Apparatus and method for user equipment positioning and network node using the same

Abstract

According to the present disclosure, a first data set comprising CSI measurements associated with a first UE and an actual position of the first UE within a coverage area of a network node is received. Then, a second data set comprising at least one CSI measurement associated with at least one second UE whose position within the coverage area of the network node is to be estimated is received. Thereafter, the at least one CSI measurement associated with the at least one second UE is assigned to a first UE among the first UEs that is closest to where the at least one second UE appears based on the first data set and the second data set. Finally, the position of the at least one second UE within the coverage area of the network node is estimated by using a semi-supervised machine learning algorithm that receives the first data set and the second data set and the actual position of the closest first UE.

Description

TECHNICAL FIELD

[0001] The present disclosure relates generally to positioning, and in particular to machine learning assisted techniques for user equipment (UE) positioning in wireless networks. BACKGROUND

[0002] Due to the tremendous progress in processing power and availability of machine learning (ML) tools, such as open source Python ML libraries, ML algorithms have recently attracted high interest in UE positioning in wireless networks. However, the available ML assisted techniques for UE positioning mainly rely on supervised ML algorithms, which require a large amount of labeled data for their correct operation. This requirement makes the use of supervised ML algorithms for UE positioning infeasible, simply due to the fact that the wireless environment is dynamically changing and any ML model needs to be frequently retrained with a large amount of newly acquired labeled data. Unfortunately, frequent access to a large amount of relevant labeled data is a huge challenge. On the other hand, ML models trained with only a small amount of labeled data, like neural networks, are prone to generalization errors, i.e., they make poor predictions on unseen inputs.

[0003] There are also semi-supervised ML assisted techniques for UE positioning, which use a limited amount of labeled training data. However, these semi-supervised ML assisted techniques either are only used for indoor UE positioning (and require multiple access points), or are not specifically designed for high-accuracy UE positioning, and therefore, the resulting position estimates are very inaccurate. SUMMARY

[0004] The summary is provided to introduce a selection of concepts in a simplified form, which will be further described below in the detailed description. This summary is not intended to identify key features or essential features of the disclosure, nor is it intended to be used to limit the scope of the disclosure.

[0005] It is an object of the present disclosure to provide a technical solution that enables high-frequency UE positioning by using a small amount of labeled data.

[0006] The above object is achieved by the features of the independent claims attached hereto. Further embodiments and examples are evident to a person skilled in the art from the dependent claims, the detailed description and the figures.

[0007] According to a first aspect, there is provided an apparatus for UE positioning. The apparatus comprises at least one processor and a memory coupled to the at least one controller. The memory is configured to store processor-executable instructions. When executing the processor-executable instructions, the at least one processor is configured to:

[0008] - receiving a first data set comprising channel state information (CSI) measurements associated with a first UE and an actual position of the first UE within a coverage area of a network node;

[0009] - receiving a second data set comprising at least one CSI measurement associated with at least one second UE, a position of the at least one second UE within the coverage area of the network node to be estimated;

[0010] - based on the first data set and the second data set, assigning the at least one CSI measurement associated with the at least one second UE to a first UE among the first UEs that is closest to an occurrence of the at least one second UE; and

[0011] - based on the first data set and the second data set and the actual position of the closest first UE, estimating the position of the at least one second UE within the coverage area of the network node by using a semi-supervised ML algorithm.

[0012] With the thus configured apparatus, a high-precision UE position estimation can be obtained with minimal supervision, i.e. by using a limited amount of labeled training data. Moreover, unlike existing semi-supervised ML-assisted techniques, the apparatus according to the first aspect of the disclosure is not limited to be applied to indoor scenarios, i.e. it can also be used for UE positioning within a coverage area of a network.

[0013] In one example embodiment of the first aspect, the first UE comprises a radio beacon and / or an unmanned vehicle, the radio beacon being fixedly arranged within the coverage area of the network node, the unmanned vehicle being configured to move along a fixed trajectory within the coverage area of the network node. By using the radio beacon and / or the unmanned vehicle, a sufficient amount of labeled data, i.e. a sufficient first data set, can be provided. Moreover, unmanned vehicles seem to be particularly attractive for office / industrial environments, in which the unmanned vehicles, such as maintenance robots, can participate in collecting the first data set along their planned routes.

[0014] In one example embodiment of the first aspect, the at least one processor is configured to perform said assigning by using a neural network classifier. This can improve the efficiency of the semi-supervised ML algorithm.

[0015] In another example embodiment of the first aspect, the at least one processor is configured to perform said assigning by using a Siamese network. This can improve the efficiency of the semi-supervised ML algorithm.

[0016] In another example embodiment of the first aspect, the at least one processor is configured to perform said assigning by solving a Euclidean distance minimization problem. This can improve the efficiency of the semi-supervised ML algorithm.

[0017] In one example embodiment of the first aspect, the semi-supervised ML algorithm is based on using a pre-trained autoencoder. This can make the apparatus according to the first aspect more flexible in use.

