Data processing device, data analyzing device, data pro ^¬ cessing system and method for processing data

Technical field

The present invention relates to a data processing device, a data analyzing device, a data processing system and a method for processing data.

Background

US 2016/110478 Al relates to blocking and featurization of time-series data gathered from at least one sensor. The input time-series data is divided into blocks with common attrib ^¬ utes (features) according to feature models that describe patterns in the data. The blocks may be overlapping or non- overlapping. The resultant feature blocks are annotated with feature information and semantic meaning.

Although applicable in principle to any data processing sys ^¬ tem employing a data model, the present invention and its un ^¬ derlying problem will be described hereinafter in combination with an industrial system processing sensor data which are locally acquired.

An intelligent control of an industrial system requires de ^¬ vices which can improve automatically their performance over time. For example, local devices perceive their environments by measuring data and automatically determine an appropriate action based on the measured data. In order to optimize the system, machine learning can be employed on the controllers of the local devices. However, due to limitations of the de- vices the machine learning on the local devices is also lim ^¬ ited, for instance by the available memory, the computational power or the energy consumption. Cloud infrastructures like Siemens Mindsphere typically have large amounts of available resources. Accordingly, these resources enable enhanced ma ^¬ chine learning. However, to perform this enhanced machine learning, data from the local devices have to be provided to the cloud infrastructure, the machine learning has to be per- formed in the cloud infrastructures and successively, the re ^¬ sults of the machine learning have to be transmitted back to the local devices. Accordingly, a huge amount of data has to be transferred between the local devices and the cloud.

Against this background, a problem addressed by the present invention is to provide a smart machine learning. Especially, the present invention aims to provide an improved generation of parameters for a data processing model in the environment of locally distributed devices.

Summary

The present invention provides a data processing device with the features of claim 1, a data analyzing device with the features of claim 6, a data processing device with the fea ^¬ tures of claim 10 and a method for processing data with the features of claim 11.

In a first aspect, a data processing device is provided, com- prising a sensor database, a model database, a controller, a model generator, a data filter and a receiver. The sensor database is adapted to store sensor data measured by a number of sensors. The model database is adapted to store model pa ^¬ rameters of a system model. The controller is adapted to pro- cess the sensor data measured by the number of sensors. Espe ^¬ cially, the sensor data are processed based on model parame ^¬ ters stored in the model database. The model generator is adapted to compute model parameters of the system model based on the sensor data stored in the measurement database. The model parameters are computed by applying a first model scheme. The computed model parameters of the system model are stored in the model database by the model generator. The data filter is adapted to filter the sensor data stored in the sensor database. Further, the data filter is adapted to for ^¬ ward the filtered sensor data to an external data analyzing device. The receiver is adapted to receive further model pa ^¬ rameters provided by the external data analyzing device. The receiver is further adapted to store the received further model parameters in the model database of the data processing device .

In a second aspect, a data analyzing device is provided com- prising a receiver, a sensor database, a first model genera ^¬ tor, a second model generator and a data analyzer. The re ^¬ ceiver is adapted to receive sensor data from an external da ^¬ ta processing device. The sensor database is adapted to store the received sensor data. The first model generator is adapted to compute first model parameters of a system model. The first model parameters are computed based on the sensor data stored in the sensor database by applying a first model scheme. The second model generator is adapted to compute sec ^¬ ond model parameters of the system model. The second model parameters are computed by the second model generator based on the sensor data stored in the sensor database by applying a second model scheme. The data analyzer is adapted to com ^¬ pare the first model parameters with the second model parame ^¬ ters. Further, the data analyzer is adapted to send the sec- ond model parameters to the data processing device if a dif ^¬ ference between the first model parameters and the second model parameters exceeds a predetermined threshold.

In a third aspect, a data processing system is provided com- prising a data processing device according to the first as ^¬ pect and a data analyzing device according to the second as ^¬ pect .

In a fourth aspect, a method for processing data is provided comprising the steps of measuring sensor data; storing measured sensor data in a sensor database of a data processing device; computing, in the data processing device, online mod ^¬ el parameters of a system model based on sensor data stored in the measurement database by applying a first model scheme; storing the computed online model parameters of the system in a model database of the data processing device; filtering, by the data processing device, the sensor data stored in the sensor database; forwarding the filtered sensor data from the data processing device to an data analyzing device; receiving, by the data processing device, further model parameters sent from the data analyzing device; storing the received further model parameters in the model database of the data processing device; and processing, by a data processing device, measured sensor data based on the model parameters stored in the model database.

