Computer-implemented method and system for operating a technical device by a client
The method enhances federated learning by selectively updating client models within a federated learning system based on accuracy and type, resulting in improved global model training and device operation precision.
Patent Information
- Application Number
- PCT/EP2024/067031
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-06-29
- Filing Date
- 2024-06-19
- Publication Date
- 2025-01-02
AI Technical Summary
Current federated learning systems update the entire client model in each round, lacking partial learning of the global model, which is time-consuming and inefficient, especially after system maintenance or when adapting to new production system parts, and do not allow focused training on specific sub-segments of the model.
A method where the server provides a global model to clients, determining client model accuracy and type, updating only client models with sufficient accuracy and matching type, and training the global model using these weights, ensuring precise operation of technical devices.
This approach improves the global model by using only accurate client models, benefiting previously inaccurate clients and enabling more precise device operation, facilitating predictive maintenance and efficient production.
Smart Images
Figure EP2024067031_02012025_PF_FP_ABST
Abstract
Description
[0001] Computer-implemented method and system for operating a technical device by a client
[0002] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identity are included.
[0003] The invention relates to a computer-implemented method and a system for operating a technical device by a client of a system having a server and at least two clients with a respective client identifier.
[0004] Furthermore, the invention relates to a computer program product.
[0005] Nowadays, industrial systems often use machine learning (ML for short) and artificial intelligence (AI for short) to control production machines, for example, or to analyze operational data collected from production plants and use this to draw conclusions about future events.
[0006] Training an ML model can be very time-consuming and it may be necessary to repeat the training after, for example, a system maintenance.
[0007] It can also be very complex to take account of adapted operating requirements, for example expanding an existing production plant with new plant components.
[0008] Federated learning (FL for short) can be used to train ML / AI models across multiple participants in a client-server system and to make a common global model available to new clients.
[0009] However, in federated learning, only the entire client model is updated in each federated round and partial learning of the global model is not currently planned.
[0010] Accordingly, the object of the invention is to provide a solution that allows specific training of a global model within an FL system. This object is achieved by a method of the type mentioned above, wherein the following steps are carried out: a) providing a global model to the server, b) providing the global model to the respective client as a respective client model by the server, c) determining a class of the respective client model with a respective accuracy within a respective predetermined value range by the respective client, d) specifying a respective model type identifier for the respective client model with the respective class by the respective client, e) specifying a set of weights as an empty set by the respective client,f) Determining respective model weights for the class in the respective client model as a neural network based on at least one respective relevance parameter determined by the layer-wise relevance propagation algorithm by the respective client, g) Linking the respective model weights for the class of the respective client model with the set of weights by the respective client, h) Providing the respective client identifier, the respective model type identifier, the respective accuracy and the respective set of weights to the server by the respective client, i) Training the global model with the set of weights from those clients that have at least a predetermined accuracy and match a predetermined model type as a trained global model by the server, j) Providing the trained global model to the at least one client as a respective client model,and operating the respective technical device with the respective client model.,
[0011] This ensures that a specific training of a model is carried out within an FL system, namely using only those client models that have a specified minimum accuracy and a match in the model type, whereby an improvement of the global model can be achieved.
[0012] In other words , only models of clients with sufficient accuracy are used to form the global model .
[0013] Subsequently, even clients who were previously classified as too inaccurate will benefit from the application of the updated global model.
[0014] The criterion for inclusion of clients is a predetermined level of accuracy , which is provided in step i ) .
[0015] There is no possibility in the state of the art to focus the FL task on a specific sub-segment of the Kl model, which can, however, be achieved by the invention.
[0016] Operating or controlling a device with a more accurate local model leads to improved, more accurate operation of the device.
[0017] When operating the device, for example, the operator or a control system can be signaled with a current accuracy in the operation of the device, and predictive maintenance can be carried out. When controlling the device, the more accurate model is used in the operation of the device, whereby the device can be operated more efficiently or a product produced by the device can be manufactured more accurately and efficiently.
[0018] In a further development of the invention, it is provided that the determination of the class of the respective client model is carried out with an accuracy within a predetermined value range using a confusion matrix.
[0019] This makes it easy to achieve and implement class assignment.
[0020] In a further development of the invention, it is provided that the layer-wise relevance propagation algorithm applies the weights in the respective client model and activations of neurons at the input in the neural network through a forward pass to the output, and propagates them back to the input in the neural network in a backward pass.
[0021] This makes it easy to determine the relevance parameter, which indicates which input variables at the input have a high influence on output variables at the output.
[0022] In a further development of the invention, it is provided that the at least one relevance parameter is defined by the following relationship:
[0023] Rj = y ajWjk R k K 1 So,j a j w jk with
[0024] Rj, R k ... relevance , ( consecutive ) layers a ... activation of a neuron , w lk ... weight between two neurons, k counters .
[0025] This makes it easy to implement the relevance parameter.
[0026] In a further development of the invention, it is provided that the respective technical device is operated with the respective client model using control commands which use the client model.
