JP6061838B2 - Communication quality degradation factor analyzer - Google Patents

Description

  The present invention relates to a communication quality deterioration factor analyzing apparatus, and more particularly to a communication quality deterioration factor analyzing apparatus for evaluating a communication quality state of an IP network.

  As communication networks are widely used, there is an increasing demand for communication quality assurance. Accordingly, it has become important to measure the communication quality to grasp the current communication quality state, and to identify the factor when quality deterioration is detected. As a direct method for grasping quality such as end-to-end delay time, packet loss rate, and throughput, there is a method of actively measuring quality by transmitting and receiving test packets between end-ends. For example, there are round-trip delay measurement using a ping command, other pathchars, and the like (for example, see Non-Patent Document 1).

  In addition, the round-trip time and packet loss rate of test packets and the bottleneck bandwidth on the path are measured between end hosts of interest, and communication is actually performed using TCP (Transmission Control Protocol). A method of analyzing in detail whether or not the TCP throughput increases when the TCP window size is gradually increased has been proposed (for example, see Non-Patent Document 2).

  Also, instead of measuring the quality status for each individual component (link, router, etc.) in the network, if test packets are sent between multiple source / destination node pairs and the packets do not reach their destination There is a method of determining that any node or link on the communication path is out of order (for example, see Non-Patent Documents 3 and 4). In the method, when a failure is detected in a plurality of node pairs, for example, a link that can cover all of the failure detection pairs with as few links as possible is extracted as a failure link candidate. If the failed link cannot be identified, perform as few additional measurements as possible.

  There has also been proposed a method for estimating the quality of each link in the network from the quality measurement (packet loss rate and delay) of the test packet at the source and destination node pairs. For example, when test packets are measured by unicast communication between multiple node pairs, the quality of each link may not be uniquely determined. The problem is solved by sending (see, for example, Non-Patent Document 5).

  Further, assuming that at most k of the n links have deteriorated in quality, studies on link quality estimation and identification of quality deteriorated links are being conducted (for example, Non-Patent Documents 6, 7, 8, 9). Non-Patent Document 6 derives a condition for routing of a network that can identify a quality degradation link particularly when k = 1. In Non-Patent Document 7, the number of necessary measurement paths is derived.

  As another study, there is a method of estimating the delay distribution of each link from the delay measurement of the arrival / departure node pair (see, for example, Non-Patent Document 10). Here, it is assumed that the delay of link i is an exponential distribution of parameters λi (the case of mixed distribution is also handled), and the value of λi of each link is assumed to be different, and parameter estimation is performed using an EM (Expectation Maximization) algorithm. In addition, there is a method for estimating the packet loss rate of a link from multicast measurement results on a plurality of trees (see, for example, Non-Patent Document 11).

http://www.caida.org/tools/utilities/others/pathchar/ Kazufumi Matoba, Shingo Ata, Masayuki Murata, "A Bottleneck Identification Method Based on Measurements on the Internet," The Institute of Electronics, Information and Communication Engineers, Telecommunications Management Study Group, pp. 65-70, November 2000. R. R. Komplella et al., "Detection and Localization of Network Black Holes," Infocom 2007. P. Barford et al., "Network Performance Anomaly Detection and Localization," Infocom 2009. M. Coates et al., "Internet tomography," IEEE Signal Processing Magazine, May 2002. M. H. Firooz et al., "Network tomography via compressed sensing," Globecom 2010. W. Xu et al., "Compressive sensing over graphs: how many measurements are needed?", Forty-Eighth Annual Allerton Conference, 2010. T. Matsuda et al., "Link quality classifier with compressed sensing based on l1-l2 optimization," IEEE Communication Letters, Vol. 15, No. 10, Oct. 2011. Takemoto, Matsuda, Takine, "Low quality link detection method by network tomography using compressed sensing," IEICE Tech. Y. Xia et al., "Inference of Link Delay in Communication Networks," IEEE JSAC, Dec. 2006. T. Bu, N. Duffield, F. L. Presti, and D. Towsley, "Network tomography on general topologies," SIGMETRICS 2002.

  The above method is mainly intended to estimate the link quality of the network using the information obtained by sending the test packet over the network and the routing information through which link the test packet passes in the network. Yes.

  On the other hand, with the recent increase in functionality and diversification of terminals and the diversification of network usage forms, communication quality using mobile terminals in particular depends on the application, wireless communication environment, terminal location, access destination server, terminal at that time The OS type and version, and the access network used (such as WiFi or 3G) are considered to depend greatly.

  In the existing method (quality estimation method closed in the network) described above, based on such various communication conditions, it is not possible to grasp the communication quality status of each terminal and analyze the cause when quality deteriorates There was a problem.

  The present invention has been made in view of the above points, and collects communication quality information and attribute information (terminal model, OS, application, terminal location, access network, etc.) of each terminal and simultaneously observes the communication quality state. An object of the present invention is to provide a communication quality deterioration factor analyzing apparatus that identifies a deterioration factor when quality deterioration is detected.

