Computer-implemented method and system for operating a technical device
Patent Information
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- SIEMENS AG
- Filing Date
- 2024-07-10
- Publication Date
- 2026-05-20
Smart Images
Figure EP2024069537_06032025_PF_FP_ABST
Abstract
Description
[0001] Computer-implemented method and system for operating a technical device
[0002] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.
[0003] The invention relates to a computer-implemented method and a system for operating a technical device by a client of a client-server system with a model based on artificial intelligence, wherein the client-server system has a set of clients, and the clients each provide a local model for operating a connected technical device, and the local models of the respective clients are independent of one another.
[0004] The invention further relates to a computer program product.
[0005] Federated learning describes a machine learning technique in which a model is trained on multiple devices. Each participating device has its own local dataset that is not shared with other participants. In contrast, traditional machine learning uses a central dataset.
[0006] Using this technique, multiple participants build a common, robust model that addresses critical issues such as privacy, data security, access rights, and access to heterogeneous data. The technique is applied in many areas, including defense, telecommunications, the Internet of Things, and healthcare.
[0007] In other words, federated learning (FL) is a machine learning (ML) technique in which an artificial intelligence (AI) model is trained across multiple decentralized edge devices of a client-server system using local data samples, without exchanging the AI model itself. To ensure good performance of a final, centralized ML model, FL relies on an iterative process divided into an atomic set of client-server interactions called an FL round. Each round of this process consists of propagating the current global model state to participating edge devices, training local models on these local edge devices to generate a set of potential model updates on each edge device, and then aggregating and processing these local updates into a single global update and applying it to the global model.
[0008] State-of-the-art often assumes synchronized model updates. In large, distributed environments, edge devices with different hardware capabilities can typically participate in FL rounds. The speed at which distributed FL training converges depends on the slowest client device. Therefore, the entire process could be slowed down because a single device might have poor hardware resources compared to its peers participating in FL training. In practice, FL training can fail due to slow response times, timeouts, and the blocking of hardware resources for long periods of time.
[0009] State-of-the-art distributed training scenarios often assume the use of homogeneous devices with the same hardware capabilities. This is especially true for cloud-based data centers.
[0010] It is known in the art to integrate client contributions as soon as they are available through asynchronous federated learning. Therefore, the object of the invention is to provide a solution for systems with different hardware resources that allows for efficient training of an ML / AI model.
[0011] The object of the invention is achieved by a system of the type mentioned at the outset, wherein the following steps are carried out: a. Determining a respective performance parameter for each client from the set of clients, which describes the technical performance capability of the respective client, b. Determining the mean value of the performance parameters of all clients from the set of clients, and determining a first client group from the set of clients who have a performance parameter at a level at least equal to the mean value of the performance parameters, c. Determining the standard deviation of the performance parameters of all clients, and determining a second client group from the set of clients whose standard deviation of the performance parameters lies in a predefined value range and additionally have a performance parameter at a level below the mean value of the performance parameters, d.Determining a respective quality parameter which describes the technical parameters of the respective client for training the respective local model, for at least those clients which are not assigned to the first client group and not to the second client group, and determining a third client group from the set of clients which have a respective quality parameter at a level of at least one predetermined threshold value, e . Generating and training a global model with model parameters of the respective local models of clients of the first, the second and the third client group according to the principle of federated learning, and operating the technical device by the connected client using the global model.
[0012] This ensures that edge devices can be advantageously selected which form a training group with which efficient training is ensured.
[0013] The clients each maintain a local model for operating a connected technical device, which has been trained independently by another client.
[0014] In other words, the respective groups are trained separately from each other in order to take into account a specific adaptation to the respective requirements of an individual client and thus to obtain a particularly accurate local model.
[0015] This ensures that the individual client models of the respective clients are more accurate, and subsequently the global model is also more accurate by merging the local models into the global model in accordance with the invention.
[0016] The improved accuracy can, for example, lead to increased training efficiency, such as by avoiding waiting times for slow clients.
