Confirm community upgrade

Abstract

Embodiments described herein relate to methods and apparatus for determining a cell upgrade, such as whether to upgrade a Long Term Evolution (LTE) cell to massive multiple-input multiple-output (M-MIMO) technology. A method for determining a cell upgrade metric is provided. The method receives samples (405, 410, 415, 420) including user session data from a user device communicating within a cell. Training data (425) is obtained from the samples by applying filtering rules and an autoencoder (430) is trained using the samples. The encoder layer and code layer of the trained autoencoder are combined with a classifier network to generate a combined classification network (435). The combined classification network (440) is trained using the training data. The trained combined classifier network is used to classify user session data within the same or different cells (505), and a plurality of the classifications are used to determine a cell upgrade metric (525) for the cell.

Description

Technical Field

[0001] Embodiments disclosed herein relate to methods and apparatus for determining a cell upgrade, for example, whether to upgrade a Long Term Evolution (LTE) cell to a massive multiple-input multiple-output (M-MIMO) technology. Background Art

[0002] As the use and demand for cellular communication services grows, some existing cellular devices may not be able to support the increased traffic or service demands. Therefore, operators need to determine which cells can benefit from equipment upgrades to provide newer and / or improved services. For example, operators may wish to determine which of their LTE cells will benefit from upgrading to equipment that supports M-MIMO technology.

[0003] The current LTE cell qualification method used to decide on investments in enhanced M-MIMO features is based on performance management (PM) counters and strict rule-based decisions. PM counters can be used to determine various metrics such as: call setup success rate, which measures the overall accessibility of the user; access network retention, which assesses the ability of the network to maintain the service requested by the user; handover queues and CPU load; number of connections, and many other metrics. Different PM counters can be specified for different devices and devices from different vendors. Summary of the invention

[0004] The PM counter approach aggregates at the cell level and therefore represents the entire user equipment (UE) ecosystem within a given cell using statistical metrics such as mean, median and quantiles. However, this does not allow for consideration of more fine-grained UE-specific metrics such as quality vs. distance, or UE-to-cell metrics such as UE throughput vs. cell load. Furthermore, not all cells will benefit equally from a cell upgrade, and this may depend on the characteristics of the propagation environment within the cell, such as whether there are multiple propagation paths between the base station (e-Node B or eNB) and the UE (e.g., a smartphone).

[0005] According to certain embodiments described herein, there is a method for determining a cell upgrade metric. The method receives samples including user session data from a user equipment communicating within a cell. Training data is obtained from the samples by applying filtering rules and an autoencoder is trained using the samples. The encoder layer and the code layer of the trained autoencoder are combined with a classifier network to generate a combined classification network. The combined classification network is trained using the training data. The trained combined classifier network is used to classify user session data within the same or different cells, and a plurality of the classifications are used to determine a cell upgrade metric for the cell.

[0006] The use of user session data allows for improved and more fine-grained cell characterization by characterizing each end-user radio environment individually. This allows for appropriate correlation with other metrics (e.g. aggregated cell metrics) or geospatial considerations (e.g. user angle distribution analysis). Large amounts of such data can be efficiently processed using appropriately trained and architected classification networks.

[0007] According to certain embodiments described herein, a method for recommending a cell upgrade is provided. The method includes: receiving user session data and cell metrics during an evaluation period; and generating samples by characterizing user session data during each sampling period and correlating the characterized user session data with the cell metrics time, wherein the user session data is from a user device communicating within the cell. A trained classifier network is used to classify the samples as being able to be improved by an upgrade, together with a confidence score. The cell upgrade recommendation is made based on the proportion of samples classified as being able to be improved and the confidence score.

[0008] According to some embodiments described herein, corresponding apparatus and computer programs are also provided, the computer programs comprising instructions, which, when executed on a processor, cause the processor to perform the methods described herein. The computer programs may be stored on a non-transitory computer-readable medium. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] For a better understanding of the embodiments of the present disclosure and to show how they may be implemented, reference will now be made, by way of example only, to the accompanying drawings, in which:

[0010] Figure 1 is a schematic diagram illustrating sample collection in a communication system according to some embodiments;

[0011] Figure 2 is a schematic diagram showing filtering samples to generate training data according to some embodiments;

[0012] Figure 3 is a schematic diagram illustrating construction and training of a classification network according to some embodiments;

[0013] Figure 4 is a flowchart illustrating construction and training of a classification network according to some embodiments;

[0014] Figure 5 is a flow chart of using a trained classifier network to determine a cell upgrade metric or rank cells according to some embodiments;

[0015] Figure 6 shows an example cell probability metric according to some embodiments;

[0016] FIG. 7A to FIG. 7Dshows a representation of some MIMO characterization values ​​at a cell level according to some embodiments;

[0017] Figure 8 is a schematic diagram showing the architecture of an apparatus according to an embodiment; and