[0018] In one example embodiment of the first aspect, the semi-supervised ML algorithm is based on using a three-term loss function, and wherein the three-term loss function comprises a first term representing a reconstruction error loss related to the autoencoder, a second term representing a mean squared error loss of the first data set, and a third term representing a maximum absolute deviation loss of the second data set. This can further improve the efficiency of the semi-supervised ML algorithm.

[0019] In one example embodiment of the first aspect, the autoencoder comprises an encoder stage, a decoder stage, and a hidden layer therebetween, and wherein each of the encoder stage and the decoder stage comprises one or more convolutional layers and / or dense layers, and the hidden layer is implemented as a two-dimensional layer or a three-dimensional layer of neurons. This configuration of the autoencoder can provide UE position estimates with sufficient accuracy.

[0020] According to the second aspect, a method for UE positioning is provided. The method starts with a step of receiving a first data set comprising CSI measurements associated with a first UE and an actual position of the first UE within a coverage area of a network node. Then, the method proceeds with a step of receiving a second data set comprising at least one CSI measurement associated with at least one second UE, the position of the at least one second UE within the coverage area of the network node to be estimated. Thereafter, the method proceeds with a step of assigning, based on the first data set and the second data set, the at least one CSI measurement associated with the at least one second UE to the first UE that is closest to the at least one second UE. The method ends with a step of estimating, by using a semi-supervised machine learning algorithm, the position of the at least one second UE within the coverage area of the network node, taking into account the first and second data sets and the actual positions of the closest first UEs. By so doing, a high-accuracy UE position estimation can be obtained with minimal supervision, i.e. by using a limited amount of labeled training data. Moreover, unlike existing semi-supervised ML-assisted techniques, the method according to the second aspect of the present disclosure is not limited to application in indoor scenarios.

[0021] In one example embodiment of the second aspect, the first UE comprises a radio beacon and / or an unmanned vehicle, the radio beacon being fixedly arranged within the coverage area of the network node, and the unmanned vehicle being configured to move along a fixed trajectory within the coverage area of the network node. By using the radio beacon and / or the unmanned vehicle, a sufficient amount of labeled data, i.e. a sufficient first data set, can be provided. Moreover, unmanned vehicles appear to be particularly attractive for office / industrial environments, in which they, like maintenance robots, can participate in collecting the first data set along their planned routes.

[0022] In one example embodiment of the second aspect, the step of assigning is performed by using a Siamese network. This can improve the efficiency of the semi-supervised ML algorithm.

[0023] In one example embodiment of the second aspect, the step of assigning is performed by using a Siamese network. This can improve the efficiency of the semi-supervised ML algorithm.

[0024] In another example embodiment of the second aspect, the step of assigning is performed by solving a Euclidean distance minimization problem. This can improve the efficiency of the semi-supervised ML algorithm.

[0025] In another example embodiment of the second aspect, the semi-supervised machine learning algorithm is based on using a pre-trained autoencoder. This can make the method according to the second aspect more flexible in use.

[0026] In one example embodiment of the second aspect, the semi-supervised machine learning algorithm is based on using a tripartite loss function, and wherein the tripartite loss function comprises a first term representing a reconstruction error loss related to the autoencoder, a second term representing a mean squared error loss of the first data set, and a third term representing a maximum absolute deviation loss of the second data set. This can further improve the efficiency of the semi-supervised ML algorithm.

[0027] In one example embodiment of the second aspect, the autoencoder comprises an encoder stage, a decoder stage, and a hidden layer therebetween, and wherein each of the encoder stage and the decoder stage comprises one or more convolutional layers and / or dense layers, the hidden layer being implemented as a two-dimensional layer or a three-dimensional layer of neurons. This configuration of the autoencoder can provide UE position estimation with sufficient accuracy.

[0028] According to a third aspect, a computer program product comprising a computer readable medium is provided. The computer readable medium stores a computer program. The computer program, when executed by at least one processor, causes the at least one processor to perform the method according to the second aspect of the present disclosure. This can simplify the implementation of the method according to the second aspect of the present disclosure.

[0029] According to a fourth aspect, a network node is provided. The network node according to the first aspect of the present disclosure comprises a baseband unit (BBU) and means for UE positioning. The BBU is configured to:

[0030] - perform a CSI measurement associated with a first UE and at least one CSI measurement associated with at least one second UE;

[0031] - generate a first data set by combining the CSI measurement associated with the first UE and an actual position of the first UE within a coverage area of the network node;

[0032] - generating a second data set based on at least one CSI measurement associated with at least one second UE; and

[0033] - providing the first and second data sets to an apparatus for UE positioning according to the first aspect of the disclosure.

[0034] With the network node thus configured, high-precision UE position estimates can be obtained with minimal supervision, i.e. by using a limited amount of labeled training data. Moreover, unlike existing semi-supervised ML-assisted techniques, the network node according to the fourth aspect of the disclosure is not limited to be applied to indoor scenarios. It should also be noted that the network node according to the fourth aspect of the disclosure can independently provide high-precision UE position estimates, i.e. without having to cooperate with any other network node (as is the case for existing semi-supervised ML techniques).

