PROCESS, SERVER, AND CONTROL ARRANGEMENT

Abstract

The disclosure relates to a process performed by a server (120), the process comprising: receiving (210) parameters defining a first plurality of trained models (111-113) from a second plurality of vehicles (101-103), generating (220) training data (231) using the first plurality of trained models (111-113) defined by the received parameters, training (230) parameters for a substitute model (122) using at least the generated training data (231), and sending (240) the trained parameters of the substitute model (122) to each vehicle of the second plurality of vehicles (101-103).

Description

TECHNICAL AREA

[0001] The present invention relates to updates of trained models in a vehicle fleet. BACKGROUND

[0002] Various computer systems can employ machine learning algorithms to perform required tasks such as anomaly detection, object recognition, and so on. Examples of machine learning algorithms include regression, random forest, neural networks, decision trees, Bayesian algorithms, and others.

[0003] Such machine learning algorithms can process training data to build, train, or generate a model. Machine learning models typically have parameters that are trained using various techniques, such as supervised learning, unsupervised learning, and / or reinforcement learning.

[0004] The complexity of a machine learning model generally increases with the precision and accuracy of the machine learning model.

[0005] Machine learning systems can typically be centralized or distributed.

[0006] Document US20230162487A1 describes a centralized machine learning system where a global model is trained on a primary dataset at a central or main site, and how the global model is then distributed to local sites. This solution has the disadvantage that transferring local data from these sites to the primary dataset for training the global model requires large amounts of data traffic, consuming communication resources.

[0007] Document CN117035061 describes a distributed federated learning, FL, and weighting aggregation method. Local models are merged by weighted averaging of the model parameters, and the averaged model is then used to generate training data on the server. Finally, a new global model is trained on the generated data and distributed to the local sites.

[0008] One problem with locally updating or retraining distributed models is that improvements to a local model in one vehicle do not propagate to other local models in other vehicles. Therefore, the machine learning models used in distributed systems are at least regularly aggregated into a global model and distributed to the heavy vehicles.

[0009] A disadvantage of conventional methods for aggregating distributed models is that they typically require either direct aggregation (e.g., averaging) of model parameters or the transfer of training data between local sites and a central node. Direct aggregation can lead to global model divergence or—in the case of local data not originating from IID—to an underfit, while data transfer requires significant use of communication resources. Averaging model parameters can easily result in a global model that is not a good fit to the actual local distributions (also known as "underfitted").

[0010] Therefore, there is a need for improved methods for disseminating model improvements in a distributed environment. TASKS OF INVENTION

[0011] One objective of embodiments of the present invention is to provide a solution that mitigates or eliminates the disadvantages described above. SUMMARY OF THE INVENTION

[0012] The aforementioned and further tasks are achieved by the subject matter described herein. Further advantageous embodiments of the invention are described herein. The invention is set forth in the accompanying claims. The scope of the invention is defined by the claims, which are incorporated into this section by reference.According to a first aspect of the invention, the object of the invention is achieved by a method carried out by a server, the method comprising: receiving parameters defining a first plurality of trained models from a second plurality of vehicles, generating training data using the first plurality of trained models defined by the received parameters, training parameters for a substitute model using at least the generated training data, and sending the trained parameters of the substitute model to each vehicle of the second plurality of vehicles.

[0013] In an embodiment according to the first aspect, the first plurality of trained models and the substitute model are Gaussian mixture models.

[0014] In an embodiment according to the first aspect, each model of the first plurality of trained models is trained in a single vehicle of the second plurality of vehicles using local sensor data.

[0015] In an embodiment according to the first aspect, the training of the parameters for a substitute model is furthermore carried out using historical training data generated in previous iterations of generating training data.

[0016] In an embodiment according to the first aspect, each model of the first plurality of trained models is configured to perform anomaly detection.

[0017] The advantages of the first aspect include at least higher application performance, more stable generation of training data, and the ability to easily implement schemes for prioritizing between old and new data, thus facilitating continuous learning.