[0017] In a further development of the invention, it is provided that the respective performance parameter of the respective client describes the number and type of processors and / or the size of the RAM and / or the size of the hard disk storage and / or the transmission properties of the data connection between the respective client and the server. This makes it easy to determine a static performance parameter for a client that appropriately represents the technical performance of the client.
[0018] In a further development of the invention, it is provided that the determination of the respective performance parameter of the respective client takes into account a current utilization of the technical resources of the respective client.
[0019] This makes it easy to determine a dynamic performance parameter for a client that appropriately represents the client’s technical performance
[0020] In a further development of the invention, it is provided that the respective quality parameter of the respective client describes the size and / or the variance and / or the number of labels of the data of the respective local model.
[0021] This makes it easy to determine a quality parameter for a client's ML / AI model that appropriately represents the technical complexity of the client's local model.
[0022] In a further development of the invention, it is provided that the determination of the respective quality parameter in
[0023] Step d) is carried out for all clients, and at least two clients selected from the first client group and / or the second client group and / or the third client group are replaced by a client without a group assignment if the quality parameters of the selected at least two clients together correspond at least to the quality parameter of the client without a group assignment.
[0024] The quality parameters of the selected at least two clients can, for example, be added together.
[0025] This makes it easy to find a client with low
[0026] Performance or high complexity in the local model in the system must be taken into account when determining a global model according to the federated approach.
[0027] In a further development of the invention, it is provided that the predefined value range lies at the simple standard deviation from or around the mean value.
[0028] This makes it easy to create an advantageous criterion for group assignment of clients, which identifies and excludes inefficient clients in federated learning.
[0029] The object according to the invention is also achieved by a system for operating a technical device by a client of a client-server system with a model based on artificial intelligence, wherein the client-server system has a number of clients, each with a processor and each with a memory, wherein the clients are each set up to keep a local model for operating a connected technical device, and the system is further set up to carry out the method according to the invention.
[0030] The object according to the invention is also achieved by a computer program product having machine-readable instructions stored therein which, when executed by a processing unit, cause the latter to carry out the method according to the invention.
[0031] The invention is explained in more detail below with reference to an embodiment shown in the accompanying drawings. In the drawings:
[0032] Fig. 1 shows an example of a client-server system architecture,
[0033] Fig. 2 shows an embodiment of the inventive
[0034] Procedure, Fig. 3-4 a statistical representation of client response times.
[0035] Fig. 1 shows an example of a client-server system architecture.
[0036] The system is used to operate a technical device TD1 by a client CI of the client-server system.
[0037] The operation is carried out with an LM1 model based on artificial intelligence.
[0038] The client-server system comprises a set of clients C1-C3, each with a processor and a memory.
[0039] The clients C1-C3 are each configured to maintain a local model LM1-LM3 for operating a connected technical device TD1-TD3.
[0040] The local models LM1-LM3 of the respective clients C1-C3 are independent of each other, i.e. they were each trained separately with individual training data, which allows for specific adaptation to the respective requirements of an individual client, for example special operating parameter ranges when controlling or monitoring during the operation of the connected technical device TD1-TD3.
[0041] The individual training data can be derived and provided from the ongoing operation of a respective technical device TD1-TD3.
[0042] The system is further configured to carry out the method described in more detail below.
[0043] Fig. 2 shows a simplified embodiment of the method according to the invention, which can be carried out on the system according to the previous figure.
