Method, communication system and computer program for clustering data streams established over wireless bridge

Abstract

The invention relates to a method for clustering data streams established by a wireless bridge of a communication system within a time-sensitive network. The method comprises the following steps: obtaining indexes of sequential communication performance and concurrent communication performance of each data stream for a plurality of data streams at a functional entity of a core network component of the wireless network bridge; the data streams are clustered based on the obtained indexes, so that if the difference between the sequential communication performance and the concurrent communication performance is lower than a preset threshold value, the two given data streams are independent, otherwise, the two given data streams are dependent, and any two data streams belonging to two different clusters are independent; and any two data streams in the same cluster are dependent. The invention also relates to a corresponding communication system and a corresponding computer program.

Description

Technical Field

[0001] The present disclosure belongs to the field of telecommunications. Background Art

[0002] The present disclosure more particularly relates to a method for clustering data flows established by wireless bridges of a communication system within a time sensitive network. The present disclosure also relates to a corresponding communication system and a corresponding computer program.

[0003] The present disclosure addresses the use of 5G networks as a means of communication for Time Sensitive Networking (TSN), which was originally developed to provide guaranteed communications over Ethernet wired networks for industrial applications.

[0004] As is customary in current TSN standards, when a flow passes through a TSN bridge, the end-to-end delay is calculated by taking into account the cumulative delays of the transmission links between the previous bridges on the path (external links) and the internal links of the bridge between the input (ingress) and output (egress) ports. In a typical TSN network, these cumulative delays are taken into account in a central scheduler or central network configuration node (CNC), which ensures that no packet collisions occur on the transmission links in the time domain.

[0005] In the context of a one-to-many bridge, the bottleneck is the external transmission link, on which a TDMA (Time Division Multiple Access) strategy is applied as a result of CNC scheduling. Even though the internal links of the bridge can be used simultaneously, they are actually used sequentially as a result of TDMA on the input transmission links. When the capacity of the internal links is much higher than the capacity of the transmission links, the impact on performance is small. However, when the internal link capacity is comparable to or worse than the external link capacity, the convention of taking into account the accumulated delays leads to a significant performance degradation. In fact, the bridge capacity is limited by not taking advantage of the possibility of using the internal links simultaneously.

[0006] This situation occurs when considering a 5G TSN bridge, where the internal link capacity involves the wireless transmission capacity, which is typically worse than the input link capacity that relies on the wired network. By using the prior art model and representing the 5GS as a one-to-many topology, a TDMA strategy is applied between all packets flowing through the 5G TSN bridge. Unfortunately, TDMA is generally not the optimal multiplexing scheme for wireless networks, since other multiplexing dimensions can be exploited with improved performance. Examples of multiplexing dimensions include frequency (using several sub-bands in the frequency domain), space (using multiple antennas), and site (frequency multiplexing between several non-interfering sites). Therefore, the prior art definition of the model of the 5G TSN bridge cannot exploit the multiplexing capabilities of the 5GS, which results in suboptimal performance.

[0007] The time division multiplexing capabilities of 5GS are generally much lower than those of Ethernet fixed networks. For example, the smallest time unit of 5GS is a time slot, which lasts at least one hundred microseconds with a maximum frame duration of about 12 microseconds on a Gigabit Ethernet link. Therefore, the low time granularity of 5GS is taken into account in the statement of independent delay as the price paid for any frame transmission. This is necessary because the guaranteed delay is calculated using dependent delays and independent delays in the CNC. However, 5GS relies on a wireless interface that naturally allows broadcasting. This involves sending several data streams simultaneously with variable capacity depending on the deployment, terminal location and channel conditions, the number of data streams that share the wireless channel at the same time, etc. When considering the prior art 5GS TSN bridge representation, this multiplexing capability of 5GS is not inherently taken into account. In fact, with this bridge model, following the low internal time granularity of 5GS, the CNC presents packets at the bridge input in a TDMA manner.

[0008] In order to improve system performance and exploit the multiplexing capability of wireless bridges such as 5GS and compensate for their low time granularity, a better TSN bridge model for designing such wireless bridges is needed. Summary of the invention

[0009] The present disclosure improves the situation.

[0010] A method is proposed for clustering data flows established via a wireless bridge, which is formed by a wireless communication system and is integrated in a time-sensitive network.

[0011] -Wireless bridge handles multiple user terminals,

[0012] -Time-sensitive network includes core network side terminal,

[0013] -The time-sensitive network also includes a plurality of device-side end stations, the device-side end stations being attached to the user terminals,

[0014] - at least two data flows are established in the time-sensitive network and forwarded by the wireless bridge, wherein one data flow is established between the core network side end station and a single device side end station,

[0015] The method comprises:

[0016] - obtaining, at a functional entity of a core network component of the wireless bridge, for each of at least two data flows, an indicator of sequential communication performance of the data flow corresponding to a situation where the at least two data flows are sequentially established,

[0017] - for each of the at least two data flows, obtaining an indicator of concurrent communication performance of the data flow corresponding to a situation where the at least two data flows are established simultaneously,

[0018] - clustering the at least two data flows into a cluster of dependent data flows based on the obtained indicators, such that:

[0019] o if the difference between the sequential communication performance and the concurrent communication performance is below a predetermined threshold, then the two given data flows are independent, otherwise the two given data flows are dependent,

[0020] oAny two data flows belonging to two different clusters are independent, and

[0021] oAny two data flows in the same cluster are dependent.

[0022] The specific clustering defined above allows the automatic discovery of the virtual topology of the wireless bridge (the so-called TSN model) from data flow measurements as a collection of independent channels, where each channel corresponds to a cluster and is associated with a pair of end-to-end guaranteed delays (i.e., a dependent delay and an independent delay).

[0023] Additionally, as described by the TSN model topology, the end-to-end guaranteed latency of each independent channel can be specifically measured by independently activating the data flows within each cluster, and can be provided together with the TSN model to a central network configuration entity (CNC) that handles data flow scheduling in TSN.

