CN110011833A - A kind of online Reliability Prediction Method of service system based on deep learning - Google Patents
A kind of online Reliability Prediction Method of service system based on deep learning Download PDFInfo
- Publication number
- CN110011833A CN110011833A CN201910179581.2A CN201910179581A CN110011833A CN 110011833 A CN110011833 A CN 110011833A CN 201910179581 A CN201910179581 A CN 201910179581A CN 110011833 A CN110011833 A CN 110011833A
- Authority
- CN
- China
- Prior art keywords
- layer
- service system
- reliability
- lstm
- time
- Prior art date
- Legal status (The legal status 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 status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/147—Network analysis or design for predicting network behaviour
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of online Reliability Prediction Methods of the service system based on deep learning, and depth confidence neural network is combined with shot and long term Memory Neural Networks, carry out feature extraction first with reliability sequence of the depth confidence neural network to history;Then using shot and long term Memory Neural Networks and improved two-way shot and long term Memory Neural Networks, to treated, reliability sequence is predicted, improves the accuracy of prediction.The present invention can effectively realize the on-line prediction of service system reliability, it can reflect the irregular feature of upheaval of service system reliability in time, it can be provided with the early stage information in terms of system reliability for the quality assurance of the service system based on Services Composition, provide support for the effectively mistake of prevention service system and abnormal quality.
Description
Technical field
The present invention relates to a kind of technologies for carrying out on-line prediction come the reliability to service system using deep learning, belong to
Service computing technique field.
Background technique
Active defect management is a kind of for improving the more effective means of system reliability.Its core concept is logical
The defect of forecasting system is crossed, and system is adjusted in advance, to achieve the purpose that avoidance system mistake.It is serviced due to constituting
Each Component service of system operates under environment complicated and changeable, and the quality of Component service can fluctuate at any time.In order to
The active running quality management for realizing service system, to evade the abnormal performance of service system, it would be desirable to can predict to take
The reliability of Component service in business system.
Under highly dynamic running environment, the QoS(Quality of Service of each service) it can occur at any time
Fluctuation.Software Quality Assurance when in order to support effective services selection and service system to run, the reliability of service system are pre-
Surveying must be a kind of online prediction, i.e., what we needed to predict is single service within the not far following period
Real-time reliability.Therefore, the online reliability time series forecasting problem that the present invention is studied service system different from the past
The reliability prediction of system.Prediction of the invention is it is intended that the selection and quality assurance of service system components service provide related service
Early stage information in terms of reliability provides support for the effectively mistake of prevention service system and abnormal quality.
Service system is a kind of cross-platform, loose coupling software application.Unlike traditional system component-based,
Each of service system Component service is usually provided by third party, is disposed in a network, is called by network remote.Due to
Multiple services are combined with, service system can satisfy the increasingly complex demand of user.However, service system operate in it is highly dynamic
In the environment of (such as network environment variation, service quality itself fluctuation it is even abnormal), the variation individually serviced may
Cascading is generated, causes whole system that cannot work.In order to construct highly reliable service system, and ensure service system
Effectively operation, reliability prediction is extremely necessary.But it is complicated and changeable and unstable due to Component service running environment
Property, so that the generation of the mistake of Component service does not have apparent regularity in time, this causes traditional online mistake pre-
Survey method is difficult to adapt to the online reliability prediction problem of service system towards uncertain error event.
Currently, on the problems such as research to service system is concentrated mainly on the building and optimization of system, it is pre- to its reliability
The research of survey is then fewer.However, reliability prediction problem there has been centainly as a hot issue in field of service calculation
Research foundation, also produce some more representative research methods.For example, personalized Reliability Prediction Method, poly-
Class Reliability Prediction Method, both methods are all that the performance of Web service is assessed with the mean reliability of history, and reflection is
From the point of view of the relatively long period, the overall reliability of each Component service.Therefore, these methods can not support that service is
The forecasting problem of the online reliability of system.In addition, there are also some online error prediction technologies towards traditional computer system
Research.For example, the detection of error tracking, sign, the error reporting detected, the mistake audit not detected etc., these methods are all
Analysis system inner parameter is needed, the more difficult acquisition for service system of these parameters.In addition there are also certain methods such as bases
In the Bayesian forecasting of conditional probability, nonparametric technique prediction, curve-fitting method, semi-Markov model, SVM model etc.,
The generation that mistake can only be modeled meets the error event of Poisson distribution in time, and in fact, service system running environment is multiple
Miscellaneous changeable, Component service itself has unstability, thus causes the generation of mistake in time and does not have apparent rule.
