CN102013148A - Multi-information fusion fire hazard detection method - Google Patents

Abstract

The invention discloses a multi-information fusion fire hazard detection method, comprising the following steps: preprocessing is carried out on a fire hazard detection signal sequence obtained by site period sampling, and unreasonable data generated by normal environmental change is removed; a gray model GM (1, 1) is built according to the preprocessed original fire hazard detection signal sequence, and fire hazard detection signal data at follow-up time points is predicted, so as to obtain an equal dimensionality new information gray prediction model; the original hazard detection signal sequence is utilized to carry out posterior check on the fire hazard detection signal data obtained by prediction, so as to check whether the fire hazard detection signal data generated by prediction on gray prediction model is qualified or not; and a diagnosis neural network is utilized to diagnose the qualified fire hazard detection signal time sequence data, so as to obtain a fire hazard detection result. By means of the invention, reliable fault diagnosis can be provided and the false alarm rate of fire detection result can be reduced.

Description

Many information fusion fire detecting method

Technical field

The invention belongs to the fire detection technology field, be specifically related to the fire recognition methods of many information fusion, the fire identification of particularly poor information, weak information scene.

Background technology

Fire have a dual nature, existing its randomness one side has its determinacy one side again.Therefore to detect be a kind of very difficult input to fire detection signal, and it requires signal processing algorithm can adapt to the variation of various environmental baselines, and adjusting parameter automatically can the quick detection fire to reach, and very low rate of false alarm is arranged again.Therefore need a kind of Function Estimation and dynamical system with non-mathematical model that quantize to realize detection.

Utilize neural network and with the fuzzy system fusion method carry out detection in 90 years since its self study, adaptivity, self-organization characteristic caused the very big concern of fire engineering circle and obtained significant progress.Wherein Ri Ben Y.Okayama proposes a kind of detection algorithm of three layers of feedforward BP neural network, having must self study and adaptivity, but it considers comprehensive inadequately to the characteristics of sensor signal, and only adopt simple thresholding directly to adjudicate, be unfavorable for reducing the rate of false alarm of fire.People such as S.Nakanishi utilize the composite signal of fuzzy logic method smoke treatment temperature, flue gas concentration and CO (carbon monoxide) concentration, neural network algorithm has also been adopted in the adjusting of system, actual result shows that rate of false alarm has lowered 50%, and the fire alarm time also shifts to an earlier date to some extent.But it does not introduce grey algorithm and thresholding algorithm, makes that the anti-environmental interference of this algorithm and the ability of reporting to the police are in advance limited to.

The defective of existing fire detecting system is, the low and false alarm rate height of detection sensitivity, and lack of wisdom, and can not play due effect at the place of information, poor information a little less than some, even produce the situation of not reporting to the police.Therefore, need a kind of method to address the above problem.

Summary of the invention

Purpose of the present invention is intended to one of solve the aforementioned problems in the prior at least, particularly solves to fail to report, report by mistake and at weak information, poor information fire scenario detection problem.

For this reason, embodiments of the invention propose a kind of accurate, reliable many information fusion fire detecting method.

According to an aspect of the present invention, the embodiment of the invention has proposed a kind of many information fusion fire detecting method, said method comprising the steps of: the fire detection signal sequence that obtains from on-the-spot periodic sampling is carried out pre-service, get rid of because home changes the unreasonable data that produce; Carry out gray model GM (1,1) modeling according to pretreated original fire detection signal sequence, the fire detection signal data of follow-up time point are predicted, to obtain waiting the fresh information grey forecasting model of dimension; The fire detection signal data of utilizing original fire detection signal sequence that prediction is obtained are carried out the check of posteriority difference, and whether the fire detection signal data of utilizing described grey forecasting model prediction to generate with check are qualified; And utilize the diagnosis neural network that qualified fire detection signal time series data is diagnosed, detect to obtain the detection result.

The further embodiment according to the present invention, underproof time series data is set up the correction of residual error model, and then predict that until qualified the step of wherein setting up the correction of residual error model comprises: the sampling time sequence according to the fire detection signal sequence obtains the sequential residual sequence corresponding with the fire detection signal sequence; Described sequential residual sequence is carried out gray model GM (1,1) modeling, to obtain corresponding sequential residual sequence predicted value; And utilize described sequential residual sequence predicted value to obtain

, so that underproof time series data is revised.

The further embodiment according to the present invention, the step of described grey GM (1,1) modeling comprises: described fire detection signal sequence is carried out one-accumulate; On the sequence basis behind the one-accumulate, set up the differential equation of albefaction form; Go out next predicted value constantly of the sequence correspondence behind the one-accumulate according to this differential equation; And described predicted value carried out once tiredly subtracting computing, obtain corresponding next fire detection signal data prediction value constantly.