[0024] In a particular cluster defined above, communication performance for a given channel may be defined by one or more delays associated with one or more communications conducted over the given channel.

[0025] In some examples, these delays may be, for example, end-to-end delays between two user terminals connected via the channel, as described above, where these delays may differ depending on the channel activation scheme. Alternatively, in some examples, these delays may include radio transmission delays associated with a radio link between a user terminal and its serving end station, where activating a given radio link between a given user terminal and its serving end station is equivalent to activating a given channel between a given user terminal and another user terminal in the communication network.

[0026] The TSN model is accurate, taking into account any type of multiplexing capabilities of the wireless bridges, and the corresponding end-to-end guaranteed delay allows the CNC to calculate the routing and scheduling of packets in the TSN network using wireless bridges with improved data flow performance relative to known methods.

[0027] In another aspect, a computer software or program is proposed, which includes one or more instructions to implement at least a part of the method defined herein when the software is executed by a processor. In another aspect, a computer readable non-transitory recording medium is proposed, on which the software is registered to implement the method as defined herein when the software is executed by a processor.

[0028] On the other hand, a communication system in a time-sensitive network is provided, the communication system comprising a wireless bridge for processing a plurality of user terminals, the time-sensitive network comprising a core network side terminal station, the time-sensitive network further comprising a plurality of device side terminal stations, the device side terminal stations being attached to the user terminals, at least two data flows being established in the time-sensitive network, wherein a data flow is established between the core network side terminal station and a single device side terminal station, the communication system being configured to:

[0029] - at a functional entity of a core network component of the wireless bridge, for each of the at least two data flows, obtaining an indicator of sequential communication performance of the data flow corresponding to a case where the at least two data flows are sequentially established, and for each of the at least two data flows, obtaining an indicator of concurrent communication performance of the data flow corresponding to a case where the at least two data flows are simultaneously established,

[0030] - clustering the at least two data flows into a cluster of dependent data flows based on the obtained indicators, such that:

[0031] o if the difference between the sequential communication performance and the concurrent communication performance is below a predetermined threshold, then the two given data flows are independent, otherwise the two given data flows are dependent,

[0032] oAny two data flows belonging to two different clusters are independent, and

[0033] oAny two data flows in the same cluster are dependent.

[0034] The following features may optionally be implemented individually or in combination with one another.

[0035] In one example, the communication performance of a data stream established between a pair of end stations is related to at least one element in a list, the list comprising:

[0036] - a first value of an end-to-end delay for uplink communications between the pair of end stations,

[0037] - a second value of said end-to-end delay for downlink communications between said pair of end stations, and

[0038] - a maximum value between the first value and the second value.

[0039] The first value and / or the second value may also be related to best effort communication between end stations.

[0040] In this example, latency is selected as the metric for determining communication performance due to its relevance to determining the end-to-end guaranteed delay according to a number of alternative definitions. However, any known QoS metric may be selected instead of or in conjunction with latency.

[0041] When the communication performance of the data streams all relate only to the value for uplink communication, then the clustered data streams all have the same direction as uplink. Conversely, when the communication performance of the data streams all relate only to the value for uplink communication, then the clustered data streams all have the same direction as downlink. This means that a separate set of clustered data streams can be determined and provided to serve as two separate TSN models for uplink and for downlink of the wireless bridge.

[0042] In one example, a given user terminal attached to a device-side end station is active when receiving or transmitting a given data stream, and inactive otherwise:

[0043] - The indicator of sequential communication performance of the first data stream of the at least two data streams may include:

[0044] o a first measure of the communication performance of the first data flow, and

[0045] o a first activity indicator, the first activity indicator at least indicating that, at the time of the first measurement, a first user terminal attached to a device-side end station participating in receiving or sending the first data stream is active, while a second user terminal attached to a device-side end station participating in receiving or sending the second data stream is inactive, and

[0046] -The concurrent communication performance indicator of the first data flow may include:

[0047] o a second measure of the communication performance of the first data flow, and

[0048] o A second activity indicator, the second activity indicator at least indicating that at the time of the second measurement, the first user terminal and the second user terminal are both active.

[0049] The activity indicator as defined above allows characterizing a measure of communication performance by determining whether such measure is to be interpreted as sequential communication performance or concurrent communication performance.

[0050] The activation indicator can be determined from a binary matrix, the binary matrix including a list of binary variables, one binary variable corresponding to one device-side end station, the binary variable having a first value when the user terminal attached to the corresponding device-side end station is active, and the binary variable having a second value when the user terminal attached to the corresponding device-side end station is inactive.

[0051] Such a binary matrix is ​​an efficient way to store the activity indicators, has no limit on the number of UEs handled by the wireless bridge, and has minimal data storage size requirements.

[0052] The method may further include:

[0053] - before obtaining an indicator of sequential communication performance and an indicator of concurrent communication performance of the at least two flows, activating the at least two flows according to a sequence of a plurality of predetermined activation topologies, and

[0054] - wherein, for each of the at least two data flows, an indicator of sequential communication performance of the data flow and an indicator of concurrent communication performance of the data flow are each obtained by measurements performed under different activation topologies in a sequence of activation topologies.

[0055] The sequential activation flow according to the above activation topology corresponds to injecting traffic for the specific purpose of quickly performing robust clustering by quickly identifying the relative dependencies of a series of data flows, so as to allow the accurate virtual topology of the wireless bridge to be quickly provided to the CNC.

[0056] Different ways of optimizing the accuracy of the virtual topology while minimizing the amount of injected traffic are presented in the following examples.

[0057] In one example, the sequence of activation topologies can be optimized to maximize the variability of the communication performance of at least two flows between two consecutive activation topologies. For example, a given activation topology of a sequence can be iteratively inferred by machine learning based on previous activation topologies of the sequence and associated measured communication performance.