[0035] According to a fifth aspect, an apparatus for UE positioning is provided. The apparatus comprises means for receiving a first data set comprising channel state information (CSI) measurements associated with a first UE and an actual position of the first UE within a coverage area of a network node. The apparatus further comprises means for receiving a second data set comprising at least one CSI measurement associated with at least one second UE. The apparatus further comprises means for assigning the at least one CSI measurement associated with the at least one second UE to a first UE among the first UEs that is closest to an occurrence of the at least one second UE based on the first data set and the second data set. The apparatus further comprises means for estimating a position of the at least one second UE within the coverage area of the network node by using a semi-supervised ML algorithm that receives the first data set and the second data set and the actual position of the closest first UE as input. With the apparatus thus configured, high-precision UE position estimates can be obtained with minimal supervision, i.e. by using a limited amount of labeled training data. Moreover, unlike existing semi-supervised ML-assisted techniques, the apparatus according to the fifth aspect of the disclosure is not limited to be applied to indoor scenarios, i.e. it can also be used for UE positioning within a network coverage area.

[0036] Other features and advantages of the present disclosure will be apparent from reading the following detailed description, and viewing the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0037] The essence of the disclosure is explained below with reference to the drawings, in which:

[0038] Figure 1 a block diagram of an apparatus for UE positioning according to one example embodiment is shown;

[0039] Figure 2 a flowchart of a method for UE positioning according to one example embodiment is shown.

[0040] Figure 3 a flowchart illustrating a method for training an autoencoder used in a semi-supervised ML algorithm according to one example embodiment is shown;

[0041] Figure 4 a block diagram of an autoencoder according to one example embodiment is shown;

[0042] Figure 5 a block diagram of a network node according to one example embodiment is shown;

[0043] Figure 6 a flowchart illustrating a method for operating Figure 5 the network node shown;

[0044] Figure 7 a hypothetical coverage area for simulation purposes is shown;

[0045] Figure 8A and Figure 8B simulation results obtained by using Figure 2 the method shown and an autoencoder pre-trained according to Figure 3 the method shown;

[0046] Figure 9A and Figure 9B simulation results obtained by using one semi-supervised ML assisted technique from the prior art are shown. DETAILED DESCRIPTION

[0047] Various embodiments of the present disclosure are described in further detail with reference to the drawings. However, the present disclosure can be embodied in many other forms and should not be construed as being limited to any specific structure or function discussed in the following description. Rather, these embodiments are provided so that the description of the present disclosure will be more thorough and complete.

[0048] From the detailed description, it will be apparent to those skilled in the art that the scope of the present disclosure includes any embodiment of the present disclosure disclosed herein, regardless of whether the embodiment is implemented independently or in conjunction with any other embodiment of the present disclosure. For example, the apparatuses and methods disclosed herein can be implemented in practice by using any number of embodiments provided herein. Also, it should be understood that any embodiment of the present disclosure can be implemented using one or more elements presented in the appended claims.

[0049] The word "example" is used herein in the sense of "serving as an example." Unless otherwise stated, any embodiment described herein as "example" should not be interpreted as being superior or having advantage over other embodiments.

[0050] According to the example embodiments disclosed herein, a first user equipment or simply first UE can refer to a special UE that is in a known location in a predefined area of a wireless network (e.g., a radio access network (RAN)) and transmits continuous or periodic wireless signals with limited information (e.g., its identity or location). The first UE can include a radio beacon that is fixedly arranged within the predefined area, and / or an unmanned vehicle that is configured to move along a fixed trajectory within the predefined area. Some examples of the unmanned vehicle include a maintenance robot, a delivery drone, other different unmanned ground and air vehicles, etc.

[0051] According to the example embodiments disclosed herein, at least one second user equipment or simply second UE can refer to a UE that is in the same predefined area as the first UE but whose location is unknown and should be estimated. Examples of the second UE include a mobile device, a mobile station, a terminal, a subscriber unit, a mobile phone, a cellular phone, a smartphone, a cordless phone, a personal digital assistant (PDA), a wireless communication device, a laptop computer, a tablet computer, a gaming device, a netbook, a smartbook, an ultrabook, a medical appliance or device, a biosensor, a wearable device (e.g., a smartwatch, smartglasses, a smart wristband), an entertainment device (e.g., an audio player, a video player, etc.), a vehicle component or sensor, a smart meter / sensor, an industrial manufacturing device, a global positioning system (GPS) device, an Internet of Things (IoT) device, a machine type communication (MTC) device, a group of massive Internet of Things (MIoT) or massive MTC (mMTC) devices / sensors, or any other suitable device configured to support wireless communications. In some example embodiments, the second UE can refer to at least two collocated and interconnected UEs as defined by this.