The further embodiment according to the present invention, described posteriority difference checking procedure comprises: first mean value and the first variance of calculating the original fire detection signal sequence of gathering; Calculate residual error between the corresponding predicted value of each original fire detection signal with it, and second mean value and the second variance of all residual errors of whole original fire detection signal sequence correspondence; And utilize described first variance and described second variance to obtain posteriority difference ratio, and utilize described residual error, described second mean value and described first variance to obtain little error frequency, to carry out the check of posteriority difference.

The further embodiment according to the present invention, the step of utilizing the diagnosis neural network that qualified fire detection signal time series data is diagnosed comprises: utilize the learning sample of predetermined quantity as the input and output sample neural network to be carried out off-line training, until network convergence to obtain described diagnosis neural network, wherein said input sample is the detection characteristic signal that comprises flue-gas temperature, flue gas concentration and carbon monoxide CO concentration, and described output sample comprises naked light probability, the fire risk corresponding with described input sample that defines and the probability that glows; And described qualified fire detection signal time series data is input to described diagnosis neural network.

The further embodiment according to the present invention, also comprise the step that the fire probability of the fuzzy set among the detection result that the diagnosis neural network is detected carries out pattern-recognition, wherein said pattern-recognition step comprises: the normal distribution membership function value that calculates detection result's fire fuzzy set and non-fire fuzzy set respectively; Compare the normal distribution membership function value of fire fuzzy set and the normal distribution membership function value size of non-fire fuzzy set; And determine the identification of final fire according to comparative result.

The present invention sets by adopting lower threshold value and trend; got rid of the interference that home changes; owing to wait of the prediction of reform grey information model to later stage fire signal development; make this algorithm be applicable to the place of poor information, weak information; the adaptivity of fuzzy neural network, learning ability, fault-tolerant ability and parallel processing capability; the characteristic signal value that makes network to make full use of to provide provides reliable fault diagnosis, and the fire detection technology of many information fusion has reduced rate of false alarm simultaneously.Therefore, applicable scope more extensively and can provide, more accurately alerting signal more Zao than other detection algorithms.

Aspect that the present invention adds and advantage part in the following description provide, and part will become obviously from the following description, or recognize by practice of the present invention.

Description of drawings

Above-mentioned and/or additional aspect of the present invention and advantage are from obviously and easily understanding becoming the description of embodiment below in conjunction with accompanying drawing, wherein:

Fig. 1 is many information fusion fire detecting method process flow diagram of the embodiment of the invention.

Embodiment

Describe embodiments of the invention below in detail, the example of described embodiment is shown in the drawings, and wherein identical from start to finish or similar label is represented identical or similar elements or the element with identical or similar functions.Below by the embodiment that is described with reference to the drawings is exemplary, only is used to explain the present invention, and can not be interpreted as limitation of the present invention.

With reference now to Fig. 1,, this figure is many information fusion fire detecting method process flow diagram of the embodiment of the invention.May further comprise the steps:

Step 102, the detection characteristic signal that the on-the-spot cycle is gathered are carried out for example pre-service of trend and threshold decision, get rid of because home changes the unreasonable data that produce.

The data that presence detector is gathered judge through threshold value and trend algorithm that earlier the setting of threshold value and trend here will be far below adopting threshold-type or trend type detector alarm value.This sets fundamental purpose is to filter out to survey in the place because home changes the unreasonable data that produce.

Step 104 is carried out gray model GM (1,1) modeling according to pretreated original fire detection signal sequence, the fire detection signal data of follow-up time point is predicted, to obtain waiting the fresh information grey forecasting model of dimension.

If rejecting the fire detection signal sequence that the cycle collects after the unreasonable data is x ⁽⁰⁾: x ⁽⁰⁾={ x ⁽⁰⁾(1), x ⁽⁰⁾(2) ..., x ⁽⁰⁾(i) ..., x ⁽⁰⁾(n) }, it is carried out one-accumulate (1-AGO) with formation sequence x ⁽¹⁾: x ⁽¹⁾={ x ⁽¹⁾(1), x ⁽¹⁾(2) ..., x ⁽¹⁾(i) ..., x ⁽¹⁾(n) }, wherein K=1,2 ..., n.

For the sequence x that generates ⁽¹⁾Can set up the differential equation of albefaction form, it becomes the single order grey differential equation, is designated as GM (1,1):

Wherein a and u are undetermined parameter.This equation separate for

This formula is called equation time response.