[0018] Fig. 9 The cell ranking with estimated improvements by upgrading is shown. DETAILED DESCRIPTION

[0019] Generally, unless clearly given and / or different meanings are implied from the context, all terms used in this article will be interpreted according to their common meaning in the relevant technical field. Unless otherwise clearly stated, all references to "one / an / element, equipment, component, device, step, etc." should be openly interpreted as referring to at least one example in element, equipment, component, device, step, etc. Unless a step must be clearly described as after or before another step and / or a step must be after or before another step implicitly, the steps of any method disclosed herein do not have to be performed in the exact order disclosed. Where appropriate, any feature of any embodiment disclosed herein may be applied to any other embodiment. Similarly, any advantage of any embodiment may be applicable to any other embodiment, and vice versa. By the following description, other purposes, features and advantages of the attached embodiments will be apparent.

[0020] Specific details are set forth below, such as specific embodiments or examples for purposes of explanation rather than limitation. It will be appreciated by those skilled in the art that, in addition to these specific details, other examples may be adopted. In some instances, detailed descriptions of known methods, nodes, interfaces, circuits and devices are omitted to avoid blurring descriptions with unnecessary details. It will be appreciated by those skilled in the art that the described functions may be implemented in one or more nodes using hardware circuits (e.g., interconnected to perform analog and / or discrete logic gates, ASICs, PLAs, etc. of specialized functions), and / or implemented using software programs and data in conjunction with one or more digital microprocessors or general-purpose computers. Nodes using air interface communications also have suitable radio communication circuits. Moreover, the technology may also be considered to be implemented completely in any form of computer-readable memory (e.g., solid-state memory, magnetic disk or optical disk) containing a suitable set of computer instructions that will cause a processor to perform the technology described herein.

[0021] Hardware implementations may include or contain, but are not limited to, digital signal processor (DSP) hardware, reduced instruction set processors, hardware (e.g., digital or analog) circuits (including, but not limited to, application specific integrated circuits (ASICs) and / or field programmable gate arrays (FPGAs)), and state machines capable of performing such functions, if applicable. Memory may be employed to store temporary variables, maintain and transfer data between processes, non-volatile configuration settings, standard message formats, etc. Any suitable form of volatile and non-volatile memory may be employed, including random access memory (RAM) implemented as a metal oxide semiconductor (MOS) or integrated circuit (IC), and storage devices implemented as hard drives and flash memory.

[0022] Embodiments described herein relate to methods and apparatus for determining cell upgrade metrics, ranking cells for upgrade, and / or recommending cell upgrades. Analysis of user data sessions allows for improved characterization of cells, and therefore improved upgrade metrics, rankings, and recommendations. However, analysis of such large amounts of data presents challenges that are addressed using improved classification network construction and training.

[0023] Figure 1 1 is a schematic diagram illustrating sample collection in a communication system according to an embodiment. The communication system 100 may be a wireless communication cell, such as a Long Term Evolution (LTE) cell including a base station or eNodeB 105 and a plurality of user equipments (UEs) 110 connected to the eNodeB via respective wireless connections 115. Each UE may generate user session data 120 including parameters and metrics related to its connection with the eNodeB 105. Similarly, the eNodeB 105 may generate cell metrics related to the overall performance of the eNodeB 105 in serving the various UEs 110.

[0024] The user data session 120 and the cell metrics 125 may be in the form of a cell tracking record (CTR) binary file specified by the 3GPP standard, and is a collection of specific events that represent a mix of performance metrics and user capabilities, and the specific events are collected periodically. Some metric values ​​may change quickly, such as radio measurements, and some may change slowly or be fixed for a session, such as user capabilities. Table 1 below shows some example CTR events 120, including UE session related events, cell or eNodeB events, UE session control events, and UE geolocation events. These examples are shown for illustrative purposes and are not intended to be limiting.

[0025]

[0026] Table 1

[0027] The user session data 120 and the cell metrics 125 are correlated by a correlator 130 to generate a sample or data unit 135. The correlator may be implemented by a processor and a memory, and time correlates the various data 120, 125 to generate periodic samples 135. For example, the data may be aggregated and correlated to generate a sample every 60 seconds, and may include UE capability metrics that do not change during the period and an average of measurements taken every 1 second during the 60 second period. Similarly, data updated less frequently than 60 seconds may be interpolated to determine appropriate values ​​for the current sampling period. Other metrics to be included in the sample may be calculated based on data received from the UE 110 and / or eNodeB. Data from each UE 115 and eNodeB may be collected and time correlated using a system timestamp. Specific data may be selected to be included in the sample and in specific calculated metrics, such as data related to MIMO usage and performance, while other data may be omitted.

[0028] Table 2 below shows an example of sample 135.

