WO2013075489A1 - 检测报文心跳周期的自适应方法和装置 - Google Patents

检测报文心跳周期的自适应方法和装置 Download PDF

Info

Publication number
WO2013075489A1
WO2013075489A1 PCT/CN2012/077723 CN2012077723W WO2013075489A1 WO 2013075489 A1 WO2013075489 A1 WO 2013075489A1 CN 2012077723 W CN2012077723 W CN 2012077723W WO 2013075489 A1 WO2013075489 A1 WO 2013075489A1
Authority
WO
WIPO (PCT)
Prior art keywords
heartbeat
heartbeat period
neural network
network model
sample set
Prior art date
Application number
PCT/CN2012/077723
Other languages
English (en)
French (fr)
Inventor
姜龙
Original Assignee
中兴通讯股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中兴通讯股份有限公司 filed Critical 中兴通讯股份有限公司
Publication of WO2013075489A1 publication Critical patent/WO2013075489A1/zh

Links

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L43/00Arrangements for monitoring or testing data switching networks
    • H04L43/10Active monitoring, e.g. heartbeat, ping or trace-route
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14Network analysis or design
    • H04L41/142Network analysis or design using statistical or mathematical methods

Definitions

  • the present invention relates to the field of communications technologies, and in particular, to an adaptive method and apparatus for detecting a heartbeat period of a message.
  • network link detection is an essential feature.
  • the network element device (referred to as the network element) ensures that the link is unobstructed by actively responding to the polling check of the network management server (referred to as the network management system) or by periodically sending the heartbeat period of the detection packet to the network management system.
  • the network management system periodically detects the heartbeat period of the packet, and the NE sends the heartbeat period of the packet to the NMS. If the link is normal, the NMS receives the test packet. The heartbeat period is considered to be normal between the network management system and the network management unit. If the network management system does not receive the heartbeat period of the detection packet of the network element, the communication between the network element and the network management system is indicated. The link has failed.
  • the existing heartbeat detection mainly has the following problems.
  • the heartbeat cycle is generally set to a fixed length, but this cycle is often difficult to meet a variety of different application scenarios.
  • the heartbeat cycle algorithm lacks self-learning function.
  • Some network management systems use heartbeat cycle dynamic calculation methods, and often use fixed calculation formulas, which are difficult to adapt to complex and variable network environments.
  • the technical problem to be solved by the embodiments of the present invention is to provide an adaptive method and apparatus for detecting a heartbeat period of a packet, so as to solve the problem of poor adaptability of the calculation method of the heartbeat period of the existing detection packet.
  • An adaptive method for detecting a heartbeat period of a packet includes: a construction step of constructing a neural network model, wherein the input layer node of the neural network model is a heartbeat cycle parameter affecting a heartbeat cycle, and the output layer node is a heartbeat cycle, and the sample set of the neural network model includes a sample parameter of the heartbeat cycle and a heartbeat The mapping relationship of periodic sample values;
  • the self-learning triggering step when determining that the triggering condition meets the self-learning, triggering the neural network model to perform self-learning according to the sample set;
  • a heartbeat cycle update step when determining a trigger condition that meets a heartbeat cycle update, updating a current heartbeat cycle according to the heartbeat cycle calculation value, and determining a trigger condition that does not meet a heartbeat cycle update, performing a sample set update step;
  • the sample set update step updates the sample set and returns to the self-learning step.
  • the parameter includes one or more of the following: a number of network elements, a network congestion coefficient, a network traffic, a device usage rate, a CPU usage rate, an average heartbeat period, a current heartbeat period, and an adjustment value.
  • the process of the self-learning of the neural network model includes:
  • the process ends. If the error value is less than the preset threshold, the process ends. If the error value is greater than the preset threshold, the connection weight of the neural network model is modified according to the error value until the calculated value of the recalculated heartbeat period sample and corresponding The error between the sample values of the heartbeat period is less than the preset threshold.
  • the neural network corrects the connection weight according to a direction in which the error gradient decreases, and superimposes the previous connection weight change amount in a ratio column.
  • the mapping relationship between the heartbeat period calculation parameter and the heartbeat period calculation value is updated into the original sample set, or the original sample set is weighted with the newly obtained mapping relationship. Add the original sample set after the merger.
  • the triggering condition of the heartbeat period update is: the deviation between the calculated value of the heartbeat period and the previously obtained n-1 heartbeat period calculation values is greater than the first threshold, and the heartbeat period is The deviation between the calculated value and the average value of the previously obtained n-1 heartbeat cycle calculation values is greater than the second threshold.
  • the trigger condition for calculating the heartbeat period, the trigger condition for self-learning, or the trigger condition for updating the heartbeat period is a timing trigger or an event trigger.
  • an embodiment of the present invention further provides an apparatus for detecting a heartbeat period of a message, the apparatus comprising:
  • the neural network model module is configured to calculate a heartbeat period, the input layer node is a heartbeat period parameter affecting the heartbeat period, and the output layer node is a heartbeat period, and the sample set of the neural network model includes a sample parameter of the heartbeat period and a sample of the heartbeat period
  • the mapping relationship of values, the method includes:
  • the self-learning triggering module is configured to trigger the neural network model to perform self-learning according to the sample set when determining that the self-learning trigger condition is met;
  • the calculation triggering module is configured to: when determining a trigger condition that meets the calculation of a heartbeat period, input a heartbeat period calculation parameter into the neural network model, and trigger the neural network model to calculate a heartbeat period calculation value;
  • a heartbeat cycle update module configured to: when determining an update condition that meets a heartbeat cycle, update a current heartbeat cycle according to a heartbeat cycle calculation value output by the neural network model;
  • a sample set update module configured to update a sample set of the neural network model.