[0027] In this way, a technical effect of the improved global model or the client model can be achieved in a simple way, since the control commands can be viewed as a representation of the model.
[0028] In a further development of the invention, it is provided that operating data are recorded by the respective technical device from a recording means which is comprised by the respective client, and an operating state of the respective client is determined by the respective client using the respective client model.
[0029] In this way, a technical effect of the improved global model or the client model can be achieved in a simple way, since the operating data can be regarded as input data when applying the model.
[0030] The object according to the invention is also achieved by a system for operating a technical device, comprising a system with a server and at least two clients, which are each connected to a respective technical device, and the system is set up to carry out the method according to the invention.
[0031] The object of 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 processing unit to carry out the method according to the invention. The invention is explained in more detail below with reference to an exemplary embodiment shown in the accompanying drawings. In the drawings:
[0032] Fig. 1 shows an example of a “Federated Learning” system,
[0033] Fig. 2 shows an example of classes of model data with a confusion matrix,
[0034] Fig. 3 shows an example of determining relevance-
[0035] parameters,
[0036] Fig. 4 shows an embodiment of the method according to the invention.
[0037] Fig. 1 shows an exemplary embodiment of a “federated learning” system according to the state of the art.
[0038] Such FL systems retrieve the AI model updates from the connected clients, calculate the overall average of the obtained model weights, and redistribute the averaged weights among the clients.
[0039] A system for operating a technical device comprises a server S and three clients C1-C3, each of which is connected to a respective technical device TD1-TD3.
[0040] The clients C1-C3 each have client identifiers CID1-CID3 and store client models CM1-CM3 with which the technical devices TD1-TD3 can be controlled using control commands that can be determined on the basis of the client models CM1-CM3.
[0041] Fig. 2 shows an example of classes of model data with a confusion matrix.
[0042] This presents data on diagnoses that can be determined using a model. The figure compares a predicted diagnosis (PD) with an actual diagnosis (TD).
[0043] Predetermined value ranges are plotted on both axes of the confusion matrix, such as a division of the entire value range into a serious state SS, a moderate state MS and a healthy state HS.
[0044] The combinations of the above-mentioned states SS, MS and HS can be assigned to different classes, such as a class CG.
[0045] Confusion matrices make it easier to determine how accurate a model's results are likely to be by revealing when the model repeatedly confuses two classes.
[0046] Confusion matrices evaluate the performance of a classification model and allow business users to determine which data their model may not be able to classify correctly.
[0047] In other words, a confusion matrix relates a classification of data with given classes in a classification prediction and an achieved classification.
[0048] For example, a client may have a classification model with three classes {CA, CB, CC}.
[0049] In this example, the prediction of classes CA and CB works perfectly for the customer with high accuracy - the customer does not need to improve or refine any model parameters related to CA or CB.
[0050] However, the model may not perform well in terms of classification accuracy for class C.
[0051] Predicted labels belonging to class CC are often confused with class CA or CB.
[0052] Therefore, the customer would benefit from improving the accuracy of Class CG. However, this scenario is not currently supported by FL systems.
[0053] Fig. 3 shows an example of determining relevance parameters, which indicate which input variables at the input have a high influence on output variables at the output.
[0054] The relevance parameters can be used in the well-known layer-wise relevance propagation algorithm.
[0055] The weights in the respective client model and activations of neurons at the input in the neural network are applied through a forward pass to the output, and in a backward pass back to the input in the neural network and an explanation EX (English "explanation") is provided as to which input neurons have the greatest influence on output neurons.
[0056] The relevance parameter can be defined by the following relationship: with
[0057] Rj, R k ... relevance , ( consecutive ) layers aj ... activation of a neuron , w ik - Weight between two neurons, k ... counter .
[0058] Fig. 4 shows an embodiment of the method according to the invention.
[0059] A system for operating a technical device comprises a server S and three clients C1-C2, each of which is connected to a respective technical device TD1-TD3.
[0060] The following method steps are carried out: a) Providing a global model GM to the server S, b) Providing the global model GM to the respective client C1-C3 as the respective client model CM1-CM3 by the server S, c) Determining a class C of the respective client model CM1-CM3 with a respective accuracy AG within a respective predetermined value range by the respective client C1-C3, d) Determining a respective model type identifier MT-ID for the respective client model CM1-CM3 with the respective class C by the respective client C1-C3, e) Determining a set of weights W as an empty set by the respective client C1-C3, f) Determining respective model weights wi for the class C in the respective client model CM1-CM3 as a neural network based on at least one respective relevance parameter Ri, Rj, Rk, which is determined by the layer-wise Relevance propagation algorithm LRP is determined,by the respective client C1-C3, g) connecting the respective model weights for the class C of the respective client model CM1-CM3 with the set of weights W, by the respective client GIGS, h) providing the respective client identifier CID1-CID3, the respective model type identifier MT-ID, the respective accuracy AG and the respective set of weights W to the server S, by the respective client C1-C3, i) training the global model GM with the set of weights W from those clients C1-C3, which have at least a predetermined accuracy, and match a predetermined model type, as a trained global model by the server S, j) providing the trained global model to the at least one client C1-C3 as the respective client model CM1-CM3, and operating the respective technical device TD1-TD3 with the respective Client model CM1-CM3 .,
[0061] Determining the class C of the respective client-
[0062] Model CM1-CM3 can be determined with an accuracy within a predetermined range of values using a confusion matrix.