[0029]

[0030]

[0031]

[0032]

[0033] Table 2

[0034] Figure 2 is a schematic diagram showing filtering samples to generate training data according to an embodiment. Figure 1The correlator generates a large number of samples for each UE 110. Some of the sample set 205 may support upgrading, while some may not. Some samples of each category (upgrade or not) may be selected using expert rules 210 applied to the samples 135 to select a smaller training data set 215 including samples that support upgrading and samples that do not support upgrading. For example, a specific metric above a threshold may indicate that the use of massive multiple-input multiple-output (M-MIMO) will improve the user session, while other metrics below the threshold may indicate that M-MIMO will not provide benefits. The combination of comparisons of these metrics with the corresponding thresholds can be used to indicate whether the user session as a whole will benefit from the M-MIMO upgrade. For example, if a sample indicates that the UE does not support MIMO, the sample can be considered useless for training data, although it is retained in the entire sample set 205. However, samples where MIMO_2x2_cap is true, Event+Param_Rank_Reported is above the orthogonal layer threshold, and SINR_MEAS_PUSCH is within a specific range may indicate suitable M-MIMO upgrade samples for the training set.

[0035] The purpose of applying expert rules 210 to filter the samples is to identify which of the samples 135 represent the best case of two categories (e.g., M-MIMO will improve and M-MIMO will not improve). The expert rules can be determined by radio network design and optimization (NDO) experts and effectively provide low noise samples for each category.

[0036] Figure 3 Shown is a classifier network that is suitable for processing samples and is constructed and trained. In the first step, an autoencoder network 310 is trained using a sample set 205. Sample sets from one or more UEs 110 can be used, and additional sample sets 210 from other cells can be used in addition. An autoencoder is an artificial neural network for learning effective data representations, vectors or codes in an unsupervised manner. The autoencoder network includes multiple layers, including an input layer 315I, an encoder layer 315He, a code layer 315LS, a decoder layer 315Hd and an output layer 315RI. The autoencoder network 310 is trained to reconstruct the input at the output. Once trained, the code layer 315LS can be used to effectively represent the input using dimensionality reduction, effectively training the network to ignore noise and focus on the most important aspects. The autoencoder network 310 is trained to generate a reconstructed sample 205R from a reduced encoding that is sufficiently close to the input sample 205.

[0037] The trained autoencoder 310 effectively converts the CTR samples into a compact vector representation or code and can be used as input to the classification network. Using a trained autoencoder as part of the classifier network provides many advantages over using and training only the classifier network, including training the autoencoder on all samples, rather than using only the training data on the normal classifier network. By using all samples for training, these are not restricted by expert rules. By training the autoencoder on all samples, the compact representation takes into account all different scenarios of the samples, so the compact vector or code is more data-intensive and therefore more meaningful. This also provides a way to check the progress and generalization of the training algorithm by checking the reconstruction loss, so problems can be solved faster.

[0038] Once trained, the encoder and code layers 320 of the autoencoder network 310 are combined with the classifier network 325 to generate a combined classifier network 330. Any suitable multi-layer binary artificial neural network can be employed. The layers can include convolution, pooling, and non-linear functions such as RELU activation and Sigmoid activation.

[0039] After the combined classifier 330 is assembled, it is trained using the training set 215 to determine samples indicating M-MIMO upgrade improvement and samples indicating no M-MIMO upgrade improvement. The use of this two-stage classifier improves the generalization ability of the classifier while focusing on more important features of the sample to classify the sample.

[0040] Further training may be used. For example, the autoencoder may be further trained using sample sets 210 from other UEs 110 in the cell 100 and UEs in other cells to help further generalize the encoder 320 portion of the combined classifier network 330. In addition, expert rules may be applied to other sample sets 210 to obtain additional training data, for example, from other UEs 110 in the same cell 100 or UEs in other cells. The additional training data may be used to further optimize the combined classifier network 330.

[0041] Figure 4 4 is a flow chart illustrating a method of building and training a classifier network for determining a cell upgrade metric for a cellular communication network. At step 405, the method receives user session data, such as CTR data, from a UE. As described above, CTR is defined by the 3GPP standard and will be readily available from a UE that complies with the standard. At step 410, the user session data is characterized, for example, by averaging measurements within a sampling period, calculating a metric within the sampling period based on one or more measurements. The characterization provides values ​​for each sampling period, which are added to the corresponding samples.

[0042] At step 415, the method may also receive cell metrics from the eNodeB, which generally characterize the performance of the cells in the aggregation. At step 420, these may be time correlated with the user data in the sample to generate a more comprehensive sample.

[0043] At step 425, samples are filtered using expert rules as previously described. For example, the sample set 210 is reduced to a smaller training data set 220 with samples that clearly fall into one of two categories. Using a smaller, well-characterized sample set makes it easier to train the classifier.

