Method for predicting a variation in the quality of service in a cellular telecommunication network, corresponding prediction device and corresponding computer program

Abstract

The invention relates to a method for predicting a variation in the quality of service (QoS) in a cellular telecommunication network comprising at least one base station to which a user device is connected. Such a method comprises: - collecting (DUE) at least one item of location information (IL) comprising an identifier of the base station; - selecting (S_MOD), from among a set of key performance indicators (KPI) of the network, at least one indicator of interest representative of the service quality; - predicting (C_MOD), via machine learning, a future value of the indicator of interest on the basis of at least one temporal trajectory of the indicator of interest, the temporal trajectory being representative of past values of the indicator of interest collected by the prediction device; - transmitting (ADAPT), within the network, a message for adjusting at least one parameter of the network on the basis of the predicted future value.

Description

[0001] DESCRIPTION

[0002] Method for predicting a variation in quality of service in a cellular telecommunications network, corresponding prediction device and computer program.

[0003] Field of invention

[0004] The field of the invention is that of cellular radio communications. In particular, the proposed solution applies notably, but not exclusively, in the context of LTE / 4G (in English "Long Term Evolution") or 5G NR (in English "New Radio") mobile networks.

[0005] More particularly, the invention relates to the use of a machine learning model for predicting the quality of service (QoS) in a cellular telecommunications network, also called a mobile network.

[0006] Previous Art

[0007] Historically, cellular telecommunications networks, or mobile networks, were designed to provide a basic level of service, such as voice and text communication. Over the past few decades, technological advances in user equipment (e.g., smartphones, tablets, laptops, etc.), as well as evolving user expectations, have led to a shift in how users interact with their devices and the Internet.

[0008] Indeed, users now expect mobile networks to provide a diverse range of services, including high-speed internet, streaming video, and real-time gaming.

[0009] One consequence of this evolution is the adaptation of mobile networks. In particular, to meet new user expectations, mobile network operators must be able to differentiate mobile network configuration based on specific services and user needs. This means that mobile network operators must be able to adapt network performance based on the type of service provided and the user's needs. For example, a mobile network may need to prioritize a video streaming service over voice communication or provide faster speeds for online gaming applications.

[0010] To enable mobile networks to adapt, new technologies have emerged, such as real-time analytics capabilities based on machine learning and new radio communication technologies such as 5G. These new technologies have enabled mobile networks to provide new capabilities (e.g., data storage or analysis) and new services to users (e.g., broadband Internet).

[0011] One of the key enablers of this transformation in mobile networks is the ability to store and analyze large volumes of data in real time using machine learning. Machine learning models enable mobile networks to detect patterns and anomalies in network traffic and adjust their parameters accordingly. In other words, by analyzing network data at the time of past anomalies, an artificial intelligence model can infer the factors causing them and therefore predict future anomalies. Furthermore, the ability to differentiate mobile network parameters based on specific service and user needs has become a key requirement for mobile network operators.

[0012] Furthermore, the emergence of new capabilities (such as edge computing) linked to the deployment of 5G networks is expected to bring significant improvements to mobile networks. For example, 5G networks are designed to provide ultra-reliable low-latency communication (URLLC), which enables new use cases such as autonomous vehicles, remote surgery, and real-time monitoring of critical infrastructure (e.g., automated factories).

[0013] Furthermore, the implementation of 5G networks is also leading to the development of network slicing, which allows mobile network operators to create multiple virtual mobile networks on a single physical mobile network infrastructure. Each mobile network slice can be customized with specific features and service levels to meet the needs of different users and use cases.

[0014] Thus, the implementation of 5G networks opens new opportunities for businesses and individuals to develop and use a variety of innovative applications. However, the most demanding 5G network use cases will require Quality of Service (QoS) control to ensure they meet the required performance levels for each user and for each service.

[0015] Quality of service control encompasses several critical parameters, including latency, jitter, throughput, and radio signal quality metrics such as RSRP (Reference Signal Received Power) and RSSI (Received Signal Strength Indicator). These parameters are representative of network performance and therefore determine the level of quality of service that can be delivered over the network. As such, it is essential that they are carefully managed and monitored. The implementation of network slicing technology can enable quality of service control in 5G networks, by providing different levels of QoS for different network slices. In this way, network operators can ensure that the most demanding use cases receive the required level of service while optimizing the use of radio resources.

[0016] However, one of the significant challenges mobile network operators face in providing QoS control for 5G networks is the constantly changing radio conditions. As users move to different areas (or cells) of the mobile network, they experience different radio conditions. Mobile network operators must then adjust network parameters to provide the required level of service.

[0017] However, it is extremely difficult to accurately predict future radio conditions, and mobile network operators may not be able to guarantee Service Level Agreements (SLAs) without a significant loss of radio resources.

[0018] Thus, despite the latest advances in radio communication technology or machine learning, it is still not possible today to individually adapt mobile network characteristics based on a prediction of future quality of service. This means that mobile networks can still experience outages or performance issues that can impact the user experience. While machine learning and real-time capabilities have helped mobile networks become more efficient and effective, there is still work to be done to ensure that mobile networks can anticipate and adapt to user needs in real time.

[0019] There is therefore a need to improve techniques for adapting to a change in QoS in a cellular telecommunications network context in order to ensure the best possible QoS for users. In particular, it is important to allow mobile networks to adapt in advance of a QoS variation in order to guarantee the best possible QoS for the user.