According to one aspect, in a communication network, a communication quality degradation factor analyzer that collects communication quality information of each terminal and identifies a degradation factor when detecting communication quality degradation,
End-end communication quality information of each terminal (hereinafter referred to as “communication quality information”) y and attribute information related to communication x → (where x → is a vector

を表す）からなる観測データを収集し、該通信品質情報yと該属性情報x→の組(x→,y)の集合を集合Sとし、記憶手段に格納する情報解析手段と、
前記記憶手段から集合Sを読み出して、前記通信品質情報yが予め定めた目標値を満たしているかを判定し、満たしている場合には「正常状態」とし、満たしていない場合は「劣化状態」とする品質状態把握手段と、
前記「正常状態」にある通信品質情報と属性の組（x→,y）を抽出し、その集合を集合Soとし、該集合Soを用いて、正常状態における品質yと属性x→の関係式y=f(x→)を推定する正常時品質推定手段と、
前記集合Sの中に、劣化状態にある観測データの組（x→,y）が存在する場合、該データと同時期に観測された他の観測データを抽出し、その集合を集合S_tとし、前記f(x→)により正常状態品質を推定し、前記通信品質情報yと該f(x→)を用いて、該集合S_tに属する各データに対して、品質劣化増加分y'を計算し、以上の手順で得られた(x→,y')の組の集合を用いて、各属性に起因する品質劣化要因を推定する品質劣化要因推定手段と、を有する通信品質劣化要因分析装置が提供される。

Information analysis means for collecting the observation data consisting of the communication quality information y and the attribute information x → (x →, y) as a set S, and storing in the storage means;
The set S is read from the storage means, and it is determined whether the communication quality information y satisfies a predetermined target value. Quality status grasping means, and
A set (x →, y) of communication quality information and attributes in the “normal state” is extracted, and the set is set as a set So. a normal quality estimation means for estimating y = f (x →);
In the set S, the set of observation data in the deteriorated state (x →, y) if exists, extracts the other observation data observed in the data the same time, and the set as a set S _t the f (x →) by estimating the normal state quality, by using the communication quality information y and the f (x →), for each data belonging to the set S _t, a quality deterioration increment y ' Quality degradation factor estimation means for estimating quality degradation factor due to each attribute using a set of (x →, y ') pairs obtained by the above procedure and calculated, and communication quality degradation factor analysis An apparatus is provided.

  According to one aspect, communication quality information and attribute information (terminal model, OS, application, terminal location, access network, etc.) of each terminal are collected, the communication quality state is observed, and the deterioration factor is detected when quality deterioration is detected. It is possible to estimate.

It is a figure which shows the relationship of embodiment of this invention. It is an example of the basic composition of the communication network in the 1st Embodiment of this invention. It is an example of a structure of the communication quality degradation factor analyzer in the 1st Embodiment of this invention. It is a flowchart of the communication quality degradation factor analyzer in the 1st Embodiment of this invention. It is an example of estimation of the quality degradation increment vector in the 1st Embodiment of this invention. It is an example of a structure of the communication quality degradation factor analyzer in the 2nd Embodiment of this invention. It is a structural example of the communication quality degradation factor analyzer in the 3rd Embodiment of this invention. It is an example of a structure of the communication quality degradation factor analyzer in the 6th Embodiment of this invention. It is an example of the function F value in the 6th Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  In the following, as shown in FIG. 1, the estimation of the quality deterioration increment vector is performed in the first embodiment, the evaluation value of the estimated vector is calculated in the second embodiment, and the extension ( The quality deterioration factor vector is estimated from the quality deterioration increment vector), and the evaluation value of the estimated vector is calculated. In the fourth embodiment, the extension (quality deterioration factor vector is estimated from the quality deterioration increment vector) is performed. In the embodiment, observation data to be input (communication quality information and attribute pair x →, y)) (where x → is a vector

を表す）の構成方法を、第６の実施の形態では、評価値を用いたパラメータの調整（複数の推定ベクトルの候補から良いものを選択）を説明する。

In the sixth embodiment, adjustment of parameters using evaluation values (selecting a good one from a plurality of estimated vector candidates) will be described.

[First Embodiment]
FIG. 2 shows an example of a basic configuration of a communication network according to the first embodiment of the present invention.

  FIG. 3 shows a configuration example of the communication quality deterioration factor analyzer in the first embodiment of the present invention.

  The communication quality degradation factor analysis device 1A shown in FIG. 1 includes an information analysis unit 11, an information management unit 12, a quality state grasping unit 13, a normal quality estimation unit 14, and a quality degradation factor estimation unit 15. Further, the information management unit 12 has a management table 121.

  An outline of the present embodiment will be described.

  First, as shown in FIG. 2, after each terminal 2 performs communication, it is assumed that the quality information of the communication (for example, data is downloaded from a certain server 3, and end / Measurement of end quality (delay etc.)) and attribute information related to the communication are transmitted to the communication quality degradation factor analyzer 1A. Examples of attributes include the terminal model, OS, application, terminal location, access network (WiFi or 3G, etc.), server 3 location, and the like.