[0044] At step 430, the autoencoder network is trained using the entire sample set as described above. The autoencoder is trained to replicate the sample across the entire sample range. At step 435, the encoder and code layers of the trained autoencoder are combined with the classifier network to generate a combined classifier network as described above. At step 440, the combined classifier network is trained using the training data determined at step 425.

[0045] Once trained, the combined classifier network can be used to classify samples from other UEs in the same and different cells.As described above, some further training of the autoencoder and / or the combined classifier network can be performed using samples from other UEs and / or cells.

[0046] The input to the trained combined classifier is a sample of unique UE sessions within the cell, such as shown in Table 2 above. The output of the trained combined classifier is the CONFIDENCE and UPGRADE cell upgrade metrics. CONFIDENCE is the probability output of the combined classifier for each unique user session. UPGRADE compares the probability to a threshold and is 1 (improved by upgrading) if above the threshold and 0 (no improvement by upgrading) otherwise. For example, some user sessions may be classified as suitable for M-MIMO and therefore will benefit from an M-MIMO upgrade, while other UE sessions may be classified as not suitable for M-MIMO and therefore will not benefit from an M-MIMO upgrade.

[0047] Figure 5 A method for recommending cell upgrades is shown. At step 502, the method receives user session data 120 and cell metrics 125 over an evaluation period that can be several hours or even days. The user session data and cell metrics are the same type of information used to train the combined classifier network and are used to generate samples at step 503 as described above. Samples are generated for each sampling period (e.g., every 60 seconds), and the user session data within that time is characterized and time-correlated with the cell metrics. Each user session can last from several seconds to several hours and will be associated with multiple samples.

[0048] At step 505, the method uses the trained combined classifier network 330 to classify each user session in the cell during the evaluation period. The network determines the cell upgrade metrics (CONFIDENCE and UPGRADE) for each user session. At step 510, the method determines whether there are more sessions to classify, and if so, returns to step 505, otherwise proceeds to step 515.

[0049] At step 515, the method determines additional cell upgrade metrics related to the aggregation of user session classifications for the cell within the evaluation period. Qualified and unqualified user sessions are grouped and analyzed for the cell. The aggregated cell upgrade metrics calculated in an embodiment are the cell qualification score (CQS) and the distribution confidence index (DCI), although other metrics may be used alternatively. CQS represents the percentage of user sessions with a CONFIDENCE greater than a threshold (threshold1). In this example, threshold1 is the same as the threshold used to determine UPGRADE, and CQS compares the number of UE sessions that benefit from the upgrade to the total number of UE sessions.

[0050] The DCI indicates how close the probability distribution is to the ideal case of upgrading, which is determined by the expert rules and the generalization of the classifier network. The DCI can also be used to monitor machine learning performance. For example, if the DCI is always low for a cell, this may indicate that more training is needed. The CONFIDENCE output of the user sessions of the cell is divided into three boxes or ranges: 0 to 30%, 30% to 70%, and 70% to 100%. The DCI is the percentage of samples in the first (0-30%) range and the third (70-100%) range. The higher the DCI, the greater the confidence of the method about whether to upgrade, because most user sessions clearly fall into the upgrade group or the non-upgrade group. For example, if the CQS is high, indicating that most user sessions support upgrading, and the DCI is also high, indicating that most user sessions are obviously upgraded or not upgraded (a small amount is in an inconspicuous middle area), then these cell upgrade metrics indicate that upgrading to M-MIMO will provide significant improvements to overall cell performance. On the other hand, if the DCI is low, most of the probabilities or CONFIDENCE outputs are in the 30-70% range, so most user sessions are between desired and undesired radio scenarios, and it is therefore less clear whether an upgrade will be beneficial.

[0051] Figure 6Example CONFIDENCE or probability metrics are shown as histograms showing the number of user sessions (y) versus the probability range (x). The first graph (left) shows a high confidence level because most of the user session classifier outputs indicate that a cell upgrade is recommended. The second graph (right) shows a mix of high CONFIDENCE outputs and mostly low CONFIDENCE outputs, with a small number in the middle range. Although not as obvious as the first graph, this second graph shows with a fairly high confidence level that the cell will not benefit from an upgrade in this case.

[0052] Return to reference Figure 5 At step 520, the method determines whether there are additional cells to analyze, and if so, returns to step 505. Once all cells have been analyzed, the method proceeds to step 525, where a preliminary ranking of cells is generated using the aggregated cell upgrade metrics. In an embodiment, the preliminary ranking is based on the product of the CQS and DCI of each cell. Cells with high CQS and high DCI have a higher ranking than cells with high CQS but low DCI.

