Mobile device positioning method

Abstract

According to one embodiment, a method for mobile terminal positioning is described, which includes the steps of: training a positioning machine learning model to output positioning information in response to mobile terminal state information by federated learning using a plurality of positioning components as federated learning clients; and determining the locations of a plurality of mobile terminals in the category using each of the trained positioning machine learning models; and determining the locations of the plurality of mobile terminals using the trained positioning machine learning models.

Description

Technical Field

[0001] This disclosure relates to a method for positioning a mobile terminal.

Background Art

[0002] Mobile terminal positioning is a task necessary for the efficient operation of a mobile communication system, for example, for efficient beamforming. This task can be solved using a machine learning approach. However, there are problems in training, such as local collection of training data in positioning components and the existence of multiple categories in mobile terminals. Therefore, an efficient positioning approach is desirable, particularly with regard to the overhead of training data transmission and the training efficiency of the machine learning model used for positioning.

Summary of the Invention

[0003] According to one embodiment, a method for mobile terminal positioning is provided. The method includes training a positioning machine learning model to output positioning information in response to mobile terminal state information by means of federated learning that uses a plurality of positioning components as federated learning clients. For each mobile terminal category among a plurality of mobile terminal categories, each positioning machine learning model outputs positioning information in response to mobile terminal state information by means of federated learning that uses each of the plurality of positioning components as a federated learning client, and determining the positions of a plurality of mobile terminals of the category using each trained positioning machine learning model; and determining the positions of a plurality of mobile terminals using the trained positioning machine learning model.

Brief Description of the Drawings

[0004] In the figures, like reference numerals generally represent the same parts throughout different figures. The drawings are not necessarily to scale, but rather are generally emphasized when explaining the principles of the present invention. In the following description, various aspects are described with reference to the following drawings. [Figure 1]This indicates a mobile communication system. [Figure 2] This document describes an architecture for training an ML (machine learning) model in a scenario using multiple LMFs (location management functions). [Figure 3] This indicates registration of the base model provider. [Figure 4] This indicates registration of an FL (Federal Learning) server. [Figure 5] This indicates LMF registration. [Figure 6] A flowchart illustrating the base model training is shown. [Figure 7] This flowchart illustrates the training of an ML model using associative learning. [Figure 8] This flowchart shows the process for using a positioning model in inference. [Figure 9] A flowchart illustrating a mobile terminal positioning method according to one embodiment is shown. [Modes for carrying out the invention]

[0005] The following detailed description will refer to the accompanying drawings illustrating specific details and embodiments in which the invention may be carried out. Other embodiments may be utilized, and structural, logical, and electrical modifications may be made without departing from the scope of the invention. The various embodiments of this disclosure are not necessarily mutually exclusive, and some embodiments of this disclosure may be combined with one or more other embodiments of this disclosure to form new embodiments.

[0006] Various examples corresponding to the aspects of this disclosure are described below.

[0007] Example 1 is a method for mobile terminal positioning, which includes training a positioning machine learning model to output positioning information according to mobile terminal state information by federated learning using multiple positioning components as federated learning clients, and determining the positions of multiple mobile terminals using the trained positioning machine learning model.

[0008] Example 2 is the same method as in Example 1, in which a positioning component and / or one or more other positioning components determine the locations of multiple mobile terminals using a trained positioning machine learning model.

[0009] Example 3 is a method of Example 1 or 2, which includes training each positioning machine learning model to output positioning information according to mobile terminal state information by using federated learning with multiple positioning components as federated learning clients for each of the multiple mobile terminal categories among multiple mobile terminal categories, and determining the positions of multiple mobile terminals in the category using each trained positioning machine learning model.

[0010] Example 4 is the method of Example 3, wherein the categories differ in terms of the mobility of the mobile terminal and / or the location of the mobile terminal.

[0011] Example 5 is the method described in any one of Examples 1-4, which involves registering positioning components that support associative learning in a positioning component repository using a positioning component profile that demonstrates associative learning capability, and selecting multiple positioning components from the registered positioning components.

[0012] Example 6 is the method described in any one of Examples 1 to 5, which includes selecting a federated learning server to support the training of a positioning machine learning model and controlling the federated learning by the selected federated learning server.

[0013] Example 7 is the same method as in Example 6, which includes registering a federative learning server in the federative learning server repository using a federative server profile that indicates federative learning capability, and selecting a federative learning server from the registered federative learning servers.