[0020] Statement of the invention

[0021] The present technique makes it possible to propose a solution aimed at remedying certain drawbacks of the prior art. According to one aspect, the present technique relates in fact to a method for predicting a variation in a quality of service in a cellular telecommunications network comprising at least one base station to which at least one user equipment is connected. According to the general principle of the proposed technique, said method is implemented by a prediction device and it comprises: a collection of at least one location information comprising an identifier of said at least one base station to which said at least one user equipment is connected, a selection, from a set of key performance indicators of said network, of at least one key performance indicator of interest representative of said quality of service provided by said at least one base station to said at least one user equipment,a prediction by machine learning of a future value of said at least one key performance indicator of interest from at least one time trajectory of said at least one key performance indicator of interest, said at least one time trajectory being representative of past values ​​of said at least one key performance indicator of interest, collected by said prediction device, a transmission, within said cellular telecommunications network, of a message for adjusting at least one parameter of said cellular telecommunications network as a function of said predicted future value.,

[0022] Thus, the invention proposes a completely new and inventive approach to predicting the variation of the quality of service (QoS) within a cellular telecommunications network (or mobile network) of an operator. More particularly, the invention proposes to predict the variation of QoS within the mobile network, solely on the basis of metric values ​​collected by the different nodes of the operator's mobile network and on the location of a user equipment at the cellular level. In other words, there is no need to resort to geolocation information of user equipment, but only to know their position in the network, for example by knowing the base station to which the user equipment is connected. Indeed, the location of the user equipment is based on the identification of the base station to which the user equipment is connected.In other words, the method according to the invention is based on the location at the cellular level of the user equipment and therefore on the detection of the change of cells (for example, the operator's network knows if the user equipment is present in a given cell because it exchanges with the base station of the cell and towards which cell this user equipment moves when it connects to another base station). It is thus possible to overcome the variation in QoS when the user equipment moves within the network.

[0023] The metrics collected are representative of indicators (for example: throughput, latency, congestion, etc.) that influence network performance and therefore directly or indirectly QoS. In particular, among these indicators, also called key performance indicators or KPIs, we find: KPIs not linked to the service used (for example SINR ("Signal to Interference plus Noise Ratio" in English), or linked to the service used (for example: pixelation for an online video service).

[0024] The ability to predict mobile network performance for each user device is an important development in mobile network optimization. Traditionally, mobile network performance could only be measured in real time or after the fact, making it difficult to optimize mobile network behavior to deliver the best possible user experience. However, with the emergence of machine learning and predictive analytics, it is now possible to develop models capable of predicting mobile network performance for each user.

[0025] In order to make a personalized prediction, one or more key performance indicators of interest, or KPIs, representative of the QoS provided by the telecommunications network to the user equipment are selected from a set of network KPIs. The past values ​​of this or these KPIs of interest are collected by the prediction device, then represented in the form of time trajectories. These time trajectories are therefore representative of the evolution in time and space of the KPIs of interest.

[0026] From these time trajectories, it is possible to predict, depending on the location at the cellular level of the user equipment, the future values ​​of the selected KPI(s) of interest and therefore the variation of the QoS of the end user.

[0027] Depending on these future values ​​of the KPI(s) of interest, it is therefore possible to transmit a message to the operator to adapt the parameters of the telecommunications network (for example: increase the priority of the user equipment, reduce the quality of the video, etc.) to guarantee the best quality of service.

[0028] Using this model, network operators can deliver a higher level of quality by predicting QoS degradation and taking steps to avoid it. For example, if network performance degradation is predicted, the network operator can adjust network parameters to ensure the user continues to receive the expected level of service. This proactive approach can significantly improve the user experience, leading to greater customer satisfaction and loyalty. In addition, network operators can create offers tailored to the specific needs and usage patterns of each user. By predicting network performance, network operators can better understand each user's needs and offer services that meet those needs.For example, a user who uses many jitter- and latency-sensitive applications may require a different level of service than a user who primarily uses voice services.

[0029] Thus, the ability to predict the performance of a mobile network for each user is an important development that offers numerous advantages to mobile network operators and users. By using machine learning, mobile network operators can offer a higher level of quality to users, create offers tailored to the needs of each user, and improve user satisfaction and loyalty. In a particular embodiment, said method further comprises determining a user profile based on said at least one location information and at least one service nature information.

[0030] In this way, for each user, it is possible to determine a user profile based on the one hand on the location information of the users (for example: move a lot from one cell to another, quickly, etc.) and on the other hand on the service information (for example: use of the network for online games, voice, online videos, etc.). This information thus defines a "behavior" of the user, characteristic of his habits in terms of movement and use of the mobile network. By taking into account a user profile, it is thus possible to make a prediction of the variation of QoS centered on the user.

[0031] Indeed, when there is good mobile network coverage, one of the elements that can cause QoS to decrease is the fact that the user equipment moves within the mobile network. Thus, the variation in QoS is closely linked to the user's movement habits. These habits are characterized in particular by its location at the cellular level over time (for example: presence in a cell every day at the same time, daily movements between the same cells, etc.).

[0032] QoS also depends on the type of service used by the user. Not every service requires the same level of network performance, and the user may also have signed a contract that guarantees a certain level of performance for a given service, for example.