[0018] According to a second aspect of the invention, the objective of the invention is achieved by a server comprising: a processor and a memory, wherein the memory contains instructions that can be executed by the processor, enabling the server to perform the method according to the first aspect.

[0019] According to a third aspect of the invention, the object of the invention is achieved by a method carried out by a control arrangement of a vehicle, the method comprising: receiving local sensor data indicating features of the vehicle, training parameters for a trained model using at least the received local sensor data, sending the parameters of the trained model to a server, receiving parameters of a substitute model, and replacing parameters of the trained model with parameters of the substitute model.

[0020] In an embodiment according to the third aspect, the trained model and the substitute model are Gaussian mixture models.

[0021] In an embodiment according to the third aspect, the trained model is configured to perform anomaly detection.

[0022] According to a fourth aspect of the invention, the object of the invention is achieved by a control arrangement comprising: a processor and a memory, wherein the memory contains instructions that can be executed by the processor, enabling the control arrangement to perform the method according to the third aspect.

[0023] According to a fifth aspect of the invention, the object of the invention is achieved by a computer program which contains instructions which, when the program is executed by a computer, cause the computer to execute the method according to one of the first or third aspects.

[0024] According to a sixth aspect of the invention, the object of the invention is achieved by a computer-readable medium containing instructions which, when executed by a computer, cause the computer to perform the method according to one of the first or third aspects.

[0025] The advantages of the second to sixth aspects are at least the same as the advantages of the first aspect.

[0026] Reference is made to the attached drawing sheets, which are briefly described first. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 shows a distributed machine learning environment with vehicles according to one or more embodiments of the present disclosure. Fig. Figure 2 shows a global training round according to one or more embodiments of the present disclosure. Fig. Figure 3 shows a flowchart of a process according to one or more embodiments of the present disclosure. Fig. Figure 4 shows a flowchart of a process according to one or more embodiments of the present disclosure. Fig. Figure 5 shows a control arrangement according to one or more embodiments of the present disclosure.

[0027] A more complete understanding of the embodiments of the invention will be made possible for the person skilled in the art by considering the following detailed description of one or more embodiments, as well as by recognizing their additional advantages. It should be noted that the same reference numerals are used to identify identical elements depicted in one or more of the figures. DETAILED DESCRIPTION

[0028] This disclosure relates to the sharing of improvements to locally trained models in a distributed machine learning environment with vehicles, particularly when using federated learning (FL).

[0029] As a rule, it is desirable that if local models evolve / improve, e.g., are able to detect new anomalies, this improved capability can be transferred to all local models and corresponding vehicles.

[0030] The present disclosure is based on the understanding that the creation of a global model is achieved by training a new model using training data generated by each local model in a distributed system.

[0031] The advantages include improved application performance, particularly for applications based on Gaussian Mixing Models (GMMs) and Functional Models (FLs). In the case of GMMs, the increased performance is achieved, for example, by avoiding the explicit merging of GMM components. Furthermore, the generation of training data on the server becomes more stable. Additionally, the proposed solution allows for the simple implementation of prioritization schemes for old and new data, thus facilitating continuous learning. Since this is a general solution, it is not possible to precisely quantify the impact; rather, it depends on the specific use case.

[0032] Some conventional distributed solutions, such as the one described in CN117035061A, are based on training a global model using training data generated from a model created by merging local models, or more precisely, by merging parameters that define local models.

[0033] A disadvantage of conventional methods for aggregating distributed models is that they typically require either direct aggregation (e.g., averaging) of model parameters or the transfer of training data between local sites and a central node. Direct aggregation can lead to global model divergence or—in the case of local data not originating from IID—to an underfit, while data transfer requires significant use of communication resources. Averaging model parameters can easily result in a global model that is not a good fit to the actual local distributions (also known as "underfitted").

