CN112702224B

CN112702224B - Method and device for analyzing quality difference of home broadband user

Info

Publication number: CN112702224B
Application number: CN202011437131.8A
Authority: CN
Inventors: 邢祥宇
Original assignee: Beijing Zznode Technology Co ltd
Current assignee: Beijing Zznode Technology Co ltd
Priority date: 2020-12-10
Filing date: 2020-12-10
Publication date: 2023-03-31
Anticipated expiration: 2040-12-10
Also published as: CN112702224A

Abstract

A method and a device for analyzing the quality difference of a home broadband user analyze the relation between KPI and KQI by taking a user perception index KQI as a medium, and analyze the relation between KQI and QoE to obtain the relation between KPI and QoE, thereby being beneficial to better and more practical evaluation of the quality difference condition of the home broadband user and early warning with higher accuracy.

Description

Method and device for analyzing quality difference of home broadband user

Technical Field

The invention relates to a family broadband technology, in particular to an analysis method and a device for the quality difference of a family broadband user.

Background

Operators often see user experience through network quality, and the user experience is not synchronously improved although the network quality is improved, so that the method for evaluating and optimizing the network quality is urgently changed. The traditional network evaluation and optimization method of an operator mainly uses a driver Test/Call Quality Test (DT/CQT) to evaluate the network Quality, and the network Quality and the user service perception have larger and larger scissors difference along with the high-speed development of the mobile internet service. Operators began to step towards introducing user awareness in network evaluation and optimization. In recent years, a DPI (Deep Packet Inspection) method has been emphasized, but there is a limitation that does not depart from the inherent mode of quality from the network.

The user perception is important, but the current means for acquiring the real user perception has limitations, and one of the most basic principles is that the closer and the more accurate the user is, so how to acquire the perception from the user side is particularly critical, namely, the user (terminal) looks at the network. Therefore, some operators try to perform network quality evaluation based on the large sensing data collected by the terminal side and list the network quality evaluation as an assessment index. The method has the advantages that on the basis of accurately mastering the user perception, the network service quality and the user satisfaction can be effectively improved, the user stickiness is improved, and the accurate marketing and customer service are further guided. In another method, the reasons of the whole network level with poor browsing service perception of the user are analyzed by tracking a plurality of key indexes influencing the user experience, and analysis and positioning delimitation are carried out, so that the bearing capacity of the mobile network to the browsing service is improved by optimizing network related parameters, algorithms, protocols and the like, and reference and basis are provided for improving the browsing service perception of the user. However, these methods cannot describe the perception difference of the user well, and the determination of the threshold value has difficulty. How to perform delimited positioning on the perception difference problem and guide network adjustment to improve service perception still lacks a systematic method. The current thought is to perform network evaluation by using the sensing big data collected by the terminal side, and a method for positioning and optimizing the network problem by using the sensing big data collected by the terminal side is not seen. And network problem positioning and optimization promotion mainly centers on utilizing DT/CQT and DPI.

The method for analyzing and optimizing the service perception degradation reasons mainly comprises a method for problem recurrence and deep troubleshooting based on DT/CQT, a data analysis method based on DPI and a method for analyzing the quality difference reasons by tracking a plurality of key indexes influencing user experience at present.

The DT/CQT-based method for problem recurrence and deep troubleshooting has the following defects: (1) The problem of non-fixed appearance is difficult to effectively reproduce, and the perception degradation problem of non-wireless network factors is difficult to find. At present, a large amount of DT/CQT dialing tests are carried out at problem sites of user complaints, on one hand, relevant parameters of a network at the position are collected, and meanwhile, the complaints of the user (such as call failure) are expected to be reproduced by the method. There is no good way to try to reproduce except a large number of dial tests. (2) high labor and time costs are incurred.

The data analysis method based on the DPI has the following defects: (1) The DPI deployment and operation cost is huge, and the DPI system cannot be developed in areas without the DPI system. (2) End-to-end delay sensing information cannot be accurately acquired and evaluated, and service sensing information cannot be acquired under the condition of no network interaction. And therefore, the perceptual degradation caused by such factors cannot be effectively analyzed and localized.

The method for analyzing the cause of the quality difference by tracking a plurality of key indexes influencing the user experience has the following defects: (1) The quality difference behavior is defined by adopting a fixed threshold value, and the perception difference of different users cannot be considered. And (2) extracting the quality difference index from the aspect of the characteristics of the service.

Disclosure of Invention

Aiming at the defects or shortcomings in the prior art, the invention provides a method and a device for analyzing the quality difference of a home broadband user, which take a user perception index KQI as a medium, analyze the relation between KPI and KQI, and analyze the relation between KQI and QoE, so as to obtain the relation between KPI and QoE, thereby being beneficial to better and more practical evaluation of the quality difference condition of the home broadband user and early warning with higher accuracy.

The technical solution of the invention is as follows:

an analysis method for home broadband user quality difference is characterized by comprising the following steps: step 1, obtaining network performance index KPI data and user perception index KQI data of a family broadband user, and establishing a relation between KPI and KQI; step 2, establishing a relation between the KQI and the QoE according to the QoE experience data of the user and the KQI data in the relation between the KPI and the KQI; and 3, indirectly determining the relationship between the KPI and the QoE by taking the KQI as an intermediary.