[0052] According to the example embodiments disclosed herein, a network node can refer to a fixed communication point of the first and second UEs in a predefined area of a wireless network. The network node can refer to a base transceiver station (BTS) according to 2G communication technology, a Node B according to 3G communication technology, an evolved Node B (eNodeB) according to 4G communication technology, a gNB according to 5G new radio (NR) communication technology. The predefined area can be represented by a coverage area served by the network node. In turn, the coverage area can be represented by different cells, such as macro cells, micro cells, pico cells, femto cells, and / or other types of cells. A macro cell can cover a relatively large geographical area (e.g., a radius of at least a few kilometers). For example, a micro cell can cover a geographical area with a radius of less than two kilometers. A pico cell can cover a relatively small geographical area, such as an office, a shopping mall, a train station, a stock exchange, etc. A femto cell can cover an even smaller geographical area (e.g., a home). Accordingly, a network node serving a macro cell can be referred to as a macro node, a node serving a micro cell can be referred to as a micro node, and so on.

[0053] According to example embodiments disclosed herein, UE positioning can refer to a ML- assisted technique for estimating an unknown position of a second UE by using specific radio measurements associated with a first UE whose position is known to be in a coverage area of a network node. The measurements are intended to be made by the network node and can include channel state information (CSI) measurements. The CSI measurements can be made in time or frequency domain and characterize the actual communication channel (i.e., channel coefficients) between a transmit antenna at the UE and a receive antenna array at the network node.

[0054] Existing ML-assisted techniques for UE positioning mainly rely on supervised ML algorithms that require a large amount of labeled (or in other words, training) data for their correct operation. Although supervised ML algorithms can potentially provide sufficient accuracy of UE position estimates, their practical implementation for UE positioning faces challenges because any wireless network is a dynamic changing environment and any supervised ML model (e.g., neural network) needs to be frequently retrained with a large amount of newly acquired labeled data. It is not trivial to quickly and frequently access such a large amount of labeled data. At the same time, intentionally reducing the amount of labeled data can lead to situations where the supervised ML algorithm makes poor predictions of UE positions.

[0055] There are also some existing semi-supervised ML-assisted techniques for UE positioning that require much less labeled data than supervised ML-assisted techniques. However, these semi-supervised ML-assisted techniques have the following drawbacks:

[0056] — they are specific to indoor scenarios and require multiple access points.

[0057] — they require the target UE to actively participate in the positioning process (i.e., the UE needs to report specific radio measurements to a central server);

[0058] — they are not precisely designed for high-precision UE positioning but for obtaining a logical map of UEs, and thus the resulting UE position estimates are very inaccurate.

[0059] The example embodiments disclosed herein provide a technical solution that allows to mitigate or even eliminate the above-mentioned drawbacks typical for the prior art. In particular, the technical solution relates to UE positioning using semi-supervised ML algorithms based on CSI measurements. Since CSI is a function of the UE position, the idea is to exploit it to estimate the UE position at the network node. Unlike existing supervised ML-assisted techniques, the technical solution disclosed herein focuses on achieving high accuracy of UE position estimation with a limited amount of labeled data. Unlike existing semi-supervised ML-assisted techniques, the technical solution disclosed herein is not limited to be applied to indoor scenarios and only requires one network node (of course, multiple network nodes can also be used, which can even lead to higher accuracy, but at the cost of increased complexity). Moreover, unlike existing semi-supervised ML-assisted techniques, the technical solution disclosed herein does not require the target UE (i.e., the second UE according to the present disclosure) to actively participate in the positioning procedure itself: the target UE does not need to report any radio measurements (e.g., Time of Arrival (ToA), Direction of Arrival (DoA), Received Signal Strength (RSS), etc.) to the network node, since CSI measurements can be made at the network node when the target UE transmits the pilot sequence.

[0060] Figure 1 A block diagram of an apparatus 100 for UE positioning according to one example embodiment is shown. As shown, the apparatus 100 includes a processor 102 and a memory 104. As described later, the memory 104 is coupled to the processor 102 and stores processor-executable instructions 106 that, when executed by the processor 102, cause the processor 102 to perform various aspects of the present disclosure. It should be noted that the number, arrangement, and interconnection of the constituent elements that make up the apparatus 100 shown in Figure 1 Figure 1 The number, arrangement, and interconnection of the constituent elements that make up the apparatus 100 shown in FIG. 1 are not intended to limit the present disclosure in any way, but merely to provide an overall sense of how the constituent elements can be implemented within the apparatus 100.

[0061] The processor 102 can be implemented as a central processing unit (CPU), a general purpose processor, a single purpose processor, a microcontroller, a microprocessor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a complex programmable logic device, etc. It should also be noted that the processor 102 can be implemented as any combination of one or more of the above. As an example, the processor can be a combination of two or more microprocessors.

[0062] ​The memory 104 can be implemented as non-volatile or volatile memory used in modern electronic computers. For example, non-volatile memory can include read-only memory (ROM), ferroelectric random-access memory (RAM), programmable ROM (PROM), electrically erasable PROM (EEPROM), solid-state drive (SSD), flash memory, magnetic disk storage such as hard drives and magnetic tape, optical disk storage such as CDs, DVDs, and Blu-ray discs, and the like, examples of which include dynamic RAM, synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and static RAM, among others.