[0053] The method then proceeds to step 530 where the preliminary ranking may be enhanced with a representation of session characteristics based on further analysis of specific user session parameters or metrics relevant to the upgrade, such as those related to MIMO characterization. FIG. 7A to FIG. 7D Representations of relevant features for MIMO characterization at the cell level (in this example, high-ranked and low-ranked cells) are shown. For example, the first parameter shown is active_ue_dl_avg, which is the average number of active users (with data to be transmitted in the eNB buffer) in the downlink during any sampling period for both high-ranked cells (left) and low-ranked cells (right). The representation or box consists of the parameter values ​​for each user session, with the 25th and 75th quartiles defined to generate the boxes shown. It can be seen that for the first parameter, the high-ranked cells have a range of 5 to 8.5, while the low-ranked cells have a range of 0.5 to 1.6. In a similar manner, other parameters for high and low-ranked cells can be analyzed. These representations can be used to support ranking by representing the values ​​of various relevant session features in different cells, and therefore confidence can be added to the ranking of users.

[0054] At step 535, the method determines a geolocation metric for each cell. Table 3 below shows example metrics that may be calculated.

[0055]

[0056]

[0057] Table 3

[0058] At step 540, the method further groups and ranks the cells using geolocation analysis based on the above or other geolocation metrics. The following geolocation analysis criteria can be used:

[0059] - x_angle_valid_samples > thresh2, there is a valid statistical representation of the cell traffic

[0060] - x_angle_spread_qualify > thresh3, the 50ile of the cell traffic is within the angular range (usually 60° horizontally and 8° vertically, which is the most common beamwidth in the deployed antenna array)

[0061] - h_ang_spread_kurt < thresh4, this criterion represents that the traffic is evenly distributed across the angular range. Based on the spatial orthogonality obtained by using beamforming, geographical isolation of concurrent users in a cell is a desirable characteristic for maximizing multi - user MIMO (MU - MIMO) performance.

[0062] Then the cells that meet these criteria can be grouped, sorted, or ranked as follows:

[0063] i) UE_classifier_criteria_fulfillment = True, CQSxDCI >= Thres5

[0064] i.i Sort by CQSxDCI AND h_angle_criteria_fulfilment = True

[0065] I.ii Sort by CQSxDCI AND h_angle_criteria_fulfilment = False

[0066] ii) UE_classifier_criteria_fulfillment = False, CQSxDCI < Thres5

[0067] Ii.i Sort by CQSxDCI AND h_angle_criteria_fulfilment = True

[0068] Ii.ii Sort by CQSxDCI AND h_angle_criteria_fulfilment = False

[0069] Cells that do not meet the criteria are retained in their own group.This approach allows for deeper analysis of cells, including consideration of geolocation information where available, and therefore facilitates improved upgrade decisions.

[0070] At step 545, the method recommends a cell upgrade. This may be in the form of a ranking in a geo-located group as described above and / or a simple preliminary ranking of all cells. Various metrics may be associated with cells as well as graphs (e.g. Figures 6 to 7D those) are displayed or available together.

[0071] Some embodiments may also provide a gain estimation tool that supplements the recommendation by providing additional information about the benefits of deploying the recommended solution (e.g., M-MIMO). The gain estimation tool or module takes input from the trained classifier at the session level, cell level ranking, and cell geolocation metrics as previously described, and may consider the following non-limiting aspects.

[0072] Digital Sectorization (DS) with Active Antenna System (AAS) Frequency Division Duplex (FDD) - DS with massive MIMO allows multiple cells with the same antenna array to be implemented on the same carrier frequency using beamforming. The DS solution increases capacity without the need for additional hardware. The main benefit is an increase in the total site throughput of the M-MIMO system in both uplink and downlink, and improved spectral efficiency (SE) in physical resource block (PRB) utilization compared to non-AAS (active antenna system) cells. This enables capacity gains when the penetration of Rel-13 / 14 (3GPP Release 13 or 14) terminal devices is low.

[0073] Transmission Mode 9 (TM9) with AAS FDD - Cell shaping is a form of cell-specific beamforming that follows the UE distribution within the cell, rather than creating narrow beams that follow the users. The coverage area is reshaped to meet and optimize traffic needs. The use of available spectrum becomes more efficient and cell capacity increases because the available RF energy is more tightly focused on the users within the cell. Cell-specific beamforming is compatible with today's FDD UEs.

[0074] UE DL Achievable Throughput on Cell Carrier (UE_DATCC) - When a UE session ends, an estimate of the downlink (DL) user throughput that could be hypothetically achieved if a new session were immediately established under the same air interface conditions and using the available free PRB resources on that cell. This is an approximation of the active throughput.

[0075] UE_DATCC is a function of: DL CQI, modulation, ranking indicator, PRB utilization, cell bandwidth, cell CFI pattern, UE capabilities. For each UE session recorded on the CTR collection file, these features are input into the model and a dl throughput value is calculated for the user session. Two UE_DATCC values ​​are calculated for each session:

[0076] 1) Modeling user session throughput using current cell configuration, cell quality and cell load conditions as reported on CTR events;

[0077] 2) Introduce a new user throughput value when calculating the M_MIMO gain.