  • the self-learning method and device for detecting a heartbeat period of a packet can dynamically adjust the heartbeat period of the detection packet according to the current environmental condition, so as to prevent the network management or the network element from detecting the packet when the network load is too large.
  • the heartbeat cycle sending period is set incorrectly, which causes important services to be affected. At the same time, avoiding bandwidth and system resource waste caused by incorrect heartbeat cycle setting, and avoiding the impact on network management performance.
  • Figure 1 is a schematic diagram of a heartbeat link detection mechanism
  • Figure 2 shows the BP neural network model structure for the heartbeat cycle calculation
  • FIG. 3 is a schematic flowchart of an adaptive method for detecting a heartbeat period of a packet according to an embodiment of the present invention
  • FIG. 4 is a schematic diagram showing a module structure of an apparatus for detecting a heartbeat period of a packet according to an embodiment of the present invention
  • an adaptive method for detecting a heartbeat period of a packet includes the following specific steps:
  • Step S101 construct a neural network model
  • BP Back Propagation neural network model
  • the method of the invention is not limited to the BP neural network model.
  • the neural network model is built when the network management system starts.
  • the input layer nodes (also called input layer neurons) of the neural network model are parameters (XrX n ) that affect the calculation of the heartbeat period, and the number of network elements is mainly n.
  • Network congestion coefficient ⁇ network traffic 1, device usage s, CPU occupancy m, average heartbeat period v , current heartbeat period t, adjustment value z, and so on.
  • the number of NEs corresponds to the number of actual NEs in the network.
  • the network congestion coefficient ⁇ can be calculated by the following formula:
  • a, b respectively represent the current load status of the network management system and the network element, a, b, respectively, the value of the load status of the network management and the network element when the heartbeat period of the last detection packet is received, 0 a, b, a, , b, ⁇ 1 , for the server weight, 1 10.
  • Network traffic 1 is the current network traffic divided by the maximum network traffic that can be carried.
  • the device usage rate s is the number of NE devices working in the current network divided by the total number of NE devices.
  • CPU usage m is the current server CPU usage.
  • the average heartbeat period V is the average value of the heartbeat periods of the network elements after the server is started.
  • the current heartbeat period t is a current heartbeat period setting value corresponding to the network element to be calculated.
  • the adjustment value Z is a dynamic adjustment parameter, which is manually adjusted according to the overall calculation of the deviation.
  • the output layer nodes (also referred to as output layer neurons) of the neural network model are heartbeat cycles, and the sample set of the neural network model includes a mapping relationship between sample parameters of the heartbeat cycle and heartbeat cycle sample values.
  • the hidden layer is responsible for adaptively adjusting the heartbeat period.
  • the number of hidden layer nodes is affected by the number of input and output units and the complexity of the problem to be solved. If the number of performance parameters of the input layer is n, according to the statistical analysis result, the number of hidden layer nodes (also called hidden layer neurons) is calculated by the following formula, and the effect is better:
  • n is the number of nodes in the input layer
  • m is the number of nodes in the output layer
  • z is the number of nodes in the hidden layer
  • k is the adjustment value
  • k is an integer ranging from 1 to 5.
  • the output of the jth node is:
  • W 3 ⁇ 4 is the connection weight of the first node of the input layer to the jth node of the hidden layer.
  • the output layer node is the calculated heartbeat period.
  • the output of the output layer is:
  • W M is the connection weight of the pth node of the hidden layer to the qth node of the output layer.
  • Self-learning triggers can be divided into timing triggers and event triggers, such as initialization triggers, manual triggers, or when network parameter values change drastically.
  • self-learning can also be performed.
  • Step S103 triggering a neural network model for self-learning
  • the network element to be learned is placed in the queue to be learned, and the other network elements are adaptively learned after self-learning.
  • the neural network model is trained according to the learning algorithm described below.
  • Self-learning and heartbeat cycle calculations for each network element can be interspersed.
  • S101 to S107 may be adaptive to the heartbeat period of each network element to which the gateway belongs, or may be adaptive to the heartbeat period of the network element itself.
  • the neural network model needs to be self-learned when the network is initialized or when the set self-learning cycle arrives.
  • the existing sample set is used for learning. That is to say, an empirical sample set is given, and each sample of the sample set serves as a specific input and output value of the neural network model, that is, a desired output heartbeat period when each performance parameter is a certain value.
  • the self-learning cycle arrives, it is necessary to use the sample set at the time of initialization, and input and calculate the output value as a new learning sample, and use the learning algorithm to re-update the weights of the neural network model. That is, the mapping relationship obtained by the subsequent calculation is added to the original sample set to implement the update of the sample set.
  • the mapping used for self-learning may be randomly or sequentially selected from the sample set, or may be the latest mapping relationship, or a weighted average of some or all of the mapping relationships.
  • the heartbeat period calculation value tends to be consistent with the heartbeat period sample value as much as possible when the input parameters are the same, thereby improving the accuracy of the heartbeat period calculation.
  • the self-learning process is as follows: First, the error function that reflects the error between the expected output and the calculated output of the neural network model needs to be defined as:
  • the neural network model transfer function can take many forms as needed.
  • the implicit layer transfer function can use the hyperbolic tangent sigmoid function Tan-sigmoid function:
  • the parameters of the neural network model are corrected layer by layer from the output layer through each hidden layer. Value, up to the input layer.
  • the adjustment amount ⁇ and ⁇ of the pair are proportional to the negative derivative of the parameter by the square error function, and the adjustment amount ⁇ of the pair can be expressed by the following formula:
  • the connection weight corresponding to the minimum error is determined, and learning can be stopped.
  • the learned neural network model can process the input information and calculate the heartbeat cycle.
  • the neural network corrects the connection weight according to the direction in which the error gradient falls, and superimposes the previous connection weight change amount by the ratio column, that is, the influence of the weight change is transmitted through a momentum factor.