The step 1 comprises the following steps: a1, data acquisition, namely collecting historical data, wherein the historical data comprises machine performance data, network performance data and service data of four services, namely home broadband video, webpage, game and television, and is divided into KPI and KQI; step A2, data preprocessing, namely cleaning and preprocessing the historical data, wherein the preprocessing comprises the following steps: missing value processing, abnormal value processing, discrete data encoding and normalization processing; step A3, regression analysis, namely, for any first KQI index, a corresponding first KPI index set exists, the mapping relation between the first KQI index and the first KPI index set is determined through the regression analysis, the gradient lifting decision tree GBDT algorithm is adopted in the regression analysis, and the parameter in the regression analysis is adjusted to be smallerOptimal solution, and the design model evaluation process adopts a decision coefficient R ² (ii) a Step A4, sorting the contribution degrees of a first KPI index set aiming at the better solution of the better solution, and deleting KPI indexes with the contribution degrees lower than a first threshold value from the first KPI index set to form a new first KPI index set; step A5, performing regression analysis again by using the new first KPI index set and the first KQI index, and adjusting model parameters to a better solution, wherein the decision coefficient is a first decision coefficient; step A6, sorting the contribution degrees of the first KPI index set in the step A5, and deleting the index with the lowest contribution degree in the first KPI index set to be used as a new first KPI index set; step A7, carrying out regression analysis again by using the new first KPI index set and the first KQI index, and adjusting the model parameters to a better solution, wherein the decision coefficient is a second decision coefficient; step A8, if the second decision coefficient is smaller than the first decision coefficient and the absolute value of the difference is larger than a second threshold value, rejecting the deletion of the KPI, and turning to step A9, otherwise, receiving the deletion of the KPI, and turning to step A5; step A9, the processed new index set is a first core KPI index set CKPI corresponding to a first KQI index; step A10, performing regression analysis by using the first CKPI index set and the first KQI index set, adjusting parameters to a better solution, and evaluating the better solution by using a determination coefficient R ² The regression analysis model serves as a first user perception prediction model UPPM; and step A11, repeating the steps for each KQI until the goal of constructing the relationship between the KPI and the KQI is completed.

The KPI data in step A1 includes hardware related information, software related information, web page related information, service related information, and/or user related information, where the hardware related information includes one or a combination of the following: manufacturer, CPU utilization rate, memory utilization rate and CPU model; the software-related information includes one or a combination of: soft probe middleware version number and interface protocol version number; the webpage related information comprises one or a combination of the following: ping average time delay, download rate, first packet time delay, packet loss rate and the like; the service-related information comprises one or a combination of the following: bandwidth, WIFI status; the user related information: cell, location, star, regional code.

The KQI data in the step A1 comprise a video service KQI, a web service KQI, a game service KQI and/or a television service KQI, and the video service KQI comprises one or a combination of the following: video buffering duration ratio, video downloading rate, video HTTP first packet response delay and video HTTP response success rate; the webpage service KQI comprises one or a combination of the following components: average webpage opening time delay, webpage downloading speed, webpage HTTP first packet response time delay and webpage HTTP response success rate; the game service KQI comprises one or a combination of the following: the game packet loss rate, the game downloading rate and the game HTTP first packet response time delay; the television service KQI comprises one or a combination of the following: the playing success rate of the television program, the ratio of the television pause time and the television pause times per hour.

The data sources for data acquisition in the step A1 comprise one or a combination of the following: the DPI data is analyzed by the home network soft probe, the set-top box soft probe and the deep packet.

The step 2 comprises the following steps: step B1, utilizing the KPI and the KQI in the step A1 and user quality difference complaint data, wherein the user quality difference complaint data comprises complaint reasons, complaint events and whether complaints are repeated or not of the home broadband quality difference; and B2, preprocessing data, wherein one or a combination of the following indexes is constructed: the method comprises the steps of busy hour and idle time marking, a statistical period, a webpage service weight, a video service weight, a game service weight, a television service weight, historical complaint times and a user star level, wherein the statistical period takes days as a unit, centralized granularity data is determined and constructed according to a first parameter, under each statistical period, a KPI index and a KQI index calculate an average value, one complaint record is realized corresponding to a first parameter data record, and the sum of the webpage service weight, the video service weight, the game service weight and the television service weight is 1, which represents the frequency proportion of different types of services used by a user; step B3, data association, namely associating the performance data with the complaint data; step B4, randomly extracting a non-complaint user data set, wherein the data volume is multiple of a second parameter of the complaint user data set, and forming a complete data set; step B5, qoE is set, qoE is designed according to whether complaint exists or not, if complaint behaviors exist, qoE is set to be yes, and if complaint behaviors do not exist, qoE is set to be no; and B6, establishing a two-class model for the KQI and the QoE, taking the first QoE as output data and the first KQI index set as input data, adjusting parameters to a better solution by adopting a GBDT algorithm for the two-class model, taking the accuracy rate for evaluating the better solution as a standard, and taking the two-class model as a first complaint early warning model CAM to finish the aim of establishing the relationship between the KQI and the QoE.

The device is characterized by comprising a core index determining module, a user perception predicting module, an OLT quality difference defining module and a user complaint and perception module, wherein the core index determining module determines a core KPI index set according to KPI/KQI data, a full-quantity KPI index set and a KQI index set, the user perception predicting module outputs post-complaint diagnosis or pre-complaint warning according to the core KPI index set and the KQI index set, the OLT quality difference defining module outputs post-complaint diagnosis, and the user complaint and perception module outputs pre-complaint warning according to the complaint data and the KQI index set.

The post-complaint diagnosis comprises the following steps: step 1, determining an available real-time acquisition KPI/KQI data set; step 2, preprocessing data; step 3 is divided into two paths, one path is a step 3A, a user perception index quality difference data set, the other path is a step 3B, an OLT dimension data set; step 4 is divided into two paths, one path is step 4A, a user perception prediction model, the other path is step 4B, an OLT quality difference definition model; step 5 is divided into two paths, one path is step 5A, whether the threshold value is exceeded or not is judged, if yes, the home network quality is poor, if not, the step 6 is carried out, the other path is step 5B, whether the quality is poor is judged, if yes, the load-bearing network quality is poor, and if not, the step 6 is carried out; step 6, the quality of the suspected content source is poor; step 7, content source dial testing; and 8, judging whether the content source is poor in quality, if so, determining that the content source is poor in quality, and if not, determining that the content source is poor in quality due to unknown reasons.

The pre-complaint early warning method comprises the following steps: step 1, obtaining a poor quality user data set; step 2 is divided into two paths, one path is step 2a for filtering, and the other path is step 2b for complaining the early warning model; step 3 is divided into two paths, wherein one path is step 3a, a quality difference cell or a quality difference OLT or a quality difference area is determined, and the other path is step 3b, a potential quality difference complaint user is determined; step 4, unifying quality difference alarm and deriving the work order.