The present invention is based on the fact that locally dis- tributed devices in an intelligent industrial control system usually only have limited computational resources. According ^¬ ly, machine learning such as computing or improving a data model for processing sensor data on local devices is limited. Furthermore, the use of computational resources, for instance the computational resources of a cloud computing system, re ^¬ quires transferring a huge amount of data between the local devices and the cloud computing system.

Starting from this fact, the present invention tries to im- prove machine learning by optimizing a system model in the local distributed devices by a trade-off of a local computa ^¬ tion of parameters for the system model in the local devices and a reduced transfer of data to an external computation system providing huge computational resources. Accordingly, the system model in the local devices can be improved almost in real time with the limited computational resources of the local devices. In addition, improved machine learning can be performed by computing an enhanced system model on an external device like a cloud computing system. For this purpose, relevant data of the local devices are transmitted from the local devices to the cloud computing system. By limiting the transmitted data, the required data transmission can be mini- mized and accordingly, the load of a communication network can be reduced.

The sensor data which are stored in the sensor database may be provided by a number, i.e. one or more, of sensors. Espe ^¬ cially, the sensors may be any kind of sensors. For example, the sensor may measure a temperature, humidity, pressure, force, acceleration, direction, speed, flow, or any other parameter which may be sensed by a sensor. It is understood that the sensor may provide its measurement results by digi ^¬ tal data or by an analogue signal. If the signal is provided in an analogue form, the analogue signal may be converted to a digital signal by an analogue to digital converter. Fur ^¬ thermore, it may be possible to apply a further processing on the measured sensor data. For example, the measured sensor data may be scaled, limited to a predetermined range, fil ^¬ tered, etc. The measured sensor data may be received, for ex ^¬ ample, by a sensor data interface. Accordingly, the sensors and the sensor database may be communicatively coupled by an appropriate communication line, for example a network like a bus system, especially an industrial bus system, an Ethernet network or any other appropriate communication network. The sensor interface may receive the sensor data and forward the received sensor data to the sensor database and/or any other related device, especially a controller which performs a con ^¬ trol based on the received sensor data.

The controller may receive the measured sensor data and pro ^¬ cess the sensor data for a control of a related system. For example, the related system may be an industrial system like a gas turbine or any other system which may be controlled based on the received sensor data. The controller may use a numerical model of the system in order to analyze the meas ^¬ ured sensor data. For example, the controller may apply the measured sensor data to the model of the system to compute predictions of the system. Further, it may be also possible to compute one or more parameters for controlling the system by applying the measured sensor data to the model of the sys ^¬ tem.