[0014] Example 8 is the method described in any one of Examples 1-7, which includes training a base model and using the trained base model as a starting point for training a positioning machine learning model.

[0015] Example 9 is the method described in Example 8, including adding one or more neural network layers to the base layer to build a positioning machine learning model.

[0016] Example 10 is the method described in Example 8 or 9, including training the base model by a base model provider and obtaining the trained base model from the base model provider.

[0017] Example 11 is the method described in Example 10, including registering a base model provider in a base model provider repository using a base model profile indicating the base model training ability, and determining a base model provider for obtaining a trained base model from the registered base model provider.

[0018] Example 12 is the method described in any one of Examples 8 to 11, including training the base model using training data obtained from a plurality of positioning components and / or one or more additional positioning components.

[0019] Example 13 is the method described in any one of Examples 8 to 12, including sampling, aggregating, and / or filtering local training data with respect to the amount of local training data available in a plurality of positioning components and / or one or more additional positioning components, and training the base model using the training data with the reduced amount.

[0020] Example 14 is the method described in any one of Examples 8 to 13, including training the base model using the training data with the reduced amount with respect to the amount of training data used by an association learning client to train a positioning machine learning model.

[0021] Example 15 is the method according to any one of Examples 1 to 14, wherein each positioning component among a plurality of positioning components is configured to perform local training in federated learning using local training data obtained from positioning using one or more other positioning machine learning models and / or one or more other non-machine learning positioning methods.

[0022] Example 16 is a mobile terminal positioning system configured to execute the method according to any one of Examples 1 to 15.

[0023] According to a further embodiment, there is provided a computer-readable medium including a computer program and instructions that, when executed by a computer, cause the computer to execute the method according to any one of the above examples.

[0024] In the following, various examples will be described in more detail.

[0025] FIG. 1 shows a mobile communication system 100.

[0026] The mobile communication system 100 includes a mobile terminal 101 (also referred to as User Equipment (UE)) served by a radio access network 102 implemented by a communication network (i.e., the network side of the communication system 100), particularly a plurality of base stations 103 (which may be of different generations, such as eNB or gNB). The radio access network 102 is connected to a core network 104, in this example, a 5G core network including network functions, particularly an Access and Mobility Management Function (AMF) 105.

[0027] One task that needs to be performed by the network when providing services to the mobile terminal 101 is to determine the position of the mobile terminal 101, i.e., to enable beamforming when transmitting data to the mobile terminal 101. For this purpose, a positioning component, in particular a Location Management Function (LMF) 106, is provided within the core network 104. The core network 104 further includes a Network Data Analytics Function (NWDAF) 107.

[0028] NWDAF107 is responsible for providing network analysis and / or predictive information in response to requests from network functions. For example, a network function may request specific analytical information regarding the load level of a particular network slice instance. Alternatively, the network function can use a subscription service to ensure that NWDAF notifies it when the load level of a network slice instance changes or reaches a certain threshold. According to 3GPP Release 17 (Rel-17), NWDAF is broken down into two functions or categories: NWDAF AnLF (Analytics Logical Function) and NWDAF MTLF (Model Training Logical Function). NWDAF containing the Analytics Logical Function is denoted as NWDAF(AnLF), NWDAF-AnLF, or simply AnLF, and can perform estimations, derive analytical information, and expose analytical services, namely Nnwdaf_AnalyticsSubscription or Nnwdaf_AnalyticsInfo. An NWDAF that includes a model training logic function is denoted as NWDAF(MTLF), NWDAF-MTLF, or simply MTLF, and trains machine learning (ML) models and exposes new training services (e.g., providing trained models). An NWDAF that includes both is represented as NWDAF(MTLF+AnLF).

[0029] The LMF106 determines the position of the mobile terminal 101 from inputs such as positioning reference signal (NR PRS, transmitted downlink) measurements and sounding reference signal (SRS, transmitted uplink) measurements. The LMF106 can do this using a machine learning (ML) model 108, which can be one of two cases. • D-AIL (Direct): The ML model 108 outputs positioning information to determine the location of the mobile terminal 101. • A-AIL (Assistant): The ML model 108 outputs assistant positioning information, which is used as an alternative method or input to the model for determining the position of the mobile terminal 101.