The user perception prediction module performs the steps of: step 1, determining a historical KPI/KQI data set; step 2, preprocessing data; step 3 is divided into two paths, one path is step 3x to form a core KPI index data set, and the other path is step 3y to form a KQI index data set; step 4, model training, wherein core KPI index data are used as input data, KQI index data are used as output data, and a GBDT algorithm is adopted; step 5, forming a pre-training model; step 6, predicting KQI data; step 7, calculating the accuracy; step 8, predicting accuracy; the pre-training model in the step 5 receives KPI data obtained by data extraction from real-time collected quality difference data; and the calculation accuracy in the step 7 adopts the real KQI data obtained after the data extraction.

The invention has the following technical effects: the invention relates to an analysis method and device for quality difference of a family broadband user, which take a user perception index KQI as an intermediary, analyze the relation between a KPI and the KQI, and can effectively determine a core KPI influencing the KQI through twice filtering to form a core KPI index set. And (4) taking the complaint data as the standard of QoE difference, comprehensively considering the KQI indexes for a long time, and constructing indexes comprising user characteristics, time characteristics and busy-hour and idle-hour characteristics. On the basis of analyzing the relationship between the KQI and the QoE, the influence of common services of different users, the difference of network quality and the user perception difference on the QoE difference can be comprehensively considered; the content source quality difference can be divided into three aspects: poor quality of home network, poor quality of bearer network and poor quality of content source; the user attribute characteristics can be added in the model by adding the indexes capable of representing the user characteristics for describing the perception difference of the user, so that the early warning capability with higher accuracy is realized.

Compared with the prior art, the invention has the following characteristics: 1. the user perception difference can be effectively reflected. Actually, the capability of dynamic threshold is formed, and the relation between KPI and QoE is analyzed in a differentiation manner according to the information such as common service of the user, user characteristics and the like. 2. The problem that the KPI and QoE relation is simply understood as a linear relation by the traditional technology is solved. 3. The diagnosis range of the quality difference is unclear and difficult to optimize for a single KPI index diagnosis conclusion, and the root cause generated by the problem can be effectively positioned through the relation analysis of KPI and KQI. 4. The traditional method has the problem that a sudden problem can disappear along with data, and the problem can be effectively solved through a large amount of data. 5. The change of the network can be dynamically reflected, and the dynamic relationship between the KPI and the QoE is formed. The traditional method only considers fixed KPIs and cannot reflect the changes of the network in time.

Drawings

Fig. 1 is a schematic diagram of an analysis device for quality difference of home broadband users, which implements the invention. FIG. 1 illustrates a Core KPI (kernel Key Performance Indicator) set determined by a Core Indicator determining module according to KPI/KQI data (KPI, key Performance Indicator, network Performance Indicator; KQI, key Quality Indicator, user perception Indicator), the full KPI set and the KQI set; determining the diagnosis after complaint or the early warning before complaint by a User Perception Prediction Module (UPPM) according to the core KPI index set and the KQI index set; an OLT (optical line terminal) quality difference definition module (OLT) determines the diagnosis after complaint; and determining the early warning before the complaint by the user complaint and perception module according to the complaint data and the KQI index set.

FIG. 2 is a schematic flow diagram of the operation of the core index determination module of FIG. 1. Included in fig. 2 are network performance indicators (e.g., X1, X2, \8230;, xi, \8230;, xn), user-perceived indicators (e.g., Y1, \8230;, yn), a set of network performance core indicators (e.g., X1a, \8230;, X1 b), a GBDT-based computational mapping between network performance indicators and user-perceived indicators (Y = F (X)), and a minimum GBDT Tree (GBDT, gradient Boosting Decision Tree algorithm) between user-perceived indicators and the set of network performance core indicators based on evaluation and pruning.

FIG. 3 is a schematic flow chart of the post-complaint diagnostic method of FIG. 1. FIG. 3 includes step 1, determining available real-time acquisition KPI/KQI datasets; step 2, data preprocessing; step 3 is divided into two paths, wherein one path is a step 3A, a user perceives index quality difference data set, and the other path is a step 3B, an OLT dimension data set; step 4 is divided into two paths, one path is step 4A, a user perception prediction model, the other path is step 4B, an OLT quality difference definition model; step 5 is divided into two paths, one path is step 5A, whether the threshold value is exceeded or not is judged, if yes, the home network quality is poor, if not, the step 6 is carried out, the other path is step 5B, whether the quality is poor is judged, if yes, the load-bearing network quality is poor, and if not, the step 6 is carried out; step 6, the quality of the suspected content source is poor; step 7, content source dial testing; and 8, judging whether the content source is poor in quality, if so, determining that the content source is poor in quality, and if not, determining that the content source is poor in quality due to unknown reasons.

Fig. 4 is a schematic flow diagram of the operation of the user perception prediction module of fig. 1 or the user perception prediction model of fig. 3. FIG. 4 includes step 1, determining a historical KPI/KQI data set; step 2, data preprocessing; step 3 is divided into two paths, one path is 3x to form a core KPI index data set (CKPI), and the other path is 3y to form a KQI index data set; step 4, model training, wherein core KPI index data are used as input data, KQI index data are used as output data, and a GBDT algorithm is adopted; step 5, forming a pre-training model; step 6, predicting KQI data; step 7, calculating the accuracy; and 8, predicting the accuracy. And the pre-training model in the step 5 receives KPI data obtained by data extraction from real-time collected quality difference data. And the calculation accuracy in the step 7 adopts the real KQI data obtained after the data extraction.

FIG. 5 is a flow chart of the pre-complaint warning method in FIG. 1. Fig. 5 includes step 1, obtaining a poor quality user data set; step 2 is divided into two paths, wherein one path is filtered in step 2a, and the other path is filtered in step 2b, and a complaint early warning model is obtained; step 3 is divided into two paths, wherein one path is step 3a, a quality difference cell or a quality difference OLT or a quality difference area is determined, and the other path is step 3b, a potential quality difference complaint user is determined; step 4, unifying quality difference alarm and deriving the work order.

Detailed Description

The invention is described below with reference to the figures (fig. 1-5) and examples.