[0063] The processor-executable instructions 106 stored in the memory 104 can be configured to cause the processor 102 to execute computer-executable code of aspects of the present disclosure. The computer-executable code can be written in any combination of one or more programming languages, such as Java, C++, and the like. In some examples, the computer-executable code can be in the form of a high-level language or pre-compiled, and can be generated by an interpreter (also pre-stored in the memory 104) on the fly.

[0064] Figure 2 A flowchart of a method 200 for UE positioning according to one example embodiment is shown. The method 200 comprises steps S202 to S208, each of which is performed by the processor 102 comprised in the apparatus 100. The method 200 starts from step S202, where the processor 102 receives a first data set comprising CSI measurements associated with a first UE and an actual position of the first UE within a coverage area of a network node. Then, the method 200 proceeds to step S204, where the processor 102 receives a second data set comprising at least one CSI measurement associated with at least one second UE, the position of which is to be estimated within the coverage area of the network node. The first data set can be referred to as a labeled data set, while the second data set can be referred to as an unlabeled data set. It should be noted that steps S202 and S204 can be performed in parallel, if desired. As for the first and second data sets themselves, they are prepared by the network node, which will be described later. Thereafter, the method 200 proceeds to step S206, where the processor 102 assigns the at least one CSI measurement associated with the at least one second UE to the first UE that is closest to the occurrence of the at least one second UE using the first and second data sets. The method 200 ends with step 208, which estimates the position of the at least one second UE within the coverage area of the network node by using a semi-supervised ML algorithm that receives the first and second data sets and the actual positions of the closest first UEs as input.

[0065] In some example embodiments, step S206 of method 200 is performed by processor 102 of apparatus 100 by using a neural network classifier, a Siamese network or by solving a Euclidean distance minimization problem. Each of these embodiments can allow the selection of the most relevant first UE among the first UEs in the first UE to be selected as the closest neighbor of the second UE, which in turn can improve the efficiency of the semi-supervised ML algorithm which is then used in step S208 of method 200.

[0066] In one example embodiment, the semi-supervised ML algorithm used in step S208 of method 200 is based on using a pre-trained autoencoder. Furthermore, the autoencoder can be pre-trained by using a three-term loss function, where the first term represents a reconstruction error loss related to the autoencoder, the second term represents a mean squared error loss of the first dataset, and the third term represents a maximum absolute deviation loss of the second dataset. The combination of all three terms is crucial for obtaining meaningful and accurate position estimates at the autoencoder.

[0067] Figure 3 A flowchart of a method 300 for an autoencoder 302 used in training a semi-supervised ML algorithm is shown, according to one example embodiment. In this example embodiment, it is assumed that the first UEs are represented by fixed radio beacons distributed within the coverage area of the network node. The exact number of radio beacons and their fixed positions depend on the coverage area of the network node. In general, the more dense the distribution of radio beacons, the better. The whole method 300 can be divided into three steps: S1, S2 and S3. Step S1 is a data preparation step, where the network node prepares the first dataset and the second dataset, and processor 102 of apparatus 100 receives it from the network node in steps S202 and S204 of method 200, i.e., it aims to assign each unlabeled data of the second dataset to its closest beacon (the closeness is defined in terms of the Euclidean distance between the source of the unlabeled data, i.e., the second UE, and the radio beacon). Step S3 is a training step, where processor 102 of apparatus 100 trains the autoencoder 302 using the first dataset and the second dataset and the result of step S2, i.e., the actual positions of the closest radio beacons. It should be clear that steps S1, S2 and S3 should be performed in the order of their numbering. Each of steps S1, S2 and S3 will be described in further detail.

[0068] Step S1

[0069] It is assumed that represents a set of beacon indices, where N b is the total number of radio beacons 304 involved. Associated with each beacon index b is a pair (x CSI，b , y b), where X CSI，b represents the CSI 306 of beacon b measured at the network node, and y b represents the actual position 308 of beacon b. Given this, a set Constructing a labeled dataset That is, the first data set according to the present disclosure. Next, let represents an unlabeled dataset, i.e., the second dataset according to the present disclosure. Each element is a CSI measurement 310, represented by X CSI = , which is associated with a second UE (not shown), but has no specific association with the location closest to the beacon. Therefore, the set of all CSI measurements is Know combination.

[0070] It should be noted that in some cases, multiple CSI measurements {X CSI，b，i , i=1,2,…,b}. In this case, X CSI，b is the collection of these measurements. This helps mitigate time-varying effects caused by moving scatterers. Therefore, the first data set is In the following part of this disclosure, we assume that there is a single measurement per beacon, but those skilled in the art will understand that the steps of method 300 are similar for the case of multiple CSI measurements.

[0071] Once the beacon location 308 is identified, For each radio beacon, it is necessary to find all points that are closer to its location than any other radio beacon in the radio beacons 304. To do this, let R denote the coverage area of ​​the network node. In other words, is the area in which the network node attempts to find the location of the second UE. Next, the following estimates should be made:

[0072]

[0073]

[0074] In other words, execution The Voronoi tessellation is used to obtain the area of ​​each radio beacon. Each point in is closer to the location of beacon b than any other beacon, and d max (b) Yes The maximum distance from any point in the beacon b. Then the d of the radio beacon 304 max All values ​​of (b) are stored in database 312 for further use.