[0078] Then, the previous / next two UE session calculation values ​​are aggregated to the cell level, and the cell-level gain estimation ratio is calculated based on these.

[0079] The UE-DTACC calculation method can be divided into two parts:

[0080] a) Raw bit information evaluation of each scheduling block;

[0081] b) An evaluation of the resources available to each scheduling block.

[0082] a) Raw bit information evaluation of each scheduling block

[0083] Based on expert knowledge of the upgraded product, a function is defined that estimates the corresponding normalized throughput (Symbol Information - SI) in bits / symbol from the SINR (Signal to Noise Ratio) value.

[0084] On the CTR trace collected as input, the CQI (Channel Quality Index) measurement reported by the UE can be used as a DL quality indicator. The first step is to approximate the SINR value from the reported CQI. Afterwards, the SINR can be used to calculate the number of bits per resource element available on the scheduling block. This method uses CQI and modulation as input to generate the SI value that will be used to estimate the DL user throughput.

[0085] b) Resource evaluation for each scheduling block

[0086] Estimation of the PRB (Physical Resource Block) cell load and remaining free cell resources is achieved by considering the free PRB resources as the total PRBs available on the cell minus the average PRB utilization over a 60 second period when the session is established.

[0087] Based on geo-referenced samples collected by MDT (Minimization of Drive Test) events, the traffic currently served by the cell can be divided into the coverage areas of the new digital sectors. This method divides the horizontal angle UE distribution into two groups:

[0088] - samples located in the area defined by horizontal angles between -180° and 0;

[0089] - a sample located in a region defined by a horizontal angle between 0° and 180°;

[0090] The horizontal angles can be converted and referenced to the cell azimuth as given on the cell map file. Therefore, all horizontal angle samples will range from -180° to 180°.

[0091] If a potential cell that is split into two sectors by implementing digital sectorization has the largest portion of its UEs concentrated in one of the regions, it will likely show lower gains compared to a cell where the traffic is almost split in half between the two regions.

[0092] For example: if 90% of the UEs are located in -180° to 0°, the expected gain in throughput will be low. Whereas for the case where 51% of the traffic is between 0° and 180°, the possible gain will be higher. The reason is that in these cases the expected new PRB utilization on the new sector will be very different.

[0093] For cells under extremely high load conditions, the model estimates the cell gain by taking into account the additional PRB utilization after massive MIMO deployment. This avoids overestimating the gain due to the expectation that previously unserved traffic (due to high load) will be processed after the solution is implemented.

[0094] When digital sectorization is implemented, the factor to estimate the additional PRB utilization is a user input variable applied to all cells under very high load. A practical approach to find a reasonable value for this factor may be to evaluate the amount of traffic moved to other layers / cells by the load sharing mechanism.

[0095] After running two pre-processings (raw bit information evaluation per scheduling block and available resource evaluation per scheduling block), it is straightforward to estimate the achievable DL throughput since other aspects (e.g. whether MIMO 4x4 is available, rank indication and transmission mode used, etc.) are known through the collected traces.

[0096] The next step after calculating the DTACC for each UE session is to estimate what would happen to the DL throughput if massive MIMO was implemented on the cell. Based on experiments and expert knowledge, a set of “control knob” variables can be introduced to fine-tune the expected changes in the elements that define the throughput value, such as the expected SINR gain after implementation, the impact on PRB loading, etc. As the expected gain components are identified, new dl throughput values ​​are calculated taking these factors into account. Therefore, there are two DTACC values ​​for each UE session (before and after M-MIMO). Since it is planned to characterize the diagnostics at the cell level, the UE session results are aggregated at the cell level, currently using the median operator. And the percentage difference between the throughput calculated before and after M-MIMO deployment is the estimated gain for each cell.

[0097] The gain estimate according to the cell configuration can be determined as follows.

[0098] MIMO 2x2 Baseline Deployment

[0099] Although MIMO 4x4 can be implemented without an AAS site solution, the resulting DL throughput gain should be calculated as if the site baseline before the introduction of AAS was MIMO 2x2 only. The gain will be calculated as a function of:

[0100] - Expected SINR gain after introducing MIMO 4x4 with TM4.

[0101] -Percentage of UE groups that can use MIMO 4x4.

[0102] - Estimated percentage of UE sessions supporting MIMO 4x4 that currently use 2x2 will then be able to use both layer3 and layer4 at least once per session.

[0103] -Estimated percentage of rank3 and rank4 in the total ranking reported.

[0104] For example, if before AAS is deployed, there are 1000 UE sessions from UEs using 2x2 that support MIMO 4x4 in a particular cell, then for 300 of these sessions, there will be at least one sample with layer layer3 / layer4 when using M-MIMO. Therefore, the value of the estimated percentage described in the third item above will be 30%. In these 300 sessions, the average percentage reported for rank3 and rank4 can be, for example, 10% and 5%, respectively. Therefore, for this scenario, the model will consider that there will be 1000 UEs supporting 4x4, for 30% of the sessions, there will be at least one sample of layer3 or layer4, and the percentage of their utilization in each of these 300 sessions will be 10% and 5%, respectively.