[0058] In another example, the sequence of activated topologies can be designed so that for any two consecutive activated topologies in the sequence of activated topologies, a fixed number of user terminals are activated in one of the two consecutive activated topologies and inactive in the other activated topology. This is equivalent to setting a fixed distance between two consecutive activated topologies, which is the number of user terminals activated in a single activated topology in the two consecutive activated topologies.

[0059] In one example, clustering at least two data flows into a cluster of dependent data flows may include:

[0060] - obtaining an autoencoder, the autoencoder comprising an encoder for mapping an input to a code according to a compression matrix, and a decoder for mapping the code to a reconstruction of the input according to a decompression matrix, the code being a representation of the input as a set of deep feature variables in a reduced feature space,

[0061] - using the obtained metrics as input to train the autoencoder,

[0062] -After training the autoencoder, we obtain the compression matrix, decompression matrix and set of deep feature variables,

[0063] - determining a clustering scheme based at least on the obtained compression matrix, and

[0064] - clustering the at least two data streams according to the clustering scheme by using the obtained set of deep feature variables as centroids of the clusters.

[0065] Autoencoders are a powerful tool that is preferred over more classical methods such as principal component analysis for automatically identifying patterns in input data without supervision.

[0066] This particular use of an autoencoder is particularly effective for determining whether groupings of data streams are dependent or independent, and, as a result, for determining clusters of dependent data streams.

[0067] Additional features, details and advantages will be shown in the following detailed description and in the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] [ Figure 1 ]

[0069] Figure 1 It is a prior art representation of a 5G system acting as a TSN bridge.

[0070] [ Figure 2 ]

[0071] Figure 2 It is a prior art representation of a 5G system as a TSN bridge in a TSN network.

[0072] [ Figure 3 ]

[0073] Figure 3 A TSN network according to one embodiment is shown.

[0074] [ Figure 4 ]

[0075] Figure 4 Shown is the prior art integration of a wireless bridge within a TSN network.

[0076] [ Figure 5 ]

[0077] Figure 5 A flow chart illustrating a software-implemented method for clustering input data and associated data transformations according to one embodiment.

[0078] [ Figure 6 ]

[0079] Figure 6 A principal component analysis for data preprocessing according to one embodiment is shown.

[0080] [ Figure 7 ]

[0081] Figure 7 An autoencoder for data processing according to one embodiment is shown.

[0082] [ Figure 8 ]

[0083] Figure 8 A TSN model of a wireless bridge according to one embodiment is shown.

[0084] [ Fig. 9 ]

[0085] Fig. 9 is a flow chart of a software-implemented method for determining the topology of a wireless bridge according to one embodiment.

[0086] [ Fig.10 ]

[0087] Fig.10 is a flow chart of a software-implemented method for determining the topology of a wireless bridge according to one embodiment.

[0088] [ Fig.11 ]

[0089] Fig.11 According to one embodiment, a method for operating Figure 5 , Fig. 9 and Fig.10 An example of a processing circuit for any software represented above. DETAILED DESCRIPTION

[0090] The present disclosure solves the problem of black-box identification of an equivalent TSN bridge model for a wireless system (e.g., a 5G system) suitable for integrating the wireless system into a TSN network as a logical TSN bridge.

[0091] The identification of this equivalent TSN model can be done based on application layer measurements from wireless systems (e.g. 5GS) and can use a deep learning framework.

[0092] It is believed that measurements at the application layer (which is layer 7 of the OSI model) are readily available at the core network components of the wireless bridge to be used to determine an equivalent TSN bridge model for the wireless system.

[0093] Application layer measurements can be related to various communication resources of a wireless system, which are not directly available in a black-box approach. Such a priori unknown communication resources can include, for example, spatial multiplexing capabilities, broadcast capabilities, and / or frequency reuse factors.

[0094] The present disclosure presents techniques and frameworks for black-box identification of an equivalent TSN bridge model based on these application layer measurements.

[0095] <Integration of 5G Systems in TSN Networks: Prior Art>

[0096] The present disclosure focuses on the transmission of packets with latency constraints. To achieve a guaranteed end-to-end latency for the transmission of packets between a talker and a listener, it is necessary to rely on a predetermined data path or route and the scheduling of the operation of network elements along the path.

[0097] In prior art approaches for the integration of a 5G system (5GS) in a TSN network, a conventional TSN bridge model (100) is used to represent the 5GS, and the conventional TSN bridge model (100) has a one-to-many topology including ports / gates as shown in Figure 1 where the bridge has one NW (network side) port (102) and two DS (device side) ports (104, 106). The internal structure of the 5GS involves a network side translator (NW-TT) and a device side translator (DS-TT). The network side translator (NW-TT) is an element that forms an interface between the 5G user plane (UP) of the 5GS in the core network and the TSN network, and the device side translator (DS-TT) forms an interface between the 5G UP and the device.

[0098] <TSN Network>

[0099] The general structure of the TSN network is shown in Figure 3 as follows.

[0100] Both the talker (200) and the listener (202) are represented as end stations (ES) of the TSN network. A series of bridges (204) are also presented.

[0101] The bridge (204) is a packet switching device of the TSN network. Packets arriving at the ingress port of the bridge can be routed to the egress port with bounded delay according to some parameters specific to each port pair represented by internal links. The parameters are typically: a dependent delay factor (i.e., which depends on the payload and can be regarded as the reciprocal of the instantaneous throughput of the link) and an independent delay (i.e., a fixed delay for each communication).

[0102] The TSN network also includes a centralized user configuration node (CUC) (206) configured to communicate with the end stations (200, 202) to receive their flow requirements.

[0103] The TSN network also includes a centralized network configuration node (CNC) (208). The CNC is capable of receiving egress port and ingress port identifications, service classes, and QoS indicators as minimum and maximum delays for each port pair from different TSN bridges (204).

