Method and system for operating a technical device
By using performance indicators to optimize training data and processes for heterogeneous clients, the method enhances federated learning efficiency and reduces training time, addressing the challenges of varying hardware capabilities in distributed systems.
Patent Information
- Application Number
- PCT/EP2025/071985
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-08-28
- Filing Date
- 2025-07-30
- Publication Date
- 2026-03-05
AI Technical Summary
Existing federated learning (FL) methods in large-scale distributed environments with heterogeneous devices are slowed down by slow hardware resources, leading to long response times, timeouts, and resource blocking, and often fail to account for varying hardware capabilities, particularly in industrial applications.
A computer-implemented method that uses performance indicators to determine training data subset sizes and adjust training processes based on computing power, storage, communication capacity, and future availability of clients, ensuring efficient aggregation and updating of AI models across heterogeneous clients.
This approach reduces training time and improves the overall system performance by optimizing training for individual client capabilities, preventing idle processes and ensuring consistent model convergence.
Smart Images
Figure EP2025071985_05032026_PF_FP_ABST
Abstract
Description
[0001] Method and system for operating a technical device
[0002] The invention relates to a computer-implemented method for operating a technical device through a client-server system, comprising a server and clients, each of which is connected to a respective technical device.
[0003] In industry, more and more applications rely on the use of artificial intelligence (AI) or machine learning, for example in the control or operation of production plants, in industrial image processing for error detection, or in the monitoring of production processes for quality assurance or predictive maintenance of plants.
[0004] Federated learning (FL) is a machine learning technique in which an AI model is trained across multiple decentralized client or edge devices containing local data samples without exchanging them.
[0005] Especially in industrial applications, the use of edge systems is particularly important with regard to data protection and privacy requirements.
[0006] To ensure good performance of a final, centralized machine learning model, FL is based on an iterative process divided into an atomic set of client-server interactions known as 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 edges 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.
[0007] The previously discussed method assumes synchronized model updates. In large-scale distributed environments, edge devices with varying hardware capabilities can typically participate in FL sessions. The rate at which distributed FL training converges depends on the slowest client device.
[0008] Therefore, the entire process can be slowed down by a single device that may have poor hardware resources compared to its counterparts participating in FL training. In practice, FL training can fail due to long response times, timeouts, and the blocking of hardware resources for extended periods. The state of the art, particularly in distributed training scenarios, often assumes that homogeneous devices have the same hardware capabilities. This is especially true for cloud-based data centers.
[0009] Furthermore, current state of the art attempts to incorporate client contributions as soon as they become available through asynchronous federated learning.
[0010] In the publication PARK JLLWON ET AL: "AMBLE: Adjusting mini-batch and local epoch for federated learning with heterogeneous devices", JOURNAL OF PARALLEL AND DISTRIBUTED COMPUTING., Vol. 170, July 28, 2022 (2022-07-28), pages 13-23, XP093246418.NL, ISSN: 0743-7315, DOI: 10.1016 / j.jpdc.2022.07.009, a method is described which is an “Adjusting Mini-Batch and Local Epoch (AMBLE)” approach that adaptively adjusts the local mini-batch and local epoch size for heterogeneous devices in federated learning and synchronously updates the parameters, thereby increasing computational performance by removing outliers and scaling the local learning rate. to improve the convergence rate and accuracy of the model. The object of the invention is to provide a solution that efficiently improves the performance of AI models used by heterogeneous clients within a client-server system.
[0011] The problem according to the invention is solved by a computer-implemented method for operating a technical device by means of a client-server system, comprising a server and clients, each of which is connected to a respective technical device, wherein the respective technical device and / or the respective client provides a performance indicator for describing the technical performance in a predetermined future time interval, and the following steps are carried out: a) Performing a respective training for a provided respective local
[0012] A model for operating the respective technical device with at least a subset of training data provided by the respective client, the size of which is determined by the performance indicator; b) the respective client providing the local model to the server; c) the server aggregating the local models into a global model and providing the global model to the clients; d) the respective client receiving the global model and updating the local model using the previously received global model; e) the technical device operating using the previously updated local model. This reduces the training time for the global model and improves the availability of the entire heterogeneous system.
[0013] This can improve the performance of a system with heterogeneous clients.
[0014] The performance indicator describes, for example, the computing speed and / or the latency and / or the absolute computing time of the calculation, that is, of the respective training or a training epoch of the local model.