The note parameter is classified A=[a as, u] ^T, can utilize least square method to find the solution A:A=(B ^TB) ^-1B ^TY _n

In the formula:

Y _n=(x ⁽⁰⁾(2), x ⁽⁰⁾(3) ..., x ⁽⁰⁾(n)) ^T,, can calculate to such an extent that generate the estimated value of k item and k+1 item in the data rows behind the one-accumulate with parameter a and u substitution equation time response that obtains With

Do then and tired subtract generation, be calculated as follows the estimated value of k+1 item in the original fire detection signal sequence

After predicting next value constantly, for the dimension that guarantees sequence equates, need remove first data of original fire detection signal sequence, set up GM (1,1) model again, predict next value constantly, recurrence successively, reform such as formations grey information forecast model.

Step 106, the fire detection signal data of utilizing original fire detection signal sequence that prediction is obtained are carried out the check of posteriority difference, and whether the fire detection signal data of utilizing described grey forecasting model prediction to generate with check are qualified.

Press the credibility that gray model is predicted for check, need carry out the posteriority inspection and test.The mean value of the real data of original fire detection signal sequence

With variance s ₁ ²Be respectively:

&

k＝1，2...n

The raw value x of k item number certificate ⁽⁰⁾(k) with the estimated value of calculating

Difference q (k) be called k item residual error

The mean value of the residual error of then whole all data item of data rows

With variance s ₂ ²Be respectively

﹠amp;

K=1,2...n.

Carry out the check of posteriority difference by calculating posteriority difference ratio c and little error frequency p, and contrast table 1 provides judgement.Wherein

K=1,2...n.

Table 1 posteriority difference testing accuracy grade

Step 108 is set up Residual GM (1,1) model and is revised.Underproof time series data is set up the correction of residual error model, and then predict until qualified.

Because x ⁽⁰⁾There is certain error usually in-GM (1,1) model, and the one, because the precision of model own is not high; The 2nd, because handle relevant by even time interval with sequential t.x ⁽⁰⁾-GM (1,1) formula is a continuous function form, and is the function of sequential t, and therefore given arbitrary moment total energy is calculated

If suppose not exist error, that is: There is residual epsilon so suppose sequential t _t, could satisfy this requirement.Make t _ε=t+ ε _t, then

${\hat{x}}^{\left(0\right)}=x^{\left(0\right)}\⇒{\hat{x}}^{\left(1\right)}\left(t_{\mathrm{\ϵ}}\right)=x^{\left(1\right)}\left(t\right);$

${\hat{x}}^{\left(1\right)}\left(t_{\mathrm{\ϵ}}\right)=(x^{\left(0\right)}\left(1\right)-\frac{u}{a})e^{-a(t_e-1)}+\frac{u}{a}$

Calculate t again _ε, then have $t_{\mathrm{\ϵ}}=1+\frac{1}{a}\ln \frac{x^{\left(0\right)}\left(1\right)-\frac{u}{a}}{x^{\left(1\right)}\left(t\right)-\frac{u}{a}}$

So ε _t=t _ε-t gets the sequential residual sequence , utilize GM (1,1) modeling method to set up the sequential residual sequence

GM (1,1) model, here be referred to as Model.That is:

${\widehat{\mathrm{\ϵ}}}_t^{\left(1\right)}\left(k\right)=|{\mathrm{\ϵ}}_t^{\left(0\right)}\left(1\right)-\frac{u_t}{a_t}|e^{-a_t(k-1)}+\frac{u_t}{a_t};$

${\widehat{\mathrm{\ϵ}}}_t^{\left(0\right)}\left(k\right)={\widehat{\mathrm{\ϵ}}}_t^{\left(1\right)}\left(k\right)-{\widehat{\mathrm{\ϵ}}}_t^{\left(1\right)}(k-1)$

In some cases, the sequential residual sequence of trying to achieve by following formula

May be not suitable for directly setting up GM (1,1) model.Two kinds of situations are arranged: first kind of situation here

Sequence is non-negative; Second kind of situation is the sequential residual sequence The middle residual error of bearing that exists suitably adds a constant b handle usually