[0075] d max The exact value of (b) may be difficult to calculate, but in this case an approximation will suffice. One possible way to calculate this approximation is as follows:

[0076] 1) A large number of points are evenly distributed over the entire coverage area of ​​the network node. The points represent fixed locations of radio beacons 304. The larger the number of points, the smaller the Euclidean distance between adjacent points, and thus the better the accuracy of the approximation.

[0077] 2) Cluster the points by assigning each point to its closest radio beacon (proximity is related to Euclidean distance). This results in the formation of N b clusters, one per beacon.

[0078] 3) Within each cluster (identified by its associated closest radio beacon), calculate the maximum distance from the point to the closest beacon. This represents the d max Approximate value of .

[0079] Step S2

[0080] The assignment step can be performed by constructing a classifier function C that assigns each unlabeled XSI measure X CSI The beacon index assigned to the beacon whose location is closest to its unknown associated location (ie, the location of the second UE). More specifically, the classifier function C is used to classify X cSI Perform the following operations, whose associated unknown position is given by y:

[0081]

[0082] Several candidates for the classifier function C can be shown as follows:

[0083] 1. A function that returns a parameter that is used for all X CSI and X CSI，b The Euclidean distance between the appropriate components of is minimal.

[0084] 2. A neural network classifier that is trained on a separate set of labeled data for classification tasks.

[0085] 3. A Siamese network that is trained based on a separate but small set of labeled data for classification purposes. A Siamese network for image classification and recognition can be used to construct a classifier function X.

[0086] Step S3

[0087] The training step is based on training the autoencoder 302 using the results of the usage steps S1 and S2 (i.e. the first and second data sets and the actual positions closest to the radio beacons) as input.

[0088] Figure 4 A block diagram of the autoencoder 302 according to one example embodiment is shown. As Figure 4 shown, the autoencoder 302 comprises an encoder stage 402 and a decoder stage 404 with an hidden layer 406 in between. The encoder stage 402 can comprise a plurality of neuron layers 408-1, 408-2,..., 408-n. Similarly, the decoder stage 404 can comprise a plurality of neuron layers 410-1, 410-2,..., 410-n. The layers of the encoder stage 402 and the decoder stage 404 can be convolutional layers or dense layers, depending on the type of CSI measurements used. Of course, this configuration of the autoencoder 302 should not be seen as any limitation of the present disclosure, and other example embodiments are possible in which e.g. dense layers are used instead of convolutional layers, or vice versa. The hidden layer 406 is typically a two-dimensional or three-dimensional layer of neurons, whose output represents the position estimate of the second UE. More specifically, for an input X CSI to the encoder stage 402, the output of the hidden layer 406 is denoted by f θ (X CSI ), where {0} is a set of parameters / weights of the autoencoder 302. The output of the decoder stage 404 is denoted by g θ (X CSI ).

[0089] In the training process, a three-term loss function 314 as defined above is preferably used. That is, the loss function comprises a first term 316 representing a reconstruction error loss related to the autoencoder 302, a second term 318 representing a mean squared error (MSE) loss for the first data set, and a third term 320 representing a maximum absolute deviation (MAD) loss for the second data set. Given the above representations, the loss function 314 for training the autoencoder 302 can then be written as follows:

[0090]

[0091] Here, w1, w2 e (0, 1] are weights assigned to the last two terms (second and third term) of the loss function 314. These weights are hyperparameters, which need to be chosen heuristically, e.g. by grid search or by meta-learning.

[0092] In some example embodiments, in addition to the beacon location, the network node may know the locations of some other UEs that are not the first UE (one example is when these UEs know their locations using a global navigation satellite system (GNSS) and report them to the network node). Unlike the first UE, these UEs are not used in step S206 of method 200, but should be considered to form part of the MSE loss (i.e., the second term 318 of the total loss function 314).

[0093] Step S3 is repeated until the loss function 314 (ie, L) is minimized, or until a predetermined number of training cycles is reached. Thereafter, the pre-trained autoencoder 302 may be used in step S208 of the method 200 .

[0094] Figure 5 1 shows a block diagram of a network node 500 according to an example embodiment. Figure 5 As shown, the network node 500 includes a baseband unit (BBU) 502, a memory 504, and the apparatus 100. As will be explained later, the memory 504 is coupled to the BBU 502 and stores executable instructions 506 that, when executed by the BBU 502, cause the BBU 502 to perform aspects of the present disclosure. It should be noted that Figure 5 The number, arrangement, and interconnection of the constituent elements constituting the network node 500 shown in FIG. 5 are not intended to be any limitation of the present disclosure, but are merely used to provide a general idea of ​​how the constituent elements may be implemented within the network node 500 .