  • the value of the momentum factor is zero, the change in weight is only generated according to the gradient descent method.
  • the value of the momentum factor is 1, the new weight change is the change of the previous weight, and other cases. Then, each time the connection weight is corrected, the correction amount at the previous learning is added in a certain ratio.
  • the specific formula is as follows:
  • X(n+1) m*(X(n)-X(nl))-(lm)* ⁇ AF(X(n))
  • AF(X(n)) is the gradient of the objective function
  • n is the iteration
  • the number of times is the learning rate
  • m is the momentum factor, ranging from 0 to 1. Adding the momentum item is to take a part of the previous weight adjustment and add it to the current weight adjustment amount to speed up the convergence.
  • Step S104 When determining a trigger condition that meets the calculation of a heartbeat period, inputting a heartbeat period calculation parameter into the neural network model to obtain a heartbeat period calculation value;
  • the neural network model can be used to dynamically calculate the heartbeat cycle.
  • the input layer nodes of the input neural network model such as CPU usage, average heartbeat period, current heartbeat period, and adjustment values, are calculated by the neural network model to obtain the reference heartbeat period.
  • the calculating the reference heartbeat period according to the network load condition in the set time includes:
  • the heartbeat period is calculated according to the input parameters through the neural network model; the network element load situation (such as by detecting the heartbeat period of the packet, the network element message or other means) is combined in a set time.
  • the current network management load, number of network elements, network congestion factor, network traffic, device usage, CPU usage, average heartbeat period, current heartbeat period, adjustment value, etc. are calculated as input reference values of the initialized learning neural network model. ⁇ .
  • step S105 it is determined whether the trigger condition of the heartbeat period update is met, if the current heartbeat period is unchanged, the step S107 is performed; if yes, the current calculated heartbeat period is accurate, and step S106 is performed;
  • the trigger condition of the heartbeat cycle update can be set according to the specific application environment. The following only gives an example of the trigger condition judgment, as follows:
  • the first threshold the range is 1-10) Seconds
  • the deviation between the calculated heartbeat period and the average of the previously obtained n-1 heartbeat period calculation values is greater than the second threshold (the range of values is 1-10 seconds). If the above trigger conditions are met, then The heartbeat cycle is changed to the nth heartbeat cycle.
  • it can also be set to update the heartbeat period to the average value of the calculated values of the n heartbeat periods when the trigger condition is met.
  • Step S106 the current heartbeat period calculation value is enabled, that is, the network element sends a heartbeat period of the detection packet to the network management according to the current heartbeat period calculation value, and step S107 is performed;
  • Step S107 updating the sample set, and executing step S102;
  • the method of updating the sample set may also be determined according to a specific algorithm, such as updating the mapping relationship between the heartbeat period calculation parameter and the heartbeat period calculation value into the original sample set, or mapping the original sample set to the newly obtained mapping relationship.
  • the relationship is weighted and merged and added to the original sample set.
  • steps S102 to S107 are loop processing flows, and there is no strict prior sequence.
  • the self-learning method of the heartbeat period in the embodiment of the present invention can dynamically adjust the heartbeat period of the detection packet according to the current environmental condition, so as to prevent the network management system or the network element from transmitting the heartbeat period of the detection packet when the network load is too large. Incorrect cycle settings result in significant business impact. At the same time, avoiding the waste of bandwidth and system resources caused by incorrect heartbeat cycle setting, and avoiding the impact on network management performance.
  • the embodiment of the present invention further provides an adaptive device for detecting a heartbeat period of a message.
  • the device includes:
  • the neural network model module is configured to calculate a heartbeat period, the input layer node is a heartbeat period parameter affecting the heartbeat period, and the output layer node is a heartbeat period, and the sample set of the neural network model includes a sample input parameter and a heartbeat period of the heartbeat period a mapping relationship between sample values, the method comprising: a self-learning triggering module, configured to trigger the neural network model to perform self-learning according to the sample set when determining that the self-learning trigger condition is met;
  • the calculation triggering module is configured to: when determining a trigger condition that meets the calculation of a heartbeat period, input a set of heartbeat period calculation parameters into the neural network model, and trigger the neural network model to calculate a calculated heartbeat period;
  • a heartbeat cycle update module configured to: when determining an update condition that meets a heartbeat cycle, update a current heartbeat cycle according to a calculated heartbeat cycle value output by the neural network model;
  • a sample set update module configured to update a sample set of the neural network model.
  • the parameters in the embodiment of the present invention may include one or more of the following: number of network elements, network congestion coefficient, network traffic, device usage, CPU usage, average heartbeat period, current heartbeat period, and adjustment value.
  • the sample set update module updates the mapping relationship between the heartbeat cycle calculation parameter and the heartbeat cycle calculation value into the original sample set.
  • the triggering condition of the heartbeat period update is: the deviation between the calculated value of the heartbeat period and the previously obtained n-1 heartbeat period calculation values is greater than the first threshold, and the calculated value of the heartbeat period is obtained previously. The deviation of the average of the n heartbeat cycle calculation values is greater than the second threshold.
  • the trigger condition for calculating the heartbeat period, the trigger condition for self-learning, or the trigger condition for updating the heartbeat period is a timing trigger or an event trigger.
  • the self-learning method and apparatus for detecting a heartbeat period of a packet can dynamically adjust the heartbeat period of the detected packet according to the current environmental condition, so as to avoid the detection of the network management or the network element when the network load is too large.
  • the heartbeat cycle of the packet is set incorrectly, which causes important services to be affected. At the same time, avoiding bandwidth and system resource waste caused by incorrect heartbeat cycle setting, and avoiding the impact on network management performance.