FIG. 1 isThe invention discloses a schematic diagram of an analysis device for quality difference of home broadband users. FIG. 2 is a schematic flow diagram of the operation of the core index determination module of FIG. 1. FIG. 3 is a schematic flow chart of the post-complaint diagnostic method of FIG. 1. Fig. 4 is a schematic flow diagram of the operation of the user perception prediction module of fig. 1 or the user perception prediction model of fig. 3. FIG. 5 is a flow chart of the pre-complaint warning method shown in FIG. 1. Referring to fig. 1 to 5, an analysis method for home broadband user quality difference includes the following steps: step 1, obtaining network performance index KPI data and user perception index KQI data of a family broadband user, and establishing a relation between KPI and KQI; step 2, establishing a relation between the KQI and the QoE according to the QoE experience data of the user and the KQI data in the relation between the KPI and the KQI; and 3, indirectly determining the relationship between the KPI and the QoE by taking the KQI as an intermediary. The step 1 comprises the following steps: step A1, data acquisition, which comprises collecting historical data, wherein the historical data comprises machine performance data, network performance data and service data of four services, namely home broadband video, a webpage, a game and a television, and is divided into KPI and KQI; step A2, data preprocessing, namely cleaning and preprocessing the historical data, wherein the preprocessing comprises the following steps: missing value processing, abnormal value processing, discrete data encoding and normalization processing; step A3, regression analysis, namely, for any first KQI index, a corresponding first KPI index set exists, the mapping relation between the first KQI index and the first KPI index set is determined through the regression analysis, the gradient lifting decision tree GBDT algorithm is adopted in the regression analysis, the parameter in the regression analysis is adjusted to be a better solution, and the decision coefficient R is adopted in the design model evaluation process ² (ii) a Step A4, sorting the contribution degrees of a first KPI index set aiming at the better solution and the better solution, and deleting KPI indexes with the contribution degrees lower than a first threshold value from the first KPI index set to form a new first KPI index set; step A5, carrying out regression analysis again by using the new first KPI index set and the first KQI index, and adjusting the model parameters to a better solution, wherein the decision coefficient of the model parameters is a first decision coefficient; step A6, carrying out contribution degree sequencing on the first KPI index set in the step A5, and deleting the index with the lowest contribution degree in the first KPI index setExcept as a new first set of KPI indicators; step A7, carrying out regression analysis again by using the new first KPI index set and the first KQI index, and adjusting the model parameters to a better solution, wherein the decision coefficient is a second decision coefficient; step A8, if the second decision coefficient is smaller than the first decision coefficient and the absolute value of the difference is larger than a second threshold value, rejecting the deletion of the KPI, and turning to step A9, otherwise, receiving the deletion of the KPI, and turning to step A5; step A9, the processed new index set is a first core KPI index set CKPI corresponding to the first KQI index; step A10, performing regression analysis by using the first CKPI index set and the first KQI index set, adjusting parameters to a better solution, and evaluating the better solution by using a determination coefficient R ² The regression analysis model serves as a first user perception prediction model UPPM; and step A11, repeating the steps for each KQI until the aim of constructing the relationship between the KPI and the KQI is fulfilled. The KPI data in step A1 includes hardware related information, software related information, web page related information, service related information, and/or user related information, where the hardware related information includes one or a combination of the following: manufacturer, CPU utilization rate, memory utilization rate and CPU model; the software related information comprises one or a combination of the following: soft probe middleware version number and interface protocol version number; the webpage related information comprises one or a combination of the following: ping average time delay, download rate, first packet time delay, packet loss rate and the like; the service-related information comprises one or a combination of the following: bandwidth, WIFI status; the user related information: cell, location, star, regional code. The KQI data in the step A1 comprise a video service KQI, a webpage service KQI, a game service KQI and/or a television service KQI, and the video service KQI comprises one or a combination of the following: video buffering time ratio, video downloading rate, video HTTP first packet response time delay and video HTTP response success rate; the webpage service KQI comprises one or a combination of the following components: average webpage opening time delay, webpage downloading speed, webpage HTTP first packet response time delay and webpage HTTP response success rate; the game service KQI comprises one or the combination of the following: the game packet loss rate, the game downloading rate and the game HTTP first packet response time delay; the television service KQI comprises one or a combination of the following: television program broadcastingSuccess rate, ratio of TV pause time and TV pause times per hour. The data sources for data acquisition in the step A1 comprise one or a combination of the following: the DPI data is analyzed through the home network soft probe, the set top box soft probe and the deep packet. The step 2 comprises the following steps: step B1, utilizing the KPI and the KQI in the step A1 and user quality difference complaint data, wherein the user quality difference complaint data comprises complaint reasons, complaint events and whether complaints are repeated or not of the home broadband quality difference; and B2, data preprocessing, including constructing one or a combination of the following indexes: busy hour and idle time marks, statistical cycles, webpage service weights, video service weights, game service weights, television service weights, historical complaint times and user star levels, wherein the statistical cycles take days as units, centralized granularity data are determined and constructed according to a first parameter, in each statistical cycle, a KPI index and a KQI index calculate an average value to realize that one complaint record corresponds to a first parameter data record, and the sum of the webpage service weights, the video service weights, the game service weights and the television service weights is 1 and represents the frequency ratio of different types of services used by a user; step B3, data association, namely associating the performance data with the complaint data; step B4, randomly extracting a non-complaint user data set, wherein the data volume is multiple of a second parameter of the complaint user data set, and forming a complete data set; step B5, qoE is set, qoE is designed according to the complaint or not, if complaint behaviors exist, qoE is set to be yes, and if complaint behaviors do not exist, qoE is set to be no; and B6, establishing a two-classification model for the KQI and the QoE, taking the first QoE as output data and the first KQI index set as input data, adjusting parameters to a better solution by adopting a GBDT algorithm for the two-classification model, taking the accuracy for evaluating the better solution as a standard, and taking the two-classification model as a first complaint early warning model CAM to finish the aim of constructing the relationship between the KQI and the QoE.