[0104] The CNC (208) can also receive data related to the user configuration from the CUC (206) through the user / network interface. The data related to the user configuration can include the flow requirements of the end station. The data related to the user configuration can also include the TSN network topology related to the flow requirements of the end station. The data related to the user configuration can also include the capacity of the network link according to the topology.

[0105] The CNC (208), which has collected the topology and capacity of the network links and the requirements of the data flow as described above, defines the routing decision, i.e. the bridge selection, in other words, defines the data path for sending packets from one end station (200) to another end station (202) through the bridge (204). The data path is calculated by the CNC in a known manner so as to meet the flow requirements of the TSN flow.

[0106] For the purpose of satisfying the constancy of the flow requirements of the TSN flow, the CNC (208) also calculates a schedule of gate openings for making packets flow from one ES to another ES while guaranteeing their transmission delays. To this end, the CNC calculates the cumulative delay from the egress port to the egress port of two consecutive bridges on the calculated data path, where the egress port is defined at the output port of the bridge in the flow direction of the packets along the path (and the ingress port is the bridge input port).

[0107] The schedule calculated at the CNC is a table of time-gated windows according to which gates are opened and closed sequentially. In practice, for each bridge port, packets are buffered and transmitted sequentially when the corresponding gate is open. Therefore, time multiplexing is inherent to each bridge port, and concurrent transmissions between two different ports of a TSN bridge are possible.

[0108] <5G System>

[0109] This disclosure mainly focuses on the example of 5G systems, but its general principles can also be applied to other wireless systems with multiplexing capabilities in different dimensions such as time, frequency and space.

[0110] The deployment of 5G systems in factories involves the integration of 5G systems in time-sensitive networks. In the context of this integration, the 5G system (5GS) acts as at least one Ethernet bridge, which belongs to the data link layer or layer 2 of the OSI model integrated within the IEEE TSN network.

[0111] Figure 2 A general view of the internal structure of 5GS and several elements of a TSN network according to the fully centralized model of IEEE 802.1Qcc is given in . The main components of 5GC (5G Core Network) and 5G RAN (5G Radio Access Network) are given in Figure 2 It is represented in the form of nodes or functional entities.

[0112] In particular, a functional split is observed between the control plane (210) and the user plane (214).

[0113] The control plane involves PCF (Policy Charging Function) and TSN-AF (Time Sensitive Network Application Function), both of which are nodes (212) of 5GC.

[0114] The user plane involves a single UPF (User Plane Function) for routing packets from 5G-RAN to the network (NW) port. The user plane also involves 5G-RAN, and more specifically, gNBs (next generation node base stations) configured to communicate with UPF and handle UE (user equipment).

[0115] Figure 2 and Figure 3 The CUC (206) (centralized user configuration node) and CNC (208) (centralized network configuration node) presented in the FIG. Figure 2 The control element is connected to the 5GS control plane via TSN-AF.

[0116] Figure 4 A known integration of a 5GS acting as an Ethernet bridge in a TSN network according to 3GPP Technical Report TR 23.734 is shown. In this framework, the 5GS includes one or more ports on a single user plane function (UPF) (306) on the network side (NW-TT) (310), which act as user plane tunnels between the UPF and user equipment (UEs) (302) through a radio access network (304). The 5GS acting as an Ethernet bridge also includes one port per user equipment on the device side (DS-TT) (300). For each 5GS bridge of the TSN network, the port on the NW-TT side supports a connection to the TSN network, while the ports on the DS-TT side are each associated with a corresponding protocol data unit (PDU) session that provides a connection to the TSN network according to 3GPP Technical Specification TS23.501.

[0117] exist Figure 4 In the figure, it is shown that the logical bridge configuration information is transmitted to the TSN network via a trusted TSN application function (AF) (308). The configuration information is related to network status and performance parameters, which are received from the access management function (AMF) through the session management (SMF) in the unified data management (UDM) and transmitted to the AF via the network exposure function (NEF) using the N33 interface according to the 3GPP technical specification TS23.502, or directly from the policy charging function to the AF via the N5 interface if the AF is considered to be trusted.

[0118] The measurements considered in the present disclosure are end-to-end (E2E) measurements related to QoS KPIs at the application layer (i.e., the application layer or layer 7 of the OSI model). The QoS KPI may be E2E latency measured on an uplink or downlink, E2E throughput measured on an uplink or downlink, or E2E link reliability measured on an uplink / downlink, and the reference points for the E2E measurements may be, for example, DS-TT (device-side TSN converter) and NW-TT (network-side TSN converter).

[0119] In case of multiple deployments of TSN-capable UEs in the network, if the UE is sending / receiving a TSN-PDU session corresponding to a TSN data stream from a TSN-talker, the UE is activated. If the UE is not sending / receiving a PDU-session corresponding to a TSN stream from a TSN talker to a TSN listener, the UE is not activated.

[0120] The QoS KPI is measured in the 3GPP domain and on the N6 interface linking the User Plane Function (UPF) to the NW-TT on uplink and on the interface N60 for downlink transmission.

[0121] QoS KPI measurements necessarily involve data flows associated with an active (or activated) UE (hereinafter referred to as the 'reference UE' and denoted UE(0)). UE(0) is considered to be part of a deployment of N UEs handled by the 5GS, where the UE is denoted UE(i), i being a natural number in the range from 0 to Nl.

[0122] This paper proposes that for each QoS KPI measurement related to UE(0), the active UEs that are activated together with UE(0) during the measurement are recorded. Assume that other deployed UEs are inactive. Then for each QoS KPI measurement related to UE(0), the record can be, for example, a binary matrix [x1, x2, …, x N-1], where if the deployed UE(i) is activated, the binary variable x1=1, otherwise, the binary variable x1=0.

[0123] The previously defined binary matrix is ​​referred to in the following as measurement tag or activation topology tag or simply tag.The QoS KPI measurements are arranged as a tagged performance table, where the measurements of active UEs are tagged by the activation topology tag.