Abstract

一种检测报文心跳周期的自适应方法,该方法包括:构建步骤,构建神经网络模型,所述神经网络模型的输入层节点为影响心跳周期的心跳周期参数,其输出层节点为心跳周期,所述神经网络模型的样本集包括心跳周期的样本参数和心跳周期样本值的映射关系;自学习触发步骤,判断符合自学习的触发条件时,触发所述神经网络模型根据所述样本集进行自学习;计算触发步骤,判断符合计算心跳周期的触发条件时,采集心跳周期计算参数输入所述神经网络模型,得到心跳周期计算值;心跳周期更新步骤,判断符合心跳周期更新的触发条件时,根据所述心跳周期计算值更新当前心跳周期,否则执行样本集更新步骤;样本集更新步骤,更新所述样本集,返回自学习步骤。

Description

检测报文心跳周期的自适应方法和装置
技术领域
本发明涉及通信技术领域, 尤其涉及一种检测报文心跳周期的自适应方 法和装置。
背景技术
对于网络管理系统, 网络链路检测是一个必不可少的功能。 网元设备(简 称网元)通过主动响应网管服务器(简称网管) 的轮询检查、 或者网元定期 向网管发送检测报文的心跳周期的方式保证链路通畅。 如图 1所示, 网络管 理系统通过检测报文的心跳周期进行周期性的信息交流, 网元向网管发送检 测报文的心跳周期, 在链路正常的情况下, 网管接收到该检测报文的心跳周 期则认为网管与网元之间的链路正常; 若由网元异常或者链路异常导致网管 未收到网元的检测报文的心跳周期, 则表明网元与网管之间的通讯链路发生 故障。
现有心跳探测主要存在以下问题。
1、 心跳周期难以设定
心跳周期一般设定为固定的长度, 但这个周期往往很难满足各种不同的 应用场景。
2、 心跳周期算法缺乏自学习功能
某些网络管理系统釆用心跳周期动态计算方法, 也往往釆用的固定的计 算公式, 很难适应复杂多变的网络环境。
发明内容
本发明实施例要解决的技术问题是提供一种检测报文心跳周期的自适应 方法和装置, 以解决现有检测报文心跳周期的计算方法适应性差的问题。
本发明实施例的检测报文心跳周期的自适应方法包括: 构建步骤, 构建神经网络模型, 所述神经网络模型的输入层节点为影响 心跳周期的心跳周期参数, 其输出层节点为心跳周期, 所述神经网络模型的 样本集包括心跳周期的样本参数和心跳周期样本值的映射关系;
自学习触发步骤, 判断符合自学习的触发条件时, 触发所述神经网络模 型根据所述样本集进行自学习;
计算触发步骤, 判断符合计算心跳周期的触发条件时, 釆集心跳周期计 算参数输入所述神经网络模型 , 得到心跳周期计算值;
心跳周期更新步骤, 判断符合心跳周期更新的触发条件时, 根据所述心 跳周期计算值更新当前心跳周期, 判断不符合心跳周期更新的触发条件时, 执行样本集更新步骤;
样本集更新步骤, 更新所述样本集, 返回自学习步骤。
可选的, 所述参数包括以下一个或多个: 网元个数、 网络拥塞系数、 网 络流量、 设备使用率、 CPU占用率、 平均心跳周期、 当前心跳周期、 调整值。
可选的, 所述神经网络模型进行自学习的过程包括:
提取样本集中的某个样本参数作为神经网络模型的输入层节点, 得到输 出层节点的心跳周期样本计算值;
计算该心跳周期样本计算值与所述样本中对应的心跳周期样本值之间的 误差;
若该误差值小于预设阔值, 则流程结束, 若该误差值大于预设阔值, 则 根据该误差值修正该神经网络模型的连接权值, 直到重新计算的心跳周期样 本计算值与对应的心跳周期样本值之间的误差小于预设阔值。
可选的, 所述神经网路根据误差梯度下降的方向修正所述连接权值, 并 按比列叠加前次的连接权值变化量。
优选地, 所述样本集更新步骤中, 将所述心跳周期计算参数和心跳周期 计算值的映射关系更新入原样本集, 或, 将原样本集中与新得到的映射关系 相似的映射关系进行加权合并后加入原样本集。
可选地, 所述心跳周期更新的触发条件为: 所述心跳周期计算值与之前 得到的 n-1 个心跳周期计算值之间的偏差都大于第一阔值, 且所述心跳周期 计算值与之前得到的 n-1个心跳周期计算值的平均值的偏差大于第二阔值。 可选的, 所述计算心跳周期的触发条件、 自学习的触发条件或所述心跳 周期更新的触发条件为定时触发或事件触发。
为解决以上技术问题, 本发明实施例还提供了一种检测报文心跳周期的 自适应装置, 该装置包括:
神经网络模型模块, 设置为计算心跳周期, 其输入层节点为影响心跳周 期的心跳周期参数, 其输出层节点为心跳周期, 所述神经网络模型的样本集 包括心跳周期的样本参数和心跳周期样本值的映射关系, 该方法包括:
自学习触发模块, 设置为在判断符合自学习触发条件时, 触发所述神经 网络模型根据所述样本集进行自学习;
计算触发模块, 设置为在判断符合计算心跳周期的触发条件时, 釆集心 跳周期计算参数输入所述神经网络模型, 触发所述神经网络模型计算得到心 跳周期计算值;
心跳周期更新模块, 设置为在判断符合心跳周期的更新条件时, 根据所 述神经网络模型输出的心跳周期计算值更新当前心跳周期;
样本集更新模块, 设置为更新所述神经网络模型的样本集。
本发明实施例的检测报文心跳周期的自学习方法和装置, 能够根据当前 环境条件动态地调整检测报文的心跳周期, 避免在网络负载过大的情况下网 管或网元由于检测报文的心跳周期发送周期设置不正确而导致重要业务受到 影响。 同时, 避免心跳周期设置不正确导致的带宽与系统资源浪费, 还可以 避免对网管性能带来的冲击。
附图概述
图 1 为心跳链路检测机制示意图;
图 2 为心跳周期计算 BP神经网络模型结构;
图 3 为本发明实施例的检测报文心跳周期的自适应方法的流程示意图; 图 4 为本发明实施例的检测报文心跳周期的自适应装置的模块结构示意 图。 本发明的较佳实施方式
下文中将结合附图对本发明的实施例进行详细说明。 需要说明的是, 在 不冲突的情况下, 本申请中的实施例及实施例中的特征可以相互任意组合。
本实施例检测报文心跳周期的自适应方法, 如图 3所示, 包括以下具体 步骤:
步骤 S101, 构建神经网络模型;