[0095] The BBU is also referred to as a baseband processor and may refer to an electronic device (ED) or an ED pool, in which baseband processing is performed on the communication channel between the network node 500 and the first and second UEs. Some practical examples of a BBU include a front-end processor, a communication controller, or any other device with similar functions. Generally speaking, the BBU is responsible for digital signal processing and may, for example, perform the following functions: CSI measurement, digital intermediate frequency (IF) to baseband conversion / digital baseband to IF conversion, modulation / demodulation, constellation mapping / demapping, scrambling / descrambling, and / or encoding / decoding.

[0096] As for the memory 504 and the executable instructions 506 , they may be implemented in the same or similar manner as the memory 104 and the processor-executable instructions 106 in the apparatus 100 , respectively.

[0097] Figure 6A flow chart of a method 600 for operating the network node 500 according to one example embodiment is shown. The method 600 comprises steps S602 to S608, each of which is performed by the BBU 502 comprised in the network node 500. The method 600 starts with step S602, where the BBU 502 performs a CSI measurement associated with a first UE and at least one CSI measurement associated with at least one second UE. Then, the method 600 proceeds with step S604, where the BBU 502 generates a first data set by combining the CSI measurement associated with the first UE and an actual position of the first UE within a coverage area of the network node 500. Thereafter, the next step S606 is initiated, where the BBU 502 generates a second data set based on the at least one CSI measurement associated with the at least one second UE. It should be noted that steps S604 and S606 can be performed in parallel by the BBU 502, if desired. Generally, steps S602 to S606 constitute step S1 (data preparation) of the method 300. The method 600 ends with step S608, where the BBU 502 provides the first data set and the second data set to the apparatus 100 for use thereof in the UE positioning according to the method 200.

[0098] Simulation results

[0099] Figure 7 A hypothetical coverage area for simulation purposes is shown. In particular, the coverage area is formed by 16384 points arranged in a conical fan, as shown in Figure 7 (where each of the X- and Y-axes represents a distance in meters). As with the network node 500, the network node is located at the sector origin (labeled "Rx" in Figure 7 ). 400 radio beacons are evenly spaced over the conical sector (these radio beacons are represented by the black dots in Figure 7 ). A simple line-of-sight model with 16 antennas at the network node is used to construct this conical sector.

[0100] Figure 8A and Figure 8B Simulation results obtained by using the method 200 and the autoencoder 302 pre-trained according to the method 300 are shown. In particular, Figure 8A the above-described Figure 7 , i.e. the actual distribution of beacon positions within the conical fan, are repeated. For ease of comparison, the Figure 8A are repeated here. Figure 8BThe reconstructed distribution of beacon positions within the same conical sector is shown. During the simulation, each radio beacon was alternatively considered as the second UE. In other words, the position of the radio beacon was estimated by using the method 200 and the autoencoder 302 to check whether it matches the radio beacon actual position. As a result of this simulation, the achieved average positioning accuracy is 11.2 m.

[0101] Figure 9A and Figure 9B The simulation results obtained by using the semi-supervised ML-assisted technique disclosed in the following article are shown: C. Studer, S. Medjkouh, E. T. Goldstein, and O. Tirkkonen, “Channel Charting: Locating Users Within the Radio Environment Using Channel State Information,” in IEEE Access, vol. 6, pp. 47682-47698, 2018. Similarly, for the sake of comparison, Figure 9A repeated Figure 7 . Figure 9B The reconstructed distribution of beacon positions within the same conical sector is shown. As a result of this simulation based on the state-of-the-art semi-supervised ML technique, the achieved average positioning accuracy is 25.7 m. Therefore, the selected semi-supervised ML-assisted technique provides a worse positioning accuracy than the one of the method 200. As a safeguard, it should be noted that the selected semi-supervised ML-assisted technique is not really designed for actual UE positioning, but only to highlight the advantages of the method 200.

[0102] It should be noted that each block, step or operation of methods 200, 300 and 600, or any combination of these blocks, steps and operations, can be implemented by various means, for example, hardware, firmware, and / or software. As an example, one or more of the blocks, steps or operations above can be implemented by processor-executable instructions, data structures, program modules and other suitable data representations. Moreover, the processor-executable instructions implementing the blocks, steps or operations above can be stored on a corresponding data carrier and executed by at least one processor that respectively implements the functionality of apparatus 100 and network node 500. The data carrier can be implemented as any computer readable storage medium configured to be readable by said at least one processor to perform the processor-executable instructions. Such computer readable storage medium can include volatile and non-volatile media, removable and non-removable media. By way of example, and not limitation, computer readable media comprise media implemented in any method or technology for storing information. More specifically, computer readable media include, but are not limited to, information delivery media, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD), holographic media or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage and other magnetic storage devices.

[0103] Although example embodiments of the present disclosure are described herein, it should be noted that any variations and modifications thereof can be made without departing from the legal protection defined by the appended claims. In the appended claims, the word "comprising" does not exclude other elements or operations, the indefinite article "a" or "an" does not exclude a plurality, and the conjunction "or" is not intended to be exclusive. In the dependent claims, the reference sign introduced by "one" does not preclude a plurality. The use of relative terms in the description herein such as "one" or "another" carries no preference or requirement for a single or plural number.