[0034] Other centralized conventional solutions, such as the one described in US20230162487A1, rely on training a global model using local training data transferred from distributed local nodes to a central node. These solutions have the disadvantage of consuming large amounts of network traffic and thus communication resources to transfer local data from local locations to the primary dataset for training the global model. Furthermore, the transfer of data raises privacy and integrity concerns.

[0035] An example of distributed machine learning is federated learning (FL). FL allows multiple devices or clients to jointly train a shared / global machine learning model without having to exchange their raw data. One example of FL is used when data is collected from IoT (Internet of Things) devices, also known as edge devices. In a vehicle, an edge device could be an electronic control unit (ECU).

[0036] In the FL approach, the training data is not sent to a central server to train a global model; instead, the devices keep their training data locally and only send the trained updates or model parameters to a central node, e.g., a server, and only these updates and / or model parameters are transmitted.

[0037] FL is particularly useful in scenarios where the data is sensitive or where it is not practical or possible to transfer large amounts of training data to a central location. For example, in the field of autonomous driving, data collected by individual vehicles could be used to improve the safety of all vehicles without compromising individual privacy.

[0038] Fig. Figure 1 shows a distributed machine learning environment 100 with vehicles 101-103 according to one or more embodiments of the present disclosure.

[0039] The distributed machine learning environment 100 includes a server 120, e.g. a general-purpose computer or a cloud computing platform.

[0040] The distributed machine learning environment 100 also includes a plurality of vehicles 101-103. Each vehicle includes one or more sensors configured to generate local sensor data. The sensors can include video cameras, vibration sensors, temperature sensors, pressure sensors, or any other suitable sensor. Each vehicle also includes one or more control arrangements (CAs) that are communicatively connected to the respective sensors. In a non-restrictive example, the CA is configured to perform anomaly detection.

[0041] At least one CA of each vehicle 101-103 maintains a locally trained model 111-113 and regularly retrains the model over time using local sensor data.

[0042] The at least one CA of each vehicle 101-103 also transmits parameters defining the respective trained models 111-113 to the server 120 at regular intervals.

[0043] The server receives the parameters that define the respective trained models 111-113 and uses them to generate training data (not shown).

[0044] In one example, the trained models are GMMs that have one or more components. Each component represents a normal distribution, which is in Fig. 1 is represented by black-filled ovals. There is also a weight for each component of the models.

[0045] Transmitting only parameters, and not training data, to server 120 offers the advantage of improved data protection, as training data is more sensitive and therefore less suitable for sharing than model parameters. Furthermore, the amount of training data is typically several orders of magnitude larger than the amount of data representing the model parameters, meaning that transmitting data displaying model parameters requires far fewer communication resources than transmitting training data.

[0046] In other words, Server 120 generates a set of data points. This set of data points represents a mixture of the local data acquired by the sensors in the vehicles during a global training lap. In this context, a global training lap is considered a cycle in which a vehicle receives a global output model, retrains the model using local training data, sends parameters defining the newly trained version of the global output model to the server, and receives an updated global model from the server.

[0047] The received / local models are a way to send information about the local data distributions of the local sensor data / training data to the server 120 without sending the training data itself.

[0048] In one example, the components of the received trained models 111-113 are all incorporated into a larger GMM model 121, which is used solely for the purpose of generating training data. It is understood that this incorporation involves only the addition of components and not the merging or aggregation of parameters from the received trained models 111-113.

[0049] To generate a data point from a given GMM, one of the components is first randomly selected such that the probability of selecting a component is proportional to its weight. Then, a data sample is generated by drawing from the Gaussian distribution given by the selected component.

[0050] In this publication, data is generated from the components present in local models 111-113. In other words, a large GMM (GL) is created that includes all components of local models 121. The component weights are normalized so that they sum to 1. The desired number of training data points can then be generated according to the principle described above. The generated training data, together with the training data generated in previous rounds, constitutes a current server dataset that can be used to train the new global model / substitute model 122, which will generally have far fewer components than the GL.