具体地, 神经网络模型有多种, 本实施例以釆用 BP(Back Propagation, 反向传播)神经网络模型为例对本发明检测报文心跳周期的自适应方法进行 说明。
可理解地, 本发明方法不限于 BP神经网络模型。
网管系统启动时构建神经网络模型, 如图 2所示, 神经网络模型的输入 层节点(也称为输入层神经元)是影响心跳周期计算的各参数 ( XrXn ) , 主 要有网元个数 n、网络拥塞系数 δ、 网络流量 1、设备使用率 s、 CPU占用率 m、 平均心跳周期 v、 当前心跳周期 t、 调整值 z等。
网元个数 n对应网络中的实际网元数量。
网络拥塞系数 δ可以釆用如下公式计算:
Figure imgf000006_0001
a、 b 分别表示当前网管与网元的负载状况的数值, a,、 b,分别表示上次 检测报文的心跳周期接收时网管与网元的负载状况的数值, 0 a、 b、 a,、 b, < 1 , 为服务器权值, 1 10。
网络流量 1为当前网络流量除以最大可承载的网络流量。
设备使用率 s 为当前网络中正在工作的网元设备数量除以网元设备总数 量。
CPU占用率 m为当前服务器 CPU使用率。
平均心跳周期 V为服务器启动后个网元心跳周期的平均值。 当前心跳周期 t为对应待计算网元的当前心跳周期设定值。
调整值 Z为动态调整参数, 由人工根据整体计算偏差情况进行调整。 所述神经网络模型的输出层节点 (也称为输出层神经元)为心跳周期, 所述神经网络模型的样本集包括心跳周期的样本参数和心跳周期样本值的映 射关系。
隐含层负责自适应调节心跳周期, 一般来说隐含层节点的数目受输入、 输出单元的数目以及待解决问题的复杂度的影响。 如果输入层性能参数个数 为 n, 根据统计分析结果, 隐含层节点 (也称为隐含层神经元)个数釆用如 下公式计算所得的值, 其效果较好:
Figure imgf000007_0001
其中, n为输入层的节点个数, m为输出层的节点个数, z为隐含层的节 点个数, k为调整值, k的取值范围为 1到 5之间的整数。
对于隐含层来说, 第 j个节点的输出为:
0]=f( ^] 1-^]) = f(u) 其中令 " = ¾ ΧΧ'— f为隐含层的传递函数, X 0」分别为输入层第 i个节 点以及隐含层第 j个节点的输出, i=l, 2, … , n; j=l, 2, … , z。 为隐含 层中第 J个神经单元的阔值, W¾是输入层第 1个节点到隐含层第 j个节点的 连接权值。
输出层节点为计算后的心跳周期。 输出层的输出为:
其中令
Figure imgf000007_0002
, g为输出层传递函数, 0Ρ、 分别为隐含层第 ρ个节点 以及输出层第 q个节点的输出, p=l, 2, - , z; q=l, 2, - , m, 为输出 层第 q个节点阔值。 WM是隐含层第 p个节点到输出层第 q个节点的连接权值。 步骤 S102, 判断是否需要对神经网络模型进行自学习, 包括两种情况: 初始化自学习, 以及定时自学习。 初始化学习时, 初始化连接权值为 0-1之 间的随机数(即 0< 1 ) , 如果满足执行步骤 S103 , 如果不满足则执行步骤 S104
自学习触发可分为定时触发和事件触发, 比如初始化触发、 人工触发, 或者网络参数值发生剧烈变化时等。
自学习与计算没有必然先后关系, 例如也可最后进行自学习。
步骤 S103 , 触发神经网络模型进行自学习;
将待学习网元放入待学习队列, 当其它网元自学习完毕后进行自适应学 习。 按照下文所述的学习算法训练神经网络模型。
针对各网元的自学习和心跳周期计算可以穿插进行。
S101至 S107是可以针对网关所属的各网元的心跳周期进行自适应, 也 可以针对网元自身的心跳周期进行自适应。
理论上所有网元应该遵循同样的规律或算法设置心跳周期。
当网络初始化时或者设定的自学习周期到达时, 需要对神经网络模型进 行自学习。
初始化时根据当前网络前期运行状态, 利用已有样本集进行学习。 也就 是说给出一经验样本集, 该样本集每个样本作为神经网络模型某个特定输入 输出值, 即各性能参数为某值时的期望输出心跳周期。
自学习周期到达时, 需要利用初始化时的样本集, 与之前计算时的输入 与计算得到输出值作为新的学习样本, 利用学习算法重新更新神经网络模型 各权值。 即, 将后续计算得到的映射关系加入原始样本集, 实现对样本集的 更新。
样本集中存在多对映射关系时, 用于进行自学习的映射关系可以是随机 或依次从样本集中选取的, 也可以是最新的映射关系, 或是某几对或全部映 射关系的加权平均值。
通过学习动态更新神经网络模型的连接权值, 使在输入参数相同的情况 下, 心跳周期计算值尽可能与心跳周期样本值趋于一致, 从而提高心跳周期 计算的准确性。 具体的, 自学习过程如下: 首先需要定义反映神经网络模型期望输出与计算输出之间误差大小的误 差函数为:
E = l/ 2x(T - 0)2 其中, T为经验样本的心跳周期, 0为输出层节点的计算的心跳周期。 神经网络模型传递函数可根据需要釆用多种形式。 隐含层传递函数可以 釆用双曲正切 S型函数 Tan-sigmoid函数:
f(u)=tansig(u)=2/( 1 +exp(-2u))- 1 输出层传递函数可以釆用线性传递函数 purelin(v):
g(v)=purelin(v)=v 计算心跳周期样本值与实际输出的心跳周期计算值的误差, 沿误差梯度 下降的方向, 从输出层经各隐含层逐层修正神经网络模型各参数值, 直到输 入层。 对 、 的调整量 Δ 、 ^正比于平方误差函数对该参数的负导数, 对 的调整量 Δ 可用如下公式表示:
θ;的调整量 可用如下公式表示:
此过程反复交替进行, 直至计算出的心跳周期长度误差函数值小于给定 的极小值 ε ( ε的取值根据计算精度的需求进行选取, 10-3<ε<10-7, ε取值越小, 精度越高, 计算复杂度越高, 计算耗时也越久) , 从而确定了与最小误差相 对应的连接权值, 学习即可停止。 学习后的神经网络模型即能对输入信息自 行处理, 计算心跳周期。
