CN102749848A - Water supply pipeline safety monitoring method based on useless component projection analysis - Google Patents

Abstract

The invention discloses a pipeline system safety event alarm system, which is realized by modeling through earlier acoustics data, inputting the detected acoustics data into a model and judging the state change situation of a detected pipeline. A safety monitoring method disclosed by the invention comprises the following steps of: adopting an EM (Expectation Maximization) algorithm and a useless component resolving algorithm to obtain a general background Gauss mixture model and a projection model respectively; using a self-adaptive technology and projecting a useless component to obtain a projection vector; adopting an SVM (Support Vector Machine) training algorithm to obtain a training model; and introducing the projection vector obtained by using detected data into the training model to obtain a corresponding detection result. The method has the advantages that the EM algorithm and useless component projection analysis are adopted, so that the requirement of the general background Gauss mixture model on modeling and testing data sizes can be lowered on a large scale, the adaption time of the system on each application environment can be shortened remarkably, and time from receiving of abnormal signals for the first time to judgment and conclusion is reduced.

Description

Water supply line method for safety monitoring based on the idle component Projection Analysis

Technical field

The present invention relates to that acoustic signal is handled, area of pattern recognition, specifically, is a kind of based on the idle component Projection Analysis, to destroying the method that the water supply line activity detects.

Background technology

For guaranteeing city supply water pipeline safety, need be to destroying the timely early warning of water supply line behavior, in case water supply line leaks.Cause the principal element of pipe leakage that the artificial damage of corrosive pipeline, non-subjectivity, subjective artificial destruction and irresistible natural force etc. are arranged.Generally speaking, piping corrosion is physics, chemical process slowly, and mainly renewal addressing this problem with pipeline through regular pipeline maintenance.The content that the artificial damage of non-subjectivity comprises is very extensive, for example, during the heavy-duty machinery operation, does not know the particular location of pipeline in the operating area, unconscious destruction pipeline; When pipeline is operated,, make pipeline burst because misoperation causes pipeline pressure too high.These factors can cause the sudden change of some physical parameter in the pipeline, and pepe monitoring system can be reported to the monitor staff based on the variation of monitoring parameter, and then through action timely and effectively, the generation of prevention serious consequence.Subjective artificial destruction refers to that the lawless person passes through the malicious sabotage to pipeline, reaches certain hidden purpose.Modal failure mode is through iron apparatus, to the rigid infringement of pipeline.We are referred to as the water supply line activity that destroys with the artificial damage of non-subjectivity with subjective artificial destruction.

Leakage and the movable detection of destruction water supply line to water supply line have multiple means, and traditional method comprises artificial process, digital picture detection method, NPW method, flow equilibrium method, optical Fiber Method and acoustics signal Processing method.So-called artificial process is meant the tour through the professional, finds the pipeline damage behavior; The digital picture detection method is meant in the key area or solution space installing digital image sensor, judges through picture, alleviates the burden of manual patrol; The NPW method is meant the pressure in the pipeline is detected that situation about changing according to pressure judges whether to take place leakage phenomenon; The flow equilibrium method is that pipeline discharge and water flow are detected, and judges the pipe network state according to both differences; Optical Fiber Method is through laying optical fiber around the pipe network, and then reaches the supervisory function bit of pipeline on every side.

Gauss hybrid models (Gaussian mixture models): be the extension of single Gaussian probability-density function, can be similar to the Density Distribution of arbitrary shape smoothly.Usually, gauss hybrid models is estimated to get through EM algorithm (EM algorithm).The EM algorithm is to carry out a kind of effective ways that maximum likelihood is estimated, it is mainly used in following two kinds of non-complete data parameter estimation: 1. observation data is incomplete, and this is because the limitation of observation process causes; 2. likelihood function is not resolved, thereby perhaps the too complicated conventional estimated method of maximum likelihood function that causes of the expression formula of likelihood function lost efficacy.Second kind of situation often runs in pattern-recognition.When adopting gauss hybrid models, because the degree of freedom of model is bigger, so be difficult to the estimation that reaches sane.The idle component projection makes the parameter of estimation more sane through removing in the variable discerning disadvantageous key element.Through idle component shadow casting technique, we obtain sane gauss hybrid models.Subsequently, we adopt SVMs, and model is distinguished.SVMs is mapped to the space of a higher-dimension with input element, and finds the solution the largest interval lineoid in this space, can find the solution nonlinear problem.

Summary of the invention

Target of the present invention is to develop a kind of acoustic data modeling in advance of passing through, and uses detected acoustic data input model and draws the judgement to the state variation situation of pipeline to be detected, and then realize piping system security incident warning system.

The present invention adopts EM algorithm and idle component derivation algorithm to draw common background gauss hybrid models and projection model respectively; Draw projection vector with adaptive technique and idle component projection; Adopt the SVM training algorithm to draw training pattern then, at last to detect the projection vector substitution training pattern that data obtain and to draw the relevant detection result.