[0051] Once the set of data points has been generated, it is combined with server-side data or historical training data generated in previous global training rounds to form the current server dataset. The exact way in which recently generated training data is combined with historical training data depends on the problem at hand. For example, a forgetting scheme based on the age of the generated training data could be used to prioritize between old and new data.

[0052] Server 120 then creates or trains a new global model 122, for example by training a new GMM, using the data in the current server dataset. This training step can also be performed using the standard expectation maximization (EM) algorithm, since all data points are available on the server.

[0053] For example, when using GMMs, training a global model using the current server dataset avoids the explicit merging of the GMM components of the local models and thus also avoids problems with instability or divergence.

[0054] Once the new global GMM has been created, it will be sent to vehicles 101-103 as the new global model 122.

[0055] Fig. Figure 2 shows a global training round according to one or more embodiments of the present disclosure.

[0056] Vehicles 101-103, or a CA (Computer Assistant) of each vehicle, receive local sensor data indicating the characteristics of the respective vehicle and train parameters for a trained model 111-113 using at least the received local sensor data. Vehicles 101-103, or a CA of each vehicle, then send the parameters of the respective trained model to server 120.

[0057] Server 120 receives the parameters that define an initial plurality of trained models 111-113 from the plurality of vehicles 101-103. All components 121 of the initial plurality of trained models 111-113 are then received by server 120.

[0058] Server 120 then generates 220 training data 231 using the first plurality of trained models 111-113, which are defined by the received parameters. In other words, it generates 220 the training data 231 using all components 121 of the first plurality of trained models 111-113.

[0059] The training data for each training round is stored on the server along with a timestamp or index to allow prioritization between old and new data, e.g., by using a forgetting scheme.

[0060] Server 120 then trains 230 parameters for a substitute model 122 using at least the generated training data 231. As mentioned earlier, the substitute model 122 has fewer components than the total set of components 121 used to generate the training data 231.

[0061] Server 120 then sends the trained parameters of the replacement model 122 to each vehicle of the second plurality of vehicles 101-103 or to a CA of the respective vehicle.

[0062] Data, such as parameters, can be sent via a communication network that connects the server to vehicles 101-103 and / or the vehicles' CAs.

[0063] Fig. Figure 3 shows a flowchart of a method 200 according to one or more embodiments of the present disclosure. The method 200 is executed by a server 120. The method comprises: Step 210: Receiving parameters defining a first plurality of trained models 111-113 from a second plurality of vehicles 101-103.

[0064] In a non-restrictive example, parameters are received that define the components of the first plurality of trained GMMs 111-113.

[0065] Another example: In a GMM that models a two-dimensional normal distribution, each component has one weight value, two values ​​for the mean, and three values ​​for the symmetric covariance matrix (assuming we use complete covariance matrices). With three components, the GMM would have a total of 3 x (1 + 2 + 3) = 18 parameters.

[0066] Step 220: Generating 220 training data 231 using the first plurality of trained models 111-113 defined by the received parameters.

[0067] In one example, the components of the received trained models 111-113 are all incorporated into a larger GMM model 121, which is used solely for the purpose of generating the training data. It is understood that this incorporation involves only the addition of components and not the merging or aggregation of parameters from the received trained models 111-113.

[0068] Step 230: Training of 230 parameters for a substitute model 122 using at least the generated training data 231.

[0069] In a non-restrictive example, a new global model or replacement model in the form of a GMM with the same number of components as the first plurality of trained models 111-113, defined by the received parameters, is trained using the current server dataset / training data.

[0070] Step 240: Sending the trained parameters of the substitute model 122 to each vehicle of the second plural of vehicles 101-103.

[0071] In one example, the server sends 120 wireless signals with parameters indicating the substitute model 122 to the CAs in each of the vehicles 101-103. The CAs in each of the vehicles 101-103 then apply the substitute model 122 to a task, e.g., anomaly detection.

