Training in communication systems

By generating multiple loss function components and using data enhancement technology, combining channel state data and its enhanced version, the machine learning model is trained, and the accuracy and model robustness of user equipment positioning in the communication system is solved, and the unlabeled Better predictive power of data.

CN113128679BActive Publication Date: 2024-11-08NOKIA TECHNOLOGIES OY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110050487.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2020-01-14
Filing Date
2021-01-14
Publication Date
2024-11-08
Estimated Expiration
2041-01-14

AI Technical Summary

Technical Problem

In communication systems, the prior art is difficult to effectively solve the positioning problem of user equipment, especially due to the changing nature of the relationship between channel state information and user equipment location, the model needs to be retrained frequently, and the supervised learning algorithm relies on a large amount of labeled data. limit.

Method used

By generating multiple loss function components, combining channel state data and their enhanced versions, training parameters using machine learning models (such as neural networks) to generate positioning information of user equipment, and using data augmentation techniques such as phase rotation, amplitude transformation, and addition Noise and filtering make the model more robust to generalization errors.

Benefits of technology

It improves the accuracy of user equipment positioning and the robustness of the model, reduces the dependence on labeled data, reduces the frequency of model retraining, and enhances the prediction ability of unlabeled data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113128679B_ABST
    Figure CN113128679B_ABST
Patent Text Reader

Abstract

Embodiments of the present disclosure relate to training in a communication system. An apparatus, method, and computer program are described, including: generating a first loss function component, including comparing first position data with a first position estimate, the first position estimate is based on channel state data, the first position estimate is generated using a model, and the model includes multiple trainable parameters; generating a second loss function component, including comparing the first position data with a second position estimate, the second position estimate is based on channel state data that has undergone a first enhancement, and the second position estimate is generated using the model; generating a third loss function component, including comparing a third position estimate based on the channel state data and a fourth position estimate based on the channel state data that has undergone a second enhancement, the third position estimate and the fourth position estimate are generated using the model; and training the trainable parameters of the model by minimizing a loss function based on a combination of the first, second, and third loss function components.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of Finnish patent application No. 20205030 filed on January 14, 2020, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This specification relates to training in a communication system. Background Art

[0004] For many reasons, it may be desirable to know the position of a device such as a user equipment in space (eg, absolute position, or relative to a communication node such as a base station).There is still a need for further developments in this area. Summary of the invention

[0005] In a first aspect, the specification describes an apparatus comprising components for performing the following: generating a first loss function component, comprising comparing first location data (e.g., measured location data of a user device) with a first location estimate (e.g., a location estimate of the user device), wherein the first location estimate is based on channel state data, wherein the first location estimate is generated using a model, and wherein the model includes a plurality of trainable parameters; generating a second loss function component, comprising comparing the first location data with a second location estimate, wherein the second location estimate is based on channel state data that has undergone a first enhancement, and wherein the second location estimate is generated using the model; generating a third loss function component, comprising comparing a third location estimate based on the channel state data with a fourth location estimate based on the channel state data that has undergone a second enhancement, wherein the third location estimate and the fourth location estimate are generated using the model; and training the trainable parameters of the model by minimizing a loss function based on a combination of the first loss function component, the second loss function component, and the third loss function component. The combination may include a sum (e.g., a weighted sum) of the first loss function component, the second loss function component, and the third loss function component. The model may be configured to estimate the location of the user equipment based on the channel state information for communication between the user equipment and the communication node.

[0006] In some example embodiments, the component is further configured to perform: generating a loss function by combining the first loss function component, the second loss function component, and the third loss function component.

[0007] In some example embodiments, the component is further configured to perform: receiving, retrieving or otherwise acquiring a first data set, wherein the first data set includes tagged data, the tagged data including channel state information and associated position data, wherein the first position data is derived from the associated position data (e.g., the position data may be position data). The component may also be configured to perform: generating a first position estimate based on at least some of the channel state data of the first data set; and generating a second position estimate based on at least some of the channel state data in the first data set that has been subjected to the first enhancement.

[0008] In some example embodiments, the component is further configured to perform: receiving, retrieving or otherwise acquiring a second data set, wherein the second data set includes unlabeled data, the unlabeled data including channel state information. The component may also be configured to perform: generating a third position estimate based on at least some of the channel state data of the second data set; and generating a fourth position estimate based on at least some of the channel state data in the second data set that has been subjected to the second enhancement.

[0009] In some example embodiments, each enhancement comprises one or more of: phase rotation; amplitude amplification; amplitude reduction; addition of Gaussian noise; filtering or smoothing; neural network based enhancement; or an element replacement algorithm.

[0010] Each enhancement may provide a random or pseudo-random transformation of the relevant channel state information.

[0011] In some example embodiments, the component is further configured to perform: generating a plurality of the second loss function components by comparing the first position data with a plurality of second position estimates, the plurality of second position estimates being based on channel state data of different instances that have undergone the first enhancement; and / or generating a plurality of the third loss function components by comparing the third position estimate with a plurality of fourth position estimates, the plurality of fourth position estimates being based on channel state data of different instances that have undergone the second enhancement.

[0012] The model may be a machine learning model.