[0105] Digital sectors as an option for AAS FDD

[0106] The modeling factors that will affect the gain estimate for digital sectorization are:

[0107] - Traffic balancing between left and right sectors, based on MDT samples (described in 3.2.4).

[0108] - The amount of traffic at the boresight and its expected SINR loss due to the increase in inter-cell interference.

[0109] - An estimate of the additional PRB utilization that will come from being able to serve users that were previously unable to connect in very highly loaded cells.

[0110] After modeling the M-MIMO implementation, there is no service loss due to cell-edge UEs moving to other carriers.

[0111] Cell Shaping as an AAS FDD Option

[0112] The modeling factors that will affect the gain estimation for the TM9 beamforming scenario are:

[0113] - The percentage of traffic located at the cell edge area and its expected SINR gain when beamforming is activated.

[0114] - Expected symbol information gain expected for TM9 supporting UEs in medium quality areas.

[0115] - Expected symbol information gain expected for TM9 supporting UEs in poor quality areas.

[0116] Output of the gain estimation process

[0117] The output of the gain estimation process may be a data frame containing the following information at the cell level:

[0118]

[0119]

[0120] Table 4

[0121] like Fig. 9 As shown, this information can be used as additional criteria together with the secondary ranking to prioritize investments at the cell / site level, e.g. Fig. 9 The extent of potential performance gain for each ranked cell is shown. Figure 8 8 is a schematic diagram of an apparatus that can be used to monitor or determine an alarm condition in one or more devices according to an embodiment. The apparatus 800 includes a processor 810, a memory 820 containing computer program instructions 825, which when executed by the processor causes the processor to perform the method of the embodiment. The apparatus also includes a trained classifier network 830. The apparatus can be implemented in a computer system remotely connected to a cell and user equipment, or a portion of the apparatus can be deployed at the cell during testing and data collection, and a portion can be deployed remotely for analysis. The apparatus can be implemented in the cloud. An example method including several steps that can be performed by the apparatus 800 is shown.

[0122] At a first step 840, the method receives user data sessions and cell metrics during an evaluation period, for example as previously described with respect to Figure 1 , Figure 4 and Figure 5 At step 845, the method generates samples by characterizing the user data and time-correlating the cell metrics, for example as previously described with respect to Figure 1 , Figure 4 and Figure 5 as described.

[0123] At step 850, the method uses the trained classifier network 830 to classify the sample as being able to be improved by upgrading, along with the confidence score. The trained classifier network 830 can be the trained combined classifier network 330 constructed and trained according to the previous embodiment, such as Figure 2 , Figure 3 and Figure 4 A trained classifier network, albeit differently constructed and / or trained, may alternatively be used to classify samples generated by characterizing user session data and time correlating with cell metrics.

[0124] At step 855, the method aggregates the user session sample classifications and recommends cell upgrades (or no upgrades) based on the proportion of samples classified as being able to be improved by upgrading and the confidence scores of the classifications. Figure 5This can be supplemented or adjusted using the geolocation analysis as described and can produce a simple list of candidate cells to be upgraded or a substantial analysis associated with each cell, such as with respect to Figure 5 as described.

[0125] Embodiments may provide multiple advantages, including more granular analysis of each cell. Each end-user radio environment is characterized and analyzed individually, which allows direct correlation of relevant cell metrics and can be performed regularly (e.g., every 60 seconds). This is a significant improvement over known methods that use performance metrics (PM) that are typically calculated every 15 minutes based on aggregated cell activity.

[0126] A richer representation is analyzed using deep learning, which provides a more accurate cell / site qualification method. While the pm counter-based process is able to provide a strict, rule-based decision threshold, the method of the embodiment using CTR-based samples with the classification provides a continuous ranking across networks / regions under analysis, providing operators with flexibility to decide where and when to invest.

[0127] The described machine learning based analysis based on user session data allows for a closer correlation with geographic data analysis results, as this can be applied on a per UE session level.

[0128] Since the described semi-supervised approach is used which can start with a smaller subset of labeled examples and then use the remaining samples for training, there is no need to manually label a large amount of data, ie whether a user session / cell needs to be upgraded or not.

[0129] The described combined classifier network approach complements the knowledge of domain experts so that unlabeled data is also automatically labeled and used for recommendations, thus improving the experts' perception.

[0130] Although embodiments are described with respect to upgrading LTE cells using M-MIMO technology, embodiments may be used to consider other upgrade paths. For example, alternative embodiments may be used in the rollout of 5G cells by recommending which cells will benefit most from M-MIMO at different frequencies specified by 5G.