优选地, 神经网路根据误差梯度下降的方向修正所述连接权值, 并按比 列叠加前次的连接权值变化量, 即将权值变化的影响通过一个动量因子来传 递。 当动量因子的取值为零时, 权值的变化仅是根据梯度下降法产生的。 当 动量因子的取值为 1时, 新的权值变化量为前一次权值的变化量, 其他情况 则是在每次对连接权值进行校正时,按一定比例加上前一次学习时的校正量。 具体公式如下:
X(n+1) = m*(X(n)-X(n-l))-(l-m)*^AF(X(n)) 这里 AF(X(n))为目标函数的梯度, n为迭代次数, 是学习速率, m是动 量因子, 取值范围 0到 1。 增加动量项即从前一次权值调整中取出一部分迭 加到本次权值调整量中, 以加快收敛速度。
步骤 S104, 判断符合计算心跳周期的触发条件时, 釆集心跳周期计算参 数输入所述神经网络模型 , 得到心跳周期计算值;
初始化自学习后, 该神经网络模型可用于动态计算心跳周期。
釆集各参数, 主要有网元个数、 网络拥塞系数、 网络流量、 设备使用率、
CPU占用率、 平均心跳周期、 当前心跳周期、 调整值等输入神经网络模型的 输入层节点, 通过神经网络模型计算得到参考心跳周期。
较佳的, 所述在设定的时间内, 根据网络负载情况计算参考心跳周期, 具体包括:
在设定的时间内, 根据输入参数通过神经网络模型计算心跳周期; 在设定的时间内, 将网元负载情况(如通过检测报文的心跳周期、 网元 报文或其他途径获得) 结合当前网管负载、 网元个数、 网络拥塞系数、 网络 流量、 设备使用率、 CPU占用率、 平均心跳周期、 当前心跳周期、 调整值等 作为已初始化学习神经网络模型输入值计算得到一个参考心跳周期 τ。
步骤 S105 , 判断是否符合心跳周期更新的触发条件, 若不符合则维持当 前心跳周期不变, 执行步骤 S107; 若符合则表明当前计算的心跳周期是准确 的, 执行步骤 S 106;
心跳周期更新的触发条件可以根据具体的应用环境进行设定, 以下仅给 出一种触发条件判断的示例, 具体如下:
计算得到第 η个心跳周期计算值后, 若该第 η个心跳周期计算值与之前 得到的 n-1个心跳周期计算值之间的偏差都大于第一阔值(取值范围为 1-10 秒) , 且所述心跳周期计算值与之前得到的 n-1 个心跳周期计算值的平均值 的偏差大于第二阔值(取值范围为 1-10秒) 。 如果符合上述触发条件, 则更 改心跳周期为第 n个心跳周期计算值。
当然也可以设定为在符合触发条件时, 将心跳周期更新为将该 n个心跳 周期计算值的平均值。
步骤 S106, 启用当前的心跳周期计算值, 即网元按照当前的心跳周期计 算值向网管发送检测报文的心跳周期, 执行步骤 S107;
步骤 S 107 , 更新样本集, 转执行步骤 S 102;
更新样本集的方法也可以根据具体的算法确定, 比如将所述心跳周期计 算参数和心跳周期计算值的映射关系更新入原样本集, 或, 将原样本集中与 新得到的映射关系相似的映射关系进行加权合并后加入原样本集。
可理解地, 步骤 S102至步骤 S107是循环处理流程, 并不存在严格的先 后顺序。
本发明实施例的所述心跳周期的自学习方法, 能够根据当前环境条件动 态地调整检测报文的心跳周期, 避免在网络负载过大的情况下网管或网元由 于检测报文的心跳周期发送周期设置不正确而导致重要业务受到影响。同时, 避免心跳周期设置不正确导致的带宽与系统资源浪费, 还可以避免对网管性 能带来的冲击。
为了实现以上方法, 本发明实施例还提供了一种检测报文心跳周期的自 适应装置, 如图 4所示, 该装置包括:
神经网络模型模块, 设置为计算心跳周期, 其输入层节点为影响心跳周 期的心跳周期参数, 其输出层节点为心跳周期, 所述神经网络模型的样本集 包括心跳周期的样本输入参数和心跳周期样本值的映射关系, 该方法包括: 自学习触发模块, 设置为在判断符合自学习触发条件时, 触发所述神经 网络模型根据所述样本集进行自学习;
计算触发模块, 设置为在判断符合计算心跳周期的触发条件时, 釆集心 跳周期计算参数输入所述神经网络模型 , 触发所述神经网络模型计算得到心 跳周期计算值;
心跳周期更新模块, 设置为在判断符合心跳周期的更新条件时, 根据所 述神经网络模型输出的心跳周期计算值更新当前心跳周期; 样本集更新模块, 设置为更新所述神经网络模型的样本集。
本发明实施例所说的参数可以包括以下一个或多个: 网元个数、 网络拥 塞系数、 网络流量、 设备使用率、 CPU占用率、 平均心跳周期、 当前心跳周 期、 调整值。
具体地, 所述样本集更新模块将所述心跳周期计算参数和心跳周期计算 值的映射关系更新入原样本集。
以上只是给出了一种样本集更新的方式, 还可以将原样本集中与新得到 的映射关系相似的映射关系进行加权合并后加入原样本集。
所述心跳周期更新的触发条件为:所述心跳周期计算值与之前得到的 n-1 个心跳周期计算值之间的偏差都大于第一阔值, 且所述心跳周期计算值与之 前得到的 n个心跳周期计算值的平均值的偏差大于第二阔值。
所述计算心跳周期的触发条件、 自学习的触发条件或所述心跳周期更新 的触发条件为定时触发或事件触发。
本领域普通技术人员可以理解上述方法中的全部或部分步骤可通过程序 来指令相关硬件完成, 所述程序可以存储于计算机可读存储介质中, 如只读 存储器、 磁盘或光盘等。 可选地, 上述实施例的全部或部分步骤也可以使用 一个或多个集成电路来实现。 相应地, 上述实施例中的各模块可以釆用硬件 的形式实现, 也可以釆用软件功能模块的形式实现。 本发明不限制于任何特 定形式的硬件和软件的结合。
工业实用性 本发明实施例的检测报文心跳周期的自学习方法和装置, 能够根据当前 环境条件动态地调整检测报文的心跳周期, 避免在网络负载过大的情况下网 管或网元由于检测报文的心跳周期发送周期设置不正确而导致重要业务受到 影响。 同时, 避免心跳周期设置不正确导致的带宽与系统资源浪费, 还可以 避免对网管性能带来的冲击。