For intelligent industrial systems, the model of the system may be automatically improved over the time. For this pur ^¬ pose, the measured sensor data may be stored in the sensor database and the stored measured sensor data may be analyzed by the model generator in order to generate a model of the system or to improve an already existing model of the system. Especially, the model of the system may be characterized by a number, i.e. one or more, of parameters. In this case, the model generator may compute or adapt the respective parame ^¬ ters in order to further improve the model of the system. For this purpose, any appropriate algorithm may be based, e.g. machine learning algorithms, for instance using artificial intelligence. However, it is understood, that any other meth ^¬ od for computing or adapting the model parameters may be pos ^¬ sible, too. Since the computational resources of the model generator in a locally distributed device usually are limited, the respec ^¬ tive algorithms for computing the model parameters may be very simple. Furthermore, only a small number of measured sensor data may be taken into account for computing the model parameters. In other words, the scheme for computing the mod ^¬ el parameters in the local device applies a computation of the model parameters based on small computational resources. For example, only sensor data of a limited time period, or a limited number of measured sensor data may be taken into ac- count. Furthermore, it may be possible to limit the computa ^¬ tion of the model parameters in the local device only to par ^¬ ticular measured sensor data. For this purpose, only sensor data fulfilling predetermined conditions, like, e.g., a pre ^¬ determined deviation, a predetermined range or any other pre- determined characteristic may be taken into account. However, it is understood that any other scheme for selecting the measured sensor data which are used for computing the model parameters may be possible, too. The model generator in the local device may compute a so- called online model and store the computed online model or the parameters characterizing the respective online model in the model database. Accordingly, the controller may refer to the model parameters stored in the model database and apply a model of the system based on the model parameters stored in the model database for processing the measured sensor data. In this way, it is possible to continuously adapt the parame- ters of the system model based on the measured sensor data. However, due to the limited computational resources in the local device, the quality of the computed model parameters may be limited. Thus, to further improve the model of the system, especially the parameters for characterizing the mod- el of the system, a further model may be computed by an ex ^¬ ternal device having enlarged computational resources. For example, such an external device may be located in a cloud or a datacenter. Accordingly, the measured sensor data have to be sent from the local device to the external device, and the result of the computation of the enhanced model may be transferred back from the external device to the local device. To limit the amount of data which has to be transferred between the local device and the external device, a data filter filters the sensor data and only forwards relevant sensor data from the local device to the external device. For this purpose, any predetermined filtering scheme may be applied. Moreover, it may be also possible to forward all measured sensor data from the local device to the external device. However, to limit the amount of data, the data filter may assess the measured sensor data and identify the sensor data which may be rele ^¬ vant for computing or enhancing the model of the system. For example, the data filter may only send data to the external device, if a value of the sensor data is different from the value of previously obtained sensor data. Furthermore, it may be also possible to forward the sensor data only if the dif ^¬ ference between the current sensor data and the previously measured sensor data exceeds a predetermined threshold value. However, any other scheme for selecting/filtering the sensor data, which are to be forwarded to the external device may be possible, too. For example, it may be also possible to deter- mine, whether or not the respective sensor data may lead to a modified model parameter when using the respective sensor da ^¬ ta for computing the model parameters based on the model scheme used in the model generator of the local device. If the respective sensor data will not lead to a change of the model parameters in the online model computed in the local device, the respective sensor data may be also not forwarded to the external device. Furthermore, the respective sensor data may be also deleted in the sensor database of the local device .

The external device may receive the sensor data forwarded by the local device and store the received sensor data in a fur ^¬ ther sensor database. Accordingly, the sensor data in the further sensor database may correspond to the sensor data in the sensor database of the local device. In this way, the ex ^¬ ternal device, especially in a cloud or a datacenter, may be in the position to compute the online model parameters com ^¬ puted by the local device by applying a same scheme for com ^¬ puting the model parameters as used in the local device. Fur- thermore, the external device may apply a further scheme for computing an enhanced system model. For this purpose, en ^¬ larged computational resources may be used. Hence, the accu ^¬ racy of the system model by applying the second scheme for computing the system model, especially the parameters of the system model can be improved. If the external device recog ^¬ nizes that there is a difference, especially a significant difference exceeding a predetermined threshold, the external device may send the respective model parameters of the en ^¬ hanced system model to the local device. Accordingly, the lo- cal device may receive the enhanced model parameters and re ^¬ place or add the enhanced parameters of the system model in the model database of the local device. Accordingly, the present invention can achieve online machine learning by computing parameters of a system model locally in the device. Furthermore, an enhanced system model can be com ^¬ puted by means of huge computational resources of a cloud or a datacenter. Hence, the local device can use a model of the system which is based on intelligent machine learning algo ^¬ rithms without the need of huge an expensive computational resources in the local device.

In a possible embodiment the date filter filters the sensor data having a predetermined impact on the computation on the model parameters by the model generator. By taking into ac ^¬ count the impact of the respective sensor data on the result of the model parameters, it can be easily determined whether or not the respective sensor data may be relevant for compu ^¬ ting the model parameters. Accordingly, the amount of data which is transferred between the local device and the exter ^¬ nal device can be further reduced. Especially, it can be de ^¬ termined whether the model parameters have been changed, or whether a mathematical difference between previous model pa ^¬ rameters and the model parameters which are computed by the respective sensor data exceeds a predetermined threshold val ^¬ ue. The mathematical difference may be computed by any mathe ^¬ matical distance measure, e.g. it may be determined whether a difference between the first model parameters and the second model parameters is greater than a predetermined threshold. Further, is may be possible to determine whether a difference of calculated values, when applying the model parameters of historical data from the sensor database, is greater than a predetermined value in a mathematical norm, especially a mathematical norm. Furthermore, it may be also possible to determine if the result of a model function by applying the respective model parameters changes more than a predetermined threshold .