[0131] The described apparatus and / or methods can be instantiated in a cloud environment such as Docker, Kubenetes or Spark. Alternatively, they can be implemented in dedicated hardware or within the device being evaluated.

[0132] With the benefit of the teachings presented in the above description and the associated drawings, modifications and other variations of the described embodiments will occur to those skilled in the art. Therefore, it should be understood that the embodiments are not limited to the specific examples disclosed, and modifications and other variations are intended to be included within the scope of the present disclosure. Although specific terms may be used herein, they are used only in a general and descriptive sense and not for limiting purposes.

Claims

1. A method for determining a cell upgrade metric, the method comprising: receiving samples including user session data from user equipment communicating within the cell; obtaining training data from the samples by applying a filtering rule; training an autoencoder using the samples; combining the encoder layer and the code layer of the trained autoencoder with the classifier network to generate a combined classification network; Using the training data to train the combined classification network; The trained combined classifier network is used to classify user session data within the same or different cells, and a plurality of the classifications are used to determine the cell upgrade metric for the cell.

2. The method according to claim 1, wherein: The samples include cell metrics, the cell metrics being time-correlated with the user session data.

3. The method according to claim 2, wherein: The user session data is derived from cell tracking record (CTR) data from the user equipment.

4. The method according to claim 3, wherein: The user session data is characterized within a sampling period to generate corresponding samples.

5. The method according to claim 1, wherein: The filtering rule includes determining whether a specified metric with the sample exceeds a corresponding threshold.

6. The method according to claim 5, wherein: The specified metrics include performance parameters related to multiple-input multiple-output (MIMO) operations.

7. The method according to any one of claims 1 to 6, comprising: classifying user sessions of multiple cells using the trained combined classifier network and ranking the cells using corresponding cell upgrade metrics; The rankings are adjusted using geolocation data from the corresponding cells.

8. The method according to any one of claims 1 to 6, comprising: A performance metric improvement by upgrading the or each cell is estimated.

9. A method for recommending a cell upgrade, the method comprising: receiving user session data and cell metrics during an evaluation period; generating samples by characterizing user session data for each sampling period and time-correlating the characterized user session data with cell metrics, the user session data being from user devices communicating within the cell; using the trained classifier network to classify the sample as being capable of being improved by upscaling, along with a confidence score; The cell upgrade is recommended according to the proportion of samples classified as capable of improvement and the confidence score.

10. The method according to claim 9, wherein: The combined confidence score is the percentage of samples that fall within the upper and lower ranges of the confidence score.

11. The method according to claim 9, comprising: ranking the plurality of cells according to respective proportions of samples classified as capable of improvement and respective confidence scores; and adjusting the ranking using geolocation data from the corresponding cells.

12. The method according to any one of claims 9 to 11, comprising: An improvement in a performance metric when the or each cell is upgraded is estimated.

13. An apparatus for determining a cell upgrade metric, the apparatus comprising a processor and a memory, the memory containing instructions executable by the processor, whereby the apparatus operates to: receiving samples including user session data from user equipment communicating within the cell; obtaining training data from the samples by applying a filtering rule; training an autoencoder using the samples; combining the encoder layer and the code layer of the trained autoencoder with the classifier network to generate a combined classification network; Using the training data to train the combined classification network; The trained combined classifier network is used to classify user session data within the same or different cells, and a plurality of the classifications are used to determine the cell upgrade metric for the cell.

14. The apparatus of claim 13, operable to: The user session data within a sampling period is characterized to generate corresponding samples, wherein: The user session data is derived from cell tracking record CTR data from the user equipment; as well as The samples are time correlated with the cell metrics.

15. The device according to claim 13, wherein: The filtering rule includes determining whether a MIMO-related performance metric within the sample exceeds a corresponding threshold.

16. An apparatus according to any one of claims 13 to 15, operable to: classifying user sessions of a plurality of cells using the trained combined classifier network and ranking the cells using corresponding cell upgrade metrics; The rankings are adjusted using geolocation data from the corresponding cells.

17. Apparatus according to any of claims 13 to 15, operable to estimate an improvement in a performance metric when the or each cell is upgraded.

18. An apparatus for recommending cell upgrade, the apparatus comprising a processor and a memory, the memory containing instructions executable by the processor, so that the apparatus operates to: receiving user session data and cell metrics during an evaluation period; generating samples by characterizing user session data for each sampling period and time-correlating the characterized user session data with cell metrics, the user session data being from user devices communicating within the cell; using the trained classifier network to classify the sample as being capable of being improved by upscaling, along with a confidence score; The cell upgrade is recommended according to the proportion of samples classified as capable of improvement and the confidence score.

19. A computer program comprising instructions which, when executed on a processor, cause the processor to perform the method of any one of claims 1 to 12.

20. A computer program product comprising a non-transitory computer readable medium having stored thereon the computer program according to claim 19.

Images