Its concrete steps are following:

(1) the original acoustic data is drawn time series through the homomorphic signal processing

(2) adopt the EM algorithm to draw the common background gauss hybrid models: with the situation of EM algorithm application to mixed distribution.Below we are that example is done simulated experiment in MatLab with a two-dimentional mixed normal distribution.In this experiment, the production method of mixed distribution sample data is following: establish X and be the stochastic variable of the hybrid density parametric family P that belongs to given, it has the probability density function of shape suc as formula (1).If x is an observed value to X, Z is the classification stochastic variable of x, its value Z=i (i=1,2 ..., m); Under the condition of given Z=i, the conditional probability of X=x is p so _i(x| θ _i) (i=1,2 ..., m), and its unconditional probability is p (x| θ).

The observation sample production process of X is following:

1. by distribution P (Z=i)=λ i (i=1,2 ..., the m) observed value of generation Z: at first produce one [0,1] interval stochastic variable r, if ${\mathrm{\Σ}}_{k=1}^{i-1}{\mathrm{\λ}}_k\≤r\≤{\mathrm{\Σ}}_{k=1}^i{\mathrm{\λ}}_k,$ Z=i then;

2. as if Z=i, the corresponding distribution of i branch of then pressing X produces an observed value x, i=1, and 2 ..., m.

Repeat (1), (2) this process n time, can obtain about the pool-size of x is the simple random sampling observed reading of n.

Experimental procedure is following:

1. utilize said method to produce 2000 sample points of a two-dimentional mixed normal distribution;

2. confirm that the initial parameter value that algorithm needs: blending ratio α adopts the mean allocation strategy; The element of the expectation of each branch is between the maximal value of each sample elements and minimum value, to produce at random; The covariance matrix of each branch adopts symmetric matrix, and the initial parameter value of generation is seen shown in the table 1.

EM algorithm iteration end condition is in this experiment: double gained estimates of parameters satisfies

Iteration stopping when , this test stops through program after 81 iteration.Get one group of initial value again and carry out iteration, stop through program after 140 iteration away from the parameter actual value.

From the massive values computation of being done, carry out the density parameter estimation with the EM algorithm to mixing normal model, can converge to the maximum likelihood estimator of each parameter, and sample is big more, estimated result is more near the parameter true value.

(3) obtain high n dimensional vector n through adaptive technique mapping time series.

(4) use the idle component derivation algorithm, high dimension vector is generated projection model.

(5) after training data obtains high n dimensional vector n through above-mentioned steps, obtain its projection vector through the idle component projection, and obtain training pattern through SVMs.

(6) acoustic data that in actual detected, obtains is equally through this series of steps; The model that substitution step 2 and 4 draws when high n dimensional vector n of mapping and projection vector, the training pattern that draws with this projection vector integrating step 5 that obtains generates detecting the final judgement of data.

The advantage that the present invention is based on the voice data analysis recognition method of idle component Projection Analysis is: avoid the desired quantity of parameters of GMM; Use limited data to set up the training pattern that meets demand; Reduced requirement, shortened learning time and improved the accuracy that detects acoustic data.

Description of drawings

Fig. 1 is the process flow diagram based on the method for detecting pipeline of idle component Projection Analysis

Embodiment

According to shown in Figure 1; Voice data analysis recognition method based on the idle component Projection Analysis adopts EM algorithm and idle component derivation algorithm to draw common background gauss hybrid models and projection model respectively; Draw projection vector with adaptive technique and idle component projection; Adopt the SVM training algorithm to draw training pattern then, at last to detect the projection vector substitution training pattern that data obtain and to draw the relevant detection result.

Its concrete steps are following:

(1) the original acoustic data is drawn time series through the homomorphic signal processing

(2) adopt the EM algorithm to draw the common background gauss hybrid models: with the situation of EM algorithm application to mixed distribution.Below we are that example is done simulated experiment in MatLab with a two-dimentional mixed normal distribution.In this experiment, the production method of mixed distribution sample data is following: establish X and be the stochastic variable of the hybrid density parametric family P that belongs to given, it has the probability density function of shape suc as formula (1).If x is an observed value to X, Z is the classification stochastic variable of x, its value Z=i (i=1,2 ..., m); Under the condition of given Z=i, the conditional probability of X=x is p so _i(x| θ _i) (i=1,2 ..., m), and its unconditional probability is p (x| θ).

The observation sample production process of X is following:

1. by distribution P (Z=i)=λ i (i=1,2 ..., the m) observed value of generation Z: at first produce one [0,1] interval stochastic variable r, if ${\mathrm{\Σ}}_{k=1}^{i-1}{\mathrm{\λ}}_k\≤r\≤{\mathrm{\Σ}}_{k=1}^i{\mathrm{\λ}}_k,$ Z=i then;

2. as if Z=i, the corresponding distribution of i branch of then pressing X produces an observed value x, i=1, and 2 ..., m.