[0015] The performance indicator refers specifically to its computing power and / or storage capacity and / or communication capacity and / or availability of the technical equipment and / or the client, and not to the data to be processed for training or inference.
[0016] In a further development of the invention, it is provided that the performance indicator includes a technical parameter which describes the technical performance in the form of computing power and / or storage power and / or communication power of the respective technical device and / or the respective client.
[0017] This ensures that the performance of individual clients is taken into account within the system and that the performance of the overall system is not impaired by individual slow clients or clients who are not performing as well as others, and that idle processes in the system can be avoided or reduced.
[0018] In a further development of the invention, it is provided that the performance indicator includes a temporal parameter which describes the future temporal availability of the technical parameters.
[0019] By forecasting the future availability of individual clients, it is possible to ensure that the time-related performance of individual clients is taken into account within the system, and that the performance of the overall system is not impaired by individual slow clients or clients who are not performing as well as others, and that idle processes in the system can be avoided or reduced.
[0020] In a further development of the invention, it is provided that the future temporal availability of the technical parameters is described based on the past availability with regard to the technical parameters of the respective client.
[0021] This allows for a simple prediction of the future availability of individual clients. In a further development of the invention, the performance indicator includes a minimum size for the size of at least a subset of the training data provided for the respective client.
[0022] This ensures that slower clients are also taken into account in the system and are not suppressed by faster clients.
[0023] The problem according to the invention is also solved by a client-server system for operating a technical device, wherein the system comprises a server and clients, each of which is connected to the respective technical devices, and the method according to the invention is carried out.
[0024] The problem according to the invention is also solved by a computer program product with machine-readable instructions stored therein which, when executed by the client-server system according to the invention, cause it to execute the method according to the invention.
[0025] The invention is described in more detail in the following drawings using an exemplary embodiment. The figures show
[0026] Fig. 1 shows an embodiment of a client-server system for operating a technical device with an artificial intelligence-based model.
[0027] Fig. 2 shows an embodiment of the method according to the invention as a flowchart,
[0028] Fig. 3 shows an embodiment of the method according to the invention as an algorithm in
[0029] Form of pseudocode.
[0030] Fig. 1 shows an embodiment of a client-server system SYS for operating a technical device TD1 with a model LM1 based on artificial intelligence.
[0031] The system SYS has one server S and three clients C1-C3.
[0032] Each client C1-C3 is connected to a technical device TD1-TD3.
[0033] The procedure described below will be carried out.
[0034] The technical device TD1-TD3, for example a pump, a production machine, a robot, etc., can use integrated sensors to collect device data such as operating voltages, temperatures, or vibrations. This data is then analyzed by a monitoring device using an artificial intelligence-based model to, for example, monitor ongoing operation. The device can then be controlled by this artificial intelligence model, for example, in the form of control signals generated by the model and transmitted to a control device.
[0035] The technical device can, for example, also be an inspection device for a product produced by a production machine, whereby imaging sensors are used to detect the product and analyze its external condition, and the production machine is controlled accordingly by a control device.
[0036] The control device can have a processor and memory and, with the help of a communication module, receive and further process the calculated relevant model data.
[0037] Fig. 2 shows an embodiment of the method according to the invention as a flowchart.
[0038] The system SYS is based on the previous figure.
[0039] Each technical device TD1-TD3 provides a performance indicator to describe the technical performance within a specified future time interval.
[0040] The following steps are carried out: a) Conducting a respective training session for a provided respective local
[0041] a) Model LM1-LM3 for operating the respective technical device TD1-TD3 with at least a subset of training data provided by the respective client CI-CS, wherein the size of the at least one subset of training data is determined by the performance indicator, b) Provision of the respective local model LM1-LM3 to the server S by the respective client C1-C3, c) Aggregation of the local models LM1-LM3 into a global model GM by the server, and provision of the global model GM to the clients C1-C3, d) Receipt of the global model GM by the respective client C1-C3 and updating of the local model LM1-LM3 using the previously received global model GM, e) Operation of the technical device TD1 using the previously updated local model LM1-LM3.
[0042] The performance indicator includes a technical parameter that describes the technical performance in terms of computing power, storage capacity, and / or communication performance of the respective client C1-C3. The performance indicator also includes a temporal parameter that describes the future availability of the technical parameters.
[0043] The future temporal availability of the technical parameters can be described based on the past availability of the technical parameters of the respective client C1-C3.