Claims

1、 一种检测报文心跳周期的自适应方法, 该方法包括:
构建步骤, 构建神经网络模型, 所述神经网络模型的输入层节点为影响 心跳周期的心跳周期参数, 其输出层节点为心跳周期, 所述神经网络模型的 样本集包括心跳周期的样本参数和心跳周期样本值的映射关系;
自学习触发步骤, 判断符合自学习的触发条件时, 触发所述神经网络模 型根据所述样本集进行自学习;
计算触发步骤, 判断符合计算心跳周期的触发条件时, 釆集心跳周期计 算参数输入所述神经网络模型 , 得到心跳周期计算值;
心跳周期更新步骤, 判断符合心跳周期更新的触发条件时, 根据所述心 跳周期计算值更新当前心跳周期, 判断不符合心跳周期更新的触发条件时, 执行样本集更新步骤; 样本集更新步骤, 更新所述样本集, 返回自学习步骤。
2、 如权利要求 1所述的方法, 其中: 所述参数包括以下一个或多个: 网 元个数、 网络拥塞系数、 网络流量、 设备使用率、 CPU占用率、 平均心跳周 期、 当前心跳周期、 调整值。
3、 如权利要求 1所述的方法, 其中: 所述神经网络模型进行自学习的步 骤包括:
提取样本集中的某个样本参数作为神经网络模型的输入层节点, 得到输 出层节点的心跳周期样本计算值;
计算该心跳周期样本计算值与所述样本中对应的心跳周期样本值之间的 误差;
若该误差值小于预设阔值, 则流程结束, 若该误差值大于预设阔值, 则 根据该误差值修正该神经网络模型的连接权值, 直到重新计算的心跳周期样 本计算值与对应的心跳周期样本值之间的误差小于预设阔值。
4、 如权利要求 3所述的方法, 其中: 在所述根据该误差值修正该神经网 络模型的连接权值的步骤中, 所述神经网路根据误差梯度下降的方向修正所 述连接权值, 并按比列叠加前次的连接权值变化量。
5、 如权利要求 1所述的方法, 其中: 所述样本集更新步骤中, 将所述心 跳周期计算参数和心跳周期计算值的映射关系更新入原样本集, 或, 将原样 本集中与新得到的映射关系相似的映射关系进行加权合并后加入原样本集。
6、 如权利要求 1所述的方法, 其中: 所述心跳周期更新的触发条件为: 所述心跳周期计算值与之前得到的 n-1 个心跳周期计算值之间的偏差都大于 第一阔值, 且所述心跳周期计算值与之前得到的 n-1 个心跳周期计算值的平 均值的偏差大于第二阔值。
7、 如权利要求 1所述的方法, 其中: 所述计算心跳周期的触发条件、 自 学习的触发条件或所述心跳周期更新的触发条件为定时触发或事件触发。
8、 一种检测报文心跳周期的自适应装置, 该装置包括:
神经网络模型模块, 其设置为: 计算心跳周期, 其输入层节点为影响心 跳周期的心跳周期参数, 其输出层节点为心跳周期, 所述神经网络模型的样 本集包括心跳周期的样本参数和心跳周期样本值的映射关系;
自学习触发模块, 其设置为: 在判断符合自学习触发条件时, 触发所述 神经网络模型根据所述样本集进行自学习;
计算触发模块, 其设置为: 在判断符合计算心跳周期的触发条件时, 釆 集心跳周期计算参数输入所述神经网络模型, 触发所述神经网络模型计算得 到心跳周期计算值;
心跳周期更新模块, 其设置为: 在判断符合心跳周期的更新条件时, 根 据所述神经网络模型输出的心跳周期计算值更新当前心跳周期;
样本集更新模块, 其设置为: 更新所述神经网络模型的样本集。
9、 如权利要求 8所述的装置, 其中: 所述参数包括以下一个或多个: 网 元个数、 网络拥塞系数、 网络流量、 设备使用率、 CPU 占用率、 平均心跳周 期、 当前心跳周期、 调整值。
10、 如权利要求 8所述的装置, 其中: 所述样本集更新模块是设置为: 将所述心跳周期计算参数和心跳周期计算值的映射关系更新入原样本集,或, 将原样本集中与新得到的映射关系相似的映射关系进行加权合并后加入原样 本集。
11、 如权利要求 8所述的装置, 其中: 所述心跳周期更新的触发条件为: 所述心跳周期计算值与之前得到的 n-1 个心跳周期计算值之间的偏差都大于 第一阔值, 且所述心跳周期计算值与之前得到的 n-1 个心跳周期计算值的平 均值的偏差大于第二阔值。
12、 如权利要求 8所述的装置, 其中: 所述计算心跳周期的触发条件、 自学习的触发条件或所述心跳周期更新的触发条件为定时触发或事件触发。
PCT/CN2012/077723 2011-11-23 2012-06-28 检测报文心跳周期的自适应方法和装置 WO2013075489A1 (zh)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN201110376348 2011-11-23
CN201110376348.7 2011-11-23
CN201210037272XA CN103139008A (zh) 2011-11-23 2012-02-17 检测报文心跳周期的自适应方法和装置
CN201210037272.X 2012-02-17