Therefore, these methods can not support online reliability prediction problem proposed by the invention completely.
It can be seen that effective solution of the online reliability prediction problem of service system, will protect the quality of service system
Barrier problem provides a kind of effective solution scheme, and certain support is provided for complex services systematic difference.
Summary of the invention
A kind of method that the main object of the present invention is to provide deep learning to carry out the reliability of service system online
Prediction.This method not only can solve the online reliability prediction problem of service system, while but also reliability prediction knot
Fruit is more accurate, is suitable for the application environment that component system is complicated and changeable, unstable.
In order to achieve the above objectives, the method that the present invention uses is: a kind of service system based on deep learning is reliable online
Property prediction technique, includes the following steps:
(1), by this three groups of data of the reliability, handling capacity, response time of analytic unit service history, predict leading time it
The reliability in time serviced in an effective time sequence period afterwards;
(2), feature extraction is carried out to the time series of history by DBN, using in DBN between visible layer and hidden layer it is non-supervisory
The method that greediness successively maps carries out feature selecting to the data being input in network, to allow DBN with smaller dimension
Initial data is reconstructed;
(3), after data processing finishes, the shot and long term Memory Neural Networks LSTM model for being usually used in handling time series is selected,
Reliability time series is predicted;
(4), it is further rested and reorganized using two-way LSTM neural network to above-mentioned model.
As an improvement of the present invention, the DBN is usually made of multilayer RBM, in successively unsupervised mode pair
Model is trained, and by using there is the measure of supervision of label to be adjusted the parameter in DBN, optimizes network knot
Structure.
As an improvement of the present invention, DBN is formed method particularly includes: first it will be seen that the unit of layer hiding to upper layer
Unit carries out log probability calculating, and using the output valve of currently hiding layer unit as the input value of upper one layer of hidden layer, will be upper
Lower two hidden layers are trained as new RBM, and in an identical manner;Top layer is a BP neural network, inputs and is
The output of the hidden layer of a RBM topmost further finely tunes whole network parameter by using BP algorithm.
As an improvement of the present invention, the LSTM is made of input layer hidden layer output layer, before being different from multilayer
Neural network is presented, LSTM allows to add connection in layer other than the connection of interlayer, and the connection in layer is so that LSTM allows
Accumulated in time-domain, LSTM be unfolded in time-domain, each moment of LSTM contain before several
The information at moment.It is successively calculated in chronological order when LSTM is propagated forward, back-propagation is then the ladder from the last one moment
Degree starts successively to accumulate forward.
As an improvement of the present invention, the two-way LSTM is using two hidden layers respectively to historical information and future
Information is saved, and shares the same output layer.
The utility model has the advantages that
The present invention can effectively realize the on-line prediction of service system reliability, can compared to current already present some predictions
By the method for property, the present invention can more reflect the irregular feature of upheaval of service system reliability in time.Therefore, the present invention can be with
It is provided with the early stage information in terms of system reliability for the quality assurance of the service system based on Services Composition, for effectively
The mistake of prevention service system and the generation of abnormal quality provide support.
Detailed description of the invention
Fig. 1 is that on-line prediction technology is effectively predicted matter of time schematic diagram;
Fig. 2 is depth confidence neural network schematic diagram;
Fig. 3 is the structure chart of shot and long term Memory Neural Networks and its expanded form in time-domain;
Fig. 4 is two-way shot and long term Memory Neural Networks schematic diagram.