[0013] This model can be implemented using a neural network.

[0014] The component may include: at least one processor; and at least one memory including computer program code, the at least one memory and the computer program being configured to implement the performance of the device together with the at least one processor.

[0015] In a second aspect, the present specification describes a method comprising: generating a first loss function component, comprising comparing first position data with a first position estimate, wherein the first position estimate is based on channel state data, wherein the first position estimate is generated using a model, and wherein the model includes a plurality of trainable parameters; generating a second loss function component, comprising comparing the first position data with a second position estimate, wherein the second position estimate is based on channel state data that has been first enhanced, and wherein the second position estimate is generated using the model; generating a third loss function component, comprising comparing a third position estimate based on the channel state data with a fourth position estimate based on the channel state data that has been second enhanced, wherein the third position estimate and the fourth position estimate are generated using the model; and training the trainable parameters of the model by minimizing a loss function based on a combination of the first loss function component, the second loss function component, and the third loss function component. The combination may include the sum (e.g., weighted sum) of the first loss function component, the second loss function component, and the third loss function component. The model may be configured to estimate the positioning of a user device based on channel state information for communication between the user device and a communication node.

[0016] The loss function may be generated by combining the first loss function component, the second loss function component, and the third loss function component.

[0017] The method may include receiving, retrieving or otherwise acquiring a first data set, wherein the first data set includes labeled data, the labeled data includes channel state information and associated position data, wherein the first position data is derived from the associated position data. The method may include retrieving or otherwise acquiring a second data set, wherein the second data set includes unlabeled data, the unlabeled data includes channel state information.

[0018] The method may include: generating a plurality of second loss function components by comparing position data with a plurality of second position estimates, wherein the plurality of second position estimates are based on channel state data of different instances that have undergone a first enhancement; and / or generating a plurality of third loss function components by comparing a third position estimate with a plurality of fourth position estimates, wherein the plurality of fourth position estimates are based on channel state data of different instances that have undergone a second enhancement.

[0019] In a third aspect, the specification describes an apparatus configured to perform any method described with reference to the second aspect.

[0020] In a fourth aspect, the present specification describes computer readable instructions which, when executed by a computing device, cause the computing device to perform any of the methods described with reference to the second aspect.

[0021] In a fifth aspect, the present specification describes a computer program comprising instructions for causing an apparatus to perform at least the following operations: generating a first loss function component comprising comparing first position data with a first position estimate, wherein the first position estimate is based on channel state data, wherein the first position estimate is generated using a model, and wherein the model comprises multiple trainable parameters; generating a second loss function component comprising comparing the first position data with a second position estimate, wherein the second position estimate is based on channel state data that has undergone a first enhancement, and wherein the second position estimate is generated using the model; generating a third loss function component comprising comparing a third position estimate based on the channel state data with a fourth position estimate based on the channel state data that has undergone a second enhancement, wherein the third position estimate and the fourth position estimate are generated using the model; and training the trainable parameters of the model by minimizing a loss function based on a combination of the first loss function component, the second loss function component, and the third loss function component.

[0022] In a sixth aspect, the present specification describes a computer-readable medium (such as a non-transitory computer-readable medium) comprising program instructions stored thereon for performing at least the following operations: generating a first loss function component comprising comparing first position data with a first position estimate, wherein the first position estimate is based on channel state data, wherein the first position estimate is generated using a model, and wherein the model comprises multiple trainable parameters; generating a second loss function component comprising comparing the first position data with a second position estimate, wherein the second position estimate is based on channel state data that has undergone a first enhancement, and wherein the second position estimate is generated using the model; generating a third loss function component comprising comparing a third position estimate based on the channel state data with a fourth position estimate based on the channel state data that has undergone a second enhancement, wherein the third position estimate and the fourth position estimate are generated using the model; and training the trainable parameters of the model by minimizing a loss function based on a combination of the first loss function component, the second loss function component, and the third loss function component.

[0023] In a seventh aspect, the present specification describes an apparatus comprising: at least one processor; and at least one memory comprising computer program code, which, when executed by the at least one processor, causes the apparatus to: generate a first loss function component, comprising comparing first position data with a first position estimate, wherein the first position estimate is based on channel state data, wherein the first position estimate is generated using a model, and wherein the model comprises a plurality of trainable parameters; generate a second loss function component, comprising comparing the first position data with a second position estimate, wherein the second position estimate is based on channel state data that has undergone a first enhancement, and wherein the second position estimate is generated using the model; generate a third loss function component, comprising comparing a third position estimate based on the channel state data with a fourth position estimate based on the channel state data that has undergone a second enhancement, wherein the third position estimate and the fourth position estimate are generated using the model; and train the trainable parameters of the model by minimizing a loss function based on a combination of the first loss function component, the second loss function component, and the third loss function component.

