Techniques used to assess connection quality

Abstract

A technique for evaluating connection quality in a wireless communication network is disclosed. A method implementation of the technique includes obtaining (S202) a radio frequency (RF) channel response measurement indicating a channel gain over time and frequency observed at a location covered by the wireless communication network, and determining (S204) an estimated connection quality at the location based on the RF channel response measurement using a machine learning model trained to map features based on the RF channel response measurement to corresponding connection quality.

Description

Technical Field

[0001] The present disclosure relates generally to wireless communication networks. In particular, techniques for evaluating connection quality in wireless communication networks are presented. The techniques may be embodied in methods, computer programs, devices, and systems. Background Art

[0002] For example, in industry, wireless network connectivity is increasingly becoming an important element of factory infrastructure and can be used to connect critical machines to controllers and wireless sensors to central control entities, or can be used to connect mobile robots such as automated guided vehicles (AGVs), transporting materials and products throughout the factory. For example, such use cases often require very low latency (e.g., <1ms) and very low packet error rates (e.g., <10ms). -6 ) and therefore adopt services such as Ultra-Reliable Low Latency Communication (URLLC) in 5G networks. The network coverage or quality of these types of services cannot be quantified by traditional signal strength measurements. Instead, the network coverage or quality needs to be characterized based on the latency and loss metrics of the URLLC connection at different locations. With such stringent quality requirements, careful network planning and connection assessment become increasingly important.

[0003] Conventional methods for network planning and radio surveys are mainly targeted at mobile broadband services, where reliability guarantees and target packet error rates are more relaxed. These services are usually provided on a best-effort basis, where applications (such as web browsing, email or video) are less sensitive to channel variations and can easily adapt to varying connection qualities. Network planning and site surveys for such services target average signal strengths that can provide sufficiently good average connectivity. Planning usually takes into account shadow fading and signal attenuation with the aid of simple models or partly based on floor plans, for example assuming some fixed value attenuation through walls, glass, etc. For example, a model can give an average signal-to-noise ratio (SNR) that can be mapped to an average block error rate (BLER), where the SNR-BLER mapping model depends on the fast fading properties of the channel, which in turn can be modeled using some general statistical models such as the Rayleigh or Price fading models. In order to obtain the SNR-BLER mapping, offline simulations are used, and simulations usually require lengthy simulation runs. For example, rather than using a general statistical model, a statistical fast fading channel model can also be obtained by assuming a certain delay spread distribution of the channel (i.e., how many and how strong multiple paths are typically received), where the distribution can be obtained by taking an inverse Fourier transform of the frequency channel impulse response, which in turn can be measured from a reference signal used in an orthogonal frequency division multiplexing (OFDM) system. If a large number of such channel response measurements can be collected, it is possible to create a statistically relevant representation in the form of a fast fading model, which can then be used in offline simulations.

[0004] These conventional approaches typically suffer from several disadvantages. One disadvantage is that only an estimate of average quality occurs, which is insufficient for industrial applications that require high reliability and low latency as described above, for which planning for the worst case scenario is instead required. In other words, new generation wireless technologies (e.g., 5G and WiFi6) that support more demanding industrial use cases including URLLC have more stringent quality requirements, and average signal level measurements taken at a few selected locations are not sufficient to derive a satisfactory quality metric.

[0005] Another disadvantage is the need to either perform lengthy measurements (to obtain the statistical channel impulse response) and lengthy simulations (to obtain the SNR-BLER mapping) or to rely on some statistical channel models that do not capture the environment-specific properties of the plant and thus only lead to rough and inaccurate estimates. In other words, conventional simulation-based estimates are not suitable because they assume some generic statistical channel models that may not be very suitable for the specific plant environment (thus erroneously estimating the connection quality) or because they require the performance of lengthy measurement and simulation runs.

[0006] For delay-critical services, reasonable measurements may not be feasible at all. For delay-critical services, radio channel variability on a millisecond time scale can often impact application quality of service (QoS), necessitating highly detailed channel measurements for RF quality assessment (e.g., in orthogonal frequency division multiple access (OFDMA) systems, millisecond and subcarrier level pilot signal measurements are required). Although direct measurements of relevant key performance indicators (KPIs) like sudden and large signal drops are often possible, this is in contrast to average signal strength (e.g., reference signal received power (RSRP) type) measurements, e.g., due to the necessity of maintaining 10 -5 to 10 -6 These events are extremely rare (or "localized") in time and space, in order to reliably measure these loss metrics, it would be necessary to inject real traffic into the network to be able to actually measure the loss and delay of the data being transmitted. In practice, this means that running statistically relevant and exhaustive measurements at the site may not be feasible for these demanding types of applications. Summary of the invention

[0007] Therefore, there is a need for techniques for assessing connection quality in a wireless communication network that avoids one or more of these problems and that can be used to ensure adequate connection quality in a served area.

[0008] According to a first aspect, a method for evaluating connection quality in a wireless communication network is provided. The method comprises obtaining RF channel response measurements indicating channel gain over time and frequency observed at a location covered by the wireless communication network, and determining an estimated connection quality at the location based on the RF channel response measurements using a machine learning model trained to map features based on the RF channel response measurements to corresponding connection qualities.