In a possible embodiment the model generator is adapted to notify the data filter if the computed model parameters are different from the model parameters stored in the model data- base. In this way, the data filter can easily recognize, whether or not the respective sensor data shall be forwarded to the external data analyzing device. In a possible embodiment, the data filter is adapted to de ^¬ lete sensor data in the sensor database based on a predeter ^¬ mined deletion strategy.

In a possible embodiment, the deletion strategy may comprise identifying sensor data of an uncompleted computation of model data, sensor data exceeding a predetermined aging, i.e. the sensor data have been determined before a predetermined point of time, sensor data having an impact to the computa ^¬ tion of the model data which is lower than a predetermined threshold value and/or sensor data having a value outside a predetermined value range.

By deleting the respective sensor data in the sensor data ^¬ base, the amount of data which are taken into account for computing the model parameters in the data processing device can be limited. Furthermore, the external data analyzing de ^¬ vice may be also informed about the deletion of the sensor data in the sensor database of the data processing device. Accordingly, the sensor database in the data analyzing device may be adapted accordingly, by also deleting the respective sensor data.

In a possible embodiment of the data analyzing device, the first model scheme which is applied by the first model gener- ator corresponds to the model scheme which is applied by the model generator of the data processing device. Accordingly, the data processing device and the data analyzing device both apply the same scheme and accordingly, the data analyzing de ^¬ vice knows the model parameters which are used in the data processing device without the need that the respective model parameters have to be transferred between the data processing device and the data analyzing device. In a possible embodiment, the second model generator performs the computation of the second model parameters with predeter ^¬ mined time intervals or each time a predetermined number of sensor data are received by the receiver of the data analyz ^¬ ing device. By limiting the computation of the second model parameters to a predetermined condition, the computational load of the data analyzing device for computing the second model parameters can be also limited.

In a possible embodiment, the computational load of the com ^¬ puting the second model parameters is greater than the compu ^¬ tational load for computing the first model parameters. Ac ^¬ cordingly, the scheme for computing the second model parame ^¬ ters may be more complex, and thus more precise. Hence, the second model parameters provide a more detailed specification of the system model. On the other hand, since the scheme for computing the model parameters, which is applied in the data processing device is small, only small computational re ^¬ sources are required for the model generator of the data pro ^¬ cessing device.

In a possible embodiment the data analyzing device sends the second model parameters to the data processing device only if a predetermined condition is fulfilled. The predetermined condition may comprise comparing the mathematical difference between the replicated model parameters in the model database of the data analyzing device and the enhanced model parame ^¬ ters, which are computed based on the second model scheme. Furthermore, the predetermined condition may comprise compu ^¬ ting a mathematical difference of the model functions using the respective model parameters.

In a possible embodiment of the method, the filtering step of the sensor data may comprise identifying sensor data having an impact to the computation of the online model data.

In a possible embodiment of the method, the method may fur ^¬ ther comprise a step of deleting sensor data in the sensor database of the data processing device and/or the data ana ^¬ lyzing device. Especially, sensor data fulfilling a predetermined condition may be deleted.

Brief description of the drawings

For a more complete understanding of the present invention and advantages thereof, reference is now made to the follow ^¬ ing description taking in conjunction with the accompanying drawings. The invention is explained in more detail below us ^¬ ing exemplary embodiments which are specified in the schemat ^¬ ic figures of the drawings, in which:

Fig. 1 shows a block diagram of an embodiment of a data processing system according to the present invention; and

Fig. 2 shows a flow diagram of an embodiment of a method for processing data according to the present invention .

The appended drawings are intended to provide further under ^¬ standing of the embodiments of the invention. The illustrated embodiments and, in conjunction with the description, help to explain principles and concepts of the invention. Other em ^¬ bodiments and many other advantages mentioned become apparent in view of the drawings.