Claims

1. An apparatus (100) for user equipment (UE) positioning, comprising: at least one processor (102); and a memory (104) coupled to the at least one processor and storing processor- executable instructions (106), wherein the at least one processor (102), when executing the processor-executable instructions, is configured to: - receive (S202) a first data set comprising channel state information (CSI) measurements associated with a first UE and an actual position of the first UE within a coverage area of a network node (500); - receive (S204) a second data set comprising at least one CSI measurement associated with at least one second UE whose position within the coverage area of the network node (500) is to be estimated; - assign (S206), based on the first data set and the second data set, the at least one CSI measurement associated with the at least one second UE to a first UE among the first UEs that is closest to an occurrence of the at least one second UE, characterized in that, when executing the processor-executable instructions, the at least one processor is further configured to: - estimate (S208), based on the first data set and the second data set and the actual position of the closest first UE, the position of the at least one second UE within the coverage area of the network node (500) by using a semi-supervised machine learning algorithm, - wherein the semi-supervised machine learning algorithm is based on using a pre-trained autoencoder (302) and a tripartite loss function (314), and wherein the tripartite loss function comprises a first term (316) representing a reconstruction error loss related to the autoencoder, a second term (318) representing a mean squared error loss of the first data set, and a third term (320) representing a maximum absolute deviation loss of the second data set.

2. The apparatus (100) of claim 1, wherein the first UEs comprise radio beacons (304) and / or unmanned vehicles, the radio beacons (304) being fixedly arranged within the coverage area of the network node (500), the unmanned vehicles being configured to move along a fixed trajectory within the coverage area of the network node (500).

3. The apparatus (100) of claim 1 or 2, wherein the at least one processor (102) is configured to perform the assigning by using a neural network classifier.

4. The apparatus (100) of claim 1 or 2, wherein the at least one processor (102) is configured to perform the assigning by using a Siamese network.

5. The apparatus (100) of claim 1 or 2, wherein the at least one processor (102) is configured to perform the assigning by solving a Euclidean distance minimization problem.

6. The apparatus (100) of claim 1, wherein the autoencoder (302) comprises an encoder stage (402), a decoder stage (404) and an in-between hidden layer (406), and wherein each of the encoder stage and the decoder stage comprises one or more convolutional layers and / or dense layers, and the hidden layer is implemented as a two-dimensional or three-dimensional layer of neurons.

7. A method (200) for user equipment (UE) positioning, comprising: - receiving (S202) a first data set comprising channel state information (CSI) measurements associated with a first UE and an actual position of the first UE within a coverage area of a network node (500); - receiving (S204) a second data set comprising at least one CSI measurement associated with at least one second UE whose position within the coverage area of the network node (500) is to be estimated; - assigning (S206) the at least one CSI measurement associated with the at least one second UE to a first UE among the first UEs that is closest to where the at least one second UE appears based on the first data set and the second data set; characterized in that the method further comprises: - estimating (S208) the position of the at least one second UE within the coverage area of the network node (500) by using a semi-supervised machine learning algorithm based on the first data set and the second data set and the actual position of the closest first UE, - wherein the semi-supervised machine learning algorithm is based on using a pre-trained autoencoder (302) and a three-term loss function (314), and wherein the three-term loss function comprises a first term (316) representing a reconstruction error loss related to the autoencoder, a second term (318) representing a mean squared error loss of the first data set and a third term (320) representing a maximum absolute deviation loss of the second data set.

8. The method (200) of claim 7, wherein the first UEs comprise radio beacons (304) and / or unmanned vehicles, the radio beacons (304) being fixedly arranged within the coverage area of the network node (500), the unmanned vehicles being configured to move along a fixed trajectory within the coverage area of the network node (500).

9. The method (200) of claim 7 or 8, wherein the assigning is performed by using a neural network classifier.

10. The method (200) of claim 7 or 8, wherein the assigning is performed by using a Siamese network.

11. The method (200) of claim 7 or 8, wherein the assigning is performed by solving a Euclidean distance minimization problem.

12. The method (200) of claim 7, wherein the autoencoder (302) comprises an encoder stage (402), a decoder stage (404), and a hidden layer (406) therebetween, and wherein each of the encoder stage and the decoder stage comprises one or more convolutional layers and / or dense layers, and the hidden layer is implemented as a two-dimensional or three-dimensional layer of neurons.

13. A computer program product comprising a computer-readable medium storing a computer program, wherein the computer program, when executed by at least one processor, causes the at least one processor to carry out the method according to any one of claims 7 to 12.

14. A network node (500) comprising: a baseband unit (BBU) (502); and an apparatus (100) for UE positioning according to any one of claims 1 to 6, wherein the BBU (502) is configured to: - make the CSI measurements associated with the first UE and the at least one CSI measurement associated with the at least one second UE (S602); - generate the first data set by combining the CSI measurements associated with the first UE and the actual position of the first UE within the coverage area of the network node (S604); - generate the second data set based on the at least one CSI measurement associated with the at least one second UE (S606); and provide the first data set and the second data set to the apparatus for UE positioning (S608). ​

Images