Monitoring the operation of a communication network

Abstract

The invention relates to a method for monitoring the operation of a cellular radio communication network, characterized in that it comprises the following: - collecting (S2) for the first cell two initial time series of values ​​taken over time by two performance indicators measured for the first cell, - collecting (S2) for the second cell two further time series of values ​​taken over time by said two performance indicators measured for the second cell, - converting (S6) said two time series into first and second images (IMi; IMj) respectively, comprising pixels whose values ​​correspond to the values ​​of the two time series respectively, - based on the result of a comparison (S7) of the first image with the second image, identifying whether said first and second cells have, in operation, a common behavior or not. Fig. 3

Description

Title of the invention: Control of the operation of a communication network Scope of the invention

[0001] The field of the invention is that of cellular radio communication networks and more particularly the control of the operation of these networks, notably for the management of parameter settings and the monitoring of network conditions. Such control involves implementing a technique for identifying common behaviors across a set of radio cells, with the aim of grouping similar cells for common management.

[0002] The invention applies in particular to the treatment of problems affecting the radio cells of a set of radio cells considered in a radiocommunication network, such as resource optimization, network planning, anomaly detection, etc. Previous art

[0003] The temporal characteristics of a cellular radio communication network are metrics on cell performance measured at regular time intervals, such as, for example, the average number of users connected to the cell, the percentage of radio packet loss, the number of inter-cell transfers, etc. Such temporal characteristics are generally called "performance indicators" or KPIs (Key Performance Indicators). To monitor the operation of a cellular radio communication network, there are techniques for identifying whether the cells in a given set of cells within the network exhibit common behavior, and thus for grouping the radio cells in that set by behavior.To this end, for a performance indicator considered for each cell of a set of cells, the grouping of cells is carried out according to new values ​​calculated from the observed values ​​of the performance indicator, according to one of the following three processes: .

[0004] - based on statistics calculated on the values ​​observed at each level cell (average of values, standard deviation, etc.);

[0005] - depending on the distances between the respective performance indicators of the different different cells (Euclidean distance, DTW (in English "Dynamic Time Warping", etc.);

[0006] - depending on new characteristics extracted from the observed values ​​of the performance indicator, such new characteristics being for example seasonality, trend, etc.

[0007] The cells that are closest in terms of statistics, distance, or new characteristics are then grouped together. Generalization to several performance indicators is achieved by applying one of the three aforementioned methods to each performance indicator independently. The grouping of cells is then carried out based on the new values ​​calculated for all cells and all metrics. The drawback of these identification techniques lies in the fact that the new values ​​are generated for each of the indicators considered independently. Information on the joint values ​​of all performance indicators at the same observation points is thus lost.However, given the multitude of factors that can impact radio cell performance and the constant change in cell state, grouping based on a single characteristic or several independently taken characteristics cannot be optimal for identifying the behavior of each cell. Object and summary of the invention

[0008] One of the aims of the invention is to remedy at least one of the drawbacks of the aforementioned prior art by proposing a new technique for identifying the common behavior of cells in a set of cells, which is based, for each cell, on the joint consideration of several different performance indicators measured for each cell.

[0009] To this end, an object of the present invention relates to a method of controlling the operation of a cellular radiocommunication network by comparing, between at least the first and second cells of the network, the evolution over time of at least one performance indicator measured for the first, respectively second, cell.

[0010] Such a method is characterized in that it comprises the following:

[0011] - collect for the first cell at least two initial time series of values ​​taken over time by said at least one performance indicator and respectively at least one other performance indicator measured for the first cell,

[0012] - collect for the second cell at least two second time series of values ​​taken over time by said at least one performance indicator and respectively said at least one other performance indicator measured for the second cell,

[0013] - convert said at least two first and second time series, into res respectively at least one first and one second image, the first and second images comprising pixels whose values ​​correspond to the values ​​of at least two first and second time series respectively,

[0014] - depending on the result of a comparison of the first image with the second image, identify whether the said at least two first and second cells have, in operation, a common behavior or not.

[0015] Such a control method makes it possible to efficiently identify whether at least two cells in a cellular radio communication network are operating in the same or similar way, by jointly considering, for each cell, the evolution over time of at least two performance indicators measured for each of the two cells. Thus, it is possible to obtain a better characterization of the behavior of each of the two cells, with the aim of more efficient grouping of cells based on their behavior, particularly for the purposes of optimizing the resources of the cellular radio communication network, planning this network, detecting anomalies in this network, etc.

[0016] According to a particular embodiment, the identification that the at least two first and second cells have, in operation, a common behavior or not, is implemented according to the result of a comparison of the first image with the second image, of at least one first parameter associated with the first cell and at least one second parameter associated with the second cell.

[0017] Adding such a parameter for each of the two cells optimizes the characterization of the behavior of each cell. Such a parameter includes, for example, for each of the two cells, the radio transmission frequency, the antenna height, the azimuth, the tilt, the transmission power, the number of neighboring cells, etc.

[0018] According to another particular embodiment, prior to the comparison, the method comprises the following:

[0019] - to divide the first image into at least two areas, respectively the second image in at least two zones,

[0020] - extract at least one feature per area of ​​the first image, respectively at least one feature per area of ​​the second image,

[0021] - generate at least a first feature matrix from said at at least one feature extracted per area from the first image, respectively at least one second feature matrix from said at least one feature extracted per area from the second image,

[0022] - to divide at least the first matrix into at least two sub-matrices of character risk factors, respectively at least the second matrix in at least two feature sub-matrices,

[0023] - reduce the dimension of said at least two sub-matrices resulting from the partitioning of the first matrix, generating at least one first feature matrix of reduced size, respectively the dimension of said at least two sub-matrices resulting from the partitioning of the second matrix, generating at least one second feature matrix of reduced size,

[0024] - generate a first vector from said characteristics of at least the first reduced-size matrix, respectively a second vector from said characteristics of at least the second reduced matrix.

[0025] Such an embodiment makes it possible, by means of a processing of cutting the at least first and second images into at least two zones, then of conversion of each of these first and second images cut into zones into respectively a first data matrix characterizing the zones of the first image and a second data matrix characterizing the zones of the second image, and finally of generation respectively of a first reduced-size feature matrix from the first data matrix and a second reduced-size feature matrix from the second data matrix, to carry out an extrapolation of the features of the at least first and second cells in the form of vectors easily comparable because of a dimension much lower than that of the at least first and second images obtained initially.