[0072] In one embodiment, the first plurality of trained models or local models 111-113 and the substitute model 122 are Gaussian mixture models. It is understood that any other suitable generative model can be used, i.e., any model that models the joint distribution of the input and output data. Examples of such models are GMMs, variational autocoders (VAEs), linear discriminant analysis models (LDAs), and generative adversarial networks (GANs).

[0073] In one embodiment, each model of the first plurality of trained models 111-113 is trained in a single vehicle of the second plurality of vehicles 101-103 using local sensor data. In other words, the local model in a vehicle uses local sensor data as training data, so each trained local model 111-113 is likely to be unique. When the parameters of the locally trained models 111-113 are transmitted to the server 120, the model differences can, in some way, be transferred to the surrogate model.

[0074] In one embodiment, the training of the parameters for a substitute model 122 is still carried out using historical training data that was generated in previous iterations of training data generation.

[0075] In one example, the training data from the last five global training rounds is used to train the surrogate model. The training data from each round can be weighted according to its age. In other words, the most recently generated training data can have a greater impact on the model training than the training data from the oldest round.

[0076] In one embodiment, each model of the first plurality of trained models 111-113 is configured to perform anomaly detection.

[0077] It is understood that any suitable task can be performed with the help of the trained models 111-113 or the substitute model 122, e.g., the detection of component or system faults, the detection and classification of obstacles and other vehicles, localization and path planning, the generation of training data for autonomous systems, the detection of intruders (cybersecurity), and the classification of driving environments.

[0078] Fig. Figure 4 shows a flowchart of a method 300 according to one or more embodiments of the present disclosure. The method is carried out by a control arrangement CA of a vehicle, the CA optionally being communicatively coupled to a server 120. The method comprises: Step 310: Receiving local sensor data that displays vehicle characteristics.

[0079] In one example, data from six different sensors related to the vehicle's air pressure system were used to detect abnormal air leaks. Here, the local GMMs each used a maximum of four GMM components. Data from 10 different clients / vehicles can be used, meaning that the parameters of 10 local models are sent to the server in each global iteration. The global model is then built from training data generated from a maximum of 10 x 4 = 40 GMM components, with the resulting global model potentially consisting of up to six components.

[0080] Step 320: Training parameters for a trained model using at least the received local sensor data.

[0081] In a non-restrictive example, a new local model in the form of a GMM with the same number of components as the first plurality of trained models 111-113 is trained using the local sensor / training data.

[0082] Step 330: Sending the parameters of the trained model to server 120. Typically, the parameters are sent via wireless communication to the server. Fig. 1 shown server 120 sent.

[0083] Step 340: Receiving the parameters of a replacement model 122. Typically, the parameters of the replacement model are received via wireless communication from server 120, which is located in relation to Fig. 1 was shown, received.

[0084] Step 350: Replacing the parameters of the trained model with the parameters of the replacement model. In other words, the current task is performed using replacement model 122.

[0085] In one embodiment, the trained model and the substitute model are Gaussian mixture models.

[0086] In one embodiment, the trained model is configured to perform anomaly detection.

[0087] Fig.Figure 5 shows a control arrangement 500 according to one or more embodiments of the present disclosure. The control arrangement 500 may, for example, be in the form of an electronic control unit, a server, an on-board computer, a vehicle-mounted computer system, or a navigation device. The control arrangement 500 may include a processor or processing means 512 that is communicatively coupled to a transceiver 504 configured for wired or wireless communication. Furthermore, the control arrangement 500 may include at least one optional antenna (not shown in the figure). The antenna may be coupled to the transceiver 504 and is configured to transmit and / or broadcast and / or receive wireless signals in a wireless communication system, e.g., wireless signals containing sensor data.In one example, the processor 512 can be a selection of processing circuits and / or a central processing unit and / or processor modules and / or multiple processors configured to work together. Furthermore, the control arrangement 500 can include a memory 515. The memory 515 can contain instructions that can be executed by the processor to perform one of the procedures described herein. The memory and / or computer-readable storage medium referred to herein can be substantially any type of memory, such as a ROM (Read-Only Memory), a PROM (Programmable Read-Only Memory), an EPROM (Erasable PROM), flash memory, an EEPROM (Electrically Erasable PROM), or a disk drive. In some embodiments, the computer-readable medium can be a non-transitory computer-readable medium, such as...a tangible electronic, magnetic, optical, infrared, electromagnetic and / or semiconductor system, device and / or apparatus.