Detailed description of the drawings

Fig. 1 shows a schematic block diagram of a data processing system according to an embodiment. The data processing system comprises a data processing device 10, in the following de ^¬ noted as edge device 10, and a data analyzing device 20, in the following denoted as cloud or datacenter 20. The edge de ^¬ vice comprises a number, i.e. one or more, of sensors 19, a sensor database 11, a model database 12, a model generator 13, controller 14, a data filter 15 and a receiver 16. The datacenter 20 comprises a sensor database 21, a first model generator 23, a second model generator 24, a data analyzer 25 and a receiver 26. Accordingly, one or more sensors 19 may measure values relat ^¬ ing to sensed parameters and provide corresponding sensor da ^¬ ta. For example, sensor 19 may measure an environmental pa ^¬ rameter, a state or another parameter related to a system, especially an industrial system. For instance, sensor 19 may measure a temperature, a humidity, a pressure, a speed, an acceleration, a direction, an intensity, a volume flow, a position, an angle, or any other parameter which may be sensed. Accordingly, a value corresponding to the sensed parameter may be directly provided by a digital value. Alternatively, the sensed parameter may be provided as an analogue signal and converted to a digital signal by an analogue to digital converter. Furthermore, any other processing of the sensed parameters, for instance a filtering, averaging over a prede ^¬ termined time period, etc. may be also performed. For exam- pie, the sensed measurements may be provided by a digital connection or a network like an Ethernet network, an industrial bus system, etc. Especially, the sensed measurements may be provided as sensor data and received by a sensor in ^¬ terface (not illustrated) . For example, the sensor interface may receive the sensor data and forward the sensor data to the sensor database 11 and/or the processor 14 of the edge device 10. Especially, the sensor interface may also buffer the sensor data or perform any other operation with the sensor data.

The measured sensor data may be provided to processor 14. Ac ^¬ cordingly, processor 14 may process the measured sensor data in order to allow a control of a related system. Especially, controller 14 may use a system model like a numerical model of a related system, in order to analyze and process the sen ^¬ sor data provided by a sensor 19. For example, controller 14 may compute a prediction or estimation of one or more prede ^¬ termined characterizing parameters of the system which is to be controlled based on the measured sensor data in associa ^¬ tion with a related system model. For this purpose, control ^¬ ler 14 may refer to model parameters stored in model database 12. For example, the model parameters stored in model data- base 12 may specify the parameters of a model function de ^¬ scribing the related system which shall be controlled. In an example, the system to be controlled may be a gas turbine, and the related model may be a model which describes the tem ^¬ perature characteristics of this gas turbine. Accordingly, a prediction of the temperature behavior of the gas turbine may be determined based on the measured sensor data and the re ^¬ lated model. However, it is understood, that the present in ^¬ vention may be also applied to any other system, especially any other industrial system, which can be modeled according- ly.

The model for characterizing the related system or at least one or more characterizing parameters of the system may be any appropriate system. For example, the model may be in a very simple case a function, for example a linear function, a function with multiple coefficients, etc. describing the properties of the system in connection with the measured sen ^¬ sor data. In an example, the model may describe a correlation between one or more measured sensor data like a pressure, ro- tational speed etc. and an associated development of a tem ^¬ perature. However, it is understood that any other modeling of characteristic parameters of a system depending on meas ^¬ ured sensor data may be also possible. Especially, the model of the system may be adapted during the life time of the sys- tern. For example, the model of the system may be adapted by a machine learning algorithm or the like. For example, a prede ^¬ termined function may be used which can be adapted by modify ^¬ ing one or more parameters of the related model function. However, any other scheme for modeling the related system may be also possible.

The model of the system, especially the parameters of such a model may be regularly or continuously adapted. For this pur- pose, the measured sensor data may be analyzed and the param ^¬ eters for specifying the related model may be computed based on previously measured sensor data. For this purpose, the measured sensor data provided by sensors 19 may be locally stored in a sensor database 11 of the edge device 10. In or ^¬ der to limit the amount of data which is stored in the sensor database 11 of the edge device 10, it may be possible to lim ^¬ it the data to the data of a predetermined time period, or to a predetermined number of sensor data. However, any other scheme for limiting the sensor data stored in sensor database 11 may be also possible.

Model generator 13 of the edge device 10 may read the sensor data stored in sensor database 11. Based on the sensor data stored in the sensor database 11, model generator 13 may com ^¬ pute model parameters of the system model. Especially, a par ^¬ ticular first model scheme may be used for computing the mod ^¬ el parameters of the system by model generator 11. Since the computational resources of the edge device 10, especially of the model generator 13 are limited, model generator 13 may apply only a simple scheme for computing the model parame ^¬ ters, wherein the scheme requires only small computational resources like CPU load, memory, etc. The computed parameters of the model may be stored in model database 12 of the edge device 10. In this way, the edge device 10 may compute param ^¬ eters for the system model directly in the edge device 10 without the need of any external resources. Hence, the re ^¬ spective model parameters are immediately available. For ex ^¬ ample, the computation of the model parameters may be contin- uously performed. Thus, the model parameters may be adapted almost in real-time. Furthermore, it may be also possible to initiate the computation of the model parameters by the model generator 13 based on predetermined conditions. For example, a further computation of model parameters by model generator 13 may be initiated upon predetermined condition is met. For example, the predetermined condition may be a reception of a predetermined number of new measured sensor data, a reception of measured sensor data fulfilling a predetermined condition, for example a measured sensor data exceeding a predetermined value, etc. However, it is understood that any other condi ^¬ tion for initiating a computation of model parameters may be also applied.