[0026] According to another particular embodiment, the steps of dividing the first image, respectively the second image, into at least two zones, extracting at least one feature per zone of the first image, respectively the second image, generating at least one first feature matrix, respectively a second feature matrix, dividing said at least one first feature matrix, respectively said at least one second feature matrix, reducing, and generating a first vector, respectively a second vector, are implemented by a convolutional neural network which receives as input the first image, respectively the second image, and which delivers as output the first vector characterizing the first cell, respectively the second vector characterizing the second cell.

[0027] The use of such a convolutional neural network is particularly well-suited for jointly processing the two dimensions of the first image and the second image, respectively. Thanks to its different levels of convolutional layers, different areas of each of the first and second images can be progressively processed and assembled to obtain first and second data matrices of a size or dimension smaller than the size or dimension of the first and second images, respectively, and thus extract their features. Furthermore, such a neural network is capable of processing a large number of images, and moreover, high-dimensional images. Thus, the method for controlling the operation of the cellular radio communication network can be applied to more than two cells and to more than two performance indicators considered per cell.

[0028] According to another particular embodiment, the method further comprises the following:

[0029] - apply the first and second input vectors to a grouping module of vectors,

[0030] - generate a group of vectors as output from the vector grouping module containing the first and second vectors if the first and second vectors meet a grouping criterion.

[0031] Such an embodiment allows us to deduce from the grouping of the first and second vectors that the first and second cells have a common behavior during operation. In the case where the first and second vectors are not grouped, such an embodiment allows us to deduce that the first and second cells have different behavior during operation.

[0032] According to another particular embodiment, the first vector, respectively the second vector, is applied as input to the vector grouping module, together with the first parameter associated with the first cell, respectively together with a second parameter associated with the second cell.

[0033] Such an embodiment has the advantage of obtaining a more precise grouping of cells because it is based, for each cell, both on the temporal evolution of several different performance indicators and on one or more cell parameters. Such grouping is also more efficient in terms of homogeneity of cell behavior.

[0034] The various modes or embodiments mentioned above can be added independently or in combination with each other, to the method of controlling the operation of a cellular radiocommunication network as defined above.

[0035] The invention also relates to a device for monitoring the operation of a cellular radio communication network by comparing, between at least the first and second cells of the network, the evolution over time of at least one performance indicator measured for the first, respectively second, cell.

[0036] Such a device is characterized in that it is configured to:

[0037] - collect for the first cell at least two initial time series of values ​​taken over time by said at least one performance indicator and respectively at least one other performance indicator measured for the first cell,

[0038] - collect for the second cell at least two second time series of values ​​taken over time by said at least one performance indicator and respectively said at least one other performance indicator measured for the second cell,

[0039] - convert said at least two first and second time series, into res respectively at least one first and one second image, the first and second images comprising pixels whose values ​​correspond to the values ​​of at least two first and second time series respectively,

[0040] - depending on the result of a comparison of the first image with the second image, identify whether the said at least two first and second cells have, in operation, a common behavior or not.

[0041] The invention also relates to a computer program comprising instructions for implementing the control method according to the invention, according to any one of the particular embodiments described above, when said program is executed by a processor.

[0042] Such instructions can be stored permanently in a non-transient memory medium of the content receiving device implementing the control method according to the invention.

[0043] This program may use any programming language, and be in the form of source code, object code, or code intermediate between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0044] The invention also relates to a recording medium or information medium readable by a computer, and comprising instructions for a computer program as mentioned above.

[0045] The recording medium can be any entity or device capable of storing the program. For example, the medium can include a storage means, such as a ROM, for example a CD-ROM or a microelectronic circuit ROM, or a magnetic recording means, for example a mobile medium, a hard drive or an SSD.

[0046] On the other hand, the recording medium can be a transmissible medium such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means, so that the computer program it contains is executable remotely. The program according to the invention can, in particular, be uploaded to a network, for example, an Internet-type network.

[0047] Alternatively, the recording medium may be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the aforementioned control method.

[0048] According to one embodiment, the present technique is implemented using software and / or hardware components. In this context, the term "device" or "module" may refer in this document to a software component, a hardware component, or a set of hardware and software components. Brief description of the drawings

[0049] Other features and advantages will become apparent upon reading particular embodiments of the invention, given by way of illustrative and non-limiting examples, and the accompanying drawings, among which:

[0050] Figure 1 represents an architecture in which the method for controlling the operation of a cellular radiocommunication network is implemented, according to a particular embodiment of the invention.

[0051] Figure [Fig. 2] represents a device for controlling the operation of a cellular radio communication network, according to a particular embodiment of the invention, as implemented in the architecture of Figure [1].

[0052] Figure 3 represents the main actions implemented in the method of controlling the operation of a cellular radiocommunication network, according to a particular embodiment of the invention, as implemented in the architecture of Figure 1.

[0053] Figure 4 represents the main actions implemented during one of the steps of the process for controlling the operation of a cellular radio communication network illustrated in Figure 3.

[0054] Figure [5] represents a data matrix obtained during the implementation of the method for controlling the operation of a cellular radio communication network illustrated in Figure [3],

[0055] [Fig.6] shows details of one of the sub-steps of the step illustrated in [Fig.4]

[0056] Figure 7 shows details of another step in the process of controlling the function. operation of a cellular radio communication network illustrated in [Fig.3],

[0057] Figure 8 represents an embodiment of a neural network used in the method of controlling the operation of a cellular radio communication network illustrated in Figure 3.

[0058] Fig. 9A represents an example of an IMT image as input to the neural network illustrated in Fig. 8.

[0059] Fig. 9B represents an example of an IM'T image output from the neural network illustrated in Fig. 8.

[0060] Fig. 10 represents a diagram of the evolution of the values ​​of the same KPI extracted from the IMT and IM'T images illustrated respectively in Figures 9A and 9B.

[0061] Detailed description of an embodiment of the invention

[0062] Figure 1 represents an architecture in which a method for controlling the operation of a cellular RC radio communication network is implemented, according to an embodiment of the invention. Such an RC network is, for example, of the 3G type. 4G, 5G, etc.

[0063] Such an architecture comprises:

[0064] - an EC set of at least two radio cells CL1, CL2 among N radio cells CL1 , CL2, ..CL;, ..CLj, ..., CLn (N>2) of the RC network, for which we want to check whether the cells of this set have a homogeneous behavior or not;

[0065] - a DCF device for monitoring the operation of said at least two cells radio CL;, CLj, the DCF device being configured to deliver one or more groups GCi, GC2, ... of radio cells, a group comprising radio cells considered to have homogeneous operation over a given time interval, for example a day, one or more hours, a week, etc.