[0033] Thus, travel habits, but also service usage habits, characterize a user profile. It is therefore possible to adapt the prediction of QoS variation according to the user profile.

[0034] In a particular embodiment, said selection of said at least one key performance indicator of interest takes into account said user profile and past values ​​of said key network performance indicators, collected by said prediction device.

[0035] Advantageously, in order to refine the prediction to the user profile, it is thus possible to make a selection, among all the key network performance indicators representative of the variation of the QoS, of one or more KPIs of interest. These KPIs of interest are adapted to the user profile, that is to say they depend on the mobility of the user equipment and the type of service used.

[0036] In a particular embodiment, said method further comprises a generation of at least one prediction model associating said at least one user profile with said at least one temporal trajectory of said at least one key performance indicator of interest.

[0037] In this way, it is possible to make a prediction of the variation of the personalized QoS, that is to say adapted to the user profile. Thus, for each user profile, there is a prediction model for which we take into account key performance indicators of interest which will depend on the user profile and in particular on the service used. It is therefore possible to make a personalized prediction, centered on the user, and thus make the necessary network adaptations so that the QoS is adapted to the user's needs.

[0038] In a particular embodiment, said prediction implements an artificial intelligence module configured to implement said at least one prediction model.

[0039] In this way, through the use of such an artificial intelligence module, techniques such as, for example, deep learning techniques based on multi-layer neural networks can be implemented to perform a prediction task from a large volume of diverse input data. More particularly, the network data or information already known and stored in memory in the nodes of the operator's mobile network are collected by the prediction device to train a machine learning algorithm that uses the past data to predict the future values ​​of key performance indicator(s) of interest representative of the QoS as a function of the service used by the user equipment and its movements within the network (i.e. as a function of the user profile).Such a machine learning algorithm can provide accurate predictions of future radio conditions and therefore future QoS, allowing the network operator to adjust network parameters accordingly.

[0040] According to another aspect, the proposed technique also relates to a device for predicting a variation in a quality of service in a cellular telecommunications network comprising at least one base station to which at least one user equipment is connected, said prediction device being configured to: collect at least one location information comprising an identifier of said at least one base station to which said at least one user equipment is connected, select, from a set of key performance indicators of said, at least one key performance indicator of interest representative of said quality of service provided by said at least one base station to said at least one user equipment, predict by machine learning a future value of said at least one key performance indicator of interest from at least one time trajectory of said at least one key performance indicator of interest,said at least one time trajectory being representative of past values ​​of said at least one key performance indicator of interest, collected by said prediction device, transmitting within said cellular telecommunications network a message for adjusting at least one parameter of said cellular telecommunications network as a function of said predicted future value. In a particular embodiment, said device is further configured to determine a user profile as a function of said at least one location information item and at least one service nature information item.,

[0041] In a particular embodiment, said selection of said at least one key performance indicator of interest takes into account said user profile and past values ​​of said key network performance indicators, collected by said prediction device.

[0042] In a particular embodiment, said prediction device implements a machine learning module configured to generate and implement at least one prediction model associating said at least one user profile with said at least one temporal trajectory of said at least one key performance indicator of interest.

[0043] The invention also relates to a computer program product comprising program code instructions for implementing a prediction method as described above, when executed by a processor.

[0044] The invention also relates to a computer-readable recording medium on which is recorded a computer program comprising program code instructions for executing the steps of the method for predicting a variation in quality of service in a cellular telecommunications network according to the invention as described above, when the program is executed by a processor.

[0045] Such a recording medium may be any entity or device capable of storing the program. For example, the medium may comprise a storage medium, such as a ROM, for example a CD ROM or a microelectronic circuit ROM, or a magnetic recording medium, for example a mobile medium (memory card) or a hard disk or an SSD.

[0046] On the other hand, such a recording medium may be a transmissible medium such as an electrical or optical signal, which may be conveyed via an electrical or optical cable, by radio or by other means, so that the computer program contained therein is remotely executable. The program according to the invention may in particular be downloaded over a network, for example the Internet.

[0047] Alternatively, the recording medium may be an integrated circuit in which the program is incorporated, the circuit being adapted to perform the steps or to be used in performing the method of predicting a variation in quality of service in a cellular telecommunications network.

[0048] According to an exemplary embodiment, the present technique is implemented by means of software and / or hardware components. In this regard, the term "module" or "device" may correspond in this document to a software component, a hardware component or a set of hardware and software components.

[0049] A software component corresponds to one or more computer programs, one or more sub-programs of a program, or more generally to any element of a program or software capable of implementing a function or a set of functions, as described below for the module concerned. Such a software component is executed by a data processor of a physical entity (terminal, server, router, etc.) and is likely to access the hardware resources of this physical entity (memories, recording media, communication buses, electronic input / output cards, user interfaces, etc.). Subsequently, resources are understood to mean all sets of hardware and / or software elements supporting a function or a service, whether individual or combined.

[0050] Similarly, a hardware component is any element of a hardware assembly capable of implementing a function or set of functions, as described below for the module concerned. It may be a programmable hardware component or one with an integrated processor for running software, for example an integrated circuit, a smart card, a memory card, an electronic card for running firmware, etc.

[0051] Each component of the system described above of course implements its own software modules.

[0052] The different embodiments mentioned above can be combined with each other for the implementation of the present technique.