[0030] Inputs to the LMF106 are distributed from the RAN102 to the LMF106 via the AMF105. These can come from the base station 103 providing services to the mobile terminal 101 (e.g., in the case of SRS measurement), or from the mobile terminal 101 via the base station 103 (e.g., in the case of PRS measurement).

[0031] The LMF106 can generate (local) training data for the ML model 108 (including training data elements including inputs and labels (i.e., ground truth of location)) by performing positioning using a non-ML method or by using another ML model.

[0032] The ML model 108 may be trained by NWDAF107 (in this case, specifically including MTLF). For this purpose, LMF106 can supply the generated training data to NWDAF107.

[0033] However, since LMF106 performs inference (i.e., uses the trained ML model 108 for positioning), sending training data from LMF106 to NWDAF107 for training and then sending the trained ML model 108 back to LMF106 is considered unnecessary overhead. On the other hand, if LMF106 performs its own model training (including MTLF), it can be inefficient if there are multiple LMF106s in the communication system, as each LMF106 will train its own ML model 108, and this may involve redundant training as the ML models 108 may be similar. Thus, this could place a significant load on LMF106, which may be at least partially redundant.

[0034] However, it's important to note that mobile devices typically fall into different categories, such as the following: • Indoor UE • Outdoor UE ·Fixed UE • High-speed UE (i.e., UE that moves at high speed) • UEs that require or demand 3D positioning (e.g., due to large altitude fluctuations)

[0035] This may necessitate the use of multiple ML models. A single ML model 108 to determine positioning information for all these categories could be either extremely complex or have low accuracy. On the other hand, training separate ML models for each category can improve accuracy, but it can lead to redundant training of the ML models because the models for different categories may be relatively similar.

[0036] As mentioned above, model training on NWDAF107 leads to high signaling overhead, and model training on LMF106 can lead to redundant and inefficient training (at least when there are multiple LMFs). From this perspective, various embodiments provide an approach that allows for efficient training of ML models (or multiple ML models, one for each of the different UE categories) on the LMF side for UE positioning when multiple LMFs are deployed.

[0037] Figure 2 shows the ML in a scenario with multiple LMF203s. This shows the model training architecture.

[0038] As illustrated with reference to Figure 1, each LMF203 is connected to one or more RAN202 (each including one or more base stations) that serve the UE201. The LMF203 is responsible for positioning the UEs that are served by each RAN202 to which they are connected. The architecture further includes a network repository function (NRF) 204, a base model provider 205 (e.g., a server and / or network function, e.g., NWDAF (MTLF)), and a federated learning (FL) server 206. As described above, the LMF203 collects local training data, which will be referred to below as local training data.

[0039] According to various embodiments, the following is performed: 1) All entities involved in training (i.e., LMF203, base model provider 205, and FL server 206) register their capabilities and supported ML models with NRF204. 2) The base model provider 205 collects a portion of the local training data from the LMF 203. This means that the amount of training data collected by the base model provider 205 is less than the total amount of local training data (from all LMFs). In other words, the base model training data does not include the complete local training data (from all LMFs). This is achieved, for example, by aggregating the local training data, sampling from the local training data, or filtering from the local training data. Consequently, the overhead is lower compared to when all training data is delivered to an NWDAF, etc. 3) The base model provider 205 uses the collected training data to train a base model (e.g., a deep neural network). 4) FL Server 206 retrieves a base model and builds (configures) a specific ML model for each category of UE (for example, by extending the base ML model by adding new layers). This is called the global model (or global version of the model) for that category. The base model may already contain all layers, but it is used as a pre-trained version of the ML model to be trained. In other words, the base model is used as a starting point for a particular ML model (with or without adding layers). 5) If LMF203 requires an ML model (or an updated or retrained ML model) to position UEs of a specific category, it searches for FL server 206 via NRF204 and requests the specific ML model for that category. 6) The FL server 206 selects some of the LMF203s (searched via NRF204) and distributes the corresponding global ML models (for each category) to the selected LMF203s. 7) Each selected LMF203 locally trains a (specific) ML model using its own complete local training data, generates updates to the ML model, and sends the updates to the FL server 206. 8) FL Server 206 updates the global version of the ML model. 9) Repeat steps 6 through 8 until several termination conditions are met (for example, the ML model converges). 10) LMF203, which requires an ML model for positioning, obtains the final version of the global model from FL server 206 and uses it for inference (direct positioning (D-AIML) or assisted positioning (A-AIML)).