[0009] Channel gain can be measured in time and frequency based on reference symbols transmitted from a base station of the wireless communication network. Features based on the RF channel response measurements can represent channel gain in time and frequency, wherein determining the estimated connection quality based on the RF channel response measurements can include inputting the measured channel gain in time and frequency into a machine learning model. Moreover, a multipath resolution algorithm can be applied to the RF channel response measurements to determine an observed multipath RF propagation pattern, wherein features based on the RF channel response measurements can represent the multipath RF propagation pattern, and wherein determining the estimated connection quality based on the RF channel response measurements can include inputting the determined multipath RF propagation pattern into the machine learning model. The multipath RF propagation pattern can include the number and strength of multiple RF propagation paths observed at the location, and optionally the directions of different RF propagation paths among the multiple RF propagation paths.

[0010] RF channel response measurements can be obtained from URLLC services transmitted in the wireless communication network. A machine learning model can be trained using training data including multiple sets of features based on RF channel response measurements as input data and corresponding connection quality measured at different locations covered by a representative wireless communication network as output data, and multiple sets of features based on RF channel response measurements are obtained from RF channel response measurements indicating channel gain over time and frequency performed at different locations covered by the representative wireless communication network. The wireless communication network can cover an industrial environment, where additional training data can be continuously obtained from one or more wireless devices performing industrial tasks in the industrial environment, and where the additional training data can be used to enhance the machine learning model.

[0011] The machine learning model may be one of a plurality of machine learning models trained to map features based on RF channel response measurements to corresponding connection qualities, wherein each of the plurality of machine learning models may be trained using training data obtained from measurements performed at locations covered by different representative wireless communication networks, each representing a different type of RF environment. A machine learning model may be selected from the plurality of machine learning models based on the RF channel response measurements using a classifier trained to map features based on RF channel response measurements to corresponding types of RF environments. Moreover, the plurality of machine learning models may be used as a set in which each of the plurality of machine learning models may be used to estimate corresponding connection qualities based on RF channel response measurements, and wherein the estimated connection quality may be determined based on the corresponding connection quality.

[0012] Obtaining RF channel response measurements and determining estimated connection qualities may be performed for a plurality of locations covered by a wireless communication network, wherein the resulting estimated connection qualities may be used to generate a connection quality map for an area covered by the wireless communication network. The wireless communication network may cover an industrial environment, wherein measured connection qualities may be continuously obtained from one or more wireless devices performing industrial tasks in the industrial environment, and wherein the connection quality map may be updated based on the measured connection qualities. The connection quality map may be segmented into different areas depending on the level of connection quality that may be provided in the respective areas.

[0013] When the estimated connection quality in an area covered by a wireless communication network is lower than a desired connection quality value, one or more mitigation actions may be triggered to improve the connection quality in the area. The one or more mitigation actions may include at least one of the following: deploying one or more additional base stations in the area, allocating one or more additional base stations to serve cells in the coverage area, adapting the RF reflection properties of at least one metasurface provided in the area, and determining the radio link configuration required to achieve the connection quality value in the area and applying the determined radio link configuration in the wireless communication network. Determining the radio link configuration may include adding a safety margin to a channel quality value reported from a wireless device in the wireless communication network and selecting a modulation coding scheme (MCS) depending on the channel quality value with the safety margin, the safety margin addressing a possible drop in channel gain.

[0014] According to a second aspect, there is provided a method for training a machine learning model configured to evaluate connection quality in a wireless communication network. The method comprises training the machine learning model using training data comprising as input data a plurality of sets of features based on RF channel response measurements and as output data corresponding connection qualities measured at different locations covered by a representative wireless communication network, the plurality of sets of features based on RF channel response measurements being obtained from RF channel response measurements indicating channel gain over time and frequency performed at different locations covered by the representative wireless communication network.

[0015] The method according to the second aspect defines a method dedicated to training a machine learning model which, once trained, can be employed in the method according to the first aspect for the purpose of assessing the quality of a connection in a wireless communication network. The machine learning model of the second aspect may therefore correspond to the machine learning model described above in relation to the first aspect.

[0016] According to a third aspect, a computer program product is provided. The computer program product includes a program code portion for performing at least one of the methods of the first aspect and the second aspect when the computer program product is executed on one or more computing devices (e.g., a processor or a distributed group of processors). The computer program product can be stored on a computer-readable recording medium such as a semiconductor memory, a DVD, a CD-ROM, etc.

[0017] According to a fourth aspect, there is provided a computing unit for evaluating a connection quality in a wireless communication network. The computing unit comprises at least one processor and at least one memory, the at least one memory containing instructions executable by the at least one processor, such that the computing unit is operable to perform any of the method steps presented herein with respect to the first aspect.

[0018] According to a fifth aspect, there is provided a computing unit for training a machine learning model configured to assess connection quality in a wireless communication network. The computing unit comprises at least one processor and at least one memory, the at least one memory containing instructions executable by the at least one processor, such that the computing unit is operable to perform any of the method steps presented herein with respect to the second aspect.