[0088] In a further embodiment, the control arrangement 500 may also include and / or be coupled to one or more sensors which are configured, for example, to receive and / or obtain and / or measure physical properties of the control arrangement 500 or the vehicle 100.

[0089] In one or more embodiments, the control arrangement 500 may further comprise an input device 517 configured to receive inputs or information from a user and to send a user input signal indicating the user inputs or information to the processor or processing means 512.

[0090] In one or more embodiments, the control arrangement 500 may further comprise a display 518 configured to receive a display signal from the processor or processing means 512, which displays rendered objects such as text or graphical user input objects, and to display the received signal as objects such as text or graphical user input objects.

[0091] In one embodiment, the display 518 is integrated into the user input device 517 and configured to receive a display signal from the processing means 512 that shows rendered objects, such as text or graphic user input objects, and to display the received signal as objects, such as text or graphic user input objects, and / or is configured to receive inputs or information from a user and to send a user input signal to the processing means 512 that shows the user input or information.

[0092] In embodiments, the processing means 512 is communicatively connected to a selection of the memory 515 and / or the communication interface and / or the transceiver and / or the input device 517 and / or the display 518 and / or the one or more sensors. In embodiments, the transceiver 504 communicates via wired and / or wireless communication techniques. The wired or wireless communication techniques may include a CAN bus, Bluetooth, Wi-Fi, GSM, UMTS, LTE or LTE Advanced communication network, or any other wired or wireless communication network known in the art.

[0093] According to one aspect of the present revelation, a Server 120 is provided. The server includes: a processor and a memory, the memory containing instructions that can be executed by the processor, enabling the Server 120 to perform each of the procedures described herein.

[0094] According to another aspect of the present disclosure, a tax order is provided, the tax order comprising: a processor and a memory, wherein the memory contains instructions that can be executed by the processor, enabling the control arrangement CA to perform one of the procedures described herein.

[0095] According to another aspect of the present disclosure, a computer program is provided which contains instructions which, when the program is executed by a computer, cause the computer to perform the procedures described herein.

[0096] According to another aspect of the present disclosure, a computer-readable medium is provided which contains instructions which, when executed by a computer, cause the computer to perform the procedures described herein.

[0097] In certain embodiments, the communication network communicates via wired or wireless communication technologies, which may include at least one of the following: Local Area Network (LAN), Metropolitan Area Network (MAN), Global System for Mobile Network (GSM), Enhanced Data GSM Environment (EDGE), Universal Mobile Telecommunications System, Long Term Evolution, High Speed ​​Downlink Packet Access (HSDPA), Wideband Code Division Multiple Access (W-CDMA), Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Bluetooth®, Zigbee®, Wi-Fi, Voice over Internet Protocol (VoIP), LTE Advanced, IEEE 802.16m, Wireless MAN-Advanced, Evolved High-Speed ​​Packet Access (HSPA+), 3GPP Long Term Evolution (LTE), Mobile WiMAX (IEEE 802.16e), Ultra Mobile Broadband (UMB) (formerly Evolution-Data Optimized (EV-DO) Rev.).C), Fast Low-latency Access with Seamless Handoff Orthogonal Frequency Division Multiplexing (Flash-OFDM), High Capacity Spatial Division Multiple Access (iBurst®) and Mobile Broadband Wireless Access (MBWA) (IEEE 802.20) systems, High Performance Radio Metropolitan Area Network (HIPERMAN), Beam-Division Multiple Access (BDMA), World Interoperability for Microwave Access (Wi-MAX) and ultrasound communication, etc., but is not limited to these.