The communication quality degradation factor analysis device 1A sets the set of the observation data (communication quality information and attribute pair (x →, y)) collected from each terminal 2 as S, and the quality state grasping unit 13 Among them, it is checked whether or not the quality y satisfies a predetermined target value, and the case where the quality is satisfied is regarded as a normal state, and the case where it is not satisfied is regarded as a deteriorated state. A set of quality and attributes in a normal state (x →, y) is extracted, and the set is set as _So and transferred to the normal quality estimation unit 14. For example, S _o is a set of data observed in the same most busy time zone (time width T _n , eg, T _n = 1 hour) on the same day of the week, and all the observed values in that time zone seem to be in a normal state. (A set of data that can be regarded as a steady state in constructing a quality estimation formula of y = f (x →) below). The normal quality estimation unit 14 estimates the relational expression y = f (x →) between the communication quality information y and the attribute x → in the normal state using _So.

Next, when there is data (x →, y) in a deteriorated state in S, the quality deterioration factor estimating unit 15 receives other data observed at the same time (time width T _r ) with the data. to extract, to the set and S _t. S _o corresponding to the time zone when the set _St was observed (for example, if S _t is 1:00 p.m. on Friday of Y month Z), S _o at 1 p.m. every Friday in Y month) The normal state quality is estimated from the constructed f (x →), and the quality deterioration increment y ′ is calculated for each data belonging to _St using y and f (x →). For example, if the quality is delay, y ′ = y−f (x →). Using the set of (x →, y ′) pairs obtained by the above procedure, the quality degradation increment due to each attribute is estimated.

  FIG. 4 shows a flowchart of the communication quality deterioration factor analyzer in the first embodiment of the present invention.

  When the quality / attribute information arrives from the terminal 2 (step 1 in FIG. 4), the information analysis unit 11 reads the quality information and attribute information, and constructs an attribute vector from the attribute information in the following procedure (step 2 in FIG. 4). ).

Number attribute and d pieces there, classifies the possible values of the attribute i in m _i number of categories. For example, if the attribute i is an access network and the categories are WiFi (registered trademark) and 3G, m _i = 2. At this time, the attribute vector x → of the terminal is expressed by Expression (1).

ここで、x(,j)は、該端末の属性iのカテゴリがjのとき"1"、そうでなければ"0"をとる変数とする。例えば属性i=アクセスネットワークの例で、WiFiを1番目のカテゴリ、３Gを2番目のカテゴリとする。ある端末のアクセスネットワークがWiFiの場合、該端末２のx(i,j)の値は、 x(i,1) =1、 x(i,2)=0となる。上記属性ベクトルx→と該端末２の通信品質yの組(x→,y)を情報管理部１２へ送信する。

Here, x (, j) is a variable that takes “1” when the category of the attribute i of the terminal is j and “0” otherwise. For example, in the example of attribute i = access network, WiFi is the first category and 3G is the second category. When the access network of a certain terminal is WiFi, the value of x (i, j) of the terminal 2 is x (i, 1) = 1 and x (i, 2) = 0. A set (x →, y) of the attribute vector x → and the communication quality y of the terminal 2 is transmitted to the information management unit 12.

  The information management unit 12 adds a time stamp to the observation data (x →, y) and then enters the management table 121 (step 3 in FIG. 4).

  After a certain period of time, the quality state grasping unit 13 determines the quality y in advance from the set of (x →, y) entered in the management table 121 in the information management unit 12 (this set is S). Whether the target value is satisfied is checked. If the quality is satisfied, the normal state is set, and if not, the deteriorated state is set (step 4 in FIG. 4). When the state classification is completed, this is notified to the normal quality estimation unit 14.

The normal quality estimation unit 14 extracts a set (x →, y) of communication quality information and attributes in a normal state, and sets the set as So. So is, for example, a set of data observed in the same most busy time of the same day (time width T _n , eg, T _n = 1 hour), and the observed values in that time zone are all in a normal state. A set of data (a set of data that can be regarded as a steady state in constructing a quality estimation formula of y = f (x →) below). Using So, the relational expression y = f (x →) between the communication quality information y and the attribute x → in the normal state is estimated by the following procedure (step 5 in FIG. 4). That is, for example, So is configured for each day of the week and time period, and f (x →) corresponding to each So is constructed.

  f (x →) is defined as the following equation (2).

ここで、a₀， a_i,jは係数とする。集合Soに属する観測データ(x→,y)がN個あるとし、k番目の観測値を(x→_k,y_k)とし、最小二乗法により該係数を推定する。つまり、式(3)を最小にする該係数を計算する。

Here, a ₀ , a _{i, j} are coefficients. Assume that there are N pieces of observation data (x →, y) belonging to the set So, the k-th observation value is (x → _k , y _k ), and the coefficient is estimated by the least square method. That is, the coefficient that minimizes Equation (3) is calculated.