[0066] The DCF device is for example a server, a platform, etc., the DCF device being installed in the RC network or in another communication network not shown.

[0067] In the embodiment shown, the DCF device comprises:

[0068] - a data collection CI module that is configured to collect:

[0069] — for cell CLI, a number p (p>l) of time series In, I2k.... Ipi of each n values ​​associated respectively with p performance indicators or KPIs as measured for cell CLI, and possibly at least one static parameter PSki from a number Q of static parameters PSn, PS2kPSki, ..., PSqi, (l <kl<Q),

[0070] — for cell CL2, a number p (p>l) of time series I12, I22, Ip2 of each n values ​​associated respectively with p performance indicators or KPIs as measured for cell CL2, and possibly at least one static parameter PSk2 from a number Q of static parameters PS[2, PS22 PSk2, ..., PSQ2, (l <kl<Q),

[0071] -...,

[0072] — for cell CL;, a number p (p>l) of time series In, I2i, Ipi of each n values ​​associated respectively with p performance indicators or KPIs as measured for cell CL; and possibly at least one static parameter PSki from a number Q of static parameters PSH, PS2i.... PSki, ..., PSQi, ( 1 <ki<Q),

[0073] -...,

[0074] — for cell CLj, a number p (p>1) of time series 1^, I2j>, Ipj of each n values ​​associated respectively with p performance indicators or KPIs as measured for cell CLj, and possibly at least one static parameter PSkj from a number Q of static parameters PSy, PS2j>PSkj, ..., PS^, ( 1 <kj<Q),

[0075] -...,

[0076] — for the CLN cell, a number p (p>1) of time series I[N, I2N>.... IpN of each n values ​​are associated respectively with p performance indicators or KPIs such as those measured for the CLN cell, and possibly at least one static parameter PSkN from a number Q of static parameters PSiN, PS2N> PS]^, • • - , PS™, (l <kN<Q).

[0077] Examples of such KPIs are the average number of users connected to the cell in question, the percentage of radio packet loss, the number of inter-cell transfers, etc.

[0078] The static parameters PSk[, PSk2, ..., PSki, ..., PSkj, ..., PSkj, ..., PSkN are of the same type, for example the radio transmission frequency or the height of the antenna or the azimuth or the tilt or the transmission power or the number of neighboring cells, or others, etc.

[0079] The DCF device further comprises:

[0080] - a CN data processing module configured to convert:

[0081] — the number p of time series In, I2k, Ipi in an image IMi comprising pxn pixels whose values ​​correspond to the n values ​​of each of the p time series In, Li, Ipi,

[0082] — the number p of time series I[2,122>.... Ip2 in an image IM2 comprising pxn pixels whose values ​​correspond to the n values ​​of each of the p time series 112, 122,.... Ip2,

[0083] -...,

[0084] — the number p of time series IH, I2i.... Ipi in an image IM; comprising pxn pixels whose values ​​correspond to the n values ​​of each of the p time series Ili, l-2i, .... Ipi,

[0085] -...,

[0086] — the number p of time series ly, I2j> , Ipj in an image IMj comprising pxn pixels whose values ​​correspond to the n values ​​of each of the p time series Ilj, l'i Ipj,

[0087] - ...,

[0088] — the number p of time series I1N, I2N, IpN in an image IMN comprising pxn pixels whose values ​​correspond to the n values ​​of each of the p time series I1N, I2N, IpN-

[0089] In a preferred embodiment, the CN data processing module is a convolutional neural network. Alternatively, the CN module is an algorithmic computer.

[0090] The DCF device further includes a CM module for comparing images according to a similarity criterion, enabling:

[0091] - to group the IMi to IMN images into one or more GCi, GC image groups 2, ..., each group comprising images whose content has been deemed similar according to the aforementioned similarity criterion,

[0092] - and to deduce by analogy, among the CLi to CLN cells, those which, in function operation, have a homogeneous behavior, and those that do not have a homogeneous behavior.

[0093] In the example shown:

[0094] - the GCi group comprises two cells, CLi and CL5, having been determined as having, in operation, a homogeneous behavior,

[0095] - the GC2 group comprises CL3, CL4, ..., CL cells; having been determined as having, in operation, a homogeneous behavior,

[0096]

[0097] Although in [Fig. 1] the CI, CN and CM modules are integrated into the DCF device for controlling the operation of said RC network, as an alternative, all or part of these modules could be external and controlled by the DCF device as a control server for these modules.

[0098] We will now describe, with reference to [Fig.2], the simplified structure of the DCF control device.

[0099] The DCF device comprises:

[0100] - a COM communication interface configured to communicate, via the network RC of [Fig. 1], with CLi to CLN cells,

[0101] - the CI module for collecting KPIs per cell, and possibly at least one static parameter PS relating to the cell in question,

[0102] - the CN module for converting KPIs into IMi to IMN images,

[0103] - the CM module for comparing IMi to IMN images, to deduce a or several groups of GCi, GC2, ... cells, as defined above.

[0104] In the example shown, the CI, CN and CM modules are installed in the DCF device. Alternatively, all or part of these modules can be made accessible by the DCF device, via the RC network or any other wireless or non-wired network, depending on the implementation envisaged, for example if the DCF device does not have the necessary hardware and software resources to store, for each of the cells, the KPIs, the static PS parameter(s), the IMi to IMN images, the GC i, GC2, ...group(s) of generated images.

[0105] At initialization, the code instructions of the computer program PG are, for example, loaded into RAM (not shown) before being executed by the PROC processor. The PROC processor of the UTR processing unit implements, in particular, the following actions, within the framework of the RC network control process described below, according to the instructions of the computer program PG:

[0106] - collect, for a cell CL;, at least two time series STa;, STb; of each n values ​​taken over time by respectively two per- indicators formance, among IH to Ipi, measured for the CL cell;

[0107] - collect, for at least one other CLj cell, at least two time series STaj, STbj of each n values, taken over time by respectively two performance indicators, from ly to Ipj, measured for cell CLj,

[0108] - convert said at least two time series STa;, STb;, into an image IM; of 2xn pixels whose values ​​correspond to the respective values ​​of at least two time series STa;, STb;,

[0109] - convert said at least two time series STa, STbj, into an image IMj of 2xn pixels whose values ​​correspond to the respective values ​​of at least two time series STaj, STbj,

[0110] - depending on the result of a comparison of image IM with image IMj, identify whether the CL; and CLj cells have, in operation, a common behavior or not.