An analysis device for quality differences of household broadband users comprises a core index determination module, a user perception prediction module, an OLT quality difference defining module and a user complaint and perception module, wherein the core index determination module determines a core KPI index set according to KPI/KQI data, a full KPI index set and a KQI index set, the user perception prediction module outputs post-complaint diagnosis or pre-complaint warning according to the core KPI index set and the KQI index set, the OLT quality difference defining module outputs post-complaint diagnosis, and the user complaint and perception module outputs pre-complaint warning according to the complaint data and the KQI index set. The post-complaint diagnosis comprises the following steps: step 1, determining an available real-time acquisition KPI/KQI data set; step 2, preprocessing data; step 3 is divided into two paths, one path is a step 3A, a user perception index quality difference data set, the other path is a step 3B, an OLT dimension data set; step 4 is divided into two paths, one path is step 4A, a user perception prediction model, the other path is step 4B, an OLT quality difference definition model; step 5 is divided into two paths, wherein one path is step 5A, whether the threshold value is exceeded or not is judged, if yes, the home network quality is poor, if not, the step 6 is carried out, the other path is step 5B, whether the quality of the integrated network is poor is judged, if yes, the load-bearing network quality is poor, and if not, the step 6 is carried out; step 6, the quality of the suspected content source is poor; step 7, dialing and testing the content source; and 8, judging whether the content source is poor in quality, if so, determining that the content source is poor in quality, and if not, determining that the content source is poor in quality due to unknown reasons. The pre-complaint early warning method comprises the following steps of: step 1, obtaining a poor quality user data set; step 2 is divided into two paths, one path is step 2a for filtering, and the other path is step 2b for complaining the early warning model; step 3 is divided into two paths, wherein one path is step 3a, a quality difference cell or a quality difference OLT or a quality difference area is determined, and the other path is step 3b, a potential quality difference complaint user is determined; and 4, unifying quality difference alarms and deriving work orders. The user perception prediction module performs the steps of: step 1, determining a historical KPI/KQI data set; step 2, preprocessing data; step 3 is divided into two paths, one path is step 3x to form a core KPI index data set, and the other path is step 3y to form a KQI index data set; step 4, model training, wherein core KPI index data are used as input data, KQI index data are used as output data, and a GBDT algorithm is adopted; step 5, forming a pre-training model; step 6, predicting KQI data; step 7, calculating the accuracy; step 8, predicting accuracy; the pre-training model in the step 5 receives KPI data obtained after data extraction from real-time collected quality difference data; and the calculation accuracy in the step 7 adopts the real KQI data obtained after the data extraction.

The application provides an analysis method for a relationship between Quality of Experience (QoE) and network Performance Indicator (KPI) of a family broadband user. The method takes a user perception index (KQI) as a medium, analyzes the relationship between KPI and KQI, and then analyzes the relationship between KQI and QoE. Therefore, the relation between the KPI and the QoE is analyzed. The application provides a family broadband quality difference diagnosis method which is realized based on a relationship between KPI and KQI. The application provides a home broadband user quality difference complaint early warning method which is realized based on a KQI and QoE relationship.

The specific invention content of the application is as follows: a method of KPI and QoE analysis, the method comprising two parts: a first part, KPI and KQI relational analysis, the method comprising: the method comprises the following steps: and (6) data acquisition. And collecting historical data including machine performance data, network performance data, service data and the like of four services of home broadband videos, web pages, games and televisions. The collected historical data is artificially divided into KPIs and KQIs. The KPI indicator data may adopt any service-related indicator, which is not limited in this application. Including, but not limited to, hardware related information: manufacturer, CPU utilization rate, memory utilization rate and CPU model; software related information: soft probe middleware version number, interface protocol version number, etc.; and webpage related information: ping average time delay, download rate, first packet time delay, packet loss rate and the like; service related information: bandwidth, WIFI state; the user related information: cell, location, star, area code, etc. In particular, the KQI of the video service includes: video buffering duration ratio, video downloading rate, video HTTP first packet response delay and video HTTP response success rate; the KQI of the webpage service comprises: average webpage opening time delay, webpage downloading speed, webpage HTTP first packet response time delay and webpage HTTP response success rate; the KQI of the game service comprises: the game packet loss rate, the game downloading rate and the game HTTP first packet response time delay; the KQI of the television service comprises: the playing success rate of the television program, the ratio of the television pause time and the television pause times per hour. Particularly, the data sources to be collected include a home network soft probe, a set-top box soft probe and DPI data, and the data sources are not limited in the present application.

Step two: and (4) preprocessing data. Cleaning and preprocessing the collected historical data, wherein the preprocessing comprises the following steps: missing value processing, abnormal value processing, discrete data encoding, normalization and other necessary preprocessing processes. Step three: and (4) performing regression analysis. For any first KQI index, there is a corresponding first set of KPI indices. The process is to determine the mapping relation between the first KQI index and the first KPI index set through regression analysis. In particular, the regression analysis employs GBDT (Gradient Boosting Decision Tree algorithm). And adjusting the parameters of the regression analysis method to a better solution. The evaluation problem of design model in adjustment adopts a determination coefficient R ² . Step four: and sequencing the contribution degrees of the first KPI index set on the better solution obtained in the third step. The contribution degree is as follows: average purity reduction (MDI) was used as the contribution of the KPI index set. And deleting the KPI with the contribution degree lower than the first threshold value from the first KPI set to form a new first KPI set.

Step five: and performing secondary regression analysis by using the first KPI index set and the first KQI index, and adjusting the model parameters to a better solution. The determination coefficient is the first determination coefficient. Step six: and D, sorting the contribution degrees of the first KPI index set in the step five, and deleting the index with the lowest contribution degree in the first KPI index set to serve as a new first KPI index set. Step seven: and performing regression analysis again by using the first KPI index set and the first KQI index, and adjusting the model parameters to a better solution. The determination coefficient is the second determination coefficient. Step eight: and if the second decision coefficient is smaller than the first decision coefficient and the absolute value of the difference is larger than a second threshold value, rejecting the deletion of the KPI, and turning to the ninth step, otherwise, receiving the deletion of the KPI, and turning to the fifth step.

Step nine: the new Indicator set processed as described above is a first Core KPI Indicator set (CKPI) corresponding to the first KQI Indicator. Step ten: according to the first CKPI index set and the first KQI indexThe set was subjected to regression analysis. Adjusting parameters to a better solution, and evaluating the better solution by adopting a decision coefficient R ² . The regression analysis Model serves as a first User Perception Prediction Model (UPPM).