[0019] According to a sixth aspect, there is provided a computing unit according to the fourth aspect and, optionally, a computing unit according to the fifth aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Implementations of the techniques presented herein are described herein below with reference to the accompanying drawings, in which:

[0021] Figure 1a and Figure 1b Exemplary compositions of a computing unit for evaluating connection quality and a computing unit for training a machine learning model according to the present disclosure are described;

[0022] Figure 2 A method performed by a computing unit according to the present disclosure that may be used to assess connection quality is described;

[0023] Figure 3 An exemplary resource block comprising reference symbols distributed in a time-frequency grid of the resource block is illustrated;

[0024] Figure 4 An example of a measured RF channel response in the frequency domain is illustrated;

[0025] Figure 5 An example of a measured RF channel response in the time domain is illustrated;

[0026] Figure 6 A two-dimensional (2D) graph illustrating channel gain in both time and frequency;

[0027] Figure 7 An exemplary environment of an area with different multipath propagation patterns is illustrated;

[0028] Figure 8 A metasurface for controlling the direction of reflected signals is described;

[0029] Fig. 9 Exemplary SNR versus BLER curves for different MCSs are illustrated;

[0030] Fig.10 Methods performed by a computing unit that may be used to train a machine learning model according to the present disclosure are described; and

[0031] Fig.11 A flow chart providing an overview of relevant steps that may be performed by the techniques presented herein is illustrated. DETAILED DESCRIPTION

[0032] In the following description, for purposes of explanation rather than limitation, specific details are set forth in order to provide a thorough understanding of the present disclosure. It will be apparent to one skilled in the art that the present disclosure may be implemented in other embodiments that depart from these specific details.

[0033] Those skilled in the art will further appreciate that the steps, services and functions explained herein below may be implemented using separate hardware circuits, using software running with a programmed microprocessor or general purpose computer, using one or more application specific integrated circuits (ASICs), and / or using one or more digital signal processors (DSPs). It will also be appreciated that when the present disclosure is described in terms of methods, it may also be embodied in one or more processors and one or more memories coupled to the one or more processors, wherein the one or more memories are encoded with one or more programs that, when executed by the one or more processors, perform the steps, services and functions disclosed herein.

[0034] Figure 1a Schematically illustrated is an exemplary composition of a computing unit 100 for evaluating a connection quality in a wireless communication network. The computing unit 100 comprises at least one processor 102 and at least one memory 104, wherein the at least one memory 104 contains instructions executable by the at least one processor 102, such that the computing unit 100 is operable to perform the method steps described herein below with reference to the computing unit for evaluating a connection quality.

[0035] Figure 1b An exemplary composition of a computing unit 110 for training a machine learning model configured to assess connection quality in a wireless communication network is schematically illustrated. The computing unit 110 comprises at least one processor 112 and at least one memory 114, wherein the at least one memory 114 contains instructions executable by the at least one processor 112, such that the computing unit 110 is operable to perform the method steps described herein below with reference to the computing unit for training a machine learning model.

[0036] For example, it will be appreciated that each of computing units 100 and 110 may be implemented on a physical computing unit or a virtualized computing unit such as a virtual machine. For example, it will be further appreciated that each of computing units 100 and 110 may not necessarily be implemented on a separate computing unit, but rather each of computing units 100 and 110 may be implemented as a component (implemented in software and / or hardware) that also resides on multiple distributed computing units (such as in a cloud computing environment).

[0037] Figure 2 A method according to the present disclosure that can be performed by the computing unit 100 is described. The method is dedicated to evaluating the connection quality in a wireless communication network. In step S202, the computing unit 100 may obtain an RF channel response measurement indicating the channel gain over time and frequency observed at a location covered by the wireless communication network. In step S204, the computing unit 100 may determine the estimated connection quality at the location based on the RF channel response measurement using a machine learning model trained to map features based on the RF channel response measurement to corresponding connection quality.

[0038] Connection quality assessment can therefore be performed by predicting connection quality from RF channel response characteristics using a machine learning model specifically trained to perform such mapping. Rather than the conventional average signal strength based approach described above, the techniques presented herein can therefore use RF channel response measurements with high accuracy in time and frequency to derive estimated channel quality at the observation location.

[0039] As will be appreciated by those skilled in the art, the RF channel response (or "frequency domain channel response") may correspond to a measured (complex-valued) channel gain in time and frequency, which may essentially represent a measure of the channel distortion observed with a fine granularity in both time and frequency of the received signal. For example, fine granularity (or "high precision") in time and frequency may be achieved by measuring reference symbols (e.g., reference symbols transmitted in designated time-frequency resource blocks such as physical resource blocks (PRBs)) transmitted by a base station (or "access point" (AP)) of a wireless communication network to the wireless device being measured. Channel gain may thus be measured in time and frequency based on reference symbols transmitted from a base station of the wireless communication network. Figure 3An exemplary time-frequency resource block 300 including reference symbols 302 distributed in a time-frequency grid of the resource block 300 is shown in FIG. The RF channel response may be measured at each of the reference symbols 302 in the resource block 300. More specifically, the reference symbols 302 may be transmitted at regular short time intervals (e.g., 1 ms) and at predefined frequency subcarriers in the time-frequency grid, and since the transmitted signal is typically known, the channel distortion may be determined by comparing the received signal with the transmitted signal to obtain a measure of the channel gain. The channel distortion may correspond to the difference between the received signal and the transmitted signal in both the amplitude and phase of the signals, thus yielding a complex value of the RF channel response.