[0111] We now describe, in relation to [Fig.3], together with Figures 1 and 2, the process of controlling the operation of the RC cellular radiocommunication network, according to a particular embodiment of the invention.

[0112] Such a method is applied to the aforementioned set of EC cells, such a set being arranged in a given perimeter, such as a city, several cities, a country, etc.

[0113] During an SI step, communication is established between the set of EC cells and the DCF device via its COM communication interface.

[0114] During an S2 step, the CI collection module of the [Fig.2] collection:

[0115] - for cell CLI, a number p (p>l) of time series In, UnIpi of each n values ​​are associated respectively with p performance indicators or KPIs as measured for the CLI cell,

[0116] - for cell CL2, a number p (p>l) of time series I12, I22 ..... Ip2 of each n values ​​are associated respectively with p performance indicators or KPIs as measured for cell CL2,

[0117] -...,

[0118] - for cell CL;, a number p (p> 1 ) of time series IH, I2i , Ipi of each n values ​​associated respectively with p performance indicators or KPIs as measured for cell CL;

[0119] -...,

[0120] - for cell CLj, a number p (p>l) of time series ly, I2j>, Ipj of each n values ​​associated respectively with p performance indicators or KPIs as measured for cell CLj,

[0121] -...,

[0122] - for the CLN cell, a number p (p>l) of LN time series, I2N>.... IpN of each n values ​​are associated respectively with p performance indicators or KPIs as measured for the CLN cell.

[0123] Each of the aforementioned time series represents n values ​​of a network indicator that are measured at different times. In other words, for example, cell CL;, a time series Ivi (l <v<p) est un vecteur d’observations xvii, xvi2, xvin, indexé par le temps. L’intervalle de temps qui sépare deux observations correspond à la granularité du KPI et peut être quelconque : une journée, une heure, 15 min, etc.

[0124] According to the invention, at least two time series of two different indicators are obtained per cell. When a time series of a single indicator is obtained per cell, it is called a univariate time series. Since, according to the invention, at least two time series are obtained per cell, these are called multivariate time series, which group several univariate series per cell. Obtaining at least two time series per cell allows for a better characterization of the cell whose operation one wishes to monitor. Thus, according to the invention, a number p of univariate time series are measured on each of the N cells C1 to CLN, with pxN univariate time series each corresponding to a vector of n observations.

[0125] During an S3 step, the CI collection module of [Fig.2] can also collect:

[0126] - for the cell CLI, at least one static parameter PSki among a number Q of pa Static ramometers PSn, PS21,..., PSki, ..., PSqi, (l <kl<Q),

[0127] - for cell CL2, at least one static parameter PSk2 among a number Q of pa Static ramometers PS[2, PS22 PSk2, ..., PSQ2, (l <k2<Q),

[0128] -...,

[0129] - for the cell CL;, at least one static parameter PSki among a number Q of pa Static rammeters PSn, PS2i PSki, ..., PSQi, (l <ki<Q),

[0130] -...,

[0131] - for cell CLj, at least one static parameter PSkj from a number Q of static parameters PSy, PS2j.....PSkj, ..., PSQj, (l <kj<Q),

[0132] -...,

[0133] - for the CLN cell, at least one static parameter PSkx among a number Q of pa Static ramometers PS1N, PS2N,... PSkN, ..., PSQN, (l <kN<Q).

[0134] Since this step S3 is optional, it is represented by a dotted line on [Fig.3].

[0135] During an S4 step, the KPIs obtained in step S2 may undergo processing. Such an S4 step is optional. For this reason, it is represented by a dashed line in [Fig. 3]. Such processing is implemented, for example, when certain values ​​in the obtained time series are missing, which can occur when maintenance is being carried out on one of the cells of the EC assembly, when a malfunction has occurred in the RC network, etc. To avoid any problems with Consistency between the values ​​of the obtained time series is ensured, as a missing value can be considered zero, for example, across all p KPIs and N cells. Only the values ​​of the time series that are completed to a certain percentage are retained. Thus, step S4 involves filtering the values ​​of the p time series obtained per cell, as well as selecting certain cells from among the N cells. Such filtering is implemented according to, for example, the following algorithm:

[0136] 1. For each KPI, Iu, (l <u<p) calculer le pourcentage de valeurs manquantes (values ​​not specified or equal to Null), i.e., vmu

[0137] If vmu > 30% then

[0138] Move Z„

[0139] Otherwise:

[0140] If vm, > 0% then

[0141] Identify the cells that have at least one missing value

[0142] For a CL cell; having at least one missing value, vmu linear interpolation of the missing value from the previous value and the next value

[0143] Otherwise

[0144] Consider the values ​​of the KPI in question as they are

[0145] End of "If" loop

[0146] Fsi

[0147] 2. For each cell, CL;, (l <i<N) vérifier que l’ensemble des KPIs retenus à Step 1 is well documented.

[0148] If yes then

[0149] Keep the CL cell;

[0150] Otherwise

[0151] Clear the CL cell;

[0152] Following the implementation of this algorithm, p'xN' univariate time series are obtained, each corresponding to a vector of n observations, p' (p'< p) is the new number of KPIs kept per cell and N' (N' < N) is the new number of cells kept.

[0153] During a step S5, a selection of some of the p' KPIs obtained in step S4 may be implemented. Such a selection is optional. It is a selection step of at least two KPIs, from among the p' KPIs, based on the correlation between them. From a set of highly correlated KPIs, i.e., different KPIs representing the same characteristic, only one KPI is retained for cell grouping. This avoids any redundancy of information that could bias the results obtained at the end of the method for monitoring the operation of the cellular radio communication network according to the invention. Such a selection S5 is implemented by a KPI selection algorithm, which includes, for example, the following steps:

[0154] - calculate a correlation matrix between the p' KPIs considering the N' cells,

[0155] - identify one or more groups of KPIs that are strongly correlated with each other (correlation (> 90% for example),

[0156] - for each identified group, keep only one KPI (random choice).

[0157] At the end of the S5 selection step, p”xN' univariate time series are delivered, each corresponding to a vector of n observations. M” (M” < M') is the new number of KPIs retained.

[0158] During a step S6, in the case where steps S4 and S5 are not performed, the pxN univariate time series are converted into N images respectively by the CN conversion module. Each of the N images comprises pxn pixels whose values ​​correspond to the respective values ​​of the p univariate time series associated respectively with the p KPIs.