The above procedure was repeated for each KQI. Thus, the aim of constructing the relationship between the KPI and the KQI is fulfilled.

A second part, KQI vs QoE analysis, the method comprising:

the method comprises the following steps: the data required is as described in the first part step one, in particular, the user quality complaint data is required. The complaint reason of the complaint data is poor household broadband quality. The complaint data comprises the following contents: cause of complaints, complaint events, whether complaints are repeated, etc.

Step two: and (4) preprocessing data. New indexes need to be constructed according to data. The following indexes are added: busy hour and idle hour marking, counting period, webpage service weight, video service weight, game service weight, television service weight, historical complaint times and user star level. And the statistical period is day-by-day, the data of concentrated granularity is determined and constructed according to the first parameter, and the average value of the KPI index and the KQI index is calculated in each statistical period. One complaint record is realized corresponding to the first parameter strip data record. The sum of the webpage service weight, the video service weight, the game service weight and the television service weight is 1, which represents the frequency proportion of different types of services used by the user.

Step three: and data association, namely associating the performance data with the complaint data.

Step four: and randomly extracting a non-complaint user data set, wherein the data volume is the second parameter times of the complaint user data set. A complete data set is formed.

Step five: and setting QoE. And designing the QoE according to the complaint or not, if complaint behaviors exist, setting the QoE to be yes, and if the complaint behaviors do not exist, setting the QoE to be no.

Step six: and establishing a two-classification model for the KQI and the QoE. And taking the first QoE as output data and the first KQI index set as input data. The algorithm adopted by the binary model is GBDT. And adjusting parameters to a better solution, wherein the accuracy is adopted as a standard for evaluating the better solution. This binary Model serves as the first Complaint early warning Model (Complaint Alert Model, CAM). Thus, the goal of constructing the relationship between the KQI and the QoE is completed.

As described above, the KPI and QoE relationship is indirectly analyzed by analyzing the KPI and KQI relationship using KQI as an intermediary and analyzing the KQI and QoE relationship.

A home broadband user quality difference complaint early warning method comprises the following steps:

the method comprises the following steps: and extracting a real-time quality difference user data set, and preprocessing data. And constructing a first parameter data record for each first quality difference user.

Step two: the first data record for each first user of poor quality, as input data to the CAM, is output as a first suspected complaint. The values for the first suspected complaint are yes and no.

Step three: and counting the first suspected complaints of each first data record, and if the percentage is greater than a third threshold value, determining that the first user with poor quality has a higher probability of complaint, namely the suspected user with poor quality.

By the method, the suspected complaint early warning is realized for the user.

A home broadband quality differential diagnostic method, the method comprising:

the method comprises the following steps: and extracting real-time quality difference user data. The poor quality user is a user known to have poor quality, and the data of the poor quality user may be: the user complaint system or the home broadband user quality difference complaint early warning method can early warn complaint users. And extracting real-time acquisition CKPI data and KQI data of the first poor quality user. The KQI data is first real KQI data. The KQI data of a first OLT (optical line terminal) where the first quality-difference user is located is extracted.

Step two: and (4) preprocessing data. Cleaning and preprocessing the acquired real-time data, wherein the preprocessing comprises the following steps: missing value processing, abnormal value processing, discrete data encoding, normalization and other necessary preprocessing processes.

Step three: the CKPI data of the first poor quality user is input data serving as input data of the UPPM, and the output data is first predicted KQI data.

Step four: and calculating the prediction accuracy. And calculating the pre-accuracy rate by using the first predicted KQI data and the first real KQI data.

Step five: and if the accuracy is higher than the first minimum accuracy, determining that the home network quality is poor. And if the accuracy is less than the threshold value, the suspected content source is considered to be poor in quality.

Step six: and if the first KQI of the first OLT exceeds the threshold value and the first OLT quality difference threshold value, outputting that the quality difference of the bearer network exists. Otherwise, the suspected content source quality is output.

Step seven: if the suspected content source is poor in quality in the fifth step and the sixth step, content source dial testing is conducted, the dial testing result is a first CKPI index set, and if any first CKPI index exceeds the first content source threshold limit, the index is poor in quality.

The method can realize the quality difference of the content source into three aspects: poor quality of home network, poor quality of bearing network and poor quality of content source.

The following further description is made with reference to the accompanying drawings:

the overall process of the invention is shown in figure 1:

first, data of KPI and KQI related indicators are collected, and for each KQI indicator, a core KPI indicator set is determined by a core indicator determination module, where the indicators in the core indicator set are from the full KPI indicator set, and the flow of the core indicator determination module is as described in embodiment 1 (see fig. 1).

And secondly, taking the screened core KPI index set and the screened KQI index set as the input and the output of a user perception prediction module. User perception is understood by training the user perception prediction module. The KQI index set can be effectively understood through the core KPI index set through the user perception prediction module.

Then, it is used by the OLT quality difference definition module to determine the OLT quality difference behavior. The user complaint and awareness module can deal with problems in the data.

Finally, post-complaint diagnostic functions can be constructed by the OLT quality difference definition module and the user perception prediction module, as described in example 2 (see fig. 3 and 4). The user perception prediction model and the user complaint and perception module may constitute a pre-complaint warning function, as described in example 3 (see fig. 5).

The scheme of example 1 is shown in FIG. 2.

S101: and calculating the mapping relation based on the GBDT algorithm. And performing regression analysis aiming at all network performance indexes, wherein an algorithm adopted by the regression analysis is GBDT. The input is a network performance index, and the data sources of the network performance index are a home gateway soft probe and a set top box soft probe.

S102: the minimum GBDT tree is determined based on evaluation and pruning. Through repeated evaluation and pruning, an optimal tree structure is achieved, and sequencing is performed according to the contribution degree of the indexes in the tree. And taking the index with the contribution degree larger than the first threshold value as a core KPI index set of the KQI index.

By repeating the above operations for each KQI index, several sets of core KPI indices may be generated.

In particular, the set of core KPI indicators may vary with changes in network performance data, which is a true reflection of objective network conditions. As part of the conditions in the network change, elements that affect the KQI index are also migrating.

The scheme of example 2 is shown in FIG. 3.