[0098] Furthermore, it is clear to those skilled in the art that the control arrangement 500 may include the necessary communication capabilities in the form of, for example, functions, means, units, elements, etc., for implementing the present solution. Examples of other such means, units, elements, and functions are: processors, memory, buffers, control logic, encoders, decoders, rate adjusters, de-rate adjusters, mapping units, multipliers, decision units, selection units, switches, interleavers, deinterleavers, modulators, demodulators, inputs, outputs, antennas, amplifiers, receiver units, transmitter units, DSPs, MSDs, encoders, decoders, power supply units, power injectors, communication interfaces, communication protocols, etc., which are arranged together in a suitable manner to implement the present solution.

[0099] In particular, the processor and / or processing device of the present disclosure may be one or more instances of processing circuits, processor modules and multiple processors configured to work together, a central processing unit (CPU), a processing unit, a processing circuit, a processor, an application-specific integrated circuit (ASIC), a microprocessor, a field-programmable gate array (FPGA), an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a system-on-chip (SoC), a programmable logic unit, a microprocessor, an application-specific integrated circuit or any other device capable of performing operations electronically in a particular manner, or other processing logic capable of interpreting and executing instructions.The terms "processor" and / or "processing means" can thus represent a processing circuit comprising multiple processing circuits, such as one, some, or all of those mentioned above. The processing means can further perform data processing functions for input, output, and processing of data, including data buffering and device control functions, such as call processing control, user interface control, or the like.

[0100] Finally, it should be understood that the invention is not limited to the embodiments described above, but also relates to and includes all embodiments within the scope of the attached independent claims. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 20230162487A1 [0006, 0034] CN 117035061

[0007] CN 117035061A

[0032]

Claims

[1] A procedure performed by a server (120), the procedure comprising: Receiving (210) parameters defining a first plurality of trained models (111-113) from a second plurality of vehicles (101-103), Generating (220) training data (231) using the first plurality of trained models (111-113) defined by the received parameters, Training (230) parameters for a surrogate model (122) using at least the generated training data (231), Sending (240) the trained parameters of the replacement model (122) to each vehicle of the second plurality of vehicles (101-103). [2] Method according to claim 1, wherein the first plurality of trained models (111-113) and the substitute model (122) are Gaussian mixture models. [3] Method according to any of the preceding claims, wherein each model of the first plurality of trained models (111-113) is trained in a single vehicle of the second plurality of vehicles (101-103) using local sensor data. [4] Method according to one of the preceding claims, wherein the training of the parameters for a substitute model (122) is further carried out using historical training data generated in previous iterations of generating training data. [5] Method according to any of the preceding claims, wherein each model of the first plurality of trained models (111-113) is configured to perform anomaly detection. [6] Server (120), where the server comprises: a processor and a memory, wherein the memory contains instructions that can be executed by the processor, enabling the server (120) to perform the method according to any one of claims 1-5. [7] Procedure carried out by a control unit (CA) of a vehicle, the procedure comprising: Receiving local sensor data that indicates vehicle characteristics, Training parameters for a trained model using at least the received local sensor data, Sending the parameters of the trained model to a server (120), Receiving parameters of a substitute model (122), Replacing parameters of the trained model with parameters of the replacement model. [8] Method according to claim 7, wherein the trained model and the substitute model are Gaussian mixture models. [9] Method according to one of claims 7-8, wherein the trained model is configured to perform anomaly detection. [10] Control arrangement (CA), wherein the control arrangement (CA) comprises: a processor and a memory, wherein the memory contains instructions that can be executed by the processor, enabling the control arrangement (CA) to perform the method according to any one of claims 7 to 9. [11] Computer program with instructions which, when the program is executed by a computer, cause the computer to execute the method according to any one of claims 1 to 5 or 7 to 9. [12] Computer-readable medium containing instructions which, when executed by a computer, cause the computer to execute the method according to any one of claims 1 to 5 or 7 to 9.

Images