[0040] Figure 4 An example of a measured channel response in the frequency domain is shown, illustrating the channel gain at one instant, where the X-axis indicates frequency and the Y-axis indicates channel gain. As can be seen, for example, there are frequencies (i.e., frequency subcarriers) where the channel is significantly weaker, which may be caused by severe signal cancellation due to multipath signal propagation patterns. On the other hand, Figure 5 Channel gain over time for several selected frequency bins (ie subcarriers) is shown, where the X-axis indicates time and the Y-axis indicates channel gain, and different grey shaded curves represent different subcarriers, ie several selected frequency bins. Figure 6 Another possible visualization of the RF channel response is shown in , which illustrates the channel gain over both time and frequency in a 2D plot, where different color values ​​indicate different values ​​of the channel gain.

[0041] As described above, based on such RF channel response measurements observed at the location, a machine learning model that is specially trained to map features based on RF channel response measurements to corresponding connection qualities can be used to determine the estimated connection quality. Features based on channel response measurements can be any features corresponding to or derived from RF channel response measurements. In the simplest case, features based on RF channel response measurements can directly correspond to channel gain over time and frequency of RF channel response measurements. In other cases, features based on RF channel response measurements can be features derived from RF channel response measurements, such as by applying a multipath resolution algorithm to the measured channel response data. As will be appreciated by those skilled in the art, for example, the signal strength dynamics in advanced wireless systems such as OFDMA systems can be highly determined by multipath propagation patterns, i.e., by the number, strength, and direction of individual signal paths (so-called "taps"). These patterns (for example, together with observed noise and interference) can determine KPIs that are important for key services. Patterns can depend on local environments such as reflective objects, walls, and other obstacles, so that points that are close in physical space may generally be susceptible to the same or similar multipath propagation characteristics.

[0042] Again, in the simplest case, the features based on the RF channel response measurements may represent channel gain in time and frequency. In this case, the step of determining the estimated connection quality based on the RF channel response measurements may include inputting the measured channel gain in time and frequency into a machine learning model. The machine learning model may therefore map the measured channel gain in time and frequency (provided as input to the model) to the corresponding estimated connection quality (provided as a corresponding output from the model). The machine learning model may be any type of machine learning model suitable for being trained for such mapping, including a neural network, a support vector machine (SVM), and the like. For example, the neural network may be given, in particular, by a convolutional neural network (CNN), a simple deep neural network (DNN), or a recurrent neural network (RNN) like a long short-term memory (LSTM).

[0043] The connection quality determined by the machine learning model based on the input may correspond to a KPI related to a (wireless) connection established with the wireless device being measured, such as a connection of the wireless device to a base station of a wireless communication network or an application-level connection of a service provided to the wireless device being measured by the wireless communication network. For example, the KPI itself may be related to a specific application such as a specific industrial application. Exemplary KPIs may include at least one of BLER, packet delay, packet delay budget, probability of a packet delay exceeding a predetermined threshold, probability of packet discard, and level of connectivity (e.g., from the perspective of the application). The level of connectivity may indicate different types of connection quality. By way of example only, for example, three categories of connection quality may be defined, the first category representing connections suitable for hard real-time applications (e.g., low-level real-time robot motion control), the second category representing connections suitable for soft real-time applications (e.g., AGV job control), and the third category representing connections suitable for monitoring and / or data collection purposes. For example, the wireless device being measured may be an RF measurement device attached to an indoor positioning device. As a cost-effective solution with acceptable accuracy, a device with advanced camera systems and augmented reality (AR) software support may be used to locate the device in a site area with sub-decimeter accuracy.

[0044] As mentioned, in one variation, the features based on the RF channel response measurements may represent channel gain in time and frequency, and in other variations, the features based on the RF channel response measurements may be features derived from the RF channel response measurements. In one such variation, the features based on the RF channel response measurements may be obtained by applying a multipath resolution algorithm (or "technique" / "model") on the measured channel response data that can calculate the number and strength of multiple signal paths and optionally calculate the direction of different paths. Exemplary multipath resolution algorithms may include Fourier transforms or multiple signal classification (MUSIC) algorithms. Similar to the input based on the measured channel gain in time and frequency, the determined multipath propagation mode may also be used as an input to the machine learning model, as long as the machine learning model has been trained to operate on such input. The multipath resolution algorithm may therefore be applied to the RF channel response measurements to determine the observed multipath RF propagation mode, wherein the features based on the RF channel response measurements may represent the multipath RF propagation mode, and wherein the step of determining the estimated connection quality based on the RF channel response measurements may include inputting the determined multipath RF propagation mode into the machine learning model. The machine learning model may also produce an estimated connection quality as a corresponding output. A multipath RF propagation pattern may include the number and strength of multiple RF propagation paths observed at a location, and optionally the directions of different ones of the multiple RF propagation paths.

[0045] The wireless communication network can be any network that allows RF channel response measurements to be determined in the manner described above. In particular, for example, the wireless communication network can be a network that supports services for reliable low-latency communication, such as a 5G or WiFi6 network. The supported services can be in particular URLLC services. RF channel response measurements can therefore be obtained from URLLC services delivered in the wireless communication network. For example, a KPI that relates to a connection and represents the quality of the connection can therefore correspond in particular to a URLLC-related KPI, or more generally, to an industrial target KPI (or "industry-related KPI").