Specific embodiment
With reference to the accompanying drawing (table) the present invention is described in detail.
Firstly, for input data, the invention solves reliability prediction problem be a kind of online prediction, i.e., we
Need to predict is real-time reliability of the Component service within the not far following period.As shown in Figure 1, i.e. pre-
Measure and (time be effectively predicted) reliability in the period.It wherein, is leading time (leading time) that the period begins
In current timet, until the Component service is called.In general, length needs to be greater than one combined system of construction or replaces
Change the time required for undesirable Component service;For time (prediction period) is effectively predicted, which should be able to
Cover the execution period after service is called.Since each service called execution period has with short, in order to meet
The demand of most users, length be defined as greater than most services called periods;For data window time (data
Window time), indicate a nearest historical time section.
Through investigating, it has been found that the reliability of service system is mainly influenced by the following aspects: (1) network bandwidth
Fluctuation and Component service to the bearing capacity of current throughput;(2) Component service server working condition (free memory,
Cpu load, magnetic disc i/o, function call etc.);(3) Component service sole mass (Memory recycle mechanism, support platform stability,
System bugs etc.);(4) system caller calling scale (if a Component service has invoked a large amount of other Component services,
The reliability that so caller is showed will also decrease).
Due to the present invention study be service system component level reliability of service, so we are primarily upon (1) (2)
Two factors.However, different Component services is voluntarily managed by each self-organizing for service system, we are difficult from clothes
Business end is gone to obtain the parameter of each Component service be respectively distributed and being managed by different tissues.The reliability of Component service
Influence factor (1) (2) will mainly have an impact the handling capacity of Component service and response time in terms of reflection to QoS, and gulp down
The amount of spitting and this two groups of parameters of response time are easier to obtain.Therefore, the quasi- reliability by analytic unit service history of the present invention,
Handling capacity, this three groups of data of response time, predict leading time after an effective time sequence period in service it is reliable
Property.
Secondly, we first carry out feature extraction to data in order to improve predictablity rate, what we selected herein is deep
It spends confidence neural network (DBN), depth confidence network achieves huge success in each application field at present, such as schemes
As classification, speech recognition, time series forecasting etc..Since DBN is using the unsupervised method successively recycled to being input to network
In data analyzed, enable this method to indicate the feature of initial data with less information content.DBN is usually by multilayer
Limited Boltzmann machine (RBM) composition, is trained model in successively unsupervised mode, and by using the prison for having label
It superintends and directs method to be adjusted the parameter in DBN, to achieve the purpose that optimize network structure.
Specifically, as shown in Fig. 2, DBN is composed of multiple RBM, first it will be seen that the unit of layer is to the hidden of upper layer
It hides unit and carries out log probability calculating, and using the output valve of currently hiding layer unit as the input value of upper one layer of hidden layer, it will
Upper and lower two hidden layers are trained as new RBM, and in an identical manner;Top layer is a BP neural network, input
For the output of the hidden layer of a RBM topmost, whole network parameter is further finely tuned by using BP algorithm.
This method carries out pre-training to initial data using the greedy algorithm of unsupervised mode, and DBN network is continuous in training
The high-order learnt in initial data indicates feature, this makes network not only possess preferable dimensionality reduction performance, while also overcoming biography
System shallow structure needs the shortcomings that largely having label data and being easily trapped into local minimum point;And then using the BP for having monitor mode
Algorithm (can also change other network models into) is finely adjusted the parameter of whole network, and final acquisition reconstructed error is the smallest defeated
Data out.This method is equivalent to the network to the Training stage to the unsupervised learning preprocessing process of original input data
Parameter is initialized, and initial parameter is maintained at a preferable range, this makes DBN network overcome conventional network structure
It is easy to be influenced by initial parameter, handles the slow-footed disadvantage of high dimensional data, and also improve classification application and prediction application
Accuracy.