[0024] In an eighth aspect, the present specification describes an apparatus comprising: a first loss function module for generating a first loss function component by comparing first position data with a first position estimate, wherein the first position estimate is based on channel state data, wherein the first position estimate is generated using a model, and wherein the model comprises a plurality of trainable parameters; a second loss function module for generating a second loss function component by comparing the first position data with a second position estimate, wherein the second position estimate is based on channel state data that has undergone a first enhancement, and wherein the second position estimate is generated using a model; a third loss function module for generating a third loss function component by comparing a third position estimate based on the channel state data with a fourth position estimate based on the channel state data that has undergone a second enhancement, wherein the third position estimate and the fourth position estimate are generated using a model; and a training module for training the trainable parameters of the model by minimizing a loss function based on a combination (e.g., the sum of the components, such as a weighted sum) of the first loss function component, the second loss function component, and the third loss function component. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Example embodiments will now be described by way of non-limiting examples with reference to the following schematic drawings, in which:

[0026] Figures 1 to 4 is a block diagram of a system according to an example embodiment;

[0027] Figure 5 is a flow chart illustrating an algorithm according to an example embodiment;

[0028] Figure 6is a block diagram of a system according to an example embodiment;

[0029] Figure 7 is a flow chart illustrating an algorithm according to an example embodiment;

[0030] Figure 8 is a block diagram of a neural network that may be used in some example embodiments;

[0031] Figures 9 to 12 is a plot according to an example embodiment;

[0032] Fig.13 is a block diagram of components of a system according to an example embodiment; and

[0033] Fig.14A and Fig. 14B Tangible media are shown, respectively, a removable memory unit and a compact disk (CD) storing computer readable code that, when executed by a computer, performs operations according to the embodiments. DETAILED DESCRIPTION

[0034] The scope of protection sought by various embodiments of the present invention is set out by the independent claims. Embodiments and features described in the specification that do not fall within the scope of the independent claims, if any, are to be construed as examples for understanding various embodiments of the present invention.

[0035] Like reference numerals refer to like elements throughout the specification and drawings.

[0036] Figure 1 1 is a block diagram of a system (generally indicated by reference numeral 10) according to an example embodiment. The system 10 includes a base station 12, a first user device 14, and a second user device 15. For many reasons, it may be desirable to know the location (in absolute terms or relative to the base station 12) of the first user device 14 and / or the second user device 15. It should be noted that although the embodiments described herein generally relate to user devices and base stations, the principles may be applicable to any device that communicates with any suitable communication node.

[0037] Figure 2 is a block diagram of a system 20 for providing an estimate of the location of a device according to an example embodiment. System 20 includes an input for receiving channel state data (such as channel state information) and an output for providing location data. For example, system 20 may be provided at base station 12 (or some other communication node). Upon receiving channel state data related to communications between a device (such as user equipment 14) and a base station, system 20 attempts to provide an estimate of the location of the device.

[0038] Machine learning (ML) is a candidate technique for converting channel state information into device location estimates. However, many known ML algorithms rely on supervised learning and therefore require large amounts of labeled data for training purposes. Not only is the availability of suitable training data limited, but due to the changing nature of the relationship between channel state information (CSI) and user device location, any such model will need to be repeatedly retrained, requiring the generation of new training data.

[0039] Figure 3 is a block diagram of a system 30 according to an example embodiment. The system 30 receives both labeled data and unlabeled data for training a machine learning (ML) model. The ML model may be implemented using a neural network, as discussed further below.

[0040] As discussed in detail below, the labeled data includes channel state data (such as channel state information) and location data associated with the channel state data. The unlabeled data includes channel state data (such as channel state information) but does not include associated location data.

[0041] Figure 4 is a block diagram of a system, generally indicated by reference numeral 40, in accordance with an example embodiment. System 40 provides a high-level view of an algorithm in accordance with an example embodiment.

[0042] The system 40 includes a first module 41 that acquires raw channel measurements (e.g., channel state data) and associated user equipment location data. The data acquired by the first module 41 is organized into labeled data 42 including data pairs (x, y), the data pair (x, y) including channel state data x and location data y.

[0043] The system 40 also includes a second module 43 that acquires raw channel measurements (eg, channel state data) without acquiring associated user equipment location data. The data acquired by the second module 43 is organized into unlabeled data 44 including channel state data x.

[0044] The labeled data 42 and the unlabeled data 44 are inputs to the above-described system 30. Figure 4 As shown, the system 40 includes a first data enhancement module 46 , a second data enhancement module 47 and a training module 48 .

[0045] The first data enhancement module 46 generates first enhanced data, the first enhanced data comprising an enhanced version of the channel state data of the labeled data 42. Similarly, the second data enhancement module 47 generates second enhanced data, the second enhanced data comprising an enhanced version of the channel state data of the unlabeled data 44. The first data enhancement module and the second data enhancement module may implement the same function (in fact, in some embodiments, a single data enhancement module may be provided), and may be used to convert the channel state data x into enhanced data

[0046] The training module 48 generates an ML model (ie, an output of the system 30 ) based on the labeled data 42 , the unlabeled data 44 , the first augmented data, and the second augmented data.

[0047] For the purpose of user equipment positioning, the raw channel measurements received by the system 40 may be channel state information (CSI) data estimated at the relevant communication node (e.g., the base station 12 described above). These CSI data, which may be provided in the time domain or the frequency domain, may, for example, represent the channel between the (multiple) transmit antennas at the user equipment and the receive antenna array at the base station. For some CSI data, the actual positioning of the user equipment may be available; if so, the labeled data 42 may be provided to the system 40. For the remaining (unlabeled) data 44, the user equipment location is not available. In many example embodiments, only a small portion of the available training data is labeled data.