[0046] Before the machine learning model can be used for inference to determine the estimated connection quality based on the RF channel response measurements, the machine learning model can be trained to perform such a mapping. The machine learning model can be any type of machine learning model suitable for being trained for such a mapping, which can in particular include models that employ supervised learning. During the training phase, similar measurements as described above can be performed in a representative environment (e.g., at least one representative wireless communication network that is different from the wireless communication network in which the inference is performed) to generate corresponding training data. In order to train the model, the radio environment covered by at least one representative wireless communication network (i.e., to determine the RF channel response) can be measured at different locations with high accuracy in time and frequency together with relevant target KPIs that can be used as labels for supervised learning. For example, the training data thus collected can correspond to a time series of RF channel response measurements obtained at different locations and can then be used to train the model using conventional techniques such as back propagation. In other words, a machine learning model can be trained using training data comprising multiple sets of features based on RF channel response measurements as input data and corresponding connection qualities measured at different locations covered by a representative wireless communication network as output data, wherein the multiple sets of features based on RF channel response measurements can be obtained from RF channel response measurements indicating channel gain over time and frequency performed at different locations covered by the representative wireless communication network.

[0047] For example, as described above, when features based on RF channel response measurements directly correspond to channel gain in time and frequency, the channel response can be stored as a vector of complex numbers, where each instance of the vector can correspond to a snapshot of the channel. The training input to the machine learning model can then be a time series of such vectors. On the other hand, when features based on RF channel response measurements represent multipath RF propagation patterns, a feature vector including the number and strength of multiple paths and optionally the direction of multiple paths can be formed and such a feature vector can then be used as a training input to the machine learning model. Through training, a mapping from channel gain in time and frequency (or, respectively, multipath RF propagation patterns) to corresponding connection quality can be learned.

[0048] For example, since an industrial environment such as an industrial plant may change from time to time due to the rearrangement of machines, containers or other large objects, it may be conceivable to continuously improve the machine learning model based on existing deployments by collecting detailed measurements from communication devices (such as AGVs or robots) in lifelong operating installations together with the corresponding connection quality. For example, such a device may run in full-time operation and thus even rare events such as interruptions in robot control communications may be detected and associated with RF channel response measurements taken at that moment. These RF channel response measurements for rare events can then be used to map the connection quality to update (or "enhance") the machine learning model, thereby continuously improving the model as more and more deployments are in operation. A wireless communication network can therefore cover an industrial environment, where additional training data can be continuously obtained from one or more wireless devices performing industrial tasks in the industrial environment, where the additional training data can be used to enhance the machine learning model.

[0049] In refinement, it may be conceivable to create multiple environment-specific machine learning models based on different deployments in order to enable the use of better fitting models in the inference phase. Different types of factories may have different equipment and furniture on the floors, which may significantly affect the radio propagation environment and the resulting connection quality. For example, fast fading channels and the propagation of radio signals in a factory where large metal parts and machines are deployed may have fundamentally different characteristics compared to a semiconductor factory or a warehouse with high shelves everywhere. Therefore, different (i.e., separate) machine learning models for mapping RF channel response measurements to connection quality can be trained and different (i.e., separate) machine learning models for mapping RF channel response measurements to connection quality can be used for inference, one for each environment type. The machine learning model can therefore be one of multiple machine learning models that are trained to map features based on RF channel response measurements to corresponding connection qualities, where each of the multiple machine learning models can be trained using training data obtained from measurements performed at locations covered by different representative wireless communication networks that each represent a different type of RF environment.

[0050] In the inference phase, a machine learning model that best fits the location where the RF channel response measurements are made can be selected for the actual inference. For example, in a radio survey of a factory, only some RF channel response measurements may be made, and based on these measurements, the best fitting machine learning model to be used for the prediction of the connection quality may be determined. Once determined, the prediction may be performed using the selected model. In order to select an appropriate machine learning model, it may be conceivable to train a separate machine learning model (e.g., a neural network) that classifies the RF channel response measurement inputs into corresponding environmental type categories, and then, the machine learning model corresponding to the determined category may be selected to predict the connection quality. In this variation, a machine learning model may be selected from a plurality of machine learning models based on the RF channel response measurements using a classifier that is trained to map features based on the RF channel response measurements to RF measurements of the corresponding type.

[0051] In another variation, for example, a set-based machine learning technique may be employed, wherein multiple machine learning models may be used in parallel to predict corresponding connection qualities based on RF channel response measurements, and wherein a final estimate of the connection quality may be made based on an aggregation of the outputs of the parallel models. Aggregation may involve selecting an output from multiple parallel models (e.g., based on some confidence or probability estimate for each output), or aggregation may involve a majority-based decision. In this variation, multiple machine learning models may thus be used as a set in which each of the multiple machine learning models is used to estimate a corresponding connection quality based on RF channel response measurements, wherein the estimated connection quality may be determined based on the corresponding connection quality. For example, a well-known set-based algorithm is the AdaBoost algorithm.