品質が遅延の場合は、観測値（通信品質情報）yをそのまま用いて上記手順を実施する。品質がパケット損失率の場合は、パケット損失率yを−log(1−y)と変換してから上記手順を実施する。

When the quality is delay, the above procedure is performed using the observed value (communication quality information) y as it is. When the quality is the packet loss rate, the above procedure is performed after converting the packet loss rate y to -log (1-y).

Next, when there is data (x →, y) in a degraded state in the set S in the information management unit 12, the quality degradation factor estimation unit 15 simultaneously with the data (time width T _r ). extract the observed other data, and the collection and S _t. Of f (x →) calculated by the normal quality estimation unit 14, So corresponding to the time zone in which the set _St is observed (for example, _St is 1:00 pm on Friday, Y month Z) if any, by using a Y month weekly built by Friday afternoon 1 o'clock So) the f (x →), for each data belonging to the S _t, estimates the quality degradation increment y 'by the following steps (Step 6 in FIG. 4).

  If the target value for observation value y is y * and observation value y is y> y * (that is, the observation value is in a quality degradation state), calculate quality degradation increment y 'by y' = y-f (x →) To do. On the other hand, y <y * (that is, the observed value in the quality maintaining state) is set to y ′ = 0.

Now, the observation data in the set S _t is the n pieces there, to the k-th y 'and x → set of (x → _k, y' and _k). Using (x → _k, y ′ _k ), a quality degradation increment vector (a vector having quality degradation increment z _{i, j} attributed to the jth category of attribute _i as an element) is estimated by the following procedure (FIG. 4 step 7).

Estimate the value of z _{i, j} as a variable that minimizes Equation (4). However, the symbols used in the formula (4) are defined by the formulas (5) to (7).

また、式(4)において、

Also, in equation (4):

はベクトルwに対するl2ノルムといい、

Is the l2 norm for the vector w,

である（w_iはベクトルwのi番目の要素）。

Where w _i is the i-th element of the vector w.

はベクトルwに対するl1ノルムといい、

Is the l1 norm for the vector w

である。またλは、上記の式（4）のl2ノルムに対するl1ノルム対する重み係数であって、l1ノルム最小化にどれだけ重みを与えるかを決めるパラメータであり、予め定める正の定数とする。

It is. Λ is a weighting coefficient for the l1 norm with respect to the l2 norm in the above equation (4), and is a parameter that determines how much weight is given to the l1 norm minimization, and is a positive constant determined in advance.

  Here, an example to which this embodiment is applied is shown. A plurality of subjects use various mobile terminals to access various sites from various locations, and use data obtained by performing experiments in which delay times at the application level at that time are totaled as quality. The number of observed communication connections is 574. The attribute vector in each communication has a total of 9 dimensions including the number of attributes and the number of categories for each attribute. The quality target value y * was set to a certain value, and the number of communication connections determined to be normal was 348. Using these 348 pieces of data, the relational expression between quality and attributes in the normal state was estimated by the procedure of step 5 in FIG. Next, 574 pieces of observation data are arranged in chronological order, and data is extracted by a sliding window (window size = 20). In other words, the i-th window includes 20 data items from i-th to i + 19-th (i = 1 to 554). In the window, it is checked whether there is any data whose quality is in a deteriorated state and data in a normal state and all the attribute vectors of each data are the same. The quality deterioration increment vector z is estimated by the procedure of step 7 in FIG. In this case, non-patent document 12 (S. -J. Kim et al., “An interior-point method for large-scale 1l-regularized least squares,” is used to calculate z that minimizes Equation (4). IEEE Journal of Selected Topics in Signal Processing, Dec. 2007.). The result is shown in FIG. In the figure, the value of each element of the quality deterioration increment vector z in each window (“time” in the figure) is plotted. Thus, for example, in the windows around 400, it is possible to estimate that the attribute corresponding to the third element of the attribute vector is a factor of quality degradation.

[Second Embodiment]
In the present embodiment, calculation of an evaluation value of an estimated vector will be described.

FIG. 6 shows a configuration example of a communication quality deterioration factor analyzer in the second embodiment of the present invention.
In the figure, the same components as those in FIG.

  The communication quality degradation factor analysis device 1B shown in the figure has a configuration in which an evaluation unit 16a is added to the configuration of FIG.

  In the present embodiment, the evaluation unit 16a performs the quality deterioration increment vector z → (where z → is the value of the quality deterioration factor estimation unit 15 of the first embodiment,

を表す）を用いて、集合S_tに属するk番目の通信に対する観測データx→_kに対して、x→_k・z→+f(x→_k)により該通信の品質を推定し、それをy_k_estとする。ここで、x→_k・z→ はx→_kとz→の内積を意味する。

For the observation data x → _k for the _kth communication belonging to the set _St , the quality of the communication is estimated by x → _k · z → + f (x → _k ) Let y _k _est. Here, x → _k · z → means an inner product of x → _k and z →.