[0048] As described above, the system 40 generates augmented data for both labeled and unlabeled data. Data augmentation refers to a transformation (e.g., a random or pseudo-random transformation) of input data (e.g., CSI data) that does not, for example, change the relationship of the data to actual user equipment positioning. The use of data augmentation is intended to make the ML model more robust to generalization error.

[0049] Some examples of data augmentation techniques are listed below. The skilled person will be aware of alternative data augmentation techniques that may be used.

[0050] Phase rotation: Some (eg, all) elements of the initial channel state (eg, CSI) vector are rotated by the same phase. The phase may be selected (eg, randomly or pseudo-randomly) between [-π, π] radians.

[0051] Amplitude upscaling or downscaling: The amplitudes of some (e.g., all) elements of the initial channel state (e.g., CSI) vector are changed by the same amount. The amount may be selected (e.g., randomly or pseudo-randomly) between [-α, α], where α is a small positive number and represents a fraction of the average amplitude of all elements of the initial channel state vector.

[0052] Gaussian noise: A Gaussian noise vector is added to the initial channel state (e.g. CSI) vector. The power of this noise vector can be a small fraction of the average power of the initial channel state vector. Filtering / smoothing: If the channel state data is in the frequency domain, a low-pass filter can be used to smooth the amplitude across subcarriers.

[0053] Neural network based augmentation: This involves using a separate neural network that is trained to find suitable augmentations for the input channel state (e.g. CSI) vector.

[0054] • Element replacement: This involves replacing a small portion (eg a fraction) of the components of each channel state (eg CSI) vector with an arbitrary value from a known range of values.

[0055] Figure 5 is a flow chart illustrating an algorithm, generally indicated by reference numeral 50 , according to an example embodiment.

[0056] The algorithm 50 starts at operation 51, in which a data set is acquired. The data set may include a first data set and a second data set, the first data set including the labeled data 42 mentioned above, and the second data set including the unlabeled data 44 mentioned above.

[0057] In operation 52, enhanced data is generated. The enhanced data may include the following data: based on at least some channel state data in the first data set that has undergone the first enhancement and / or based on at least some channel state data in the second data set that has undergone the second enhancement.

[0058] In operation 53, a position estimate is generated. The position estimate may be generated by a model, such as model 30. The model may be used to generate a position estimate from channel state data and / or enhanced channel state data, as discussed in detail below.

[0059] The data acquired or generated in operations 51 to 53 may be used in operations 54 to 58 discussed below. However, it should be noted that the data may be received, retrieved or acquired in some other manner, such that in some example embodiments, one or more of operations 51 to 53 are optional.

[0060] In operation 54, a first loss function component is generated. The first loss function component may be generated by comparing the measured position data with a first position estimate based on the channel state data, wherein the first position estimate is generated using the model (e.g., in operation 53 discussed above). As discussed further below, the model may include a plurality of trainable parameters.

[0061] In operation 55, a second loss function component is generated. The second loss function component may be generated by comparing the measured position data with a second position estimate based on the first enhanced channel state data, where the second position estimate is generated using the model (eg, in operation 53 discussed above).

[0062] In operation 56, a third loss function component is generated. The third loss function component may be generated by comparing a third position estimate based on the channel state data with a fourth position estimate based on the modified channel state data, wherein the modified channel state data is the channel state data that has been second enhanced. The third position estimate and the fourth position estimate are generated using a model (e.g., in operation 53 discussed above).

[0063] In optional operation 57, a combined loss function may be generated by combining the first loss function component, the second loss function component, and the third loss function component. This step is optional because in some example embodiments, an explicit combination operation is not required (or the operation may be performed elsewhere, such as on a server).

[0064] Finally, in operation 58, trainable parameters of a model (e.g., model 30) may be trained by minimizing a loss function based on a combination of the first loss function component, the second loss function component, and the third loss function component (e.g., by minimizing the loss function generated in operation 57). The combination may be a sum (e.g., a weighted sum) of the first loss function component, the second loss function component, and the third loss function component.

[0065] Of course, although operations 54 through 56 are shown as three separate operations in a particular order, the order of these operations may be changed; furthermore, one or more of these operations may be combined into a single operation, or some or all of the operations may be performed in parallel.

[0066] Figure 6 is a block diagram of a system, generally indicated by reference numeral 60, according to an example embodiment. System 60 may be used to implement algorithm 50 described above.

[0067] The system 60 includes the labeled data 42, the unlabeled data 44, the first data enhancement module 46, and the second data enhancement module 47 of the above-mentioned system 40. The system 60 also includes a first function block 61, a second function block 62, a third function block 63, a fourth function block 64, a first MSE loss module 66, a second MSE loss module 67, a third MSE loss module 68, and a total loss module 69. The first to fourth function blocks 61 to 64 can be functionally identical and can, for example, implement the above-mentioned model 30. Similarly, as described above, the first data enhancement module 46 and the second data enhancement module 47 can be functionally identical.