[0124] <Clustering>

[0125] Clustering is a key step aimed at learning grouping of dependent TSN data flows from QoS KPI measurements and from associated activation topology labels.

[0126] The virtual TSN bridge topology is then determined from the packets of the dependent TSN data streams.

[0127] Let us consider a pair of TSN data flows. A pair of TSN data flows is said to be "dependent" if the measured end-to-end QoS KPI of one of the pair of TSN data flows changes depending on whether the other TSN data flow of the pair of TSN data flows is activated at the same time.

[0128] More specifically, sequential communication performance and concurrent communication performance may be defined. For an activation topology in which two data flows in the pair of TSN data flows are activated together, the measured end-to-end QoS KPI of a given data flow in the pair of TSN data flows reflects concurrent communication performance. Conversely, for an activation topology in which two data flows in the pair of TSN data flows are activated one by one rather than together, the measured end-to-end QoS KPI of a given data flow reflects sequential communication performance.

[0129] For example, when the UE-related QoS KPI measurement differs according to whether x1=1 (concurrent communication performance) or x1=0 (sequential communication performance), the UE pair formed by UE(0) and UE(1) is dependent.

[0130] This principle applies more generally to any group of two or more data flows. Then for an activation topology in which no other data flow of the group is activated simultaneously, the sequential communication performance specifies the measured end-to-end QoS KPI for a given data flow. For an activation topology in which at least one other data flow of the group is activated simultaneously, the concurrent communication performance of a given data flow specifies the measured end-to-end QoS KPI for the given data flow.

[0131] The absolute value of the difference between the sequential communication performance and the concurrent communication performance of a given data flow may be calculated and referred to hereinafter as a "QoS KPI change". When a group of data flows exhibits a QoS KPI change that exceeds a predetermined threshold, then the group of data flows is said to be dependent. Otherwise, the group of data flows is said to be independent. The threshold may be preset to a non-zero value, which may be positive or negative. The sign and absolute value of the threshold are selected according to the application.

[0132] <clustering measure>

[0133] The following section presents possible options for QoS KPIs that can be used in the clustering step. QoS KPIs are described in Service Level Specifications (SLS) that are sent from the CUC to the TSN talkers and listeners and used by the CNC to calculate the schedule on the TSN network.

[0134] By way of example, the QoS KPI contemplated herein is E2E latency or delay. Examples of key QoS KPI cluster metrics corresponding to E2E latency include:

[0135] - E2E latency of the communication of TSN data streams between a DS-TT and a NW-TT (i.e., uplink TSN data stream transmission on the 5GS network),

[0136] - E2E latency of TSN data stream communication between one NW-TT and one DS-TT (i.e., uplink-downlink TSN data stream transmission on 5GS network),

[0137] -The maximum value of the E2E delay between uplink and downlink data stream transmission in the 5GS network,

[0138] - The E2E latency of the best-effort communication of TSN data streams between DS-TT and NW-TT (i.e., best-effort uplink TSN data stream transmission on 5GS),

[0139] - The E2E latency of the best-effort communication of TSN data streams between NW-TT and DS-TT (i.e., best-effort downlink TSN data stream transmission on 5GS),

[0140] -The maximum value of the best-effort E2E delay between uplink and downlink data stream transmissions in a 5GS network.

[0141] <Clustering Algorithm>

[0142] A clustering algorithm is used to learn to cluster UEs into groups of dependent QoS KPIs.The input to the clustering algorithm is the labeled performance table defined above, including the recorded QoS KPI measurements labeled and arranged by activating the topology label.

[0143] Discovery of performance tables can be accomplished in various ways.

[0144] For example, the data flows may be ordered (marked) and the UEs associated with the data flows may be activated sequentially, pair by pair.The performance of the data flows is measured for each activated topology and stored into a data table: this corresponds to a sequential discovery of pairs of data flows.

[0145] Alternatively, data flows may be marked and activated sequentially pair by pair in the deployed regions, and then the performance of the data flows may be measured and stored in a data table: this corresponds to dictionary-style discovery of data flows.

[0146] When a UE is activated for the purpose of measuring the performance of a data flow, the data flow may be associated with an active PDU session or a dummy PDU session.

[0147] Clustering algorithms belong to a general class of algorithms that are suitable for detecting features in performance tables in a similar way to how image segmentation algorithms detect regions in an image. These features show a collection of data flows with constant performance over multiple activation topologies, i.e., a collection of independent data flows.

[0148] According to the obtained set of independent data flows, a cluster of dependent data flows is constructed and outputted through a clustering algorithm.

[0149] Reference now Figure 5 , Figure 5 is a flow chart describing the general steps of a clustering algorithm in an exemplary embodiment.

[0150] These general steps include:

[0151] - data pre-processing (400) of the input data (406) (labeled performance tables),

[0152] - then, feature detection within the pre-processed data (408) combined with segmentation of the pre-processed data based on the detected features (402), and

[0153] - Finally, the segmented data (410) is post-processed (404) to obtain and output clustered data (414) (clusters formed by a collection of dependent data flows).

[0154] <Data Preprocessing>

[0155] The goal of data preprocessing is to optimize deep clustering performance by reordering the input data (i.e., performance table) relative to a specific activation topology order. This reordering allows to simplify the discovery of clusters that depend on the data flow.

[0156] The activation topology order can be determined by various criteria.

[0157] A standard example is to maximize the variability of the performance of a flow between two consecutive activation topologies. For example, a principal component analysis of the performance data of the flow can be performed. This allows finding a decentralized activation topology that maximizes the performance of the flow as an eigenvector of the covariance matrix of the performance data.

[0158] Another example of a criterion is to constrain the distance between two consecutive activated topologies. For example, a lexicographic order may be chosen in which the selected subset of flows is activated such that every two consecutive activated flow sets are disjoint and have a distance of k flows. Consider two consecutive activated topologies for a deployment of UE and The distance of k flows means that for exactly k values ​​of i from 1 to Nl, In a particular embodiment, a gray sorting may be selected in which the distance between two activated topologies is limited to one flow. In other words, the sorting activates (or deactivates) one flow at a time in a separate area of ​​the entire activated topology.