[0044] The following steps are carried out: a. Determining a respective performance parameter PP1-PP3 for each client C1-C3 from the set of clients, which describes the technical performance of the respective client C1-C3, b. Determining the mean value M of the performance parameters PP1-PP3 of all clients C1-C3 from the set of clients, and determining a first client group G1 from the set of clients who have a performance parameter PP1-PP3 at a level at least equal to the mean value M of the performance parameters, c. Determining the standard deviation SD of the performance parameters PP1-PP3 of all clients, and determining a second client group G2 from the set of clients whose standard deviation SD of the performance parameters lies in a predefined value range PR, and additionally have a performance parameter PP1-PP3 at a level below the mean value M of the performance parameters PP1-PP3, d.Determining a respective quality parameter QP1-QP3, which describes the technical parameters of the respective client C1-C3 for training the respective local model LM1-LM3, for at least those clients which are not assigned to the first client group Gl and not to the second client group G2, and determining a third client group G3 from the set of clients which have a respective quality parameter QP1-QP3 at a level of at least one predetermined threshold value, e. Creating and training a global model GM with model parameters of the respective local models LM1-LM3 of clients C1-C3 of the first, second and third client groups G1-G3 according to the principle of federated learning, and operating the technical device TD1 by the connected client CI using the global model GM.
[0045] In other words, only high-performing clients and clients with sufficient data quality are considered in federated learning.
[0046] This ensures that the performance of the system is maintained by only considering favorable clients when building the global model GM .
[0047] The performance parameters PP1-PP3 and the quality parameters QP1-QP3 can, for example, lie in a range from 1 . 0 to 9 . 9, where 1 . 0 is the lowest value and 9 . 9 is the highest possible value, and a low value stands for low performance or low quality, and a high value stands for high performance or high quality.
[0048] In other words, the respective performance parameter PP1-PP3 is directly proportional to the technical performance of the respective client.
[0049] The performance parameters PP1-PP3 are determined by the component of the respective client that has the lowest performance, such as processor, memory, etc.
[0050] The respective performance parameter PP1-PP3 of the respective client C1-C3 preferably describes the number and type of processors and / or the size of the main memory and / or the size of the hard disk storage and / or the transmission properties of the data connection D between the respective client C1-C3 and the server S .
[0051] This takes into account the static properties of clients, such as computing capabilities, storage capabilities, or communication capabilities. The determination of the respective performance parameters PP1-PP3 of the respective client C1-C3 preferably also takes into account the current utilization of the respective client C1-C3's technical resources.
[0052] This allows for the dynamic properties of clients to be taken into account, such as current processor or memory utilization. This allows for resource availability to be considered.
[0053] The respective quality parameter QP1-QP3 of the respective client C1-C3 describes the size and / or the variance and / or the number of labels of the data of the respective local model LM1-LM3.
[0054] The determination of the respective quality parameter QP1-QP3 in step d) is optionally carried out - shown as step f) - for all clients C1-C3.
[0055] At least two clients, selected from the first client group Gl and / or the second client group G2 and / or the third client group G3, are replaced by a client without a group assignment if the quality parameters of the selected at least two clients together correspond at least to the quality parameter of the client without a group assignment.
[0056] In other words, a client with low performance or low data quality in the respective local model can still be considered in federated learning if favorable clients that meet the performance and quality requirements are not considered in the global model.
[0057] This can be particularly advantageous if a client with low performance or low data quality nevertheless contains particularly relevant data and should therefore be considered. Figure 3 shows a representation of the response time of clients and criteria for assignment to the first client group.
[0058] The figure shows the statistical distribution of reaction times of clients C1-C3.
[0059] Clients with poor, i.e. long, response times have poor performance parameters.
[0060] The standard deviation SD of the performance parameters PP1-PP3 of all clients C1-C3 from the set of clients around the mean value M of the performance parameters PP1-PP3 is shown.
[0061] The first client group Gl is determined from the set of clients by assigning only those clients to the group who have a performance parameter PP1-PP3 at a level at least equal to the mean value M of the performance parameters.
[0062] Fig. 4 shows a further development of the previous figure with a representation of the reaction time of clients and criteria for assignment to the second client group.
[0063] The second client group G2 is determined from the set of clients by only assigning those clients to the group who have at least a value of one as the standard deviation SD of the performance parameters PP1-PP3 and who additionally have a performance parameter PP1-PP3 at a level below the mean value M of the performance parameters PP1-PP3.
[0064] It is assumed that a simple standard deviation SD from the mean M is still acceptable for clients C1-C3.