S201: and (4) preprocessing data. Soft probe data is collected in real time by filtering. A user perception index quality difference data set is constructed from two dimensions. Extracting data of which the KQI indexes do not meet threshold conditions; and combining the data of the same OLT to form an OLT dimension data set.

S202: a user-aware predictive model. The input of the user perception prediction model is a core KPI index set, and the output is a KQI index set. The flow of the user-perceived prediction model is shown in fig. 4.

S2021: and (4) preprocessing data. Carrying out data preprocessing on the historical KPI/KQI data set, wherein the preprocessing content comprises the following steps: discrete data encoding, abnormal value filtering, missing value processing and data normalization processing. The preprocessed output is separated into a core KPI index data set and a KQI index data set.

S2022: and (5) training a model. And (3) taking the KQI index data set as an output item and taking the corresponding KPI index data set as an input item to perform regression analysis, wherein the GBDT algorithm is adopted for the special regression analysis. And optimizing to a better solution by adjusting parameters. And storing the calculated model. Evaluation determination coefficient R of regression analysis ² 。

And after the model training is finished, the prediction is realized through real-time data calling.

S2023: and (6) data extraction. And extracting the real-time collected quality difference data into KPI data and KQI data, wherein the KQI data is real KQI data.

S2024: and taking the KPI data as input data, calling a pre-training model, and calculating predicted KQI data.

S2025: and calculating the accuracy. And calculating the accuracy according to the predicted KQI data and the real KQI data. And returning the prediction accuracy.

S203: and calculating whether the prediction accuracy exceeds a threshold value. The threshold value is set to be the determination coefficient R in the model training process in S2022 ² . If the quality of the home network is higher than the threshold value, outputting the quality difference of the home network, otherwise, determining that the suspected content source quality difference exists.

S204: the OLT quality difference defines the model.

S2041: and calculating the wanted CKPI index of the OLT.

S2042: and judging the quality difference of the CKPI index according to the threshold value, and returning the related quality difference. If the OLT is remarkably worse than the CKPI indexes of other OLTs, the OLT is considered to have poor collection quality.

S205: the poor constitution is judged. And if the set quality difference exists, outputting the set quality difference, otherwise, outputting the suspected content source quality difference.

S206: and (4) content source dial testing. And (4) carrying out dial testing on the quality difference record of the suspected content source.

S207: and analyzing the returned dialing test result. If the threshold is exceeded. The content source is deemed to be of poor quality. And if the dialing test result is not abnormal, outputting the quality difference of the unknown reason.

Thus, the second embodiment realizes the diagnosis of the content source quality. The KPI/KQI data collected in real time are analyzed. The cause of the user's poor quality is diagnosed as: poor quality of home network, poor quality of bearing network and poor quality of content source. The method can be used for diagnosis after complaints of the home broadband users.

The flow of example 3 is shown in FIG. 5.

S301: filtering the data of the poor quality users, collecting the poor quality users according to the dimensions of the cell, the OLT and the area, and sequencing according to the user rate of the poor quality to form the data of the poor quality cell, the poor quality OLT and the poor quality area.

S302: and judging potential quality difference complaint users through the complaint early warning model.

S303: and performing unified quality difference alarm and work order derivation according to the correlation analysis result. The analysis result comprises: poor quality cells, poor quality OLTs, poor quality areas, and potential poor quality complaint users.

Those skilled in the art will appreciate that the invention may be practiced without these specific details. It is pointed out here that the above description is helpful for the person skilled in the art to understand the invention, but does not limit the scope of protection of the invention. Any such equivalents, modifications and/or omissions as may be made without departing from the spirit and scope of the invention may be resorted to.

Claims

1. An analysis method for home broadband user quality difference is characterized by comprising the following steps: step 1, obtaining network performance index KPI data and user perception index KQI data of a family broadband user, and establishing a relation between KPI and KQI; step 2, establishing a relation between the KQI and the QoE according to the QoE data experienced by the user and the KQI data in the relation between the KPI and the KQI; step 3, indirectly determining the relationship between KPI and QoE by taking KQI as an intermediary;

the step 1 comprises the following steps: a1, data acquisition, including collection of historical data, wherein the historical data includes four industries of family broadband video, web page, game and televisionThe method comprises the steps of obtaining machine performance data, network performance data and business data of a business, and dividing historical data into KPIs and KQIs; step A2, data preprocessing, namely cleaning and preprocessing the historical data, wherein the preprocessing comprises the following steps: missing value processing, abnormal value processing, discrete data encoding and normalization processing; step A3, regression analysis, namely, for any first KQI index, a corresponding first KPI index set exists, the mapping relation between the first KQI index and the first KPI index set is determined through the regression analysis, the gradient lifting decision tree GBDT algorithm is adopted in the regression analysis, the parameter in the regression analysis is adjusted to be a better solution, and the decision coefficient R is adopted in the design model evaluation process ² (ii) a Step A4, sorting the contribution degrees of a first KPI index set aiming at the better solution, and deleting KPI indexes with the contribution degrees lower than a first threshold value from the first KPI index set to form a new first KPI index set; step A5, performing regression analysis again by using the new first KPI index set and the first KQI index, and adjusting model parameters to a better solution, wherein the decision coefficient is a first decision coefficient; step A6, sorting the contribution degrees of the first KPI index set in the step A5, and deleting the index with the lowest contribution degree in the first KPI index set to be used as a new first KPI index set; step A7, carrying out regression analysis again by using the new first KPI index set and the first KQI index, and adjusting the model parameters to a better solution, wherein the decision coefficient of the model parameters is a second decision coefficient; step A8, if the second decision coefficient is smaller than the first decision coefficient and the absolute value of the difference is larger than a second threshold value, rejecting the deletion of the KPI, and turning to step A9, otherwise, accepting the deletion of the KPI, and turning to step A5; step A9, the processed new index set is a first core KPI index set CKPI corresponding to the first KQI index; step A10, performing regression analysis by using the first CKPI index set and the first KQI index set, adjusting parameters to a better solution, and evaluating the better solution by using a determination coefficient R ² The regression analysis model is used as a first user perception prediction model UPPM; and step A11, repeating the steps for each KQI until the goal of constructing the relationship between the KPI and the KQI is completed.