When computing the model parameters by a model generator 13, it may be possible to determine whether or not a particular sensor data has an impact to the computed model parameters. For example, it may be determined that the computed model pa- rameters are the same or almost the same as the previously computed model parameters, even though further sensor data have been stored in sensor database 11 and the further sensor data are also used for computing the model parameters. If it is detected that the newly added sensor data do not have any impact to the computed model parameters, such sensor data may be deleted in sensor database 11. Furthermore, data filter 15 may be informed about the impact of the sensor data to the computation of the model parameters. Data filter 15 may analyzes the sensor data stored in sensor database 11 in order to determine whether the respective sen ^¬ sor data may be forwarded to the datacenter 20. For this pur ^¬ pose, any appropriate scheme for filtering or analyzing the data stored in sensor database 11 may be applied. For exam- pie, data filter 15 may only send the sensor data to datacen ^¬ ter 20 if the respective sensor data are relevant for the computation of the model parameters by model generator 13 of the edge device 10. If model generator 13 does not use the respective sensor data stored in sensor database 11 of the edge device 10, such sensor data are not sent to the datacen ^¬ ter 20 by data filter 15. Furthermore, data filter 15 may al ^¬ so not send such sensor data to edge device 20 which do not have any impact on the model parameters when computing the model parameters by model generator 13 of the edge device 10. However, it is understood that any other or further scheme for selecting the sensor data which shall be send from the edge device 10 to the datacenter 20 may be also possible. Es ^¬ pecially, data filter 15 may send all sensor data of the sen- sor database 11 to the edge device 20 which are required for obtaining the model parameters computed by a model generator 13. The following description refers to datacenter 20. Receiver 26 of the data center 20 receives the sensor data sent by da ^¬ ta filter 15 of the edge device 10. Further, receiver 26 forwards the received sensor data to the sensor data base 21 of the datacenter 20. Since data filter 15 forwards all sensor data which are required for obtaining the model parameters computed by model generator 13 of the edge device 10, sensor database 21 of the datacenter 20 also comprise all the rele ^¬ vant sensor data for computing these model parameters. Ac ^¬ cordingly, the first model generator 23 of the datacenter 20 is in the position to compute the same model parameters as computed by the model generator 13 of the edge device. For this purpose, the model generator 23 of the data center 20 applies the same model scheme as used by the model generator 13 of the edge device 10. Further to this, datacenter 20 may compute an enhanced model of the related system by a second model generator 24. Since datacenter 20 comprises huge compu ^¬ tational resources, the second model generator 24 may apply a more complex scheme for computing model parameters. Especial ^¬ ly, a larger number of sensor data may be used and/or a more complex scheme for computing the model parameters may be ap ^¬ plied to the sensor data to compute the respective model pa ^¬ rameters. Accordingly, a more detailed and precise modeling of the system can be achieved based on the enhanced model pa ^¬ rameters computed by the second model generator 24 of the datacenter 20.

In order to minimize the computational load of the data cen ^¬ ter 20, the computation of the enhanced model parameters by the second model generator 24 may be limited to predetermined conditions. For example, the computation of the enhanced mod ^¬ el parameters may be performed at predetermined time inter ^¬ vals, for example once per minute, once per hour, once per day, etc. Furthermore, the computation of the enhanced model parameters by the second model generator 24 may be also ini ^¬ tiated upon receiving a predetermined number of new sensor data. However, it is understood that any other criteria for initiating a computation of the enhanced model parameters by the second model generator 24 may be also applied.