[0044] The performance indicator may include a minimum size for the size of at least a subset of the training data provided for the respective client C1-C3.
[0045] Fig. 3 shows an embodiment of the method according to the invention as an algorithm in the form of pseudocode, in which clients are aggregated based on the size of at least a subset of the training data provided.
[0046] To achieve favorable overall system performance, the performance indicator can also be expressed and defined by applying an optimization goal to perform the calculation, i.e., the respective training or training epoch of the local model, for example, a desired total time, which is chosen so that each client has the opportunity to pass at least the minimum processed time.
[0047] In such a scenario, the desired total time can be selected as the median of the total time, for example 21 hours, which means that the slowest client can use up to 21 hours for training, which is often not enough to process the entire dataset, but at least the minimum processed size of the data.
[0048] Additionally, the standard deviation can be applied to the time per data record for each client to minimize the desired total time in combination with the standard deviation.
[0049] The batch size of the training dataset can be calculated using a solver and a solution method such as the Generalized Reduced Gradient (GRG). The GRG method is an extension of the reduced gradient method to account for nonlinear inequality constraints. This method establishes a search direction so that, with every small movement, the currently active boundary conditions remain exactly the same.
[0050] Based on knowledge of each client's batch size, the FL server can adjust the weights in the compound averaging algorithms. Instead of weighting client contributions equally, the contributions of n clients are weighted more evenly. k weighted based on the amount of training data (or batch size) used by each client (the more training data, the better): with w ... weight for a client in aggregation n ... contribution for a client k ... client
[0051] K ... number of clients t ... iteration round
[0052] The first part of the algorithm shown in the figure is executed by the server and indexes K clients by the index k.
[0053] The second part of the algorithm, which shows an update by the client, is executed by the respective client k.
[0054] The size of at least one subset of the local training is denoted by B (English "batch").
[0055] The number of epochs of local training is denoted by E.
[0056] The learning rate of the local training is denoted as 77 and describes how strongly the network adjusts the weighting of individual neurons with respect to detected errors after each iteration. The learning rate thus also determines the duration of the training process.
[0057] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.
[0058] Reference symbol list
[0059] B Batch, subset size
[0060] C1-C3, K Client
[0061] E Training period k Index
[0062] LM1-LM3 local model
[0063] GM global model
[0064] S Server
[0065] SYS Client-Server System TD1-TD3 technical device
[0066] 77 Training-learning rate
Claims
Patent claims 1. A computer-implemented method for operating a technical device through a client-server system, comprising a server (S) and clients (C1-C3), each connected to a respective technical device (TD1-TD3), wherein the respective technical device (TD1-TD3) and / or the respective client (C1-C3) provides a performance indicator to describe the technical performance at a specified future time interval, and the following steps are performed: a) Performing training for a provided respective local model (LM1-LM3) for operating the respective technical device (TD1-TD3) with at least a subset of provided training data by the respective client (C1-C3), b) Providing the respective local model (LM1-LM3) to the server (S) by the respective client (C1-C3), c) Aggregating the local models (LM1-LM3) into a global model (GM) by the server.and providing the global model (GM) to the clients (C1-C3), d) receiving the global model (GM) by the respective client (C1-C3) and updating the local model (LM1-LM3) using the previously received global model (GM), e) operating the technical device (TD1) using the previously updated local model (LM1-LM3), characterized in that the size of at least a subset of the provided training data is determined by the performance indicator.
2. Method according to the preceding claim, wherein the performance indicator comprises a technical parameter which describes the technical performance in the form of the computing power and / or the storage power and / or the communication power of the respective technical device (TD1-TD3) and / or the respective client (C1-C3).
3. Method according to the preceding claim, wherein the performance indicator comprises a temporal parameter which describes the future temporal availability of the technical parameters.
4. Method according to the preceding claim, wherein the future temporal availability of the technical parameters is described based on the past availability with regard to the technical parameters of the respective client (C1-C3).
5. Method according to any of the preceding claims, wherein the performance indicator comprises a minimum size for the size of at least a subset of training data provided for the respective client (C1-C3).
6. Client-server system (SYS) for operating a technical device (TD1) with a server (S) and clients (C1-C3), each of which is connected to respective technical devices (TD1-TD3), and the method according to one of the preceding claims is carried out.
7. Computer program product comprising machine-readable instructions stored therein which, when executed by the client-server system (SYS) according to the preceding claim, cause the latter to execute the method according to any of the preceding claims.