[0063] The layer-wise relevance propagation algorithm LRP can apply the weights in the respective client model CM1-CM3 and activations of neurons at the input IN in the neural network through a forward pass to the output OUT, and propagate back to the input IN in the neural network in a backward pass.
[0064] The at least one relevance parameter Ri , Rj , Rk can be defined by the following relationship: with
[0065] Rj, R k ... relevance , ( consecutive ) layers a ... activation of a neuron , w ik - Weight between two neurons, k ... counter .
[0066] The respective technical device TD1-TD3 can be operated with the respective client model CM1-CM3 using control commands that use the client model CM1-CM3 or are generated by applying the respective client model CM1-CM3. The respective technical device TD1-TD3 can record operating data from a recording device included in the respective client C1-C3, and the respective client C1-C3 can determine an operating state of the respective client C1-C3 using the respective client model CM1-CM3.
[0067] List of reference symbols:
[0068] accuracy
[0069] C, CA, CB, CC class
[0070] C1-C3 Client
[0071] CID1-CID3 client identifier
[0072] EX explanation
[0073] HS Healthy state
[0074] IN, IN1-IN3 input
[0075] K-ID Client ID
[0076] LRP layer-wise relevance propagation
[0077] Algorithm (layer-wise relevance propagation)
[0078] M input model
[0079] MS Moderate state
[0080] MT Model Type
[0081] MT-ID Identification of model type
[0082] OUT, OUT1-OUT3 output
[0083] PD predicted diagnosis
[0084] Ri, Rj relevance parameters
[0085] SS severe state
[0086] S Server
[0087] TD actual diagnosis
[0088] W Quantity of weights
Claims
Patent claims 1. Computer-implemented method for operating a technical device by a client of a system having a server (S) and at least two clients (C1-C3) with a respective client identifier (CID1-CID3), wherein the following steps are carried out: a) Providing a global model (GM) to the server (S), b) Providing the global model (GM) to the respective client (C1-C3) as the respective client- Model (CM1-CM3) by the server (S), c) determining a class (C) of the respective client model (CM1-CM3) with a respective accuracy (AG) within a respective predetermined value range by the respective client (C1-C3), d) specifying a respective model type identifier (MT-ID) for the respective client model (CM1-CM3) with the respective class (C) by the respective client (C1-C3), e) specifying a set of weights (W) as an empty set by the respective client (C1-C3), f) determining respective model weights (wi) for the class (C) in the respective client model (CM1-CM3) as a neural network based on at least one respective relevance parameter (Ri, Rj, Rk), which is determined by the layer-wise relevance propagation algorithm (LRP), by the respective clients (C1-C3) , g) Connecting the respective model weights for the class (C) of the respective client model (CM1-CM3) with the set of weights (W) by the respective client (C1-C3), h) providing the respective client identifier (CID1-CID3), the respective model type identifier (MT-ID), the respective accuracy (AC) and the respective set of weights (W) to the server (S) by the respective client (C1-C3), i) training the global model (GM) with the set of weights (W) from those clients (C1-C3) which have at least a predetermined accuracy and correspond to a predetermined model type, as a trained global model by the server (S), j) providing the trained global model to the at least one client (C1-C3) as a respective client model (CM1-CM3), and operating the respective technical device (TD1-TD3) with the respective client model (CM1-CM3).
2. Method according to the preceding claim, wherein the determination of the class (C) of the respective client- model (CM1-CM3) with an accuracy within a predetermined range of values using a confusion matrix.
3. Method according to one of the preceding claims, wherein the layer-wise relevance propagation algorithm (LRP) applies the weights in the respective client model (CM1-CM3) and activations of neurons at the input (IN) in the neural network through a forward pass to the output (OUT), and propagates back to the input (IN) in the neural network in a backward pass.
4. Method according to the preceding claim, wherein the at least one relevance parameter (Ri, Rj , Rk) is defined by the following relationship: with Rj, R k ... relevance, (consecutive) layers Activation of a neuron, w ik- Weight between two neurons, k ... counter.
5. The method according to one of claims 1 to 4, wherein the respective technical device (TD1-TD3) with the respective client model (CM1-CM3) is operated with control commands which use the client model (CM1-CM3).
6. Method according to one of claims 1 to 4, wherein operating data are acquired by the respective technical device (TD1-TD3) from a recording means which is comprised by the respective client (C1-C3), and an operating state of the respective client (C1-C3) is determined by the respective client (GIGS) using the respective client model (CM1-CM3).
7. System for operating a technical device, comprising a system with a server and at least two clients, each of which is connected to a respective technical device (TD1-TD3), and the system is 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 it to carry out the method according to any one of the preceding claims