[0065] Therefore, the predefined value range PR is the simple standard deviation SD from the mean M . Reference symbols:
[0066] C1-C3 clients
[0067] D Data connection
[0068] G1-G3 client group
[0069] GM global model
[0070] LM1-LM3 local model
[0071] M mean
[0072] PP1-PP3 performance parameters
[0073] PR predefined range
[0074] QP1-QP3 quality parameters
[0075] S Server
[0076] SD standard deviation
[0077] TD1-TD3 technical device
Claims
Patent claims 1. Computer-implemented method for operating a technical device (TD1) by a client (CI) of a client-server system with a model (LM1) based on artificial intelligence, wherein the client-server system has a set of clients (C1-C3), and the clients (C1-C3) each have a local model (LM1-LM3) for operating a connected technical device (TD1-TD3), and the local models (LM1-LM3) of the respective clients (C1-C3) are independent of one another, and the following steps are carried out: a. Determining a respective performance parameter (PP1-PP3) for each client (C1-C3) from the set of clients, which describes the technical performance of the respective client (C1-C3), b.Determining the mean value (M) of the performance parameters (PP1-PP3) of all clients (C1-C3) from the set of clients, and determining a first client group (Gl) from the set of clients who have a performance parameter (PP1-PP3) at a level at least equal to the mean value (M) of the performance parameters, c. Determining the standard deviation (SD) of the performance parameters (PP1-PP3) of all clients, and determining a second client group (G2) from the set of clients whose standard deviation (SD) of the performance parameters (PP1-PP3) lies in a predefined value range (PR) and additionally have a performance parameter (PP1-PP3) at a level below the mean value (M) of the performance parameters (PP1-PP3), d. Determining a respective quality parameter (QP1-QP3) which describes the technical parameters of the respective client (C1-C3) for training the respective local model (LM1-LM3), for at least those clients. which are not assigned to the first client group (Gl) and not to the second client group (G2), and determining a third client group (G3) from the set of clients which have a respective quality parameter (QP1-QP3) at a level of at least one predetermined threshold value, e. generating and training a global model (GM) with model parameters of the respective local models (LM1-LM3) of clients (C1-C3) of the first, the second and the third client group (G1-G3) according to the principle of federated learning, and operating the technical device (TD1) by the connected client (Gl) using the global model (GM).
2. Method according to the preceding claim, wherein the respective performance parameter (PP1-PP3) of the respective client (C1-C3) describes the number and type of processors and / or the size of the main memory and / or the size of the hard disk storage and / or the transmission properties of the data connection (D) between the respective client (C1-C3) and the server (S).
3. Method according to one of the preceding claims, wherein the determination of the respective performance parameter (PP1-PP3) of the respective client (C1-C3) takes into account a current utilization of the technical resources of the respective client (C1-C3).
4. Method according to one of the preceding claims, wherein the respective quality parameter (QP1-QP3) of the respective client (C1-C3) describes the size and / or the variance and / or the number of labels of the data of the respective local model (LM1-LM3).
5. Method according to one of the preceding claims, wherein the determination of the respective quality parameter (QP1-QP3) in Step d) is carried out for all clients (C1-C3), and at least two clients selected from the first client group (Gl) and / or the second client group (G2) and / or the third client group (G3) are replaced by a client without a group assignment if the quality parameters of the selected at least two clients together correspond at least to the quality parameter of the client without a group assignment.
6. Method according to one of the preceding claims, wherein the predefined value range (PR) is the simple standard deviation (SD) from the mean value (M).
7. System for operating a technical device (TD1) by a client (G1) of a client-server system with a model (LM1) based on artificial intelligence, wherein the client-server system has a set of clients (C1-C3), each with a processor and each with a memory, wherein the clients (C1-C3) are configured to each maintain a local model (LM1-LM3) for operating a connected technical device (TD1-TD3), and the system is further configured to carry out the method according to one of the preceding claims.
8. A computer program product having machine-readable instructions stored therein which, when executed by a processing unit, cause the processing unit to carry out the method according to any one of the preceding claims.