2. The method for analyzing the quality difference of home broadband users according to claim 1, wherein the KPI data in step A1 includes hardware related information, software related information, web page related information, service related information and/or user related information, and the hardware related information includes one or a combination of the following: manufacturer, CPU utilization rate, memory utilization rate and CPU model; the software related information comprises one or a combination of the following: soft probe middleware version number and interface protocol version number; the webpage related information comprises one or a combination of the following: ping average time delay, download rate, first packet time delay and packet loss rate; the service-related information comprises one or a combination of the following: bandwidth, WIFI state; the user-related information includes: cell, location, star, regional code.

3. The analysis method for the quality difference of the home broadband users according to claim 1, wherein the KQI data in step A1 comprises a video service KQI, a web service KQI, a game service KQI and/or a television service KQI, and the video service KQI comprises one or a combination of the following: video buffering duration ratio, video downloading rate, video HTTP first packet response delay and video HTTP response success rate; the webpage service KQI comprises one or a combination of the following components: average webpage opening time delay, webpage downloading speed, webpage HTTP first packet response time delay and webpage HTTP response success rate; the game service KQI comprises one or a combination of the following: the game packet loss rate, the game downloading rate and the game HTTP first packet response time delay; the television service KQI comprises one or a combination of the following: the playing success rate of the television program, the ratio of the television pause time and the television pause times per hour.

4. The analysis method for the quality difference of the home broadband users according to claim 1, wherein the data sources of the data collection in the step A1 comprise one or a combination of the following: the DPI data is analyzed through the home network soft probe, the set top box soft probe and the deep packet.

5. The analysis method for the quality difference of the home broadband users according to claim 1, wherein the step 2 comprises the following steps: step B1, acquiring the KPI and the KQI in the step A1 and user quality difference complaint data, wherein the user quality difference complaint data comprises complaint reasons, complaint events and whether complaints are repeated or not of the home broadband quality difference; and B2, data preprocessing, including constructing one or a combination of the following indexes: the method comprises the steps of busy hour and idle time marking, a statistical period, a webpage service weight, a video service weight, a game service weight, a television service weight, historical complaint times and a user star level, wherein the statistical period takes days as a unit, centralized granularity data is determined and constructed according to a first parameter, under each statistical period, a KPI index and a KQI index calculate an average value, one complaint record is realized corresponding to a first parameter data record, and the sum of the webpage service weight, the video service weight, the game service weight and the television service weight is 1, which represents the frequency proportion of different types of services used by a user; step B3, data association is carried out, and the performance data and the complaint data are associated; step B4, randomly extracting a non-complaint user data set, wherein the data volume is multiple of a second parameter of the complaint user data set, and forming a complete data set; step B5, qoE is set, qoE is designed according to the complaint or not, if complaint behaviors exist, qoE is set to be yes, and if complaint behaviors do not exist, qoE is set to be no; and B6, establishing a two-classification model for the KQI and the QoE, taking the first QoE as output data and the first KQI index set as input data, adjusting parameters to a better solution by adopting a GBDT algorithm through the two-classification model, taking the accuracy rate for evaluating the better solution as a standard, and taking the two-classification model as a first complaint early warning model CAM to finish the aim of establishing the relation between the KQI and the QoE.

6. An analysis device for quality difference of home broadband users, which is used for implementing the analysis method for quality difference of home broadband users according to any one of claims 1 to 5, and is characterized by comprising a core index determination module, a user perception prediction module, an OLT quality difference definition module and a user complaint and perception module, wherein the core index determination module determines a core KPI index set according to KPI/KQI data, a full-quantity KPI index set and a KQI index set, the user perception prediction module outputs post-complaint diagnosis or pre-complaint warning according to the core KPI index set and the KQI index set, the OLT quality difference definition module outputs post-complaint diagnosis, and the user complaint and perception module outputs pre-complaint warning according to the complaint data and the KQI index set.

7. The analysis device for home broadband user quality difference according to claim 6, wherein the post-complaint diagnosis comprises the steps of: step 1, determining an available real-time acquisition KPI/KQI data set; step 2, data preprocessing; step 3 is divided into two paths, one path is a step 3A, a user perception index quality difference data set, the other path is a step 3B, an OLT dimension data set; step 4 is divided into two paths, one path is step 4A, a user perception prediction model, the other path is step 4B, an OLT quality difference definition model; step 5 is divided into two paths, one path is step 5A, whether the threshold value is exceeded or not is judged, if yes, the home network quality is poor, if not, the step 6 is carried out, the other path is step 5B, whether the quality is poor is judged, if yes, the load-bearing network quality is poor, and if not, the step 6 is carried out; step 6, the quality of the suspected content source is poor; step 7, content source dial testing; and 8, judging whether the content source is poor in quality, if so, determining that the content source is poor in quality, and if not, determining that the content source is poor in quality due to unknown reasons.

8. The device for analyzing the quality difference of home broadband users according to claim 6, wherein the pre-complaint warning comprises the following steps: step 1, obtaining a user data set with poor quality; step 2 is divided into two paths, one path is step 2a for filtering, and the other path is step 2b for complaining the early warning model; step 3 is divided into two paths, wherein one path is step 3a, a quality difference cell or a quality difference OLT or a quality difference area is determined, and the other path is step 3b, a potential quality difference complaint user is determined; and 4, unifying quality difference alarms and deriving work orders.

9. The device for analyzing home broadband user quality difference according to claim 6, wherein the user perception prediction module performs the following steps: step 1, determining a historical KPI/KQI data set; step 2, preprocessing data; step 3 is divided into two paths, one path is step 3x to form a core KPI index data set, and the other path is step 3y to form a KQI index data set; step 4, model training, wherein core KPI index data are used as input data, KQI index data are used as output data, and a GBDT algorithm is adopted; step 5, forming a pre-training model; step 6, predicting KQI data; step 7, calculating the accuracy; step 8, predicting accuracy; the pre-training model in the step 5 receives KPI data obtained by data extraction from real-time collected quality difference data; and the calculation accuracy in the step 7 adopts the real KQI data obtained after the data extraction.