Row become non-negative

, obtain

After again it is reduced to

, promptly

f ₁： ${\mathrm{\ϵ}}_t^{\left(0\right)}\→{{\mathrm{\ϵ}}_{t_1}}^{\left(0\right)}\{{{\mathrm{\ϵ}}_{t_1}}^{\left(0\right)}={\mathrm{\ϵ}}_t^{\left(0\right)}+b\};$

f ₂： ${{\mathrm{\ϵ}}_{t_1}}^{\left(0\right)}={{\mathrm{\ϵ}}_{t_1}}^{\left(0\right)}-\mathrm{GM}\left(\mathrm{1,1}\right)\→{{\widehat{\mathrm{\ϵ}}}_{t_1}}^{\left(0\right)};$

f ₃： ${{\widehat{\mathrm{\ϵ}}}_{t_1}}^{\left(0\right)}\→{\widehat{\mathrm{\ϵ}}}_t^{\left(0\right)}\{{\widehat{\mathrm{\ϵ}}}_t^{\left(0\right)}={{\widehat{\mathrm{\ϵ}}}_{t_1}}^{\left(0\right)}-b\}$

Utilize sequential residual prediction value then Try to achieve the revised numerical value of residual error

, so that underproof time series data is revised.

Because

So will

Substitution x ⁽⁰⁾-GM (1,1) solution formula:

${\hat{x}}_{\mathrm{\ϵ}}^{\left(1\right)}\left({\hat{t}}_{\mathrm{\ϵ}}\left(t\right)\right)=|x^{\left(0\right)}\left(1\right)-\frac{u}{a}|e^{-a({\hat{t}}_{\mathrm{\ϵ}}\left(t\right)-1)}+\frac{u}{a}$

Can reduce and obtain ${\hat{x}}_{\mathrm{\ϵ}}^{\left(0\right)}\left(t\right)={\hat{x}}_{\mathrm{\ϵ}}^{\left(1\right)}\left({\hat{t}}_{\mathrm{\ϵ}}\left(t\right)\right)-{\hat{x}}_{\mathrm{\ϵ}}^{\left(1\right)}\left({\hat{t}}_{\mathrm{\ϵ}}(t-1)\right).$

In a preferred embodiment, in order further to improve the precision of revising data, can then return the precision of the residual error model of step 106 check foundation.

Step 110, the fire detection signal time series data of Sheng Chenging is transfused to the diagnosis neural network and carries out many information fusion judgements after the assay was approved.

Utilize in the present embodiment the detection characteristic signal for flue-gas temperature, flue gas concentration and CO (carbon monoxide) concentration as input, and naked light probability, the fire risk of the correspondence that defines and glow probability as output neural network is carried out off-line training.

If network has m layer (not comprising input layer), the node number of l layer is n _l,

Represent the output of l node layer k, and be expressed from the next: In the formula

Be the weight vector of articulamentum l-1 node layer to l node layer k; Y ⁽⁰⁾=X.

Given sample mode (X, Y) after, the weights of neural network will be adjusted, and make following criterion function minimum:

In the formula,

Be the output of network, and By the gradient descent method, can try to achieve the gradient of E (W) and revise weights, be i.e. weight vector Correction can try to achieve by following formula:

Wherein η is a learning rate; For output layer M, the vague generalization error of each unit is

For other layer, the vague generalization error of each unit is

L=1,2 ..., M-1.

Constantly adjust weights for given sample repeatedly according to said process, the output that makes network is near desirable output.Up to the network global error less than predefined minimal value, i.e. a network convergence.If the study number of times is greater than predefined value, network just can't be restrained.

Training about neural network can repeat no more here with reference to existing BP nerual network technique.

Then formed the diagnosis neural network through above-mentioned steps.After three the fire detection signal parameters normalization that collects when the on-the-spot cycle is transfused to this neural network, corresponding naked light probability, fire risk and the detected output of the probability that glows.

Output through above-mentioned neural network can judge tentatively that the possibility of breaking out of fire has much.

Step 112 utilizes pattern-recognition to carry out the judgement of fuzzy logic fire.。

In some cases, the output result according to above-mentioned diagnosis neural network is easy to find out, when the naked light probability greater than 0.8 the time, fire can take place certainly.And when the naked light probability less than 0.2, and the smoldering fire probability is when also very little, can think does not have fire to occur.

When the naked light probability then may relatively be difficult to interpretation near 0.5, because the blur level maximum of fire probability at this moment.Therefore in preferred above-mentioned example of the present invention, may further include step 112, so that the fuzzy set fire probability is carried out pattern-recognition.

With the naked light probability is example, and other two parameters are similar.If x is the naked light probability, A represents the fire fuzzy set, and B represents non-fire fuzzy set.A given x value will rule out whether fire is arranged, and only needs relatively μ _A(x) and μ _B(x) size gets final product.Rule of thumb with to the statistical study of fire data, adopt the subordinate function of a kind of normal distribution as A, B.Through repeatedly comparison, modification, λ and τ are defined as 0.2 and 0.4 respectively.The subordinate function of A and B is following two formulas respectively.