[0068] As described above, the labeled data 42 includes a data pair (x, y) including the channel state data x and the position data y, and the unlabeled data 44 includes the channel state data x.

[0069] The first function block 61 receives the channel state data x and generates a position estimate based on the channel state data x. The second function block 62 also generates a position estimate based on the channel state data x Thus, the first function block and the second function block may differ only in the source of the channel state data (and may be implemented as a single function block).

[0070] The third function block 63 receives enhanced channel status data Based on the enhanced channel state data Generating a Position Estimate The fourth function block 64 is also based on the enhanced channel state data Generating a Position Estimate Therefore, the third function block and the fourth function block may differ only in the source of the enhanced channel state data (and may be implemented by a single function block, and in fact the first to fourth function blocks may be implemented by a single function block).

[0071] The first MSE loss module 66 generates a first loss function component, thereby implementing operation 54 of the above-described algorithm 50. The first MSE loss module 66 is based on the position data y (obtained from the labeled data 42) and the position estimate (generated by the first function block 61) to generate a first loss function component, where the position estimate is generated using a model (such as model 30 described above).

[0072] The second MSE loss module 67 generates a second loss function component, thereby implementing operation 55 of the above-described algorithm 50. The second MSE loss module 67 is based on the position data y (obtained from the labeled data 42) and the position estimate (generated by the third function block 63) to generate a second loss function component, where the position estimate is generated using a model (such as model 30).

[0073] The third MSE loss module 68 generates a third loss function component, thereby implementing operation 56 of the above algorithm 50. The third MSE loss module 68 generates a third loss function component based on the positioning estimate (generated by the second function block 62) and the position estimate (generated by the fourth function block 64) to generate the third loss function component.

[0074] The total loss module 69 generates a combined loss function based on the outputs of the first MSE loss module, the second MSE loss module, and the third MSE loss module 66 to 68, thereby implementing the operation 57 of the above-mentioned algorithm 50. Therefore, the output of the total loss module 69 can be used to train the parameters of the model (e.g., model 30), thereby implementing the operation 58 of the algorithm 50.

[0075] Figure 7 is a flow chart illustrating an algorithm, generally indicated by reference numeral 70 , according to an example embodiment.

[0076] Algorithm 70 starts at operation 71 and moves to operation 72 .

[0077] Operation 72 is a data preparation step in which the labeled data 42 and the unlabeled data 44 are generated. Therefore, operation 72 is an example of operation 51 of algorithm 50.

[0078] Expressed mathematically, represents the labeled dataset 42. Therefore, Each element of is a pair (x, y), where x represents the channel state data (CSI) and y represents its associated location (e.g., in Euclidean space, relative to some reference point as the origin). In addition, let represents an unlabeled dataset44, which means that the actual positioning is In addition, let f θ represents a machine learning model with trainable parameters {θ} (e.g., modules 61 to 64 described above), and g represents a data augmentation function (e.g., modules 46 and 47 described above). As described above, the data augmentation function can generate random transformations of the input without changing its relationship to the actual UE positioning.

[0079] In operation 73 , a training period begins. Algorithm 70 then moves to operation 74 .

[0080] In operation 74, the input feature x is obtained. a . Input feature xa Includes channel state data, which is part of the labeled data or unlabeled data. Input feature x a is passed to operations 75 and 78 described further below.

[0081] In operation 75, x a Apply g enhancement to get g(x a ). Depending on whether the input feature is part of the labeled data or part of the unlabeled data, the enhancement may be applied to the data enhancement module 46 or the data enhancement module 47. Of course, in an example implementation, the data enhancement modules 46 and 47 may be implemented by the same module, so the system 60 may be functional rather than providing a circuit schematic.

[0082] From operation 75, the enhanced data g(x a ) is provided to operation 76, in which the augmented data is applied to the model to generate an output f θ (g(x a ). This output is provided to both operation 77 and operation 80 described below.

[0083] In operation 78, the input x is determined a Is the (received from operation 74 in operation 78) from a labeled data set. If so, algorithm 70 moves to operations 80 and 81 described further below. If not (indicating that input x a from unlabeled data), then algorithm 70 moves to operation 79.

[0084] In operation 79, the data x a Applied to the model to generate the output f θ (x a ). The output is provided to operation 77.

[0085] In operation 77, the data x has been determined a is unlabeled data. In operation 77, two inputs are received: function f θ (g(x a ) (received from operation 76, such as from the fourth function block 64 of the system 60 described above) and function f θ (x a )(received from function 79, for example, received from the second function block 62 of the above-mentioned system 60).

[0086] Operation 77 calculates the square error || f θ (xa)-f θ (g(x a))||. Operation 77 is an example of operation 56 of the algorithm 50 described above, and may be implemented by the third MSE loss module 68 of the system 60 described above. The output of operation 77 is provided to operation 83, which is described further below.

[0087] Operation 80 (in data x a In the case of a marked set, starting from operation 78 above) based on the channel state data x a and its associated positioning y a Generate output. Operation 80 calculates the squared error || y a -f θ (g(x a ))||. Operation 80 is an example of operation 55 of the algorithm 50 described above, and may be implemented by the second MSE loss module 67 of the system 60 described above. The output of operation 80 is provided to operation 83.