[0040] When performing positioning for multiple categories of UEs, the LMF203 can use an ensemble of machine learning models acquired in this way.

[0041] As mentioned above, in step 1) of the procedure described above, the relevant entities are registered in NRF204. This will be explained in more detail with reference to Figures 3 to 5.

[0042] Figure 3 shows the registration of the base model provider 301 in NRF302.

[0043] In 303, the base model provider 301 sends a registration message to the NRF 302. The registration message contains the profile of the base model provider 301, which specifies its ability to support base model training and includes a list of ML models supported for UE positioning. For example: Model 1 Direct AIML List of input data The basis for a specific model for the following: Indoor positioning Stationary UE positioning Model 2 Direct AIML List of input data The basis for a specific model for the following: Outdoor positioning High-speed UE positioning

[0044] In step 304, the NRF stores the profile of the base model provider 301, and in step 305, it sends an acknowledgment.

[0045] Figure 4 shows the registration of FL server 401 in NRF402.

[0046] In step 403, FL server 401 sends a registration message to NRF402. The registration message includes a profile of FL server 401 specifying its ability to operate as an FL server, and a list of supported ML models. For example: Model 1 Direct AIML List of input data function Indoor positioning Model 2 Indirect AIML List of input data function Outdoor positioning High-speed UE positioning

[0047] In 404, the NRF stores the profile of FL server 401, and in 405, it sends an acknowledgment.

[0048] Figure 5 shows the registration of LMF server 501 in NRF502.

[0049] In step 503, the LMF server 501 sends a registration message to the NRF 502. The registration message includes a profile of the LMF 501 specifying its ability to operate as an FL client, and a list of ML models supported for FL. For example: Model 1 Direct AIML List of input data function Indoor positioning Model 2 Indirect AIML List of input data function Outdoor positioning High-speed UE positioning

[0050] In step 504, the NRF stores the LMF501 profile, and in step 505, it sends an acknowledgment.

[0051] Figure 6 shows a flowchart 600 illustrating base model training, including the base model provider 601 and the selected LMF602.

[0052] To collect training data for training the base model, the base model provider 601 in 603 subscribes to (or requests training data from) LMF602 for training data. In the corresponding subscription message, it can specify rules on how the amount of data provided by LMF602 should be reduced relative to the (overall) local data, such as aggregation rules, filtering rules (specifying which parts of the local data should be filtered and removed), and sampling rules.

[0053] In step 604, LMF602 provides training data to base model provider 601 according to the rules.

[0054] In step 605, the base model provider 601 trains the base model using the training data provided by LMF602.

[0055] Figure 7 shows flowchart 700 illustrating the training of an ML model using associative learning.

[0056] The flow takes place between NRF701 (where entities involved in federative learning are registered as described above with reference to Figures 3 through 5), LMF702, base model provider 703 (which trained the base model as described with reference to Figure 6), FL server 704, and a set of LMF705 to which LMF702 may or may not belong.

[0057] In step 706, LMF702 determines that a specific positioning (ML) model (i.e., an ML model that generates positioning information for a mobile terminal from input data containing information about the mobile terminal) needs to be trained or retrained.

[0058] In 707, LMF702 uses NRF701 to find the appropriate FL server 704 by sending a corresponding request that provides information about the model (e.g., the UE category on which the model is trained).

[0059] At 708, LMF702 sends a request for the model (including information about the model) to the discovered FL server 704.

[0060] In step 709, the FL server 704 determines the base model to be used for training the model.

[0061] In 710, the FL server 704 finds a base model provider 703 that can provide the determined base model by sending a request containing information about the base model (the UE category that should be appropriate as a starting point).

[0062] In step 711, the FL server 704 requests a base model from the discovered base model provider 703 (for example, by sending a request that includes the identification of the requested base model).

[0063] In step 712, the base model provider 703 provides the requested base model to the FL server 704.

[0064] In step 713, the FL server 704 uses the NRF 701 to discover the appropriate FL client by sending a corresponding request that provides information about the model (e.g., the UE category on which the model is trained).

[0065] In step 714, the FL server 704 constructs a specific positioning model from the base model received from the base model provider 703 (for example, by adding one or more layers). This is optional if the base model already has the same architecture as the model to be trained and is a pre-trained version thereof. Furthermore, the FL server 704 selects the FL clients to be used for federated learning of the model. The selected FL clients are a set of LMFs 705, which may or may not include the LMF client 702.