[0052] In the above description of the inference phase, the techniques presented herein have been presented with reference to predicting the channel quality of a single location in an area covered by a wireless communication network. However, it will be appreciated that techniques may be employed to assess the entire environment (or a larger portion of the environment), and therefore, RF channel response measurements may be collected at multiple locations in the area, thereby creating a graph indicating the quality of connection in the surveyed area. Obtaining RF channel response measurements and determining an estimated connection quality may therefore be performed for multiple locations covered by a wireless communication network, wherein the resulting estimated connection quality may be used to generate a connection quality graph for an area covered by the wireless communication network. For example, as described above, since the environment may change from time to time due to the rearrangement of a machine, container, or large object, it may be conceivable to handle such variability by using continuous monitoring of the operating environment of a measuring device attached to a mobile device such as a mobile robot, AGV, or drone in the area. The graph may then be updated using the new reference data collected. The wireless communication network may therefore cover an industrial environment, wherein the measured connection quality may be continuously obtained from one or more (e.g., movable) wireless devices performing industrial tasks in the industrial environment, and wherein the connection quality graph may be updated based on the measured connection quality.

[0053] Once the connection quality graph is created, it may be segmented into different sub-areas that may be characterized by the same or at least similar connection quality. The connection quality graph may thus be segmented into different areas depending on the level of connection quality that may be provided in the respective area. The level of connection quality may correspond to a certain degree of connection quality (e.g. a degree of at least one of BLER, packet delay, packet delay budget, probability of packet delay exceeding a predetermined threshold, probability of packet drops and level of connectivity), or to a class of connection quality in the above sense. Figure 7 An exemplary connection quality graph 700 of an industrial plant with regions having different multipath propagation patterns is illustrated. In the example shown, a signal path originating from a base station 702 is propagated along multiple paths having different strengths and directions, wherein the resulting connection quality graph 700 exposes three segmented regions 704, 706, and 708, each with a different level of connection quality KPI A 、KPI B and KPIs C The connection quality map 700 may have been created from a radio survey conducted by walking around the area, recording RF channel measurement responses with a measurement device, and inferring the corresponding channel quality using a machine learning model to characterize areas 704, 706, and 708 with different RF complexities based on detected multipath characteristics.

[0054] When the predicted connection quality in any area of ​​such an area is lower than the required connection quality, for example, when an area is identified as having a low KPI value for sensitive services, corresponding countermeasures can be taken to improve the connection quality in the relevant area. When the estimated connection quality in an area covered by a wireless communication network is therefore lower than the required connection quality value, one or more mitigation actions can be triggered to improve the connection quality in the area. Such actions may include identifying weaknesses to deploy new base stations (or APs) or identifying reconfiguration actions or radio mitigation actions required to achieve the desired connection quality. One or more mitigation actions may include at least one of the following: deploying one or more additional base stations in the area, allocating one or more additional base stations to serve cells in the coverage area, adapting the RF reflection properties of at least one metasurface provided in the area, and determining the radio link configuration required to achieve the connection quality value in the area and applying the determined radio link configuration in the wireless communication network.

[0055] The action of allocating one or more additional base stations to serve cells in a coverage area may be related to a use case in which the area is covered by a dense deployment of APs (e.g., using an Ericsson Dot system). In such a use case, reducing small-scale fading variability (so-called "channel hardening") may be achieved by allocating more APs to serve the same cell. In such a deployment, each cell may be served by multiple APs and the assignment of APs to cells may be reconfigured, resulting in changes in the coverage area or boundaries of the cell. Multiple APs serving the same cell can provide transmission diversity by sending signals on multiple channels with independent implementations, which may be key to combating small-scale fading. If the identified location with low connection quality for critical (e.g., delay-sensitive) services is close to the cell edge, reconfiguring the allocation of APs serving the cell may effectively improve the channel quality in that area.

[0056] The act of adapting the RF reflection properties of at least one metasurface provided in an area may be related to a use case where a radio environment is dynamically optimized through a software-controlled metasurface whose RF reflection properties can be dynamically adapted to positively influence or optimize the radio environment in a target area. Figure 8 As exemplarily illustrated in , the metasurface 802 can control the direction of the signal it reflects (e.g., by adding a controlled phase shift / delay to the incoming signal) so that the resulting sum of the signals for the intended receiver can be affected in a favorable manner. In cases where the incoming signal is found to be highly variable and in cases where the multipath environment is found to be quite unfavorable for a high level of quality connection (e.g., as determined by a machine learning model), signal reflections and therefore the multipath environment can be improved by applying control of the metasurface 802 by the metasurface.

[0057] The action of determining the radio link configuration required to achieve a desired connection quality value in an area and applying the determined radio link configuration in a wireless communication network may be related to a use case in which a more appropriate radio link configuration is determined (e.g., changing the MCS) based on, for example, fast fading channel fluctuations. In this case, in order to achieve a desired BLER, a safety margin may be added to the channel quality indication (CQI) reported by a wireless device (e.g., a user equipment (UE)) in order to account for possible sudden changes in channel gain. If the BLER is increased by 10 -6 If the BLER is to be maintained at 10 -6 Below, a 10 dB margin may be added to the reported CQI, and a corresponding MCS may be selected accordingly. Determining the radio link configuration may therefore include adding a safety margin to a channel quality value (e.g., CQI) reported from a wireless device in the wireless communication network and selecting an MCS depending on the channel quality value with the safety margin, the safety margin accounting for possible drops in channel gain. For example, the CQI backoff to be applied on a communication channel and a specific radio environment may be determined by a machine learning model. Fig. 9 Exemplary SNR vs. BLER curves for different MCS are illustrated. If the amount of signal drop that can occur with a certain probability is known from measurements and machine learning models, then the same CQI / SNR backoff amount can be applied and "jump" to a more conservative modulation coding MCS curve. It will be appreciated that such CQI backoff estimates are merely exemplary and provide a way to estimate how performance will change if the system configuration is changed. Note that this can be estimated without taking any new measurements.