Again, after data processing finishes, just reliability time series is predicted, we are usually used in locating quasi- select
Shot and long term Memory Neural Networks (LSTM) model of time series is managed, as shown in Figure 3.This model is exported by input layer hidden layer
Layer is constituted, and is different from multilayer feedforward neural network, and LSTM allows to add connection, layer in layer other than the connection of interlayer
Interior connection is so that LSTM permission is accumulated in time-domain.LSTM is unfolded in time-domain, it may be seen that
The information at several moment before each moment of LSTM contains.It is successively counted in chronological order when LSTM is propagated forward
It calculates, back-propagation is successively accumulated forward since the gradient at the last one moment.
Finally, in order to further increase the accuracy rate of prediction, the two-way LSTM neural network (bi- of our proposed adoptions
Directional LSTM, Bi-LSTM) it is further rested and reorganized to above-mentioned model.The characteristic of LSTM is to obtain active cell
The information of all units before to this unit, it the shortcomings that be that can not obtain the information of the unit after this unit, therefore it is two-way
LSTM neural network is just come into being.As shown in figure 4, the model is using two hidden layers respectively to historical information and Future Information
It is saved, and shares the same output layer.Such setting can make model when handling specific data, use entire sequence
Data information in column has greatly improved effect to the estimated performance of model.
It is by above description as can be seen that proposed by the invention based on depth confidence neural network and shot and long term memory mind
It combines through network and based on the method that depth confidence neural network is combined with two-way shot and long term Memory Neural Networks, passes through
Offline analysis is carried out to the history parameters of service system, can efficiently realize and proposed by the invention exist to service system
Line reliability prediction, not only meets the demand of on-line prediction, also effectively improves the accuracy rate of prediction.It can be seen that this
One invention is very useful.
Claims (5)
1. a kind of online Reliability Prediction Method of service system based on deep learning, which comprises the steps of:
(1), by this three groups of data of the reliability, handling capacity, response time of analytic unit service history, predict leading time it
The reliability in time serviced in an effective time sequence period afterwards;
(2), feature extraction is carried out to the time series of history by DBN, using in DBN between visible layer and hidden layer it is non-supervisory
The method that greediness successively maps carries out feature selecting to the data being input in network, to allow DBN with smaller dimension
Initial data is reconstructed;
(3), after data processing finishes, the shot and long term Memory Neural Networks LSTM model for being usually used in handling time series is selected,
Reliability time series is predicted;
(4), it is further rested and reorganized using two-way LSTM neural network to above-mentioned model.
2. the online Reliability Prediction Method of a kind of service system based on deep learning according to claim 1, feature
Be: the DBN is usually made of multilayer RBM, is trained in successively unsupervised mode to model, and by using having
The measure of supervision of label is adjusted the parameter in DBN, optimizes network structure.
3. the online Reliability Prediction Method of a kind of service system based on deep learning according to claim 2, feature
It is, DBN composition method particularly includes: first it will be seen that the unit of layer carries out log probability calculating to the hidden unit on upper layer, and
Currently to hide the output valve of layer unit as the input value of upper one layer of hidden layer, using upper and lower two hidden layers as new RBM,
And it is trained in an identical manner;Top layer is a BP neural network, is inputted as the defeated of the hidden layer of a RBM topmost
Out, whole network parameter is further finely tuned by using BP algorithm.
4. the online Reliability Prediction Method of a kind of service system based on deep learning according to claim 1, feature
Be: the LSTM is made of input layer hidden layer output layer, is different from multilayer feedforward neural network, LSTM is in addition to interlayer
Connection is outer, while allowing to add connection in layer, and the connection in layer is so that LSTM permission is accumulated in time-domain;By LSTM
It is unfolded in time-domain, the information at several moment before each moment of LSTM contains;It is propagated forward in LSTM
When successively calculate in chronological order, back-propagation is successively accumulated forward since the gradient at the last one moment.