Publications (1)

Publication Number Publication Date
WO2013075489A1 true WO2013075489A1 (zh) 2013-05-30

Family

ID=48469070

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2012/077723 WO2013075489A1 (zh) 2011-11-23 2012-06-28 检测报文心跳周期的自适应方法和装置

Country Status (2)

Country Link
CN (1) CN103139008A (zh)
WO (1) WO2013075489A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111698132A (zh) * 2020-06-12 2020-09-22 北京字节跳动网络技术有限公司 用于控制集群中心跳事件的方法、装置、设备和介质

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103532805A (zh) * 2013-11-06 2014-01-22 中国联合网络通信集团有限公司 家庭终端周期性报文的发送方法及装置
CN107078925B (zh) * 2014-10-29 2020-07-07 华为技术有限公司 一种心跳周期的设置方法及终端
CN110417565B (zh) * 2018-04-27 2021-01-29 华为技术有限公司 一种模型更新方法、装置及系统
CN109548066B (zh) * 2018-12-27 2022-04-22 郑州天迈科技股份有限公司 一种无线通信链路状态检测方法
CN110049532B (zh) * 2019-03-12 2021-04-30 上海感悟通信科技有限公司 一种用于无线自组网的心跳方法、受控节点和控制节点
CN113162813B (zh) * 2020-01-22 2022-06-03 烽火通信科技股份有限公司 一种ptp报文丢失检测分析方法和装置
CN111953569B (zh) * 2020-08-27 2022-04-29 浪潮电子信息产业股份有限公司 状态信息上报方法、装置、设备及介质
CN112671633B (zh) * 2020-12-01 2022-08-23 重庆邮电大学 一种基于bp神经网络预测的二分探测心跳间隔系统及方法
CN116800301A (zh) * 2022-03-14 2023-09-22 华为技术有限公司 一种电力线通信系统、在网状态检测方法及设备

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6370656B1 (en) * 1998-11-19 2002-04-09 Compaq Information Technologies, Group L. P. Computer system with adaptive heartbeat
CN101594268A (zh) * 2009-06-23 2009-12-02 中兴通讯股份有限公司 链路检测方法

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1301387C (zh) * 2004-06-04 2007-02-21 广东科龙电器股份有限公司 基于神经网络的空调器噪声源识别方法
CN101115313B (zh) * 2007-08-23 2011-01-19 中兴通讯股份有限公司 一种软心跳检测方法及系统
CN101408933A (zh) * 2008-05-21 2009-04-15 浙江师范大学 基于粗网格特征提取和bp神经网络的车牌字符识别方法
CN101504790B (zh) * 2008-11-07 2010-12-22 清华大学 红外光束型火灾烟雾探测器
US8840562B2 (en) * 2009-09-24 2014-09-23 Covidien Lp Signal processing warping technique
CN102201937B (zh) * 2011-06-13 2013-10-23 刘胜利 基于心跳行为分析的快速木马检测方法
CN102223257B (zh) * 2011-07-25 2014-02-19 莫雅静 一种心跳间隔调整方法、服务器、客户端及通信系统

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6370656B1 (en) * 1998-11-19 2002-04-09 Compaq Information Technologies, Group L. P. Computer system with adaptive heartbeat
CN101594268A (zh) * 2009-06-23 2009-12-02 中兴通讯股份有限公司 链路检测方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LI, JI ET AL.: "An Momentum BP Neural Network Algorithm and Application", ELECTROTECHNICAL APPLICATION, vol. 24, no. 6, June 2005 (2005-06-01), pages 42 - 44 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111698132A (zh) * 2020-06-12 2020-09-22 北京字节跳动网络技术有限公司 用于控制集群中心跳事件的方法、装置、设备和介质
CN111698132B (zh) * 2020-06-12 2022-03-01 北京字节跳动网络技术有限公司 用于控制集群中心跳事件的方法、装置、设备和介质

Also Published As

Publication number Publication date
CN103139008A (zh) 2013-06-05

Similar Documents

Publication Publication Date Title
WO2013075489A1 (zh) 检测报文心跳周期的自适应方法和装置
CN111092823B (zh) 一种自适应调节拥塞控制初始窗口的方法和系统
Bu et al. Fixed point approximations for TCP behavior in an AQM network
CN113966596B (zh) 用于数据流量路由的方法和设备
CN112054966B (zh) 多路传输数据调度方法、装置及设备
CN110247795B (zh) 一种基于意图的云网资源服务链编排方法及系统
CN112600759A (zh) 基于深度强化学习在Overlay网络下多路径流量调度方法及系统
Liu et al. ACCP: adaptive congestion control protocol in named data networking based on deep learning
CN110995619A (zh) 一种服务质量感知的虚拟网络映射方法和装置
Godfrey et al. Incentive compatibility and dynamics of congestion control
Wang et al. A thorough analysis of the performance of delay distribution models for IEEE 802.11 DCF
CN112104563A (zh) 一种拥塞控制方法及装置
Bisoy et al. Design of an active queue management technique based on neural networks for congestion control
CN110971451B (zh) Nfv资源分配方法
Poojary et al. Analysis of multiple flows using different high speed TCP protocols on a general network
Gupta et al. Analytical modeling of TCP flow in wireless LANs
CN116055406A (zh) 拥塞窗口预测模型的训练方法及装置
Cho et al. Adaptive neural queue management for TCP networks
Dunaytsev et al. The PFTK-model revised
Kassa Analytic models of TCP performance
CN113556780B (zh) 一种拥塞控制方法及装置
Mamduhi et al. State-dependent data queuing in shared-resource networked control systems
Haider et al. PID based congestion control algorithms for AQM routers supporting TCP/IP flows
Ali et al. Efficient congestion control in communications using novel weighted ensemble deep reinforcement learning
WO2022176266A1 (ja) 制御システムに含まれるネットワークの通信帯域を最適化する方法、装置、プログラム、および制御システム

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 12852262

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 12852262

Country of ref document: EP

Kind code of ref document: A1