[0058] Fig.10 A method that can be performed by the computing unit 110 according to the present disclosure is described. The method is dedicated to training a machine learning model configured to evaluate the connection quality in a wireless communication network. In step S1002, the computing unit 110 can train the machine learning model using training data including a plurality of sets of features based on RF channel response measurements as input data and corresponding connection qualities measured at different locations covered by a representative wireless communication network as output data, the plurality of sets of features based on RF channel response measurements being obtained from RF channel response measurements indicating channel gains over time and frequency performed at different locations covered by the representative wireless communication network. Fig.10 The method defines a method specifically for training machine learning models. Once the machine learning model is trained, it can be used according to the above Figure 2 The method described above employs a machine learning model for the purpose of evaluating the quality of a connection. Figure 2The aspects described in the training of the machine learning model used in the method may also be applicable to the operation of the computing unit 110. Therefore, unnecessary repetition is omitted at this point.

[0059] Fig.11 A flow chart summarizing the above method is illustrated, providing an overview of the relevant steps that may be performed. In step S1102, during a training phase, a machine learning model may be generated and trained using RF channel response measurements and corresponding measured connection qualities (indicated in the figure by KPIs). In step S1104, site-specific RF channel response measurements and corresponding KPIs may be collected, onwards to update the machine learning model (see arrow returning to step S1102). In step S1106, during an inference phase, the trained machine learning model may be used to characterize site areas (e.g., by walking around the area and inferring corresponding KPIs). In step S1108, site area segmentation may be performed using the determined KPIs. In step S1110, coverage improvement actions may then be triggered as needed for areas exhibiting lower KPIs.

[0060] As has become apparent from the above, the present disclosure provides techniques for evaluating the quality of connections in wireless communication networks. The techniques can be directed to ensuring good / sufficient connection quality in the service area covered by the wireless communication network, addressing the need for measurement mechanisms to collect more detailed information about the radio propagation environment (e.g., reflections, multipaths, etc.) and connecting these RF attributes to higher-level (e.g., URLLC) performance estimates. The techniques can establish a machine learning model that maps the measured time-frequency channel response to a predicted (e.g., URLLC-type) KPI and can use the machine learning model to evaluate the radio environment from the perspective of industrial applications. The techniques can be considered as a two-stage (i.e., learning and inference) measurement method for reliable low-latency service coverage assessment. Based on the predicted target KPI quality, various mitigation actions can be performed, including network reconfiguration and deployment extensions to improve quality in areas with lower KPIs.

[0061] The technology can enable fast coverage estimation without the need for lengthy simulation runs or lengthy traffic measurements that would otherwise be required, i.e., in order to measure rare events in a statistically reliable manner. Machine learning models can take into account location and environment-specific propagation properties, making predictions more accurate than estimates and simulations based on general statistical channel models known from conventional techniques. The technology can enable strict industrial connection quality KPIs to be estimated based on a few time-frequency channel measurement samples. Channel assessment can be performed in real time, so that quality can be estimated online while measurements are being collected, which can enable new ways of implementing radio surveys (e.g., real-time surveys). Since the best matching machine learning model can be selected from the growing measurement group of existing deployments, the estimation can be more accurate than methods based on simulations or general statistical models. This "knowledge base" of machine learning models can continue to grow, making predictions and models more and more suitable as more deployments are measured. Using the presented technology, radio experts also do not need to plan or evaluate deployments, because virtually anyone can perform measurements while the machine learning model evaluates the robustness of radio connectivity using relevant KPIs.

[0062] It is believed that the advantages of the technology presented herein will be fully understood from the foregoing description, and it will be apparent that various changes can be made in the form, construction and arrangement of its exemplary aspects without departing from the scope of the invention or sacrificing all of its advantageous effects. Since the technology presented herein can be changed in many ways, it will be recognized that the invention should be limited only by the scope of the following claims.

Claims

1. A method for evaluating the connection quality in a wireless communication network, which is applied to the scenario of ultra-reliable low-latency communication (URLLC) in a 5G network, the method comprising: Obtaining (S202) radio frequency (RF) channel response measurements indicative of channel gain over time and frequency observed at a location covered by the wireless communication network; as well as determining (S204) an estimated connection quality at the location based on the RF channel response measurements using a machine learning model trained to map features based on the RF channel response measurements to corresponding connection qualities; wherein a multipath resolution algorithm is applied on the RF channel response measurements to determine an observed multipath RF propagation pattern; The machine learning model is trained using training data including multiple sets of features based on RF channel response measurements as input data and corresponding connection qualities measured at different locations covered by a representative wireless communication network as output data, and the multiple sets of features based on RF channel response measurements are obtained from RF channel response measurements indicating channel gains over time and frequency performed at the different locations covered by the representative wireless communication network.