[0159] Alternatively, during step S6, in the case where steps S4 and S5 are performed, the p”xN’ univariate time series are converted into N’ images respectively by the CN conversion module. Each of the N’ images comprises p”xn pixels whose values ​​correspond to the respective values ​​of the p” univariate time series associated respectively with the p” KPIs.

[0160] During an S7 step, the N or N' images are compared with each other by the CM comparison module. At the end of this comparison, one or more groups of GCi, GC2, ... cells are generated, as defined above, each of these groups containing cells having, in operation, a homogeneous behavior.

[0161] During step S7, in order to optimize cell grouping, the comparison may also take into account:

[0162] - static parameters PSn, PS21,..., PSk[, ...,PSQi to PSiN, PS2N,... PSrn, ...,PSqn collected at step S3, in the event that steps S4 and S5 are not implemented,

[0163] - static parameters PSn, PS21,..., PSk[, ...,PSQi to PSiN, PS2N'PSkN-, ...,PSQN' associated respectively with the N' cells selected in S4.

[0164] We will now describe, with reference to [Fig.4], an embodiment of the aforementioned conversion step S6.

[0165] In the example shown, and for the sake of simplifying the notation, steps S4 and S5 are not considered to be implemented. However, the conversion that will be described below applies similarly to p” time series I ii, la,.... IPi of each n values.

[0166] For a cell considered among N, for example cell CL;, the step S6 of converting cell CL; into an image IM; comprises, in S60, a modeling of the number p of time series K, F,..... Ipi of each n values ​​associated respectively with p performance indicators or KPIs of cell CL;, by a matrix MAT; two-dimensional pxn as illustrated in [Fig.5].

[0167] On [Fig. 5]:

[0168] - for the same KPI, for example Ivi, xvii, xvi2, ..xviw, ..xvin (l <w<n) représentent the n values ​​of this KPI measured at different time intervals,

[0169] - for the same time step, for example t2, the values ​​xH2, x2i2, ..., xvi2, ..., xpi2 are the respective values ​​of the p KPIs.

[0170] Such a matrix representation allows for the joint consideration of all p KPIs, unlike existing methods that only consider the row dimension. Adding information on the value of all KPIs at different time steps ti to tN allows for a better characterization of the behavior of each cell and therefore a more efficient grouping of cells based on their behavior.

[0171] The S6 conversion continues with, in S61, a scaling of the values ​​of these p KPIs.

[0172] Since the n values ​​of each of the p KPIs have different amplitudes, scaling of these n values, per KPI and per cell, is necessary for conversion to an image. The goal is to reduce all the values ​​to the same interval [0, 1] following, for example, the algorithm below:

[0173] For the MAT matrix; of the CL cell KPIs;:

[0174] For each line, for example the line of Li values, perform the scaling as follows:

[0175] 1. Scaling the values ​​of T . = [x,4i, x„;n, ... x, ; J according to the method of VI L vil* VIjl.' vinJ Robust scaling (in English, "RobustScaler"):

[0176] ■ _ xviw-median(IJ xviw " C95-Ç5

[0177] C5 being the 5th percentile and C95 the 95th percentile.

[0178] 2. Scaling the values ​​of 1^ = [xvil, X^7 ... X^ ] according to the method of MinMax scaling (in English "MinMaxScaler"):

[0179] x" = 'viw maX(Q-fmn(Q

[0180] Following this algorithm, the values ​​of all KPIs across all cells are scaled. Other known scaling methods exist and can be used. However, RobustScaler was chosen for its robustness against outliers, while MinMaxScaler was chosen for its control over the final range of values.

[0181] The S6 conversion continues with, in S62, a processing of the MAT matrix; in order to extract one or more features of the CL cell;. Such processing is an image processing according to the invention.

[0182] For this purpose, as shown in [Fig.6], the values ​​xHi to xpiN of the matrix MAT; are considered as pixel values ​​of a single-channel IM; image, i.e. an image which is not in colour and which has only one value per pixel.

[0183] The S62 processing step comprises the following substeps:

[0184] - in S620, cut the IM image into at least two zones Zn, Z2i which do not don't want to,

[0185] - in S621, extract at least one fH feature from the Zn zone and at least one ca characteristic f2i of zone Z2i,

[0186] - in S622, generate at least one MATDn matrix containing at least the following characteristics fn, f2i risks

[0187] - in S623, divide the MATDn matrix into at least two submatrices of character ristiques SMATDioi, SMATDni,

[0188] - in S624, reduce the dimension of said at least two submatrices, generating a reduced-size feature matrix, MATNu,

[0189] - in S625, generate a vector Y; from said characteristics of at least the reduced matrix MATNu, and where appropriate, from features of other reduced matrices corresponding respectively to other types of image filtering IM;, said vector Y; comprising features Du, D2i,..., DRi characterizing the cell CL;.

[0190] In [Fig.6], only two areas have been shown for clarity. of the figure. It should be noted that the areas obtained after the S620 segmentation cover the entire IM image. In the example shown, all areas are the same size. Depending on the intended implementation, the size chosen for these areas may, of course, differ.

[0191] Each of the zones Zn, Z2i contains both information on the temporal variation of at least one KPI and information on the values ​​of several KPIs at the same time. Thus, in the example shown:

[0192] - the Zn zone contains information on the temporal variation between ti and t2 of the two KPIs IH and I2i and information on the respective values ​​of IH and I2i at t1 and t2,

[0193] - the Z2i zone contains both information on the temporal variation between t3 and t4 of the two KPIs IH, I2i, and information on the respective values ​​of In, I2i at t3 and t4.

[0194] Step S620 comprises a horizontal scan of the image zone by zone, for example in the direction of arrow Fl, and a vertical scan of the image zone by zone, for example in the direction of arrows F2 and F3. Such a scan is implemented by a filtering algorithm. The other steps S621 to S625 are implemented by an algorithm for extracting data from the vector Y from the reduced-size MATNu matrix. Preferably, according to the invention and as shown in [Fig. 6], these steps are implemented by a convolutional neural network. RNC. Such a neural network first learns to implement steps S620 to S625 on a plurality of images. Then, for each image IMi to IMN, it segments the image into areas which are then progressively processed and assembled by different neural layers of the RNC network in order to extract the features of that image.

[0195] In the embodiment shown in [Fig.6], the RNC network includes, for example, at least one CNV convolution layer followed by a CMC pooling layer.