[0053] The aforementioned prediction device, computer program and corresponding recording medium have at least the same advantages as those conferred by the method for predicting a variation in the quality of service according to the present invention.

[0054] Brief description of the figures

[0055] Other aims, characteristics and advantages of the invention will appear more clearly on reading the following description, given as a simple illustrative, and non-limiting, example, in relation to the figures, among which:

[0056] Figure 1 schematically illustrates a cellular radiocommunication network to which the prediction method can be applied according to different embodiments of the invention;

[0057] Figure 2 represents in the form of a flowchart the different stages of the learning phase of the prediction method according to one embodiment of the invention;

[0058] Figure 3 represents an example of a heat map representing the evolution in time and space of the values ​​of a key performance indicator, or KPI; Figure 4 represents in the form of a flowchart the different stages of the prediction method according to an embodiment of the invention;

[0059] Figure 5 represents in the form of a flowchart an example of application of the method for predicting a variation in the quality of service according to an embodiment of the invention in a 2G, 3G or 4G type network;

[0060] Figure 6 represents in the form of a flowchart an example of application of the method for predicting a variation in the quality of service according to an embodiment of the invention in a 5G type network;

[0061] Figure 7 schematically illustrates an example of architecture of a device for predicting a variation in quality of service in a cellular telecommunications network, according to one embodiment of the invention.

[0062] Descriptions of embodiments

[0063] The general principle of the invention is based on the spatio-temporal prediction of the variation in the quality of service (or QoS) linked to the movements of user equipment within a cellular telecommunications network, or mobile network.

[0064] Indeed, when there is good mobile network coverage, one of the elements that can cause QoS to decrease is the fact that the user equipment moves within the network and changes areas, or cells. In other words, when a user equipment moves within the mobile network, it can move from one area to a second and experience different radio conditions that will influence the quality of service.

[0065] Advantageously, this method of predicting the variation in QoS does not take into account the geolocation of the user equipment. Indeed, for an operator of a cellular telecommunications network, it is very complicated, or sometimes even impossible (particularly for reasons of confidentiality, for example), to retrieve the geolocation information of a user equipment. Indeed, this geolocation data is not explicitly in the mobile network. In other words, the geographical position may be known to the user equipment if the GPS (or "Global Positioning System" in English, or "Geopositioning by satellite") is turned on, but by default this information is not sent back to the mobile network.

[0066] Thus, the prediction of the variation of the QoS according to the invention is based solely on the exploitation of data already known from the different nodes of the operator's mobile network (for example: type of service used by the user equipment, information on the metrics, or KPIs, representative of the network performances, location of the user equipment at the cellular level, etc.). This prediction of the variation of QoS then allows the operator to proactively adapt the parameters of the mobile network to compensate for this variation and guarantee the user the QoS most suited to his needs.

[0067] We first present, in connection with Figure 1, a cellular radiocommunication network 1, or mobile network, composed of a network of relay antennas (or base stations) Ni to Ni (i being an integer greater than zero), each covering a portion of territory delimited Ci to Q (i being an integer greater than zero), commonly called a cell (represented schematically in hexagonal form in Figure 1), and routing communications in the form of radio waves to and from user equipment located in the corresponding cell.

[0068] To access the services offered by the network operator (voice or mobile data), user equipment must therefore be located within the coverage area of ​​a Ni relay antenna. This has a limited range and only covers a restricted area around it, called a cell. To cover as much territory as possible and ensure that user terminals always have access to the services offered, operators deploy thousands of Ci cells, each of which is equipped with Ni antennas, ensuring that their coverage areas overlap, so as to provide as complete a coverage of the territory as possible.

[0069] It should be noted that other well-known infrastructures (hardware and / or software) constitute this mobile network (such as a base station controller, etc.) thus forming the different nodes of the mobile network (not shown).

[0070] The mobile network consists of an access network and a core network. The access network is the part of the mobile network that connects end-user devices (such as smartphones or laptops) to radio base stations and other network infrastructure. Access network data includes information such as signal strength, throughput, latency, and packet loss, which are important factors in determining network performance for individual users. The core network also plays a crucial role in determining network performance, especially for data-intensive services such as video streaming or online gaming. Core network data includes information such as routing, type of service, congestion, and load balancing, which can have a significant impact on network performance for individual users.

[0071] Therefore, to have accurate predictions, it is important to consider both access network and core network data. Indeed, by analyzing both access network data and core network data, we can gain a more complete understanding of network performance and provide a better user experience.