[0088] In operation 81, the data x a Applied to the model to generate the output f θ (x a ). This output is provided to operation 82. Operation 82 calculates the squared error || y a -f θ (x a )|| 2 Operation 82 is an example of operation 54 of the algorithm 50 described above, and may be implemented by the first MSE loss module 66 of the system 60 described above. The output of operation 81 is provided to operation 83.

[0089] Thus, operation 83 receives inputs from operations 77, 80, and 82 described above. Operation 83 computes a loss function based on the squared error outputs of operations 77, 80, and 82, and is thus an example of operation 57 of algorithm 50 described above, and may be implemented by total loss module 69 of the system described above.

[0090] The loss function generated in operation 83 can be given by:

[0091]

[0092] The loss function generated in operation 83 is used to update the parameters θ of the model 30 .

[0093] From operation 83, the algorithm moves to operation 84, where it is determined whether a stopping criterion has been reached (e.g., whether training of model 30 is complete). The stopping criterion may take a variety of forms: whether operations 73 to 83 have been completed for such number of iterations, whether the loss has reached or fallen below a certain value, whether the loss has not decreased during a fixed number of iterations, etc. A skilled person will be aware of other options for implementing operation 84.

[0094] If the stopping criteria have been reached, the algorithm terminates at operation 85. Otherwise, the algorithm returns to operation 73, where another model training session is implemented. (In some example embodiments, multiple samples are considered in one session, and the loss is averaged over these samples.)

[0095] In algorithm 70, the first MSE term is related to the labeled data. This term enables the model to learn the hidden relationship between the input channel state data and the actual device location in a manner that simulates supervised learning with a large labeled data set.

[0096] The second MSE term is based on the predicted location based on the augmented data and the associated device location. The third MSE is based on the predicted location of the unlabeled data and its augmented counterpart. The use of augmented data attempts to make the model more robust to generalization error. It is important to note that at each training epoch, the augmented data seen by the model is different from the data seen in earlier epochs due to the randomness of the augmentation. Due to the nature of the proposed training procedure, for the same amount of labeled training data, it can be expected that the learned model will be more accurate in predicting unseen data than traditional ML techniques.

[0097] In an alternative embodiment, instead of enhancing once per data sample per period (as described above with reference to Figure 7 As described above, each data sample in each period may have multiple enhancements. For example, in each period, for each sample from the labeled set, there may be k enhancements, and for each sample from the unlabeled data set, there may be l enhancements, where k and l are positive integers. The values ​​of k and l can be determined experimentally.

[0098] Figure 8 is a block diagram of a neural network, generally indicated by reference numeral 90, that may be used in some example embodiments. For example, model 30 may be a machine learning model that may be implemented using a neural network such as neural network 90.

[0099] Neural network 90 includes an input layer 91, one or more hidden layers 92, and an output layer 93. During use of neural network 90, at input layer 91, inputs such as channel state data may be received. Hidden layer 92 may include a plurality of hidden nodes, where processing may be performed based on the received input (e.g., channel state data). At output layer 93, one or more outputs (e.g., position data) may be provided that are related to the input.

[0100] The neural network 90 may be trained using inputs including the labeled data 42 and the unlabeled data 44 described above. Training may be performed at the base station 12 or elsewhere, such as on a server. The neural network 90 may be trained offline (e.g., pre-trained before starting to use the model) and / or may be trained online (e.g., training may continue and the neural network 90 may be updated based on new data).

[0101] Fig. 9 1 is a drawing (generally indicated by reference numeral 100) according to an example embodiment. Drawing 100 shows 24,576 user equipments (UEs) arranged in six clusters (generally indicated by reference numerals 101 to 106, respectively). A base station communicating with the 24,576 UEs is located at the origin of the drawing and is equipped with 16 receive antennas. Each UE has a single transmit antenna. The simulation results discussed below consider a line of sight (LoS) model of the system.

[0102] The training data (with labels) consisted of 2213 samples (less than 10% of all samples). We used a deep neural network with two hidden layers. The first hidden layer had 256 neurons, while the second layer had 32 neurons. The input was a CSI vector of 32 real dimensions (since in this case, the CSI vector was of 16 complex dimensions due to the 16 receive antennas at the base station). We used the “relu” activation function throughout, except for the “linear” activation at the output layer. We trained the model for 1000 epochs using the mean squared error loss function.

[0103] Fig.10 1 is a plot (generally indicated by reference numeral 110) according to an example embodiment. Plot 110 shows the initial clusters 101 to 106 (on the left), and shows the reconstructed positions of all 24,576 UEs (on the right). Due to the small training sample size, the reconstruction error (the average of the absolute Euclidean distance between the actual position and the estimated position) is large (14.26 m in this particular example).

[0104] Next, we trained the same model but using data augmentation as described above. For data augmentation purposes, we used random phase rotations. In each epoch and for each mini-batch (consisting of labeled and unlabeled samples), an arbitrary phase was chosen and all vector elements of all samples in the batch were rotated by that phase angle. As before, we considered the size of the labeled data to be 2213 and training was performed for 1000 epochs.