Finally provide clear and definite fire identification by the computing comparison.

Step 114 provides fire alarm/non-fire alarm judgement.

According to the fuzzy logic judged result, draw the fire identification of fire alarm/non-fire alarm.

The present invention sets by adopting lower threshold value and trend; got rid of the interference that home changes; owing to wait of the prediction of reform grey information model to later stage fire signal development; make this algorithm be applicable to the place of poor information, weak information; the adaptivity of fuzzy neural network, learning ability, fault-tolerant ability and parallel processing capability; the characteristic signal value that makes network to make full use of to provide provides reliable fault diagnosis, and the fire detection technology of many information fusion has reduced rate of false alarm simultaneously.Therefore, applicable scope more extensively and can provide, more accurately alerting signal more Zao than other detection algorithms.

Although illustrated and described embodiments of the invention, for the ordinary skill in the art, be appreciated that without departing from the principles and spirit of the present invention and can carry out multiple variation, modification, replacement and modification that scope of the present invention is by claims and be equal to and limit to these embodiment.

Claims (6)

1. information fusion fire detecting method more than a kind is characterized in that, said method comprising the steps of:

The fire detection signal sequence that obtains from on-the-spot periodic sampling is carried out pre-service, get rid of because home changes the unreasonable data that produce;

Carry out gray model GM (1,1) modeling according to pretreated original fire detection signal sequence, the fire detection signal data of follow-up time point are predicted, to obtain waiting the fresh information grey forecasting model of dimension;

The fire detection signal data of utilizing original fire detection signal sequence that prediction is obtained are carried out the check of posteriority difference, and whether the fire detection signal data of utilizing described grey forecasting model prediction to generate with check are qualified; And

Utilize the diagnosis neural network that qualified fire detection signal time series data is diagnosed, detect to obtain the detection result.

2. many information fusion fire detecting method as claimed in claim 1 is characterized in that, underproof time series data is set up the correction of residual error model, and then predicts until qualified, and the step of wherein setting up the correction of residual error model comprises:

Sampling time sequence according to the fire detection signal sequence obtains the sequential residual sequence corresponding with the fire detection signal sequence;

Described sequential residual sequence is carried out gray model GM (1,1) modeling, to obtain corresponding sequential residual sequence predicted value; And

Utilize described sequential residual sequence predicted value to obtain the revised numerical value of residual error

, so that underproof time series data is revised.

3. many information fusion fire detecting method as claimed in claim 1 is characterized in that, the step of described grey GM (1,1) modeling comprises:

Described fire detection signal sequence is carried out one-accumulate;

On the sequence basis behind the one-accumulate, set up the differential equation of albefaction form;

Go out next predicted value constantly of the sequence correspondence behind the one-accumulate according to this differential equation; And

Described predicted value is carried out once tiring out subtracting computing, obtain corresponding next fire detection signal data prediction value constantly.

4. many information fusion fire detecting method as claimed in claim 1 is characterized in that, described posteriority difference checking procedure comprises:

Calculate first mean value and the first variance of the original fire detection signal sequence of gathering;

Calculate residual error between the corresponding predicted value of each original fire detection signal with it, and second mean value and the second variance of all residual errors of whole original fire detection signal sequence correspondence; And

Utilize described first variance and described second variance to obtain posteriority difference ratio, and utilize described residual error, described second mean value and described first variance to obtain little error frequency, to carry out the check of posteriority difference.

5. many information fusion fire detecting method as claimed in claim 1 is characterized in that, the step of utilizing the diagnosis neural network that qualified fire detection signal time series data is diagnosed comprises:

Utilize the learning sample of predetermined quantity neural network to be carried out off-line training as the input and output sample, until network convergence to obtain described diagnosis neural network, wherein said input sample is the detection characteristic signal that comprises flue-gas temperature, flue gas concentration and carbon monoxide CO concentration, and described output sample comprises naked light probability, the fire risk corresponding with described input sample that defines and the probability that glows; And

Described qualified fire detection signal time series data is input to described diagnosis neural network.

6. as claim 1 or 5 described many information fusion fire detecting methods, it is characterized in that also comprise the step that the fire probability of the fuzzy set among the detection result that the diagnosis neural network is detected carries out pattern-recognition, wherein said pattern-recognition step comprises:

Calculate the normal distribution membership function value of detection result's fire fuzzy set and non-fire fuzzy set respectively;

Compare the normal distribution membership function value of fire fuzzy set and the normal distribution membership function value size of non-fire fuzzy set; And

Determine final fire identification according to comparative result.

Images