After the second model generator 24 has computed enhanced model parameters, the enhanced model parameters may be com ^¬ pared with the model parameters computed by the first model generator 23 of the datacenter 20. As already mentioned above, the model parameters computed by the first model gen ^¬ erator 23 correspond to the model parameters computed in the edge device 10. If the enhanced model parameters computed by the second model generator 24 are different from the model parameters computed by the first model generator 23, data an ^¬ alyzer 25 may send the enhanced model parameters to the edge device 10. For this purpose, data analyzer 25 may compare the enhanced model parameters computed by the second model gener ^¬ ator 24 with the model parameters computed by the first model generator 23. Alternatively, it may be also possible to com ^¬ pute the result of a model function applying the enhanced model parameters and the model parameters computed by the first model generator 23, and to compare the result of the respective functions. If the result exceeds a predetermined threshold value, the enhanced model parameters may be sent to the edge device 10. Accordingly, it is only necessary to send the enhanced model parameters from the datacenter 20 to the edge device 10 if the enhanced model parameters are different from the model parameters computed by the first generator 23, which correspond to the model parameters already stored in the model database 12 of the edge device 10. Accordingly, an unnecessary transfer of data can be avoided.

The edge device 10 may receive the enhanced model parameters from the data center 20 by receiver 16. Accordingly, receiver 16 may store the received enhanced model parameters in model database 10 of the edge device 10. The received enhanced mod ^¬ el parameters may be stored in addition to the model parame- ters computed by model generator 13 of the edge device. Al ^¬ ternatively, the received enhanced model parameters may re ^¬ place all the model parameters previously stored in the model database 13 of the edge device 10.

Fig. 2 shows a flow diagram of a method for processing data underlying an embodiment. The features of the method in Fig. 2 corresponds to the feature performed by the components of the data processing system as already described in associa- tion with this Fig. 1. Hence, all operations described above in connection with the data processing system may be also performed by the method as described in the following, and accordingly, all steps performed by the following method may be also executed by the data processing system described above. For sake of clarity, reference numerals mentioned above in connection with Fig. 1 will be maintained for the following description of the method.

In step SI sensor data are measured, especially by one or more sensors 19. In step S2, the method sensor data are stored in a sensor database 11 of the edge device 10. In step S3, the edge device 10, especially the model generator 13 of the edge device 10 computes online model parameters of a sys ^¬ tem model. The online model parameters are computed based on the sensor data stored in the measurement database 11 of the edge device 10 by applying a first model scheme. In step S4 the computed online model parameters of the system are stored in a model database 12 of the edge device 10. In step S5, the sensor data which are stored in the sensor database 11 of the edge device 10 are filtered by the edge device 10, especially by data filter 15 of the edge device 10. In step S6 the fil ^¬ tered sensor data are forwarded from the edge device 10 to the datacenter 20. In step S7, further model parameters from the datacenter 20 are received by the edge device 10, espe- cially by receiver 16 of the edge device 10. In step S8, the received further model parameters are stored in a model data ^¬ base 12 of the edge device 10. In step S9 the measured sensor data are processed based on the model parameters stored in the model database 12 of the edge device 10.

The method may further comprise the steps of receiving the forwarded filtered sensor data by the edge device 20, storing the forwarded filtered sensor data in a further model data ^¬ base 21 of the edge device 20, computing replicated model pa ^¬ rameters corresponding to the online model parameters in the edge device 10, computing further model parameters based on a second model scheme in the datacenter 20, and sending the second model parameters from the datacenter 20 to the edge device 10 if a mathematical difference between the replicated model parameters and the further model parameters is greater than a predetermined threshold.

Especially, the method may send the further model parameters to the edge device 10, if a mathematical difference between the replicated model parameters and the further model parame ^¬ ters exceeds a predetermined threshold and/or a difference of a model function using the replicated model parameters and the model function using the further model parameters exceeds a predetermined threshold.

The filtering of the sensor data in the edge device 10 may comprise identifying sensor data having an impact to the computation of the online data model. Accordingly, sensor data which do not have any impact to the parameters of the online model data are filtered out and not forwarded to the datacen ^¬ ter 20.

The method may further comprise deleting sensor data in the sensor database 11 of the edge device 10 for the sensor data- base 21 of the data center 20 if a predetermined condition is fulfilled .

Summarizing, the present invention relates to an enhanced computation of a data model for an intelligent data pro ^¬ cessing device. The data processing device may be a device having limited computational resources. Accordingly, a system model for processing the data is computed in the local de ^¬ vice. Additionally, an enhanced model may be computed in a remote device like a cloud or a data center. For this pur- pose, the cloud or datacenter is provided with filtered data for computing an enhanced model. The cloud or datacenter may compute an enhanced model and forward the respective model to the local device if the enhanced model is better than the model locally generated.