[0159] Another example is a combination or aggregation of at least two of the above orderings.

[0160] Figure 6 A principal component analysis of a set (406) of typical input data is shown, here represented as a scatter plot, where the principal components determine the directions of maximum dispersion of the streaming data set.

[0161] <Feature Detection>

[0162] Feature detection allows to perform actual clustering of the pre-processed dataset (re-ordered performance table) in order to learn some features shared by different streams for different activation topologies, i.e., similar levels of performance variability with respect to changes in activation topology.

[0163] A common technique to perform this type of clustering is to perform K-means clustering on the preprocessed dataset.

[0164] First, a subset of flows is selected arbitrarily or randomly. This means that each flow of the subset is selected arbitrarily or randomly. Furthermore, the number of flows forming the subset may also be selected arbitrarily or randomly.

[0165] Each selected flow is assigned to a different cluster as its cluster center.

[0166] Then, iteratively:

[0167] - adding to the cluster the flow that exhibits a change in its QoS performance measure closest to the center of the existing cluster center relative to the activation topology change, and

[0168] - The center of the cluster is updated to the average of the QoS performance measurements within the cluster.

[0169] Once all flows are assigned to clusters, K-means clustering converges automatically.

[0170] However, in practice, in the case of large input data sets, this standard clustering does not converge well. In particular, due to clustering initialization, this standard clustering may not provide a satisfactory cluster representation of the input data set.

[0171] To this end, it is proposed to use autoencoders for deep clustering purposes. Figure 7 Shown is a portion of a regenerative neural network trained to reconstruct a dataset.

[0172] An autoencoder is a pair of an encoder (500) and a decoder (502), such as Figure 7 As shown, the pair of encoder (500) and decoder (502) learns a deep representation of a dataset via a back-propagation algorithm that attempts to reconstruct the input data at the output of the encoder / decoder pair.

[0173] The encoder part of the autoencoder converts the following compression relationship into Figure 7 The pre-processed data (408) (x1, x2, x3, x4) is compressed into a reduced feature space (504) represented by cluster centers (z1, z2).

[0174] z=f α (Wx)

[0175]

[0176] x=(x1 x2 x3 x4) T

[0177] z=(z1,z2) T

[0178] The decoder part reconstructs the original data set from the feature space by the following reconstruction equation (506).

[0179] x=f β (Rz)

[0180]

[0181] x=(x1 x2 x3 x4) T

[0182] z=(z1,z2) T

[0183] Function f α and f βare nonlinear functions, they may be, for example, sigmoidal functions as defined below:

[0184]

[0185]

[0186] The autoencoder learns the parameters W for clustering the data and the parameters R for jointly reconstructing the data with the sigmoid curve slope a by performing the following process called back-propagation technique.

[0187] - Random values ​​are initially assigned to the entries of matrices W and R, and loose values ​​are assigned to the S-curve slopes, e.g. α = β = 0.1,

[0188] -Compression and reconstruction equations are used to find the deep feature variables (z1,z2),

[0189] - Update the entries of matrices W and R to minimize the reconstruction error (the reconstruction error is the sum of the input data x and the reconstructed data ), and in order to increase the S-curve slope of the encoder and decoder filters,

[0190] -Further iterations of finding deep feature variables and updating entries of matrices W and R are then performed until the reconstruction error is below a threshold.

[0191] Once the autoencoder parameters W, R, α, β are learned from the back-propagation technique, the encoder part of the autoencoder is used as a clustering scheme for the input data as follows:

[0192] -Compression matrix W and nonlinear function f α is used to define the clustering scheme, and

[0193] - The deep feature variables (z1, z2) are used as the centroids of the data flow dependent clusters to be constructed.

[0194] <Post-processing>

[0195] To further improve the clustering scheme, post-processing is performed.

[0196] Post-processing can rely on a K-means clustering scheme such as the one described previously, which is initialized by using the deep feature variables (z1, z2) output by the autoencoder as the centroids of the clusters that depend on the data flow. Other known clustering schemes can also be applied as long as the deep feature variables (z1, z2) output by the autoencoder are used as descriptive data for the clusters that depend on the data flow in the initialization stage.

[0197] For example, another option is to further improve the separation of clusters by calculating the probability of data being associated with a cluster and adjusting the clusters by maximizing the inter-cluster Kullback-Leibler divergence.

[0198] <Proposed TSN model for 5GS>

[0199] Reference now Figure 8 .

[0200] Once the grouping of UEs is formed and identified, a master bridge model (600) may be determined indicating whether the different UEs have competing access to time resources, or conversely, whether their concurrent activation does not impact data flow performance.

[0201] The topology of the main bridge model (600) is as follows.

[0202] The master bridge model includes as many NW ports (602) as the number of groups of identified UEs. The master bridge model also includes as many DS ports (604, 606) as the total number of UEs of all groups of identified UEs. The master bridge model also includes internal links between each NW port and all UEs belonging to the associated group.

[0203] Alternatively, the master bridge model may be determined as multiple model parts, one model part for each group of UEs, each model part having only one NW port and as many internal links and DS ports as the number of UEs in the associated group.

[0204] Any configuration of the merger of these model parts is possible as long as the configuration link between the NW port associated with the group and the DS port associated with the UE is maintained.

[0205] The capacity on the internal links of the main bridge model can be learned by measurement (i.e., sending data packets to any UE and measuring the capacity or delay). As a remark, the UEs of the same group should be learned independently, which involves that the capacity obtained for a given internal link is the capacity when the corresponding UE is the only activated UE in the group of UEs to which it belongs.

[0206] It can be assumed that the external links reaching any ingress port of the main bridge model have zero dependency and independent delays.