[0196] The CNV convolution layer takes the IM image as input and applies the aforementioned S620 zone segmentation. The CMC layer applies a convolution matrix, called a filter, to each zone. The result of applying the convolution matrix to a zone, for example Z2i, is a new zone ZD2i of the same size as the original Z2i and containing a number of data points corresponding to the number of pixels in the original Z2i. The result of applying the convolution matrix to all zones is therefore a MATDh matrix of the same size as the input IM image. In order to extract several features from the IM image, several different filters, i.e., with different parameters, can be applied to this image; four filters in the example shown. By way of non-limiting examples, one filter is dedicated to filtering the image's edge, another to filtering the image's texture, and so on.At output, the CNC layer produces as many MATDH, MATD2i, MATD3i, and MATD4i matrices as there are filters applied to the IM image; and the size of each of these matrices is equal to the size of the IM image. For clarity in the drawings, only MATDn and MATD2i are referenced in [Fig. 6].

[0197] The CMC pooling layer functions to reduce the dimensionality of each of the data matrices MATDn, MATD2i, MATD3i, and MATD4i produced by the CNV convolution layer, thereby reducing the number of variables. To this end, starting from a data matrix, for example, the MATDn matrix, the CMC layer divides this matrix into at least two submatrices, SMATDioi and SMATDiH. From these submatrices, the CMC layer calculates a new matrix, MATNn, in which a data point D is represented, as an example in [Fig. 6], said data point D having a value extracted from the submatrix SMATDiOi. Such a value corresponds, for example, to the maximum value of the data in an area of ​​the SMATDioi submatrix. Alternatively, such a value could correspond to an average of the data values ​​in the area of ​​the SMATDioi submatrix, to a variance of these values, etc.Other areas of the SMATDioi submatrix, as well as the SMATDi submatrix, are processed by the CMC layer to complete the values ​​of the new MATNü matrix. Other matrices, MATN2i, MATN3i, and MATN4i, are calculated by the CMC layer in the same way as the MATNü matrix. For clarity... of the drawings, only MATNH and MATN2i are referenced on [Fig.6]. Following the activation of this CMC layer, a plurality of data Dn, D2i, ..., DRi are produced constituting the values ​​of the vector Y;.

[0198] In order to optimize the accuracy of the data Dn, D2i, ..., DRi, the RNC neural network can include a succession of CNV layers (extracting features from the results of the previous layer) and CMC layers (reducing the dimensionality of the outputs of the previous layer). The feature vector Y is delivered as the output of this succession of layers.

[0199] The S6 conversion is implemented iteratively for the <i<n. ainsi, pour au moins une autre cellule considérée parmi n, par exemple la cj, l’étape s6 de conversion clj en image imj comprend, s60, modélisation du nombre p séries temporelles 1^, i2j,.... ipj chacune n valeurs associées respectivement à indicateurs performance ou kpis cl;, matrice matj. en s61, mise l’échelle des ces est œuvre. s62, mis œuvre un traitement matj vue d’extraire plusieurs caractéristiques clj. l’étape s62 comprend les sous-étapes suivantes :

[0200] - in S620, scan the IMj image into several zones Zy, Z2j, ... covering the entire image IMj and not overlapping,

[0201] - in S621, extract at least one fy feature from the Zy region and at least one ca characteristic f2j of zone Z2j,

[0202] - in S622, generate at least one MATDij matrix containing at least the following characteristics ristiques fij, f2j,

[0203] - in S623, divide the MATDij matrix into at least two sub-matrices of character SMATDioj, SMATDnj risks

[0204] - in S624, reduce the dimension of said at least two submatrices, generating a reduced-size feature matrix, MATNij,

[0205] - in S625, generate a vector Yj from said characteristics of at least the reduced size MATNij matrix, and where appropriate, from features of other reduced matrices corresponding respectively to other types of filtering of the IMj image, said vector Yj comprising features Dy, D2j,..., DRj characterizing the CLj cell.

[0206] Steps S620 to S625 are iterated for all N images. At the end of the S6 conversion, N vectors Yi to YN are obtained, characterizing respectively the N cells CLi to CLN.

[0207] We will now describe, with reference to [Fig.7], a mode of implementation of the aforementioned step S7 of grouping the N or N' images.

[0208] Advantageously, according to the invention, the grouping is not based directly on the IMi to IMN images, respectively IMi to IMN-, but on the characteristic vectors Yi to YN, respectively Yi to YN-, as obtained at the end of the aforementioned S6 conversion step, such vectors being of much lower dimension than the IMi to IMN images, respectively IMi to IMN,

[0209] Optionally, in order to refine the grouping step, the grouping can also take into account at least one static parameter PSkh ..., PSKX, respectively PSk[, ..., PSkn-, associated respectively with the N images, respectively N' images.

[0210] The CM module (Figures 1 and 2), which implements step S7, is configured, in the embodiment of [Fig. 7], to generate, from the vectors Yi to YN or YrPSi to Yn-PSn, respectively, the group(s) GCi, GC2, ..., GCN of radio cells, or from the vectors Yi to YN- or YrPSi to YN-PSN, respectively, the group(s) GCi, GC2, ..., GCN' of radio cells. The CM module uses, for example, a clustering algorithm, such as the K-Means algorithm, for which the number of groups to be generated is predefined and can take any value, the DBSCAN algorithm (density-based spatial clustering of applications with noise), the Gaussian Mixture Model technique, etc. Each of the groups GCi, GC2, ..., GCN, respectively GCi, GC2, ..., GCN, is associated with a respective group identifier IDGi, IDG2, ..., IDGN, respectively IDGi, IDG2, ..., IDGN.

[0211] An embodiment of the network of [Fig.8] is now described in relation to [Fig.8]. convolutional neurons RNC.

[0212] The type of neural network considered is a convolutional autoencoder. Autoencoders are an unsupervised learning technique for learning efficient data encodings, typically for dimensionality reduction. As illustrated in [Fig. 8], an autoencoder comprises two learning layers: an encoding layer (ENC) that transforms the input data of an IMT image generated from different KPIs (at least two), and a decoding layer (DEC) that recreates the input data from a DC-encoded representation of the IMT image. The goal of this learning is for the pixel values ​​of the input IMt image to be equal to the pixel values ​​of the output IM't image.