Repeat (1), (2) this process n time, can obtain about the pool-size of x is the simple random sampling observed reading of n.

Experimental procedure is following:

1. utilize said method to produce 2000 sample points of a two-dimentional mixed normal distribution;

2. confirm that the initial parameter value that algorithm needs: blending ratio α adopts the mean allocation strategy; The element of the expectation of each branch is between the maximal value of each sample elements and minimum value, to produce at random; The covariance matrix of each branch adopts symmetric matrix, and the initial parameter value of generation is seen shown in the table 1.

EM algorithm iteration end condition is in this experiment: double gained estimates of parameters satisfies

Iteration stopping when

, this test stops through program after 81 iteration.Get one group of initial value again and carry out iteration, stop through program after 140 iteration away from the parameter actual value.

From the massive values computation of being done, carry out the density parameter estimation with the EM algorithm to mixing normal model, can converge to the maximum likelihood estimator of each parameter, and sample is big more, estimated result is more near the parameter true value.

(3) obtain high n dimensional vector n through adaptive technique mapping time series.

(4) use the idle component derivation algorithm, high dimension vector is generated projection model.

(5) after training data obtains high n dimensional vector n through above-mentioned steps, obtain its projection vector through the idle component projection, and obtain training pattern through SVMs.

(6) acoustic data that in actual detected, obtains is equally through this series of steps; The model that substitution step 2 and 4 draws when high n dimensional vector n of mapping and projection vector, the training pattern that draws with this projection vector integrating step 5 that obtains generates detecting the final judgement of data.

Claims (2)

1. one kind is passed through acoustic data modeling in advance, uses detected acoustic data input model and draws the judgement to the state variation situation of pipeline to be detected, and then realize piping system security incident warning system;

The present invention adopts EM algorithm and idle component derivation algorithm to draw common background gauss hybrid models and projection model respectively; Draw projection vector with adaptive technique and idle component projection; Adopt the SVM training algorithm to draw training pattern then, at last to detect the projection vector substitution training pattern that data obtain and to draw the relevant detection result.

2. its concrete steps are following:

(1) the original acoustic data is drawn time series through the homomorphic signal processing;

(2) adopt the EM algorithm to draw the common background gauss hybrid models: with the situation of EM algorithm application to mixed distribution; Below we are that example is done simulated experiment in MatLab with a two-dimentional mixed normal distribution; In this experiment; The production method of mixed distribution sample data is following: establish X and be the stochastic variable of the hybrid density parametric family P that belongs to given, it has the probability density function of shape suc as formula (1); If x is an observed value to X, Z is the classification stochastic variable of x, its value Z=i (i=1,2 ..., m); Under the condition of given Z=i, the conditional probability of X=x is p so _i(x| θ _i) (i=1,2 ..., m), and its unconditional probability is p (x| θ);

The observation sample production process of X is following:

1. by distribution P (Z=i)=λ i (i=1,2 ..., the m) observed value of generation Z: at first produce one [0,1] interval stochastic variable r, if ${\mathrm{\Σ}}_{k=1}^{i-1}{\mathrm{\λ}}_k\≤r\≤{\mathrm{\Σ}}_{k=1}^i{\mathrm{\λ}}_k,$ Z=i then;

2. as if Z=i, the corresponding distribution of i branch of then pressing X produces an observed value x, i=1, and 2 ..., m; Repeat (1), (2) this process n time, can obtain about the pool-size of x is the simple random sampling observed reading of n;

Experimental procedure is following:

1. utilize said method to produce 2000 sample points of a two-dimentional mixed normal distribution;

2. confirm that the initial parameter value that algorithm needs: blending ratio α adopts the mean allocation strategy; The element of the expectation of each branch is between the maximal value of each sample elements and minimum value, to produce at random; The covariance matrix of each branch adopts symmetric matrix, produces initial parameter value;

EM algorithm iteration end condition is in this experiment: double gained estimates of parameters satisfies

Iteration stopping when , this test stops through program after 81 iteration; Get one group of initial value again and carry out iteration, stop through program after 140 iteration away from the parameter actual value;

From the massive values computation of being done, carry out the density parameter estimation with the EM algorithm to mixing normal model, can converge to the maximum likelihood estimator of each parameter, and sample is big more, estimated result is more near the parameter true value;

(3) obtain high n dimensional vector n through adaptive technique mapping time series;

(4) use the idle component derivation algorithm, high dimension vector is generated projection model;

(5) after training data obtains high n dimensional vector n through above-mentioned steps, obtain its projection vector through the idle component projection, and obtain training pattern through SVMs;

(6) acoustic data that in actual detected, obtains is equally through this series of steps; The model that substitution step 2 and 4 draws when high n dimensional vector n of mapping and projection vector, the training pattern that draws with this projection vector integrating step 5 that obtains generates detecting the final judgement of data.

Images