[0072] Thus, advantageously, the method for predicting the variation in QoS according to the invention takes advantage of the collection and exploitation of mobile network usage data already stored in memory in one or more nodes of the mobile network. More particularly, by user data or mobile network usage data is meant all of the so-called "network" data or information, known to one or more nodes of the network. This set of network information groups together information from both the access network and the core network. This information is in particular: so-called "location" information representative of the location over time and at the cellular level of one or more user devices (for example, presence of user equipment in cell Ci, then movement to cell C2).It is the base station of the cell in which the user equipment is present that detects the presence of the user equipment when the latter connects to the base station. The location information includes in particular an identifier of the base station to which the user equipment is connected. It is therefore understood that location information means any information making it possible to know at any time the position of the user equipment within the network; so-called "nature of service" information representative of the nature of the service(s) used by a user equipment (for example: online gaming, video, voice, etc.), metrics, or key network performance indicators (also called network KPIs), making it possible to quantify the network performance (for example: throughput, latency, SINR, etc.) and which will therefore directly or indirectly influence the QoS.Among these network KPIs, we find: o key performance indicator(s) or so-called "low-level" KPIs, i.e. not linked to the service used and which will therefore indirectly influence the QoS. These KPIs, called "low-level" KPIs, are representative, for example, of the quality of the radio signal, congestion conditions, etc. In other words, they do not allow us to directly predict the impact on the service. These are, for example, the signal-to-interference ratio or SINR; o key performance indicator(s) called "high-level", i.e. KPIs dependent on the service. In other words, for a given service (e.g. online video) we can select one or more KPIs representative of the QoS of this service (e.g.: re-buffering of a video, video freezing, pixelation). In this case, we must know in real time the service used by the user equipment.For this, either the nature of service information is directly known to the mobile network, or it can be deduced based on known network data such as throughput pattern, IP address, etc. It can also be predicted based on a history of all services used by a user equipment over a given period. Indeed, by having a sufficient volume of data over time, a machine learning model can deduce the application that a user could use over a period of time. For example, a user who watches an online video service every evening, a music application while going to work, ...The method for predicting the variation of QoS in a cellular telecommunications network according to the invention comprises two main phases: a phase of learning one or more model(s) for predicting the variation of QoS, and a phase of implementing this(these) prediction model(s). The different steps of the prediction method according to the invention are implemented by an ML machine learning module of a DISP prediction device. The DISP prediction device will be described in connection with Figure 7.

[0073] The ML machine learning module uses machine learning algorithms to predict future quality of service (QoS) from a prediction model ("low-level" prediction) or from a combination of multiple prediction models ("high-level" prediction) and thus allow the operator to adjust network parameters accordingly. This allows operators to improve network performance, as well as the user experience of that network, and provide a more personalized and responsive service.

[0074] We now present, in connection with Figure 2, a flowchart representing the different stages of the learning phase of the QoS variation prediction models. The objective of this learning phase is to train the ML machine learning module, implemented by the DISP prediction device, to identify different user profiles on the basis of a set of network information, and to associate at least one prediction model with it. More particularly, for each user profile, we have one or more prediction model(s) associated according to the previously selected metrics or KPIs of interest (“low-level” KPIs or “high-level” KPIs). The KPIs of interest are selected on the basis of the service used by the user equipment and the user profile.

[0075] For this, during the learning phase, during a step E1, the mobile network usage data, stored in memory in one or more nodes of the mobile network, are collected by the prediction device DISP. More specifically, for a group of several user equipments, the prediction device DISP collects over a period P (for example 24H) a set of information from an access network ACC and a core network CO, such as: a set of location information of the user equipments at the cellular level, a set of KPIs representative of the network performances (“low level” KPIs or “high level” KPIs) during the period P, information on the nature of services used during the period P.

[0076] User equipment does not need to send this network information because it is already known to one or more nodes in the network. For example, when a user equipment uses an online video service, the mobile network already knows the quality of the video, as well as other information such as throughput, latency, congestion, etc.

[0077] From this network information, the ML learning module creates a HIS network usage history. The HIS network usage history includes, for each user equipment in the group, a set of time trajectories representing the evolution, during period P, of a set of KPI values. In other words, this HIS history includes all possible values ​​of the metrics, or KPIs, collected during period P. These values ​​will therefore depend on: the radio conditions encountered by the user when moving around the network and the service(s) used by the user equipment. These time trajectories can be anonymized or not depending on the user's consent.

[0078] From these temporal trajectories, the ML machine learning module also generates in step E1 temporal maps representative of the evolution in time and space of the KPIs. In particular, a representation in the form of a heat map is used. In connection with Figure 3, an example of a heat map representing the evolution in time and space of the values ​​of a KPI is shown. This heat map (CTH_KPI_C_tps) represents the values ​​of the KPI (for example: a flow rate) VAL_KPI as a function of time and the attachment cell Cel l_att, i.e. the “cellular” location of the user equipment.A heat map therefore makes it possible to represent both: the evolution over time of the values ​​of a KPI, for a given cell Q, (Val_KPI_Ci_tps), at each past, present or future moment, the variation of the values ​​of a KPI on a set of cells, the evolution over time of the values ​​of the KPI for all the cells (Val_KPI_C_tps).

[0079] A heat map is therefore a graphical representation of data where the values ​​of that data are represented as colors (or grayscale, in the example in Figure 3). In network optimization, a heat map can be used to visualize different KPIs and services on a geographical map, allowing network operators to identify areas of high or low network performance. The heat map can also help network operators optimize network performance by identifying areas where improvements are needed. For example, if many users are experiencing poor network performance in a particular area, network operators can use the heat map to identify the root cause of the problem and take steps to improve network performance in that area.

[0080] During a step E2, also from the aforementioned network information, the ML learning module creates one or more user profile(s) PU. These user profile(s) PU group(s) together the users who have the same profile, i.e. the same “behavior”, in order to create predictions for each type of behavior, or user profile. The behaviors are defined by: MOB location information of the users (for example users who move a lot from one cell to another, quickly, etc.) and by SER service nature information (for example: network use for online games, voice, online videos, etc.). The different user profiles are therefore characteristic of the users’ habits. For example, we can have a group of users who make the same daily journey using the same online music application.This group therefore has the same location information (figure 1: movement from cell Ci to cell C2 then cell C3) and type of service (online music application). The prediction device therefore collects, for this group of users, network KPI measurements (latency, throughput, etc.) which will vary throughout the journey depending on radio conditions.