[0213] The tests carried out on the RNC autoencoder made it possible to verify and validate the following points:

[0214] - the images generated from the KPIs can be used in a process learning and they are not noise,

[0215] - the main characteristics of the KPIs are reconstructed by the decoder,

[0216] - taking into account the previous point, the encoded data output from the encoder can can be used as input for grouping algorithms such as those implemented in the aforementioned CM comparison module.

[0217] In one embodiment of the invention, the RNC autoencoder has the following configuration, obtained, by way of non-limiting example, from the TensorFlow library of Python. The phrases preceded by the # symbol explain the lines of code that directly follow these phrases.

[0218] model = Sequential()

[0219] # Application of 14 filters of 7x7 dimension to a 360x48 pixel image, Therefore, 14 characteristic cards were generated.

[0220] model.add(layers.Conv2D(14, kemel_size=7, padding='same', activation='relu', input_shape=(360,48,l)))

[0221] # Dimensionality reduction by multiplying each of the 14 cards by a matrix of 2x2

[0222] model.add(layers.MaxPool2D((2,2), padding='same')) #Model optimization

[0223] model. add(lay ers. Dropout(0.2) )

[0224] #Application on all 7x7 areas of 7 filters of dimension 5x5, therefore ge creation of 7 characteristic cards

[0225] model.add(layers.Conv2D(7, kemel_size=5, padding='same', activation='relu'))

[0226] # Dimensionality reduction by multiplying each of the 7 cards by a matrix of 2x2

[0227] model.add(layers.MaxPool2D((2,2), padding='same')) #Model optimization

[0228] model.add(layers.Dropout(0.2))

[0229] #Application on all 5x5 areas of 7 filters of dimension 3x3, therefore ge creation of 7 characteristic cards and issuance of Y

[0230] model.add(layers.Conv2D(7, kemel_size=3, padding='same', activation='relu')) #end encoder #decoder start #inverse of maxpool

[0231] model.add(layers.UpSampling2D((2,2)))

[0232] model.add(layers.Dropout(0.2))

[0233] model.add(layers.Conv2D(14, kemel_size=5, padding='same', activation='relu'))

[0234] model.add(layers.UpSampling2D((2,2)))

[0235] model. add(lay ers. Dropout(0.2) )

[0236] model.add(layers.Conv2D(l, kemel_size=7, padding='same', activation='relu'))

[0237] model.compile(optimizer='adam', loss='mse', metrics=['mse'])

[0238] model.summaryO

[0239] In another example, the aforementioned Python language can be replaced by the R programming language.

[0240] To train the model of this RNC autoencoder, 3140 radio cells were divided into 2 sets: 80% for training and 20% for testing. The training data were subsequently divided into 2 subsets: 80% for training and 20% for validation.

[0241] The first training runs of the model gave good results, namely a mean squared error of 0.0215 for the training data and 0.017 for the validation data.

[0242] For each of these cells, 48 ​​KPIs are considered. For each of these 48 KPIs, 360 values ​​per KPI were obtained over 24 hours for 15 days.

[0243] An example of an IMT test image, representative of these 48 KPIs, and applied as input to the RNC autoencoder, is shown in [Fig. 9A]. The vertical axis V of this image represents the number of KPIs, i.e., 48, and the horizontal axis H of this image represents the 360 ​​values ​​obtained over 24 hours for each of these KPIs.

[0244] The IM'T image delivered at the output of the RNC auto-encoder is shown in [Fig.9B]

[0245] As shown in the graph in [Fig. 10], which represents the 24-hour evolution of the values ​​(between 0 and 1) of the same KPI extracted from the two images IMT and IM'T, the behavior of the KPI is well preserved at the autoencoder output. The black curve represents the evolution of the KPI extracted from the input image IMT, which is designated by IT in [Fig. 10]. The gray curve represents the evolution of this same KPI extracted from the output image IM'T, which is designated by I'T in [Fig. 10].

[0246] It follows from these tests carried out on real data that the ENC encoder is a device well suited to be used as an RNC neural network to deliver the vector Y; whose data are of the same type as the DC coded data represented in [Fig.8].

[0247] These DC-coded data, or the Y-coded data; if we consider for example the IM image;, are therefore relevant for the grouping implemented in step S7 of [Fig.7], such grouping being obtained according to the seasonality and / or the trend which will characterize the behavior of each cell in operation.

[0248] We will now describe, in a non-exhaustive manner, various examples of possible applications of the invention.

[0249] Taking into account, for a given cell, a multitude of performance indicators Formance and possibly one or more static parameters associated with a given cell, for behavioral identification, allows for more efficient grouping in terms of cell homogeneity. This grouping is performed by the network and must be periodic to account for variations that may occur over time and that can change cell behavior, such as changes in cell parameters, the deployment of new cells, changes in user behavior, etc. Once the cells are grouped, cell management mechanisms can be defined for each group to optimize network efficiency. Among these mechanisms, we can notably mention...

[0250] - Anomaly detection:

[0251] The objective of such detection is to guarantee quality of service. Since the cells in the same group are assumed to behave similarly, anomaly detection can be done by group with the assumption that an anomaly corresponds to a behavior different from that of the group.

[0252] By focusing, for example, on identifying whether, among the last n observed values ​​of the KPIs of a set of cells, there was an anomaly or not in one or more of these cells, the steps of anomaly detection can be as follows:

[0253] — 1. Identification of cell groups according to the process described above;

[0254] — 2. On each group considered separately:

[0255] — predict, for each cell in the group, using prediction methods of time series such as for example SARIMA (“Seasonal Autoregressive In-tegrated Moving Average”), Prophet or any other method, the last n expected values ​​of one or more KPIs. The input data of the prediction method can for example be the hourly values ​​observed on one or more KPIs over the last two weeks excluding the last day (i.e. 13 days of history), and the output will be the expected values ​​of the KPI(s) on the last day, calculated based on the history;

[0256] — calculate the difference between the predicted (expected) values ​​and the observed values;

[0257] — identify as abnormal cells those whose difference between the The predicted values ​​and observed values ​​are much more important compared to all the cells in the group.