[0066] Next, the FL server 704 and LMF 705 set perform federated learning 715 to train the model. This involves a loop (checked by the FL server 704) until a stopping criterion is met, with each iteration including: 716 the FL server 704 sending a model training request to the LMF 705 including the current version of the model; 717 the LMF 705 performing local training of the model (using the current version of the model as a starting point) to decide whether to update the current version of the model; 718 the LMF 705 sending the update to the FL server 704; and 719 the FL server 704 updating the model version (also referred to as the global model).

[0067] Next, FL server 704 provides the trained model to LMF702.

[0068] Figure 8 shows flowchart 800, which illustrates the flow for using a positioning model in inference.

[0069] This flow involves UE801, RAN802, AMF803, LMF804, and FL Server 805.

[0070] In 806, UE801, RAN802, or AMF803 issues a positioning request to UE801. In 807, the AMF sends a request to LMF804 to determine the position of UE801, including specifications on which mobile terminal information (shown as UE characteristics in Figure 8) should be used as input data for positioning. This input data for positioning may include, for example, the UE's device type, the UE's mobility information (e.g., speed estimation), or the UE's signal information (e.g., PRS or SRS measurement).

[0071] In 808, LMF804 identifies one or more ML models suitable for determining the positioning information of UE801 using mobile terminal information.

[0072] In step 809, the LMF804 acquires one or more identified ML models (as described, for example, with reference to Figure 7).

[0073] In 810, the LMF804 collects inference data, i.e., mobile terminal information (of the type specified by the AMF in 806).

[0074] In 811, the LMF804 performs inference by supplying the collected inference data to each of the one or more acquired ML models.

[0075] As described above, each of the one or more ML models may directly output the UE position. To improve accuracy, in 812, the outputs of multiple ML models can be combined (e.g., averaged) using an ensemble method. In addition, each ML model may output assistant positioning information. In this case, in 813, the LMF804 supplies the assistant positioning information to another (non-ML) positioning method to determine the UE position.

[0076] In either case, at 814, LMF804 indicates the determined UE position in response to AMF803.

[0077] Note that the base model provider 205 and the FL server 206 may be combined into a single entity. The FL server 206 may be, for example, an NWDAF (which may include the base model provider 205). The FL server 206 may also be an LMF (which may include the base model provider 205).

[0078] According to one embodiment, the base model provider can construct category-specific ML models. In this embodiment, this is as follows: • Base model providers also indicate their ability to provision category-specific ML models and a list of supported category-specific ML models in their NRF registration. The FL server retrieves category-specific ML models from the base model provider and trains them using federative learning.

[0079] According to one embodiment, an ML model can be continuously updated using asynchronous federative learning.

[0080] In summary, according to various embodiments, a method is provided as shown in Figure 9.

[0081] Figure 9 shows a flowchart illustrating a mobile terminal positioning method according to one embodiment.

[0082] In 901, a positioning machine learning model for outputting positioning information in response to mobile terminal state information (i.e., the machine learning model is configured to output positioning information in response to mobile terminal state information) is trained by federated learning using multiple positioning components (in the example above, LMFs, but depending on the architecture or generation of the communication system, these may be positioning components other than LMFs (e.g., with different names)) as federated learning clients.

[0083] In 902, the locations of multiple mobile terminals are determined using a trained positioning machine learning model.

[0084] In various embodiments, in other words, a positioning (ML) model is trained by using positioning components (e.g., computer devices configured to perform mobile terminal positioning) as federated learning clients for federated learning (i.e., running or implementing federated learning clients, e.g., running software that makes them act as federated learning clients for federated learning to train a model, which they can then perform positioning). Since the positioning components are entities that have training data (for the positioning model) at their disposal (i.e., are data gateways), in particular to perform positioning without machine learning and thus obtain training data (mobile terminal information as input and examples of resulting locations as output), this speeds up training and reduces overhead (compared to approaches where the positioning components transfer training data to the training entities). Furthermore, instead of redundant ML model training in multiple positioning components, the positioning components cooperate through federated learning to achieve better performance.