[0081] In step E3, we define the temporal fineness FT, or granularity, used for setting up MOD usage models. The temporal fineness FT depends on: the nature of the KPIs taken into account (for example, some KPI values ​​can be obtained after 15 minutes, others after one hour), the service used and the mobility of the user equipment (for example: the user equipment stays 4 hours on a cell, then moves to another cell). Generally, “low-level” KPIs are reported in “real time”. For other KPIs, particularly “high-level” KPIs, a certain processing time must be allowed (for example, a few minutes or several hours). This processing time depends on the KPI and a threshold that is set. For example, for a KPI related to mobility, the processing time can be set to several hours to check whether a user is passing through a cell or not.

[0082] The FT temporal fineness therefore makes it possible to define the prediction horizon (for example, we make a prediction over the next 15 minutes or 1 hour, etc.). In other words, the FT temporal fineness depends on the scenario being treated: the mobility of the user equipment and the KPI(s) taken into account for the QoS evaluation.

[0083] In a step E4, the ML machine learning module of the DISP prediction device is therefore trained to associate with each user profile PU, time trajectories representative of the evolution, during the period P, of the values ​​of KPIs of interest selected from among the set of network KPIs.

[0084] In order to be able to select the most relevant KPIs of interest for the PU user profile and therefore associate them with the appropriate time trajectories, the heat maps generated in step E1 are used. Indeed, the generation of heat maps can help to identify the most relevant KPI / service pairs for each PU user profile by visualizing the network performance data. In the previous example of the group of users with a daily commute and the use of the same online music application, a KPI of interest is for example the throughput.

[0085] By displaying different KPIs and, possibly, services on the map, the heat map can help network operators identify areas with high or low network performance, and identify the specific KPIs that are most relevant to the user in each location and for each service.

[0086] Once the most relevant KPIs are identified, the prediction model(s), centered on a PU user profile, can use the past values ​​of the KPIs of interest to predict the future network performance for each user, based on their past usage habits and other factors (e.g., current network status, user location, etc.). In other words, from the past time trajectories of KPIs of interest and the PU user profile, it is possible to predict the future values ​​of these KPIs of interest and therefore, if necessary, adapt the network parameters to guarantee the ideal QoS for the user.

[0087] Thus, several MOD prediction models are created in step E4. These MOD prediction models depend on the PU user profile, the selected KPIs of interest, and possibly the service used at the time of the prediction. In other words, depending on the PU user profile, we have “low-level” prediction models and “high-level” prediction models. Indeed, each MOD prediction model is associated with its own list of KPIs of interest to be taken into account as input to the learning algorithm, as well as its own temporal fineness. In other words, it is not always the same KPIs of interest that are used for a new prediction.We thus define prediction models: so-called "low level" for which we only take into account so-called "low level" KPIs of interest not linked to the service used during the implementation of the prediction process, so-called "high level" for which we take into account, alternatively or in combination, so-called "high level" KPIs of interest dependent on the service (for example: for online video, we take into account KPIs specific to this type of application).

[0088] Each prediction model MOD is then stored S_MOD (see figure 4) in a memory of the prediction device DISP.

[0089] Additionally, each model is also associated with a learning rate and / or an interference rate.

[0090] More specifically, the learning rate is a parameter used when training a machine learning model. It determines how quickly the model adjusts its weights and parameters based on errors it makes when predicting outcomes. A high learning rate can allow the model to converge to a solution more quickly, but also carries a risk of missing some local optima. A low learning rate can slow down the training of the learning model, but it generally allows for better exploration of the solution space. The choice of learning rate is usually made based on the complexity of the model, the size of the dataset, and the nature of the problem to be solved. It is common to adjust the learning rate over time (e.g., by gradually decreasing it) to improve the stability and accuracy of the model.The interference rate, or regularization rate, is a measure of a model's ability to generalize from training data to new data. Interference occurs when the model fits the training data too closely, which can lead to poor performance on test data or real data. To avoid such behavior, regularization techniques are used, such as L1 / L2 regularization, dropout, etc. The interference rate controls the magnitude of these regularization techniques. A higher interference rate reduces the model's ability to fit the training data too closely, but can also reduce its ability to capture complex patterns and impair its performance on the training data.A lower interference rate may allow the model to better fit the training data, but may also result in overfitting and poor generalization.

[0091] It should be noted that the optimal choice of the learning rate and the interference rate may vary depending on the specific model, the dataset, the problem to be solved and other factors. According to a particular feature of the proposed technique, a cross-validation is performed in order to identify the best parameters (learning rate and / or interference rate) for each model used in the prediction of a given KPI. The different steps of the method for predicting the variation of the QoS according to an embodiment of the invention are now presented in connection with Figure 4, in the form of a flowchart.

[0092] After the end of the learning phase, the ML learning module of the DISP prediction device is therefore able to implement a so-called “low level” prediction model, and / or a so-called “high level” prediction model.

[0093] Thus, in a phase of implementation of the prediction models, the prediction device DISP collects DUE a new set of network information for a new user equipment or for a new group of user equipment.