[0258] - Reinforcement learning for joint management:

[0259] The goal of this learning is to enable the cellular network to learn the best management policy, which depends on the behavior of a given group of cells. Such management can be based, for example, on optimal parameters, actions to be performed depending on the network state, etc. The learning takes place at the level of the C-SON (Centralized Self-Organizing Network) by... Based on the experience of all cells within the same group, C-SON is the technology that enables the self-configuration, self-operation, and self-optimization of mobile network equipment. The learning steps may be as follows:

[0260] — 1. Identify groups of cells according to the procedure described above,

[0261] — 2. in each group, teach the C-SON the optimal settings (example: (signal strength, radio resource allocation policy, etc.) through the states and actions executed at the level of each cell in the group. In machine learning, this refers to reinforcement learning in which the cells constitute the agents. As an example, the main components of learning, namely: states, actions, rewards, and transitions, can correspond to:

[0262] — States: relating to the degree of charge of the cell (heavily charged, moderately charged) charged, slightly charged, no charge...),

[0263] — Actions: functions executed at the level of a loaded cell allowing to balance its load with neighboring cells via, for example, a transfer of users, a decrease / increase in transmission power, etc.

[0264] — Reward: measured by the fairness of the charge between cells and the quality of service provided to the user, — Transitions: the transition from one state to another is then the change in the state of a cell following the execution of an action,

[0265] — 3. apply to cells of the same group the best actions learned by the C- SOUND for a cell.

[0266] - Collaborative learning for predicting performance indicators

[0267] In this case, the cells collaborate to predict indicators in order to anticipate actions. The steps may be as follows:

[0268] — 1. Identify groups of cells according to the procedure described above,

[0269] — 2. predict, for each cell in the group, using prediction methods Using time series analysis methods such as SARIMA, Prophet, or any other method, the last n expected values ​​of one or more KPIs are used. The input data for the prediction method could, for example, be the hourly values ​​observed for one or more KPIs over the last two weeks, corresponding to 14 days of historical data, and the output would be the expected values ​​of the KPI(s) in the future (e.g., the following day), calculated based on historical data.

[0270] — 3. Each cell builds its KPI prediction model using by For example, a neural network. The input data for this neural network are the observed values ​​of the KPIs, and the output data are the predicted values ​​of the KPIs.

[0271] — 4. Each cell shares the parameters of its model, the weights of the different layers of the neural network or gradient, with the cells of its group,

[0272] — 5. each cell optimizes its parameters according to those of the cells of its group, by calculating for example a weighted average over all the weights.

Claims

Demands

1. Method of checking the operation of a cellular radio communication network from a comparison, between at least the first and second cells (CL;, CLj) of said network, of the evolution over time of at least one performance indicator measured for the first, respectively second, cell, characterized in that it comprises the following: - collect (S2) for the first cell at least two first time series of values ​​taken over time by said at least one performance indicator and respectively at least one other performance indicator measured for the first cell, - collect (S2) for the second cell at least two second time series of values ​​taken over time by said at least one performance indicator and respectively said at least one other performance indicator measured for the second cell, - convert (S6) said at least two first and second time series, into respectively at least one first (IM;) and a second (IMj) image, the first and second images comprising pixels whose values ​​correspond to the values ​​of at least the first and second time series respectively, - depending on the result of a comparison (S7) of the first image with the second image, identify whether said at least two first and second cells have, in operation, a common behavior or not.;

2. A method for controlling the operation of a cellular radio communication network according to claim 1, wherein the identification that said at least two first and second cells have, in operation, a common behavior or not, is implemented as a function of the result of a comparison of the first image with the second image, of at least one first parameter (PSki) associated with the first cell (CLj) and of at least one second parameter (PSkj) associated with the second cell (CLj).

3. A method for checking the operation of a cellular radio communication network according to claim 1 or claim 2, wherein, prior to said comparison, the method comprises the following: - cutting (S620) the first image into at least two zones, respec-

4. - tively the second image into at least two zones, - extract (S621) at least one feature per zone from the first image, respectively at least one feature per zone from the second image, - generate (S622) at least a first feature matrix (MATDH) from said at least one feature extracted per zone from the first image, respectively at least a second feature matrix (MATDij) from said at least one feature extracted per zone from the second image, - to cut (S623) at least the first matrix (MATDn) into at least two feature submatrices (SMATDioi, SMATDm), respectively at least the second matrix (MATDij) into at least two feature submatrices (SMATDiOj, SMATDnj), - reduce (S624) the dimension of said at least two sub-matrices resulting from the partitioning of the first matrix, generating at least one first reduced-size feature matrix (MATNu), respectively the dimension of said at least two sub-matrices resulting from the partitioning of the second matrix, generating at least one second reduced-size feature matrix (MATNy), - generate (S625) a first vector (Y;) from said characteristics of at least the first reduced size matrix (MATNu), respectively a second vector (Yj) from said characteristics of at least the second reduced size matrix (MATNy). Method of controlling the operation of a cellular radio communication network according to claim 3, wherein the steps of dividing the first image, respectively the second image, into at least two zones, of extracting at least one feature per zone of the first image, respectively the second image, of generating at least a first feature matrix, respectively a second feature matrix, of dividing said at least a first feature matrix, respectively said at least a second feature matrix, of reducing, and of generating a first vector, respectively a second vector, are implemented by a convolutional neural network which receives as input the first image, respectively the second image, and which delivers as output the first vector characterizing the first cell, respectively the second vector characterizing the second cell.

5. Method of controlling the operation of a cellular radio communication network according to claim 3 or claim 4, further comprising the following: - applying (S7) the first and second vectors as input to a vector grouping module, - generating (S7) as output from the vector grouping module a group of vectors containing the first and second vectors if the first and second vectors fulfill a grouping criterion.

6. Method of controlling the operation of a cellular radio communication network according to claim 5, wherein the first vector, respectively the second vector, is applied to the input of the vector grouping module, together with the first parameter associated with the first cell, respectively together with a second parameter associated with the second cell.

7. Device (DCF) for monitoring the operation of a cellular radio communication network by comparing, between at least the first and second cells of said network, the evolution over time of at least one performance indicator measured for the first, respectively second, cell, characterized in that it is configured to: - collect for the first cell at least two initial time series of values ​​taken over time by said at least one performance indicator and respectively at least one other performance indicator measured for the first cell, - collect for the second cell at least two subsequent time series of values ​​taken over time by said at least one performance indicator and respectively said at least one other performance indicator measured for the second cell, - convert said at least two initial and subsequent time series,in at least one first and second image respectively, the first and second images comprising pixels whose values ​​correspond to the values ​​of at least two first and second time series respectively, - depending on the result of a comparison of the first image with the second image, identify whether said at least two first and second cells have, in operation, a common behavior or not.

8. Computer program containing code instructions program for implementing the method for controlling the operation of a cellular radio communication network according to any one of claims 1 to 6, when executed on a computer.

9. Computer-readable information carrier, and containing instructions for a computer program according to claim 8.

Images