This is performed for each observation data x → _k , and the ratio of y _k _est ≥ y * in the observation data where y _k <y * (that is, data in the normal state)
A = P [y _k _est ≧ y * | y _k <y *],
And the ratio of y _k _est <y * among the observation data (ie, data in a degraded state) where y _k ≧ y *
B = P [y _k _est <y * | y _k ≧ y *]
, And A is calculated as a false detection rate (a rate at which quality deterioration is erroneously detected) based on the quality deterioration increment vector z →, and B is calculated as a miss rate (a rate at which quality deterioration cannot be detected) due to z →.

[Third Embodiment]
In the present embodiment, an example will be described in which the quality degradation factor vector z ′ → is estimated from the quality degradation increment vector z → estimated by the quality degradation factor estimation unit 15.

  FIG. 7 shows a configuration example of a communication quality deterioration factor analyzer in the third embodiment of the present invention. In this figure, the same components as those in FIGS. 3 and 6 are denoted by the same reference numerals, and description thereof is omitted.

  The communication quality degradation factor analysis device 1C shown in FIG. 7 has a configuration in which a quality degradation factor vector estimation unit 17 is added to the configuration of FIG.

  The quality degradation factor vector estimation unit 17 converts the value of each of the top H elements having the largest value among the elements of the quality degradation increment vector z → output from the quality degradation factor estimation unit 15 to 1, Elements other than form a vector z ′ → whose value is converted to zero. The value of H may be set to a predetermined value or, for example, the average avgz of each element of z → may be calculated, and the number of elements that exceed avgz among the elements of z → may be H. Further, as shown in a sixth embodiment described later, a plurality of patterns may be prepared and selected from them.

  The vector z ′ → is referred to as a quality deterioration factor vector. For example, if z → = [100, 0.5, 0.2, 51] and H = 2, z ′ → = [1,0, 0, 1]. An element that becomes 1 in z ′ → means an attribute that causes quality degradation. When z ′ → = [1,0,0,1], this corresponds to the fact that category 1 of attribute 1 and category 2 of attribute 2 are determined as quality degradation factors.

Evaluation unit 16b, the use of quality deterioration factor vector z '→ of, with respect to the observation data z → _k belonging to the set S _t in the first embodiment, the above formula (8) I (z → _k , z '→).

これを各x_kに対して実施し、y_k<y*である観測データのうちI(x→_k, z'→)が1となる割合

This is performed for each x _k , and the ratio of I (x → _k , z '→) is 1 in the observation data where y _k <y *

およびy_k≧y*である観測データのうちI(x→_k, z'→)が0となる割合

And the proportion of I (x → _k , z '→) is 0 among the observation data with y _k ≧ y *

を計算し、Aを品質劣化要因ベクトルz'→による誤検出率、Bをz'→による見逃し率として算出する。

A is calculated as the false detection rate due to the quality deterioration factor vector z ′ →, and B as the miss rate due to z ′ →.

[Fourth Embodiment]
In the present embodiment, a quality deterioration factor vector different from that in the third embodiment is estimated.

  The configuration of the communication quality deterioration factor analyzer is the same as that of the communication quality deterioration factor analyzer 1C of FIG.

In the present embodiment, instead of setting each element of the quality degradation factor to “0” or “1” as in the third embodiment, the quality degradation factor vector estimation unit 17 uses the first embodiment. The probability p _j that is a quality degradation factor is determined in accordance with the value of each element z _j of the quality degradation increment vector z → obtained by the quality degradation factor estimation unit 15 of, and a vector having p _j as the j-th element is determined. The quality deterioration factor vector z ′ → is configured. Here, the value of p _j is set so as to increase as the value of z _j increases. For example, if the quality is delayed, the constant in the z0, (and z0 <z _j) to an amount corresponding to the processing time required to process one packet. For simplicity, when the delay time follows M / M / 1 of the queuing model, the average delay time T is T = z0 / (1−ρ). Here, ρ corresponds to the server usage rate in the queue. On the other hand, since the probability that the number of jobs waiting to be processed on the server (including the job being processed) will be k is (1−ρ) ρ ^k , the probability that the number of customers in the system will exceed a certain threshold value k0 is about ρ ^k0. . Now, assuming that the value of the element z _j is an average delay due to the corresponding attribute, z _j = z0 / (1−ρ) and ρ = 1−z0 / z _j . Here, ρ ^k0 is used as p _j . That is, p _j = (1−z0 / z _j ) ^k0 is set. Alternatively, for a predetermined value p0 (0 <p0 <1), if (1−z0 / z _j ) ^k0 > p0, p _j = (1−z0 / z _j ) ^k0 , otherwise p _j = It may be 0.

Evaluation unit 16b, the above quality deterioration factor vector z 'with respect to the attribute vectors x _k belonging to the set S _t in the first embodiment, P of the formula _{(9) (x → k,} z' →) is calculated.