[0085] It is possible to support and train multiple models for positioning mobile devices of different categories (for example, using FL which uses a positioning component as an FL client). Instead of training (similar) models for different categories of mobile devices, a base model can be trained to avoid redundant training and speed up model training. In this way, the commonalities of a set of categories can be incorporated into one or more base models that are trained using aggregated data and used to speed up the training of category-specific ML models.

[0086] The ML model (or, if multiple ML models are trained for multiple categories, each ML model) may output the location of the mobile device directly, or it may output the assistant's positioning information.

[0087] The ML model is trained to generate mobile terminal positioning information from input data that includes PRS signal measurements (of PRS signals received by the mobile terminal), SRS signal measurements (of SRS signals transmitted by the mobile terminal), an estimate of the mobile terminal's speed, information about the mobile terminal such as the radio cell or tracking area where the mobile terminal is located, and the type of UE device.

[0088] As described above, entities can be registered in repositories such as NRF, which act as a repository for registering all of them, namely the positioning component repository, the FL server repository, and the base model provider repository. Registration is done through a profile that indicates each FL / base model training capability, for example, the training of the models they support (e.g., with respect to input and / or output types).

[0089] A method may be performed, and the components of a positioning system that perform the method may be implemented, for example, by one or more circuits. "Circuit" may be understood as any kind of logic implementation entity, which may be a dedicated circuit or processor that runs software, firmware, or any combination thereof stored in memory. Accordingly, "circuit" may be a programmable processor, such as a hard-connected logic circuit or a programmable logic circuit such as a microprocessor. "Circuit" may be a processor that runs software, such as any kind of computer program. Any other kind of implementation of each of the functions described above may be understood as "circuit".

[0090] Although specific embodiments have been described, those skilled in the art will understand that various modifications of form and detail may be made without departing from the scope of the embodiments of this disclosure as defined in the appended claims. The scope is therefore indicated by the appended claims and should encompass all modifications that fall within the equivalent meaning and scope of the claims.

Claims

1. A method for positioning a mobile terminal, A step of training a positioning machine learning model to output positioning information in response to mobile terminal state information by using federated learning, which uses multiple positioning components as federated learning clients, For each of the multiple mobile terminal categories, each positioning machine learning model outputs positioning information in response to mobile terminal state information through federated learning, using its respective multiple positioning components as federated learning clients, and the positions of the multiple mobile terminals in the category are determined using each trained positioning machine learning model. The steps include determining the locations of multiple mobile terminals using the aforementioned trained positioning machine learning model, A method that includes this.

2. The method according to claim 1, wherein the positioning component and / or one or more other positioning components determine the positions of the plurality of mobile terminals using the trained positioning machine learning model.

3. The method according to claim 1, wherein the categories are characterized by different mobility of the mobile terminal and / or different location of the mobile terminal.

4. The method according to claim 1, comprising the steps of registering positioning components that support associative learning in a positioning component repository using a positioning component profile that indicates associative learning capability, and selecting the plurality of positioning components from the registered positioning components.

5. The method according to claim 1, comprising the steps of selecting a federated learning server that supports training the positioning machine learning model, and controlling federated learning by the selected federated learning server.

6. The method according to claim 5, comprising the steps of registering a federative learning server in a federative learning server repository using a federative server profile that indicates federative learning capability, and selecting the federative learning server from the registered federative learning servers.

7. The method according to claim 1, comprising the step of training a base model and using the trained base model as a starting point for training the positioning machine learning model.

8. The method according to claim 7, further comprising the step of adding one or more neural network layers to a base layer in order to construct the positioning machine learning model.

9. The method according to claim 7, comprising the steps of training the base model with a base model provider and obtaining the trained base model from the base model provider.

10. The method according to claim 9, comprising the steps of registering a base model provider in a base model provider repository using a base model profile that indicates the base model training capability, and determining a base model provider from which to obtain a trained base model.

11. The method according to claim 7, further comprising the step of training the base model using training data obtained from the plurality of positioning components and / or one or more further positioning components.

12. The method according to claim 7, further comprising the step of training the base model with reduced training data by sampling, aggregating, and / or filtering the local training data with respect to the amount of local training data available to the plurality of positioning components and / or one or more further positioning components.

13. The method according to claim 7, further comprising the step of training the base model using training data with a reduced amount of training data, with respect to the amount of training data that the federated learning client uses to train the positioning machine learning model.

14. A mobile terminal positioning system configured to perform the method described in any one of claims 1 to 13.

Images