[0094] From this network information, the machine learning module ML then searches R_PU for a similar user profile from the set of PU user profiles created and stored in memory S_PU during the training phase. Once a similar user profile D_PU is determined, the learning module selects S_MOD based on this profile at least one prediction model associating the PU user profile with the temporal trajectories of the KPIs of interest selected based on the user profile and the service used.

[0095] The learning module then applies C_MOD the selected prediction model(s) taking into account as input the values ​​of the KPIs of interest (among the network information collected during the DUE collection) of the new user equipment or group of user equipment, and thus predicts a future value of at least one key performance indicator of interest from at least one time trajectory of the selected key performance indicator(s) of interest.

[0096] It is then possible to determine whether the predicted QoS is considered suitable (QoS_ok = O) or not (QoS_ok=N). If the predicted QoS is considered suitable (QoS_ok = O), the prediction process is stopped (STOP). Otherwise, if the predicted QoS is not considered sufficient (QoS_ok = N), the network operator can then modify (ADAP) the mobile network parameters to ensure that the end user has a QoS adapted to his needs.

[0097] To this end, the prediction device transmits a message to adjust the network parameters to the operator. For example, the prediction device sends an adjustment message to the OAM (Operations, Administration and Maintenance).

[0098] This enables network operators to take proactive measures to optimize network performance and avoid degradation of service quality, thereby providing a better user experience and increasing customer satisfaction.

[0099] We now present in connection with Figures 5 and 6 examples of prediction of the variation of QoS and adaptation of the mobile network based on models of prediction of the variation of QoS according to an embodiment of the invention.

[0100] In the example presented in connection with Figure 5, the prediction model is a so-called "low-level" model which only takes into account a KPI of interest not linked to the service, representative of the radio conditions, for example the SINR. In this example, the prediction model is applied to a 2G, 3G or 4G mobile network.

[0101] Over a period P' defining the temporal fineness of the prediction (for example a few minutes), the ML machine learning module collects a set of network information associated with a user's network usage. It should be noted that the finer the granularity, the more accurate the prediction. From the location information and the nature of the service(s) used over the period P', the ML machine learning module determines the profile of this new user and identifies, among the PU user profiles already stored in memory during the learning phase, the one that is similar to the new user.

[0102] Depending on the determined user profile, and on the basis of a so-called “low-level” prediction model, the learning module selects a KPI of interest, the SINR for example.

[0103] The ML machine learning module then identifies, based on the selected KPI of interest, in this case the SINR, an appropriate prediction model from among the prediction models stored in memory. The ML machine learning module then implements the identified prediction model to predict (DEG KPI), from a heat map, the future values ​​of the considered KPI of interest, here the SINR.

[0104] If a degradation of the SINR, and therefore a loss of throughput, is predicted based on the movements of the user equipment, then the ML machine learning module looks (Ex_SER) at which service is used and whether it requires high throughput, for example. If this is not the case (N), then the prediction process is stopped (STOP). In other words, the degradation of the SINR has little or no influence on the QoS for this user.

[0105] If, on the other hand, the service used requires high bandwidth (O), for example, online video, then if this loss of bandwidth (i.e., the degradation of SINR) is significant enough to imply a degradation of video and therefore of QoS, in this case, the operator can modify the network parameters. For example, the operator can reduce the definition of the video or order the network to increase the client's priority level (Adap_Res).

[0106] If no SINR degradation is predicted (N), then it is possible to move to a so-called "high-level" prediction model by taking into account other KPIs (KPI_ADD) in addition. In particular, for a "high-level" model it is possible to select one or more KPIs associated with the service used during the period P' and then predict the future QoS (Pred_QoS).

[0107] To do this, the ML machine learning module determines, based on information on the nature of the service, which service is used (e.g., an online video service). Based on the user profile and the nature of the service, the ML machine learning module selects one or more KPIs of interest related to the service used (e.g., video freezing, pixelation, etc.).

[0108] Based on the user profile and the selected KPIs of interest, the ML machine learning module selects the appropriate prediction model. For this purpose, the ML module determines (QoS_ok?) whether the QoS is satisfactory or not. According to this prediction model, if the QoS is satisfactory (O), the prediction process is terminated (STOP). On the contrary, if the QoS is not satisfactory (N) and the KPI(s) degrade(s) significantly enough, then the operator can modify the network parameters and reduce the video definition or can order the network to increase the client priority level (Adap_Res), for example.

[0109] In connection with Figure 6, we present an example of prediction of the variation of QoS also based on a so-called “low level” prediction model taking into account only a KPI of interest not linked to the service, representative of the radio conditions, for example the SINR, but this time applied in a 5G type network.

[0110] The steps are similar to those described in connection with Figure 5, except that the particularities of the 5G network are taken into account. Thus, if a degradation of the SINR, and therefore a loss of throughput, is predicted based on the movements of the user equipment, then the ML machine learning module looks at whether the network slice or the user's SLA is satisfactory (Ex_SLA). If this is the case (O), then the prediction process is terminated (STOP). Otherwise (N), the operator can reduce the video definition or can order the network to increase the client's priority level (Adap_Res).

[0111] If no SINR degradation is predicted (N) then the same steps are implemented as in Figure 5.