ここで、x_jはベクトルx→_kの第j要素とし、Dはベクトルx→_kの次元数であり、
D=m₁+m₂+…+m_d
である。P(x→_k, z'→)はk番目の通信が品質劣化している確率を意味する。

Here, x _j is the j-th element of the vector x → _k, D is the number of dimensions of the vector x → _k,
D = m ₁ + m ₂ +… + m _d
It is. P (x → _k , z ′ →) means the probability that the k-th communication has degraded quality.

This is performed for each x → _k , and the ratio of P (x → _k , z '→) that is greater than or equal to the predetermined threshold th in the observation data where y _k <y *

およびy_k≧y*である観測データのうちP(x→_k, z'→)がth未満となる割合

And the proportion of observation data with y _k ≧ y * that P (x → _k , z '→) is less than th

を計算し、Aをz'→による誤検出率、Bをz'→による見逃し率として算出する。

And A is calculated as a false detection rate due to z ′ → and B as an overlook rate due to z ′ →.

[Fifth Embodiment]
In the present embodiment, a configuration method of observation data input to the communication quality degradation factor analyzer 1 will be described.

  It is assumed that the communication quality degradation factor analyzer in the present embodiment has the same configuration as that in FIG. In the present embodiment, the information analysis unit 11 groups observation data having the same attribute vector.

In this embodiment, the observation data (x → _k, y _k) belonging to the set S _t in the first embodiment with respect to classify the observed data with the same attribute vector x → _k in the same group. For example, it is assumed that the number of attributes d is 2 and the attributes 1 and 2 each have two categories. At this time, the attribute vector is four-dimensional. If there are n = 5 observation data, x → ₁ = [1, 0, 0, 1], x → ₂ = [1, 0, 1, 0], x → ₃ = [1, 0, 0, 1], x → ₄ = [1, 0, 1, 0], x → ₅ = [0, 1, 1, 0]. At this time, x → ₁ , x → ₃ is classified into _three groups, x → ₂ , x → ₄ is group 2, x → ₅ is classified into three groups. The group quality representative value is calculated for each group. Typical values are average, median, 95th percentile, etc. In the above example, the representative value (average) of group 1 is (y ₁ + y ₃ ) / 2. Replace the attribute vector of group k with x → _k and the quality representative value with y → _k .

(X → _k , y _k ) obtained by the above procedure is applied to the quality state grasping unit 13 in the first to fourth embodiments to perform subsequent processing, or the second to fourth implementations It is possible to perform the subsequent processing by applying it to the evaluation unit 16 in the form.

[Sixth embodiment]
In the present embodiment, an example will be described in which a parameter is adjusted using an evaluation value (a good one is selected from a plurality of estimated vector candidates).

  FIG. 8 shows the configuration of a communication quality deterioration factor analyzing apparatus according to the sixth embodiment of the present invention.

  The configuration shown in FIG. 6 is obtained by adding a parameter adjustment unit 18 to FIGS.

  In the present embodiment, several patterns are prepared for the value of λ in the first embodiment, and for each value of λ, the evaluation units 16a and 16b in the second to fourth embodiments respectively The false detection rate A and the miss rate B are calculated, and the function F (A, B) of A and B is configured by the parameter adjustment unit 18, and F (A, B) takes a larger value as A or B becomes smaller. Then, the value of λ that maximizes F (A, B) is obtained, and the quality deterioration factor estimating unit 15 outputs the quality deterioration factor vector z → or the quality deterioration factor vector z ′ → when using the λ.

  Alternatively, in the quality deterioration factor vector estimation unit 17 in the third embodiment, several patterns of H values are prepared, and the parameter adjustment unit 18 sets the value of H that maximizes F (A, B) in the same procedure. Then, the quality deterioration factor vector estimation unit 17 outputs the quality deterioration factor vector z ′ → when H is used.

  Alternatively, the quality deterioration factor vector estimation unit 17 of any of the third and fourth embodiments forms a quality deterioration factor vector z ′ →, and converts some of the non-zero elements of the vector to 0 The vector configured as described above is set as a quality degradation factor vector candidate z ″ →, several patterns are prepared for z ″ →, and z ″ → which maximizes F (A, B) configured by the parameter adjustment unit 18 is prepared. Is output as a quality deterioration factor vector.

Here, as an example of F (A, B), there is an evaluation scale called F value (for example, see Non-Patent Document 13: http://en.wikipedia.org/wiki/Precision_and_recall). First, TP, FN, FP, and FN are defined as follows.
・ TP: Proportion of data with true positive (quality degradation) and positive estimation result ・ FN: Proportion of true positive but negative estimation (normal) ・ FP: True Ratio of negative status but positive estimation result ・ TN: Proportion of true negative and negative estimation result At this time, the false positive rate A is A = FP / (FP + TN), and the missed rate B is
B = FN / (FN + TP)
It becomes. Here, if the ratio that the true state is negative is RN = FP + TN, FP = A × RN. If the ratio that the true state is positive is RP = FN + TP, then FN = B × RP.