[0207] Prior art TSN bridge models typically include a single NW port. Therefore, for compatibility reasons, any model of a TSN bridge is also assumed to include a single NW port.

[0208] To meet this assumption, a Demultiplexing (Demux) bridge model (608) with a single NW port (610) and as many DS ports (612) as the number of packets of the identified UEs can optionally be further introduced. The Demultiplexing bridge also includes internal links with the NW port for each DS port. These internal links have zero independent and dependent delays, which corresponds to infinite capacity. The external links have the characteristics of the physical interface that links the 5GS CN to the TSN network, which is typically an Ethernet interface.

[0209] As a result, when considering the prior art model, the end-to-end delay for sending two identical packets from the upstream bridge B0 to the end station associated with the egress port of B is associated with 1 / C'l+l / C'2, while with the new model currently disclosed, the end-to-end delay is associated with l / min(C'l, C'2). Therefore, the multiplexing capacity of 5GS is well utilized and the end-to-end delay is reduced.

[0210] <Apply to uplink>

[0211] Although the above examples focus on the downlink, one of ordinary skill in the art can easily apply the same principles to the uplink, where the ingress port is a DS port and the egress port is a NW port.

[0212] Two different models are then constructed for the downlink and for the uplink.

[0213] In fact, the uplink radio access technology may be different from the downlink radio access technology.

[0214] This may result in different ways of grouping UEs and different capacity measurements for the internal links of the bridge.

[0215] Considering the uplink specifically, the master bridge model has as many ingress ports as the number of UEs handled by the gNBs of the radio bridge and as many egress ports as the number of packets for the UEs.

[0216] Optionally, the egress ports of the main bridge model are each connected to an ingress port of a multiplex (Mux) model having a single egress port.

[0217] <Algorithm and processing circuit>

[0218] Fig. 9 is a flow chart illustrating a general algorithm of a computer program, which may be stored on a storage medium and / or may be carried as a signal, for the purpose of executing a method for determining a topology of a time-sensitive network model of a wireless bridge of a communication system within a time-sensitive network for a downlink application.

[0219] The algorithm includes obtaining (700) a partitioning of a set of P user devices handled by a wireless bridge into N groups of user devices. Both N and P are natural numbers strictly greater than 1. This partitioning is clustered data (414) obtained as an output of post-processing (404) when the performance table (406) relates to a downlink application (i.e., a downlink data flow from a talker (200) to a listener (202)).

[0220] As already mentioned, the user devices are activated when transmission of at least one data stream is involved through the wireless bridge, and the data stream performance is related to a given data stream of the at least one data stream, and for any first subset of user devices belonging to the same user device group, the concurrent activation of all user devices of the first subset reduces the data stream performance at the wireless bridge relative to the activation of a single user device in the first subset, and for any second subset of user devices all belonging to different groups, the concurrent activation of all user devices of the second subset does not reduce the data stream performance at the wireless bridge relative to the activation of a single user device in the second subset.

[0221] The algorithm next includes constructing (702) a topology of a time-sensitive network model of the wireless bridge, the time-sensitive network model including at least a main bridge model, the main bridge model including N network-side ports, P device-side ports, and P internal links, the network-side ports being associated with corresponding groups of user devices processed by the wireless bridge, the device-side ports being associated with corresponding user devices, and the internal links connecting the device-side ports associated with the user devices of a given group and the network-side ports associated with the given group.

[0222] The algorithm finally includes providing the topology of the time sensitive network model to a centralized network configuration node within the communication system (704).

[0223] Fig.10 The invention is a flowchart showing a general algorithm of a computer program that can be stored on a storage medium and / or can be carried as a signal for the purpose of performing a method for determining the topology of a time-sensitive network model of a wireless bridge of a communication system within a time-sensitive network for an uplink application. This partitioning is clustered data (414) obtained as an output of post-processing (404) when the performance table (406) relates to an uplink application (i.e., an uplink data flow from a listener (202) to a talker (200)).

[0224] The algorithm includes partitioning (800) a set of P user devices into M groups of user devices such that uplink data flow performance is associated with an uplink data flow defined as data transmission from each activated user device to a wireless bridge, for any first subset of user devices belonging to the same user device group, concurrent activation of all user devices of the first subset reduces the uplink data flow performance at the wireless bridge relative to activation of a single user device in the first subset, and for any second subset of user devices all belonging to different groups, concurrent activation of all user devices of the second subset does not reduce the uplink data flow performance at the wireless bridge relative to activation of a single user device in the second subset.

[0225] The algorithm next includes constructing (802) a topology of an uplink master bridge model, which includes M network-side ports, P device-side ports, and P internal links, the network-side ports being associated with corresponding groups of user devices processed by the wireless bridge, the device-side ports being associated with corresponding user devices, and the internal links connecting the device-side ports associated with the user devices of a given group and the network-side ports associated with the given group.

[0226] The algorithm finally comprises providing (804) the topology of the uplink master bridge model, or (more generally) providing (804) the topology of the time sensitive network model of the wireless bridges including the uplink master bridge model, to a centralized network configuration node within the communication system.

[0227] Fig.11 Schematically shows a suitable Figure 5 , Fig. 9 and Fig.10 The processing circuitry comprises a memory (904) for storing a computer program comprising the one or more algorithms. The processing circuitry also comprises a processor (902) for accessing the memory and executing the computer program. The processing circuitry also comprises a communication interface (906) which can be controlled by the processor for transmitting at least one or more topologies of the time-sensitive network model to a centralized network configuration node within the communication system.

[0228] Of course, it does not matter whether the processor (902) uses a single core or multiple cores to run the computer program. Cloud computing technology can also be used to run the computer program across processors of multiple processing circuits.