[0105] Fig.111 is a plot (generally indicated by reference numeral 120) according to an example embodiment. Plot 120 shows the initial clusters 101 to 106 (on the left), and shows the reconstructed positioning of all 24576 UEs (on the right). The reconstruction error achieved is 4.86m (i.e., much better than the reconstruction error in plot 110).

[0106] Next, we train the model on only the 222 labeled data samples. Fig.12 is a plot (generally indicated by reference numeral 130) according to an example embodiment. Again, the plot 130 shows the initial cluster (on the left) and the estimated locations of all samples (on the right). Despite the small amount of labeled data, the reconstruction error achieved is 6.67m.

[0107] In the above example embodiments, enhancement data is generated based on channel state information. In some example embodiments, multiple enhancement data may be generated for at least some of the channel state information. For example, this may increase the amount of enhancement data that can be used to train a machine learning model.

[0108] For completeness, Fig.13 is a schematic diagram of components of one or more of the example embodiments described previously, which is generally referred to below as a processing system 300. The processing system 300 may be, for example, an apparatus as referred to in the following claims.

[0109] The processing system 300 may include one or more of the following: a processor 302, a memory 304 tightly coupled to the processor and consisting of a RAM 314 and a ROM 312, a user input 310 (such as a touch screen input, a hardware key and / or a voice input mechanism), and a display 318 (in some example embodiments, at least some of these components may be omitted). The processing system 300 may include one or more network / device interfaces 308 for connecting to a network / device, which may be a wired or wireless modem, for example. The interface 308 may also operate as a connection with other devices (such as devices / devices that are not network-side devices). Therefore, direct connections between devices / devices without the involvement of a network are possible.

[0110] The processor 302 is connected to each of the other components in order to control the operation of that component.

[0111] The memory 304 may include a non-volatile memory, such as a hard disk drive (HDD) or a solid state drive (SSD). Among other things, the ROM 312 of the memory 304 also stores an operating system 315, and may store a software application 316. The RAM 314 of the memory 304 is used by the processor 302 for temporary storage of data. The operating system 315 may include code that implements various aspects of the above-mentioned algorithms 50 and 70 when executed by the processor. Note that in the case of a small device / apparatus, the memory may be best suited for small size use, i.e., a hard disk drive (HDD) or a solid state drive (SSD) is not always used. The memory 304 may include a computer program code so that the at least one memory 304 and the computer program may be configured to achieve the performance of the device together with the at least one processor 302.

[0112] Processor 302 may take any suitable form. For example, it may be a microcontroller, multiple microcontrollers, a processor, or multiple processors.

[0113] The processing system 300 may be a standalone computer, server, console or network thereof. The processing system 300 and the required structural components may all be inside a device / apparatus such as an IoT device / apparatus, i.e. embedded in a very small size.

[0114] In some example embodiments, the processing system 300 may also be associated with external software applications. These may be applications stored on a remote server device / apparatus and may run partially or exclusively on the remote server device / apparatus. These applications may be referred to as cloud-hosted applications. The processing system 300 may communicate with the remote server device / apparatus in order to utilize the software applications stored therein.

[0115] Fig.14A and Fig. 14B Tangible media are shown, respectively a removable memory unit 365 and a compact disk (CD) 368, storing computer readable code, which when run by a computer can perform a method according to the above-described example embodiments. The removable memory unit 365 can be a memory stick, such as a USB memory stick, having an internal memory 366 storing computer readable code. The internal memory 366 can be accessed by the computer system via a connector 367. The CD 368 can be a CD-ROM or DVD or the like. Other forms of tangible storage media can be used. The tangible medium can be any device / device capable of storing data / information that can be exchanged between devices / devices / networks.

[0116] Embodiments of the present invention may be implemented in software, hardware, application logic, or a combination of software, hardware, and application logic. The software, application logic, and / or hardware may reside in a memory or any computer medium. In an exemplary embodiment, the application logic, software, or instruction set is maintained on any of a variety of conventional computer-readable media. In the context of this document, a "memory" or "computer-readable medium" may be any non-transient medium or component that may contain, store, transmit, propagate, or transfer instructions for use by or in conjunction with an instruction execution system, device, or device, such as a computer.

[0117] Where relevant, references to "computer-readable storage medium", "computer program product", "tangibly embodied computer program", etc. or "processor", "processing circuitry", etc. should be understood to cover not only computers with different architectures, such as single processor / multi-processor architectures and serial / parallel architectures, but also special-purpose circuits, such as field programmable gate arrays FPGAs, special-purpose circuits ASICs, signal processing devices / apparatuses and other devices / apparatuses. References to computer programs, instructions, codes, etc. should be understood to mean software for programmable processor firmware, such as programmable content of hardware devices / apparatuses, as instructions for a processor or configured for fixed-function devices / apparatuses, gate arrays, programmable logic devices / apparatuses, etc. or configuration settings therefor.

[0118] If desired, the different functions discussed herein may be executed in a different order and / or in parallel with each other. In addition, if desired, one or more of the above functions may be optional or may be combined. Similarly, it is also understood that Figure 5 and Figure 7 The flowcharts in are merely examples, and various operations depicted herein may be omitted, re-ordered, and / or combined.