RN and RP use the k-th communication quality information y _k belonging to the set St,
RN = P [y _k <y *], RP = P [y _k ≧ y *]
(These are quantities that do not depend on the estimation result).

On the other hand, Precision Precision is Precision = TP / (TP + FP), recall recall is
Recall = TP / (TP + FN)
And the harmonic mean F of these
F = 2 × Precison × Recall / (Precision + Recall)
It is defined as When F is expressed by A and B, it becomes as follows. First, Recall = 1−B.

Next, considering Precision,
TP = RP−FN = (1−B) RP
Than,
Precision = (1−B) RP / ((1−B) RP + A × RN)
It becomes. Therefore, F can be expressed as a function of A and B, and it is used as F (A, B).

Here, an example to which this embodiment is applied is shown. The same data as the application example shown in the first embodiment is used. First, the quality deterioration factor vector z ′ → is constructed by the method of the third embodiment. At this time, H used in the third embodiment calculates the average avgz of each element of z →, and the number of elements exceeding avgz among the elements of z → is H. Among the non-zero elements of the vector z ′ →, a vector configured by converting some elements to 0 is defined as a quality degradation factor vector candidate z ″ →, and several patterns of z ″ → are prepared (z ′ Of the number of combinations of non-zero elements). For each pattern, after the procedure of grouping the observation data of the fifth embodiment, the F value defined above is calculated, and z ″ → that maximizes the F value is defined as the quality deterioration factor vector. The value of the F value in each window at this time is shown in the upper part of FIG. Here, the quality degradation factor vector was calculated for a window in which both observation data where y _k <y * and observation data where y _k ≧ y * exist. This figure shows the result when the calculation of the vector is complete (the calculation of the vector is complete (specifically, the result obtained by converging the solution of Equation (4)) is 99%. there were). As comparison, (since 9-dimensional, 2 ^nine pairs) all pairs of nine-dimensional binary vector seek vector to maximize the F value respect, also shown in FIG. 9 the lower value of F value at that time ( This F value corresponds to the optimal solution). From this, it can be confirmed that the upper part of FIG. 9 is able to obtain the F value close to the lower part of FIG.

  The processing of each component of the communication quality deterioration factor analyzer 1 shown in FIGS. 3, 6, 7 and 8 is constructed as a program, and is installed and executed on a computer used as the communication quality deterioration factor analyzer. Or can be distributed via a network.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

1, 1A, 1B, 1C, 1D Communication quality degradation factor analysis device 2 Terminal 3 Server 4 Access network 5 Core network 11 Information analysis unit 12 Information management unit 13 Quality state grasping unit 14 Normal quality estimation unit 15 Quality degradation factor estimation unit 16a, 16b Evaluation unit 17 Quality degradation factor vector estimation unit 18 Parameter adjustment unit

Claims (9)

A communication quality degradation factor analysis device that collects communication quality information of each terminal in a communication network and identifies a degradation factor when detecting communication quality degradation,
End-end communication quality information of each terminal (hereinafter referred to as “communication quality information”) y and attribute information related to communication x → (where x →

Information analysis means for collecting the observation data consisting of the communication quality information y and the attribute information x → (x →, y) as a set S, and storing in the storage means;
The set S is read from the storage means, and it is determined whether the communication quality information y satisfies a predetermined target value. Quality status grasping means, and
A set (x →, y) of communication quality information and attributes in the “normal state” is extracted, and the set is set as a set So, and the set So is used to set the communication quality information y and the attribute x → in the normal state. A normal quality estimation means for estimating the relational expression y = f (x →),
In the set S, the set of observation data in the deteriorated state (x →, y) if exists, extracts the other observation data observed in the data the same time, and the set as a set S _t estimates the normal state quality by the equation f (x →), using said communication quality information y and the f (x →), for each data belonging to the set S _t, quality deterioration increment y Quality degradation factor estimation means for estimating quality degradation factors caused by each attribute using a set of (x →, y ') pairs obtained by the above procedure.
A communication quality deterioration factor analyzing apparatus characterized by comprising: The normal quality estimation means includes:
When estimating the relation y = f (x →),
Number attribute and d pieces is, the possible values of the attribute i classified into m _i number of categories, the category j for the attribute i, x (i, j) when the category of the attribute i is j 1, Otherwise, it is a variable that takes 0, and the relational expression f (x) is

(Where a ₀ , a _{i, j} are coefficients),
Assume that there are N pieces of observation data (x →, y) belonging to the set So, the k-th observation value is (x → _k , y _k ), and by the least squares method,

Including means for calculating a coefficient that minimizes
The quality deterioration factor estimating means includes
The target value for the communication quality information y is y *. (For the observed value y> y * in the quality degradation state, the quality degradation increment y ′ is calculated by y ′ = y−f (x →) to maintain the quality. the observed value y <for y * is located in the state, y 'and = 0, the set S observation data in the _t is is n a, k-th y' and x → set a (x → _k of, y ' _k ) (k = 1, ..., n), and the quality degradation increment attributed to the jth category of attribute _i is z _{i, j}