[0229] <Reference List>

[0230] For any purpose, the following non-patent elements are cited:

[0231] -nplcit1: Lu Si and Li Ruisi. DAC: Deep Autoencoder-based Clustering, a General Deep Learning Framework of Representation Learning (DAC: Clustering based on deep autoencoders, a general deep learning framework for representation learning). arXiv preprint arXiv:2102.07472,2021;

[0232] -nplcit2: Tom M. Mitchell, “Machine learning”, McGraw Hill Education; and

[0233] -nplcit3: Asperti, A., & Trentin, M. (2020). Balancing reconstruction error and Kullback-Leibler divergence in Variational Autoencoders. IEEE Access, 8, 199440-199448.

Claims

1. A method for clustering data flows established via a wireless bridge, the wireless bridge being formed by a wireless communication system and being integrated in a time-sensitive network, - the wireless bridge handles multiple user terminals, -The time-sensitive network includes a core network side terminal, - the time-sensitive network further comprises a plurality of device-side end stations, the device-side end stations being attached to the user terminals, - at least two data flows are established in the time-sensitive network and forwarded by the wireless bridge, wherein one data flow is established between the core network side end station and a single device side end station, The method comprises the following steps: - obtaining, at a functional entity of a core network component of the wireless bridge, for each of the at least two data flows, an indicator of sequential communication performance of the data flow corresponding to a situation where the at least two data flows are sequentially established, - for each of the at least two data flows, obtaining an indicator of concurrent communication performance of the data flow corresponding to a situation where the at least two data flows are established simultaneously, - clustering the at least two data flows into a cluster of dependent data flows based on the obtained indicators, such that: o if the difference between the sequential communication performance and the concurrent communication performance is below a predetermined threshold, then the two given data flows are independent, otherwise, the two given data flows are dependent, oAny two data flows belonging to two different clusters are independent, and oAny two data flows in the same cluster are dependent.

2. The method according to claim 1, wherein: The communication performance of a data stream established between a pair of end stations is related to at least one element in a list, the list comprising: - a first value of an end-to-end delay for uplink communications between the pair of end stations, - a second value of the end-to-end delay for downlink communications between the pair of end stations, and - a maximum value between the first value and the second value.

3. The method according to claim 2, wherein: The first value and / or the second value are related to best effort communication between end stations.

4. A method according to any one of the preceding claims, wherein: A given user terminal attached to a device-side end station is active when receiving or transmitting a given data stream, and inactive otherwise: - The indicator of sequential communication performance of the first data stream of the at least two data streams comprises: o a first measure of the communication performance of the first data flow, and o a first activity indicator, the first activity indicator at least indicating that, at the time of the first measurement, a first user terminal attached to a device-side end station participating in receiving or sending the first data stream is active, while a second user terminal attached to a device-side end station participating in receiving or sending a second data stream is inactive, and - The indicators of the concurrent communication performance of the first data flow include: o a second measure of the communication performance of the first data flow, and o A second activity indicator, the second activity indicator at least indicating that at the time of the second measurement, the first user terminal and the second user terminal are both active.

5. The method according to claim 4, wherein: The activation index is determined from a binary matrix, the binary matrix including a list of binary variables, one binary variable corresponding to one device-side end station, the binary variable having a first value when a user terminal attached to the corresponding device-side end station is active, and the binary variable having a second value when the user terminal attached to the corresponding device-side end station is inactive.

6. The method according to claim 4 or 5, further comprising the following steps: - before obtaining an indicator of sequential communication performance and an indicator of concurrent communication performance for at least two flows, activating the at least two flows according to a sequence of a plurality of predetermined activation topologies, and -in, For each of the at least two data flows, an indicator of sequential communication performance of the data flow and an indicator of concurrent communication performance of the data flow are each obtained by measurements performed under different activation topologies in the sequence of activation topologies.

7. The method according to claim 6, wherein: The sequence of activated topologies is optimized to maximize the variability of the communication performance of the at least two flows between two consecutive activated topologies.

8. The method according to claim 6 or 7, wherein: For any two consecutive activation topologies of the sequence, a fixed number of user terminals are activated in one activation topology of the two consecutive activation topologies and are inactive in the other activation topology.

9. A method according to any one of the preceding claims, wherein: The step of clustering the at least two data flows into a data flow dependent cluster comprises the following steps: - obtaining an autoencoder, the autoencoder comprising an encoder for mapping an input to a code according to a compression matrix, and a decoder for mapping the code to a reconstruction of the input according to a decompression matrix, the code being a representation of the input as a set of deep feature variables in a reduced feature space, - using the obtained sequential communication performance indicator and concurrent communication performance indicator as input to train the autoencoder, - after training the autoencoder, obtaining the compression matrix, the decompression matrix and the set of deep feature variables, - determining a clustering scheme based at least on the obtained compression matrix, and - clustering the at least two data streams according to the clustering scheme by using the obtained set of the deep feature variables as the centroids of the clusters.

10. A method according to any one of the preceding claims, wherein: The clustered data streams all have the same direction, either uplink or downlink.

11. A communication system in a time-sensitive network, the communication system comprising a wireless bridge for processing a plurality of user terminals, the time-sensitive network comprising a core network side terminal station, the time-sensitive network further comprising a plurality of device side terminal stations, the device side terminal stations being attached to the user terminals, at least two data flows being established in the time-sensitive network, wherein a data flow is established between the core network side terminal station and a single device side terminal station, the communication system being configured to: - obtaining, at a functional entity of a core network component of the wireless bridge, for each of the at least two data flows, an indicator of sequential communication performance of the data flow corresponding to a case where the at least two data flows are sequentially established, and obtaining, for each of the at least two data flows, an indicator of concurrent communication performance of the data flow corresponding to a case where the at least two data flows are simultaneously established, - clustering the at least two data flows into clusters of dependent data flows based on the obtained indicators, such that: o if the difference between the sequential communication performance and the concurrent communication performance is below a predetermined threshold, then the two given data flows are independent, otherwise, the two given data flows are dependent, oAny two data flows belonging to two different clusters are independent, and oAny two data flows in the same cluster are dependent.

12. A computer program comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 10.

Images