5. the online Reliability Prediction Method of a kind of service system based on deep learning according to claim 1, feature
Be: the two-way LSTM respectively saves historical information and Future Information using two hidden layers, and shares same
A output layer.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910179581.2A CN110011833B (en) | 2019-03-11 | 2019-03-11 | Service system online reliability prediction method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910179581.2A CN110011833B (en) | 2019-03-11 | 2019-03-11 | Service system online reliability prediction method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110011833A true CN110011833A (en) | 2019-07-12 |
CN110011833B CN110011833B (en) | 2022-04-22 |
Family
ID=67166810
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910179581.2A Active CN110011833B (en) | 2019-03-11 | 2019-03-11 | Service system online reliability prediction method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110011833B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111062464A (en) * | 2019-10-24 | 2020-04-24 | 中国电力科学研究院有限公司 | Power communication network reliability prediction and guarantee method and system based on deep learning |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107103359A (en) * | 2017-05-22 | 2017-08-29 | 东南大学 | The online Reliability Prediction Method of big service system based on convolutional neural networks |
-
2019
- 2019-03-11 CN CN201910179581.2A patent/CN110011833B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107103359A (en) * | 2017-05-22 | 2017-08-29 | 东南大学 | The online Reliability Prediction Method of big service system based on convolutional neural networks |
Non-Patent Citations (3)
Title |
---|
HONGBING WANG等: "《Online Reliability Prediction via Motifs-Based》", 《IEEE》 * |
磐创AI: "《使用Keras进行深度学习:(六)LS丁M和双向LS丁M讲解及实践》", 《CSDN论坛》 * |
阴法明等: "《 基于深度置信网络的语音增强算法》", 《CNKI》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111062464A (en) * | 2019-10-24 | 2020-04-24 | 中国电力科学研究院有限公司 | Power communication network reliability prediction and guarantee method and system based on deep learning |
CN111062464B (en) * | 2019-10-24 | 2022-07-01 | 中国电力科学研究院有限公司 | Power communication network reliability prediction and guarantee method and system based on deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN110011833B (en) | 2022-04-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10360517B2 (en) | Distributed hyperparameter tuning system for machine learning | |
US11693917B2 (en) | Computational model optimizations | |
CN112270547A (en) | Financial risk assessment method and device based on feature construction and electronic equipment | |
Soleimani et al. | Artificial neural network application in predicting probabilistic seismic demands of bridge components | |
CN112308288A (en) | Particle swarm optimization LSSVM-based default user probability prediction method | |
CN110781409A (en) | Article recommendation method based on collaborative filtering | |
AU2019371339A1 (en) | Finite rank deep kernel learning for robust time series forecasting and regression | |
CN111369344B (en) | Method and device for dynamically generating early warning rules | |
CN112819523B (en) | Marketing prediction method combining inner/outer product feature interaction and Bayesian neural network | |
Tomar et al. | Active learning method for risk assessment of distributed infrastructure systems | |
Talagala et al. | Meta‐learning how to forecast time series | |
Zhang et al. | Seismic response prediction of a damped structure based on data-driven machine learning methods | |
CN116484105B (en) | Service processing method, device, computer equipment, storage medium and program product | |
CN110011833A (en) | A kind of online Reliability Prediction Method of service system based on deep learning | |
Zou et al. | Intentionality-related deep learning method in web prefetching | |
CN117035038A (en) | Model pruning method, device, equipment and storage medium | |
Qin et al. | A hybrid deep learning model for short‐term load forecasting of distribution networks integrating the channel attention mechanism | |
Zang | Construction of Mobile Internet Financial Risk Cautioning Framework Based on BP Neural Network | |
Singh et al. | A feature extraction and time warping based neural expansion architecture for cloud resource usage forecasting | |
Moghadas et al. | A New Wavelet Back Propagation Neural Networks for Structural Dynamic Analysis. | |
CN112667394A (en) | Computer resource utilization rate optimization method | |
KR101982636B1 (en) | Method and apparatus for generating prediction model of company performance | |
CN111563767A (en) | Stock price prediction method and device | |
Zhu et al. | Research on system of data mining technology based on computer | |
Hou | Decision tree algorithm for big data analysis |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |