected by a sensor element (102) attached to a human or animal body, in particular an acceleration sensor, an angular velocity sensor, a velocity sensor, a magnetometer, a pressure sensor, a blood pressure sensor, a VO2 sensor, a pulse sensor, comprising: For each sensor element (102), the time series (x １，１ ...x １，ｎ , ..., x ｍ，１ ..., x ｍ，ｎ ) is provided (502), For each time series, a feature associated with the time series is determined depending on the time series. [Equation 1] is determined (504), For each time series, a plurality of features associated with the time series, each of the features depending on a respective portion of the time series. [Equation 2] is determined (506), For each time series, the features associated with that time series [Equation 3] Depending on the value associated with the time series, [Equation 4] is determined (508), For each time series, the value associated with that time series [Equation 5] Depending on the time series value associated with the time series, [Equation 6] is determined (510), For each time series, the value associated with that time series [Equation 7] Depending on a part of the time series value associated with the time series, [Equation 8] is determined (512), For at least one time series, a time series value associated with the at least one time series [Equation 9] at least one predicted value (x) of the value of the at least one time series associated with the at least one time series, depending on ｉ，ｎ＋１ , ..., x ｍ，ｎ＋１ ) is determined (514) A method characterized by:

2. a plurality of time series are associated with a single sensor including a plurality of sensor elements (102) for detecting the plurality of time series; for at least one time series, a time series value associated with the time series depending on a portion of the time series associated with the sensor; [Equation 10] is determined (512), The method of claim 1.

3. At least one sensor element (102) is provided (516) with meta-information at the time of detection of the time series by the sensor element (102), in particular with respect to the location (302) or specifications of the sensor element (102) or demographic information relating to the subject to which the sensor element (102) is attached; The meta-information is mapped (518) to a representation (306) of the meta-information; a signal (312) for the time series is determined (520) depending on the representation (306) and at least one predicted value (222) of the values ​​of the time series; 3. The method according to claim 1 or 2.

4. The signal (312) is classified (318) (522), or A future value (318) of the time series is predicted (522) depending on the signal (312); or The time series of values ​​(318) is reconstructed (522) depending on the signal. The method of claim 3.

5. The values ​​of the sensor data provided upon detection of the sensor data for the time series are normalized by a transformation (502); the time series includes normalized values; The future values ​​are normalized and predicted, and the future values ​​are determined by inverse transforming the normalized predicted values ​​(522); or The reconstructed values ​​are normalized and reconstructed, and the reconstructed values ​​are determined by inverse transforming the normalized and reconstructed values ​​(522). The method of claim 4.

6. A plurality of sets, each consisting of a time series and meta-information, are provided for each sensor element (102) (502); The sets each include sensor data detected within a different time range; The signal (312) is determined (514) for each set.

6. The method according to claim 4 or 5.

7. An artificial neural network is provided, a first portion (200, 304, 308) of the artificial neural network configured to determine the signal (312) depending on the time series (202) of the sensor data of the sensor element (102) and depending on the meta-information (302); a second part (316) of the artificial neural network configured to process (318) the signal (312), in particular to classify or to predict or reconstruct the value of one of the time series depending on the signal (312); the artificial neural network, in particular the first part (200, 304, 308) of the artificial neural network, is trained on the basis of a first part of the set to process the signal (312), in particular to classify it or to predict or reconstruct the value of one of the time series depending on the signal (312); Subsequently, the second part (316) of the artificial neural network is trained based on the second part of the set to process, in particular to classify, the signal (312) or to predict or reconstruct the value of one of the time series depending on the signal (312). The method of claim 6.

8. The predicted value (x ｉ，ｎ＋１ ) is a time series value associated with at least one time series, [0011] In addition to being dependent on [0012] (514) is determined depending on 8. The method according to any one of claims 1 to 7.

9. For each time series, a plurality of, in particular successive, time series values ​​associated with said time series are determined (512); determining 514 at least one predicted value for the time series dependent on a plurality of time series values ​​determined for each of the time series values; 9. The method according to any one of claims 1 to 8.

10. For at least one time series, a value associated with the time series [0013] is the above-mentioned feature [0014] In addition to being dependent on ｉ，１ ...x ｉ，ｎ ) (508), 10. The method according to any one of claims 1 to 9.

11. An apparatus (100) for processing sensor data detected in particular by body-worn sensor elements (102), in particular by acceleration sensors, angular velocity sensors, velocity sensors, magnetometers, pressure sensors, blood pressure sensors, VO2 sensors, pulse sensors, The device (100) comprises: at least one processor (102); At least one memory (104); Including, the at least one memory (104) includes instructions executable by the at least one processor (102); The instructions, when executed by the at least one processor (102), cause the device (100) to implement the method of any one of claims 1 to 10.

1. An apparatus (100) comprising:

12. 11. A computer program comprising computer executable instructions, the instructions causing, when executed by the computer, to perform the method of any one of claims 1 to 10.

Images