(Where λ is a positive constant determined in advance)
Estimate the value of z _ij as a variable that minimizes

Defined by
The communication quality deterioration factor analyzing apparatus according to claim 1, further comprising means for outputting a vector z → having z _{i, j} as an element as a quality deterioration increment vector z →. Using been said quality deterioration increment vector z → estimated by the quality degradation factor estimating means, with respect to the observed data x → _k for the k-th communication belonging to the set _{S t, x → k · z} → + f ( x → _k ) (where x → _k · z → is the inner product of x → _k and z →), and the quality of the observed data (ie, data in the normal state) where y _k <y * Of these, the ratio of y _k _est ≧ y * A = P [y _k _est ≧ y * | y _k <y *], and the ratio of y _k ≧ y * in the degraded state where y _k _est <y * B = P [y _k _est <y * | y _k ≧ y *] is calculated, and the false detection rate A based on the quality deterioration increment vector z → and the rate over which the quality deterioration due to z cannot be detected are calculated as the miss rate B. The communication quality deterioration factor analyzer according to claim 2, further comprising one evaluation means. Among the elements of the quality deterioration increment vector z → estimated by the quality deterioration factor estimating means, the value is converted to 1 for each of the top H elements having a large value, and the values of the other elements are set to 0. Quality degradation factor vector estimation means constituting the converted quality degradation factor vector z ′ →;
Said using a quality deterioration factor vector z '→, for the attribute vector z → _k belonging to the set S _t,

I (x → _k , z '→) is calculated for the observed data with y _k <y *, and the ratio of I (x → _k , z' →) to 1 is A = P [I (x → _k , z '→) = 1 | y _k <y *] and the proportion of observation data with y _k ≧ y * where I (x → _k , z' →) is 0 B = P [I ( x → _k , z ′ →) = 0 | y _k ≧ y *], and a false detection rate A due to the quality degradation factor vector z ′ → and a rate over which the quality degradation due to z ′ → cannot be detected B The communication quality deterioration factor analysis device according to claim 2, further comprising: a second evaluation unit that calculates as follows. A probability p _j that is a quality degradation factor is determined according to the value of each element z _j of the quality degradation increment vector z → estimated by the quality degradation factor estimation means, and a vector having p _j as the j-th element Quality degradation factor vector estimation means for constructing as a quality degradation factor vector z ′ →,
Using the quality degradation factor vector z ′ →, for the attribute vector x → _k belonging to the set St,

K th communication calculates the probabilities are quality degradation _{P (x → k, z '} →) by (wherein, x _j is the j-th element of the vector x → _k, D is the number of dimensions of the vector x → _k D = m ₁ + m ₂ +... + M _d ), among the observation data with y _k <y *, P (x → _k , z ′ →) is the ratio A = P [P (x _k , z ′ →) ≧ th | y _k <y *] and the ratio B of P (x → _k , z ′ →) less than th among observation data with y _k ≧ y * = P [P (x _k , z ′) <th | y _k ≧ y *] is calculated, and a false detection rate A due to z ′ → and a rate at which quality degradation due to z ′ → cannot be detected are calculated as a miss rate B The communication quality deterioration factor analyzer according to claim 2, further comprising a third evaluation unit. For observation data belonging to the set St (x → _k , y _k ), classify observation data having the same attribute vector x → _k into the same group, calculate a group quality representative value for each group, and 6. The communication quality deterioration factor analysis device according to claim 1, further comprising communication quality classification means for replacing the attribute vector with x → _k and the quality representative value with y _k . The quality degradation factor vector estimation means includes:
Prepare several patterns of the value of λ, take a larger value as the error detection rate A or the miss rate B is smaller, constitute a function F (A, B) of the error detection rate A and the miss rate B, 5. A means for obtaining a value of λ that maximizes the function F (A, B) and outputting the quality deterioration increment vector z → or the quality deterioration factor vector z ′ → when the λ is used. Or the communication quality deterioration factor analyzer according to 5. The quality degradation factor vector estimation means includes:
Several values of the value of H are prepared, and the function F (A, B) of the false detection rate A and the miss rate B is configured to take a larger value as the false detection rate A or the miss rate B is smaller. 5. The communication quality deterioration factor analysis device according to claim 4, further comprising means for obtaining a value of H that maximizes the function F (A, B) and outputting a quality deterioration factor vector z ′ → when using the H. The quality degradation factor vector estimation means includes:
Among the non-zero elements of the quality deterioration factor vector z ′, a vector constituted by converting some elements to 0 is defined as a quality deterioration factor vector candidate z ″ →, and several patterns of z ″ → are prepared, A function F (A, B) of the false detection rate A and the miss rate B takes a larger value as the false detection rate A or the miss rate B becomes smaller, and the F (A, B) is maximized. 6. The communication quality deterioration factor analyzing apparatus according to claim 4 or 5, further comprising means for obtaining z ″ → and outputting it as a quality deterioration factor vector z ″ →.

Images