2. The method of claim 1, wherein: The channel gain is measured in time and frequency based on reference symbols (302) transmitted from a base station (702) of the wireless communication network.

3. The method of claim 1, wherein: The features based on RF channel response measurements represent channel gain in time and frequency, and wherein determining the estimated connection quality based on the RF channel response measurements includes inputting the measured channel gain in time and frequency into the machine learning model.

4. The method of claim 1, wherein: The features based on the RF channel response measurements represent multipath RF propagation patterns, and wherein determining the estimated connection quality based on the RF channel response measurements includes inputting the determined multipath RF propagation patterns into the machine learning model.

5. The method of claim 4, wherein: The multipath RF propagation pattern includes the number and strength of multiple RF propagation paths observed at the location, and optionally the directions of different ones of the multiple RF propagation paths.

6. The method of claim 1, wherein: The RF channel response measurements are obtained from an ultra-reliable low-latency communication (URLLC) service delivered in the wireless communication network.

7. The method of claim 1, wherein: The wireless communication network covers an industrial environment, wherein additional training data is continuously obtained from one or more wireless devices performing industrial tasks in the industrial environment, and wherein the additional training data is used to enhance the machine learning model.

8. The method according to any one of claims 1 to 7, wherein: The machine learning model is one of a plurality of machine learning models trained to map features based on RF channel response measurements to corresponding connection qualities, wherein each of the plurality of machine learning models is trained using training data obtained from measurements performed at locations covered by different representative wireless communication networks that each represent a different type of RF environment.

9. The method of claim 8, wherein: The machine learning model is selected from the plurality of machine learning models based on the RF channel response measurements using a classifier trained to map features based on the RF channel response measurements to corresponding types of RF environments.

10. The method of claim 8, wherein: The multiple machine learning models are used as a set wherein each of the multiple machine learning models is used to estimate a corresponding connection quality based on the RF channel response measurement, and wherein the estimated connection quality is determined based on the corresponding connection quality.

11. The method of claim 1, wherein: Obtaining RF channel response measurements and determining estimated connection qualities are performed for a plurality of locations covered by the wireless communication network, and wherein the resulting estimated connection qualities are used to generate a connection quality map (700) of the area covered by the wireless communication network.

12. The method of claim 11, wherein: The wireless communication network covers an industrial environment, wherein measured connection qualities are continuously obtained from one or more wireless devices performing industrial tasks in the industrial environment, and wherein the connection quality map (700) is updated based on the measured connection qualities.

13. The method of claim 11, wherein: The connection quality map (700) is segmented into different regions (704, 706, 708) depending on the level of connection quality that can be provided in the respective region (704, 706, 708).

14. The method of claim 1, wherein: When the estimated connection quality in an area covered by the wireless communication network is lower than a required connection quality value, one or more mitigation actions are triggered to improve the connection quality in the area.

15. The method of claim 14, wherein: The one or more mitigation actions include at least one of the following: deploying one or more additional base stations in the area (702), allocating one or more additional base stations (702) to serve cells covering the area, adapting an RF reflection property of at least one metasurface (802) provided in the region, and A radio link configuration required to achieve the connection quality value in the area is determined and the determined radio link configuration is applied in the wireless communication network.

16. The method of claim 15, wherein: Determining the radio link configuration includes adding a safety margin to a channel quality value reported from a wireless device in the wireless communication network and selecting a modulation coding scheme MCS depending on the channel quality value with the safety margin, the safety margin accounting for a possible drop in channel gain.

17. A method for training a machine learning model configured to evaluate connection quality in a wireless communication network, which is applied to the scenario of ultra-reliable low-latency communication (URLLC) in a 5G network, the method comprising: The machine learning model is trained (S1002) using training data including multiple sets of features based on RF channel response measurements as input data and corresponding connection qualities measured at different locations covered by a representative wireless communication network as output data, the multiple sets of features based on RF channel response measurements being obtained from RF channel response measurements indicating channel gains over time and frequency performed at the different locations covered by the representative wireless communication network; wherein a multipath resolution algorithm is applied on the RF channel response measurements to determine observed multipath RF propagation patterns.

18. A computer program product comprising program code portions for performing the method of any one of claims 1 to 17 when the computer program product is executed on one or more computing devices.

19. The computer program product of claim 18, stored on a computer readable recording medium.

20. A computing unit (100) for evaluating the quality of a connection in a wireless communication network, the computing unit (100) comprising at least one processor (102) and at least one memory (104), the at least one memory (104) containing instructions executable by the at least one processor (102), such that the computing unit (100) is operable to perform the method according to any one of claims 1 to 16.

21. A computing unit (110) for training a machine learning model, the machine learning model being configured to evaluate connection quality in a wireless communication network, the computing unit (110) comprising at least one processor (112) and at least one memory (114), the at least one memory (114) containing instructions executable by the at least one processor (112), such that the computing unit (110) is operable to perform the method of claim 17.

22. A system comprising the computing unit (100) according to claim 20.

23. The system of claim 22, further comprising a computing unit (110) according to claim 21.

Images