[0119] It is to be understood that the above exemplary embodiments are illustrative only and do not limit the scope of the invention.Other variations and modifications will be apparent to those skilled in the art after reading this specification.

[0120] Furthermore, it should be understood that the disclosure of the present application includes any novel feature or any novel combination of features disclosed herein explicitly or implicitly, or any generalization thereof, and that new claims may be issued during the prosecution of the present application or any application derived therefrom to cover any such feature and / or combination of such features.

[0121] Although various aspects of the invention are set out in the independent claims, further aspects of the invention comprise other combinations of features from the described example embodiments and / or the dependent claims with features of the independent claims, including not only the combinations explicitly set out in the claims.

[0122] It is also noted herein that, although various examples are described above, these descriptions should not be viewed in a limiting sense. Rather, numerous variations and modifications may be made without departing from the scope of the present invention as defined in the appended claims.

Claims

1. An apparatus for training in a communication system, comprising: means for generating a first loss function component, the generating the first loss function component comprising comparing first position data to a first position estimate, wherein the first position estimate is based on channel state data of a first data set, the first data set comprising labeled data, the labeled data comprising channel state information and associated position data, wherein the first position data is derived from the associated position data, wherein the first position estimate is generated using a model, and wherein the model comprises a plurality of trainable parameters; means for generating a second loss function component, the generating the second loss function component comprising comparing the first position data to a second position estimate, wherein the second position estimate is based on channel state data of the first data set that has been subjected to a first enhancement, and wherein the second position estimate is generated using the model; means for generating a third loss function component, the generating the third loss function component comprising comparing a third position estimate based on channel state data of a second data set and a fourth position estimate based on channel state data of the second data set that has been subjected to a second enhancement, the second data set comprising unlabeled data, the unlabeled data comprising channel state information, wherein the third position estimate and the fourth position estimate are generated using the model; as well as means for training the trainable parameters of the model by minimizing a loss function based on a combination of the first loss function component, the second loss function component, and the third loss function component, The model is configured to estimate the location of the user equipment based on channel state information used for communication between the user equipment and a communication node.

2. The device according to claim 1, wherein the device further comprises: means for generating the loss function by combining the first loss function component, the second loss function component and the third loss function component.

3. The apparatus of claim 2, wherein the combination comprises the sum of the first loss function component, the second loss function component, and the third loss function component. The apparatus of claim 3 , wherein the sum is a weighted sum.

5. The device according to claim 1, wherein the device further comprises: Means for acquiring said first data set by receiving or retrieving.

6. The device according to claim 5, wherein the device further comprises: means for generating the first position estimate based on at least some of the channel state data of the first data set; as well as Means for generating the second position estimate based on at least some of the channel state data in the first data set that has been subjected to the first enhancement.

7. The device according to claim 1, wherein the device further comprises: Means for acquiring, by receiving or retrieving, said second data set.

8. The device according to claim 7, wherein the device further comprises: means for generating said third position estimate based on at least some of said channel state data of said second data set; as well as Means for generating the fourth position estimate based on at least some of the channel state data in the second data set that has been subjected to the second enhancement.

9. The apparatus of claim 1, wherein each enhancement comprises one or more of: Phase rotation; Amplification The amplitude is reduced; Addition of Gaussian noise; filtering or smoothing; Neural network-based augmentation; or Element replacement algorithm.

10. The apparatus of claim 1, wherein each enhancement provides a random or pseudo-random transformation of the associated channel state information.

11. The device according to claim 1, wherein the device further comprises: means for generating a plurality of said second loss function components by comparing said first position data with a plurality of second position estimates, said plurality of second position estimates being based on channel state data of different instances that have been subjected to said first enhancement; and / or means for generating a plurality of said third loss function components by comparing said third position estimate with a plurality of fourth position estimates, said plurality of fourth position estimates being based on channel state data that have been subjected to different instances of said second enhancement.

12. The apparatus of claim 1, wherein the model is a machine learning model.

13. The apparatus of claim 1, wherein the model is implemented using a neural network.

14. A method for training in a communication system, comprising: generating a first loss function component, comprising comparing first position data to a first position estimate, wherein the first position estimate is based on channel state data of a first data set, the first data set comprising labeled data, the labeled data comprising channel state information and associated position data, wherein the first position data is derived from the associated position data, wherein the first position estimate is generated using a model, and wherein the model comprises a plurality of trainable parameters; generating a second loss function component, comprising comparing the first position data to a second position estimate, wherein the second position estimate is based on channel state data of the first data set that has been subjected to a first enhancement, and wherein the second position estimate is generated using the model; Generating a third loss function component includes comparing a third position estimate based on channel state data of a second data set and a fourth position estimate based on channel state data of the second data set that has been subjected to a second enhancement, the second data set including unlabeled data, the unlabeled data including channel state information, wherein the third position estimate and the fourth position estimate are generated using the model; as well as training the trainable parameters of the model by minimizing a loss function based on a combination of the first loss function component, the second loss function component, and the third loss function component, The model is configured to estimate the location of the user equipment based on channel state information used for communication between the user equipment and a communication node.