[0112] Thus, the previously described user-centric prediction model(s) for QoS variation predictions can be applied to various types of networks, including 4G and 5G networks. However, the benefits may be greater for 5G networks due to the capabilities of the technology.

[0113] Indeed, even if the network slicing function is not necessary, it can improve the efficiency of the prediction model. Network slicing allows the creation of multiple virtual networks with different characteristics, such as different levels of quality of service. With network slicing, the prediction method according to the invention can predict the network performance for each user in different slices, allowing network operators to offer tailored services that meet the needs of each user.

[0114] However, even without network slicing, the prediction method of the invention can still be useful for optimizing network performance and providing a better user experience. By predicting network performance for each user, network operators can take proactive measures to avoid degradation and improve overall service quality.

[0115] In order to illustrate more precisely the principle of the invention, Figure 7 schematically presents the architecture of a DISP prediction device, according to one embodiment of the invention.

[0116] The DISP prediction device includes an ML machine learning module.

[0117] The prediction device DISP further comprises a RAM, a CPU processing unit equipped for example with a processor, and driven by a computer program stored in a read-only memory (for example a ROM memory or a hard disk). Upon initialization, the code instructions of the computer program are for example loaded into the RAM before being executed by the processor of the CPU processing unit.

[0118] The DISP prediction device further comprises a memory MEM for storing data from the communication network 1, such as for example the type of service used, the KPIs and metrics representative of the network performance, the mobility information of one or more user devices, etc. The DISP prediction device also comprises a communication module COM for receiving data from the nodes of the mobile network and transmitting a message for adjusting the network parameters.

[0119] Figure 7 illustrates only one particular way, among several possible ways, of implementing the prediction device DISP, so that it performs the steps of the method for predicting a variation in quality of service in a communication network detailed above, in relation to Figures 2 to 6 in its different embodiments. Indeed, these steps can be carried out indifferently on a reprogrammable computing machine (a PC computer, a DSP processor or a microcontroller) executing a program comprising a sequence of instructions, or on a dedicated computing machine (for example a set of logic gates such as an FPGA or an ASIC, or any other hardware module).

[0120] In the case where the DISP prediction device is produced with a reprogrammable computing machine, the corresponding program (i.e. the sequence of instructions) may be stored in a removable storage medium (such as for example an SD card, a USB key, a CD-ROM or a DVD-ROM) or not, this storage medium being partially or totally readable by a computer or a processor.

Claims

CLAIMS 1. Method for predicting a variation in a quality of service (QoS) in a cellular telecommunications network comprising at least one base station to which at least one user equipment is connected, characterized in that said method is implemented by a prediction device and comprises: a collection (DUE) of at least one location information (IL) comprising an identifier of said at least one base station to which said at least one user equipment is connected, a selection (S_MOD), from a set of key performance indicators (KPIs) of said network, of at least one key performance indicator of interest representative of said quality of service provided by said at least one base station to said at least one user equipment,a prediction (C_MOD) by automatic learning of a future value of said at least one key performance indicator of interest from at least one time trajectory of said at least one key performance indicator of interest, said at least one time trajectory being representative of past values ​​of said at least one key performance indicator of interest, collected by said prediction device, a transmission (ADAPT), within said cellular telecommunications network, of a message for adjusting at least one parameter of said cellular telecommunications network as a function of said predicted future value., 2. Prediction method according to claim 1, characterized in that it further comprises: a determination of a user profile as a function of said at least one location information and at least one service nature information.

3. Prediction method according to claim 2, characterized in that said selection of said at least one key performance indicator of interest takes into account said user profile and past values ​​of said key network performance indicators, collected by said prediction device.

4. Prediction method according to claim 3, characterized in that it further comprises a generation of at least one prediction model associating said at least one user profile with said at least one temporal trajectory of said at least one key performance indicator of interest.

5. Prediction method according to claim 4, characterized in that said prediction implements an artificial intelligence module configured to implement said at least one prediction model.

6. Device for predicting a variation in a quality of service in a cellular telecommunications network comprising at least one base station to which at least one user equipment is connected, characterized in that said prediction device is configured to: collect at least one location information comprising an identifier of said at least one base station to which said at least one user equipment is connected, select, from a set of key performance indicators of said network, at least one key performance indicator of interest representative of said quality of service provided by said at least one base station to said at least one user equipment, predict by machine learning a future value of said at least one key performance indicator of interest from at least one time trajectory of said at least one key performance indicator of interest,said at least one time trajectory being representative of past values ​​of said at least one key performance indicator of interest, collected by said prediction device, transmitting within said cellular telecommunications network a message for adjusting at least one parameter of said cellular telecommunications network as a function of said predicted future value., 7. Prediction device according to claim 6, characterized in that it is further configured to determine a user profile based on said at least one location information and at least one service nature information.

8. Prediction device according to claim 7, characterized in that said selection of said at least one key performance indicator of interest takes into account said user profile and past values ​​of said key network performance indicators, collected by said prediction device.

9. Prediction device according to claim 8, characterized in that it implements a machine learning module configured to generate and implement at least one prediction model associating said at least one user profile with said at least one temporal trajectory of said at least one key performance indicator of interest.

10. Computer program product comprising program code instructions for implementing a method according to any one of claims 1 to 5, when executed by a processor.

11. A computer-readable recording medium comprising program code instructions which, when executed by a processor, cause the processor to carry out a method according to any one of claims 1 to 5.

Images