CN102610055A - Wireless intelligent alarming system for automatically monitoring multivariate information of tunnel - Google Patents

Abstract

The invention discloses a wireless intelligent alarming system for automatically monitoring multivariate information of a tunnel. The wireless intelligent alarming system is characterized by being realized by the following steps of: firstly, identifying surrounding rock multivariate parameters based on a differential evolution algorithm and a support vector machine; secondly, predicating a surrounding rock multivariate information time sequence based on the differential evolution algorithm and the support vector machine; and thirdly, carrying out judgment and alarm on the safety of the surrounding rock according to filed monitoring information and an information predication result obtained in the second step. According to the system, the requirements on a field communication cable and the field processing capability are reduced, and thus an acquisition-transmission instrument of a filed part is integrated on a circuit board; a wireless data transmission technology is adopted on site and field monitoring data is comprehensively and smoothly transmitted to a data processing system for an ultra-long distance; and an analytic result alarms through short messages, emails and qq, so that the pertinence, the flexibility and the reliability of alarm are improved. The direct contact between relevant persons and complex and danger surrounding rock field can be avoided and status information of the surrounding rock can be safely and remotely acquired.

Description

Multiple information intelligent wireless warning system is monitored in the tunnel automatically

Technical field

The present invention relates to a kind of tunnel long-range automatic monitoring multiple information intelligent alarm system.

Background technology

Carrying out along with the Tunnel Engineering construction of China emerges tunnel, increasing rich pool, for example tunnel, maritime province, seabed tunnel, cross-river tunnel, subfluvial tunnel and tunnel, karst region or the like.These tunnels are built under underground water, and country rock exists complicated action of seepage-stress coupling, and its stability status shows as multiple comprehensive characterization information such as displacement, stress, flow.The on-line monitoring instrument of personal monitoring that the constructing tunnel phase is common and general single information often is difficult to obtain effective country rock status predication information; Therefore security incidents such as the landslide in tunnel, rich pool, gushing water occur repeatedly, and bring serious economy and social danger.

Present tunnel monitoring technology is used for tunnel surrounding state judgement aspect, rich pool and has sizable limitation, mainly shows: 1, general on-line monitoring instrument information is single, and tunnel, rich pool has typical many couplings, characteristics that parameter is various.Single monitoring information can not reflect the steadiness of Tunnel Engineering very all sidedly.If the 2 on-the-spot multiple sensors of laying will be referred to the transmission and the processing of complicated multiple information data.These complicated mass data processing are not only needed the very strong computer processing hardware system of processing power, more need the software systems and the algorithm of Treatment Analysis and prediction and alarm in real time, these all are that key issue to be solved is arranged.3, the tunnel site environment is relatively poor, and the space is limited, does not allow the huge computer system of installation volume, and complicated multiple information Processing tasks is that embedded system institute such as single-chip microcomputer is inefficient.The transmission range of the sound and light alarm that general instrument is sent is limited, and how warning message breaks through space constraint, and to be delivered to specific target in time be that problem to be solved is arranged.

The general tunnel monitoring system that is used for just provides data simply, has significant limitation aspect analytical approach and the pre-alerting ability.The space of tunnel field monitoring instrument volume and layout is restricted; Do not allow to lay long data cable yet; Generally can only adopt the microprocessor of single-chip microcomputer character and than short cable, be difficult for realizing complicated pluralism information complicacy fusion treatment and analyze computing and data transmit at a distance.The software and hardware system of conventional tunnel monitoring is difficult to meet the automatic prognoses system requirement of multiple information in tunnel, rich pool.Because the complicacy that tunnel, rich pool is on-the-spot; Polynary automatic monitoring system is quite necessary; Polynary automatic monitoring system is excavated and is analyzed data fully through obtaining the polynary real-time Monitoring Data of magnanimity, thereby reasonably stability of surrounding rock is estimated and predicted.In collection, analyses and prediction and the warning processing procedure of data, need to break through the on-the-spot distance limit in tunnel to the processing center.Research and development can carry out in real time multiple information automatic monitoring, and can carry out multiple information fusion treatment and analysis fast, have the tunnel warning system of overlength distance and advanced prediction function, have important economic implications and social effect.

Summary of the invention

The present invention is directed to the proposition of above problem, and develop a kind of multi-channel information merge, can advanced prediction and the automatic alarm system of wireless transmission.The technological means that the present invention adopts is following:

Multiple information intelligent wireless warning system is monitored in a kind of tunnel automatically, it is characterized in that comprising the steps:

Step 1, based on the country rock parameter recognition of difference evolution algorithm and SVMs, concrete steps are following:

(1) according to engineering problem, confirm the span of country rock mechanics parameter and hydrologic parameter, and according to the numerical procedure of Orthogonal Experiment and Design principles of construction parameter combinations;

(2) adopt the three-dimensional flow Finite Element Method that is coupled admittedly that each parameter combinations scheme of structure is calculated, obtain the physical quantity information of the corresponding observation station of each scheme, and each numerical procedure is constituted one group of learning sample with corresponding observation station information calculations value;

(3), adopt best SVMs penalty factor c and the nuclear parameter σ of difference evolution algorithm search based on above-mentioned sample;

(4) use the best SVMs parameter that obtains,, set up and treat the country rock parameter of inverting and the non-linear supporting vector machine model between the polynary monitoring information, replace flowing the finite element solving that is coupled admittedly top sample learning training;

(5) according to treating the mechanics of inverting and the supporting vector machine model between hydrologic parameter and the polynary observation information; Error with typical position observed reading and calculated value is an objective function; Adopt the search of difference evolution algorithm to treat the country rock parameter of inverting, utilize the parameter of inverting further to confirm the mechanical property of country rock;

Step 2, based on the country rock multiple information time series forecasting of difference evolution algorithm and SVMs, concrete steps are following:

(a) through the time series collection of multichannel sensor to polynary observation information, structure input and output sample data is utilized the principal component analytical method dimensionality reduction, has formed the predicted data sample of the input-output of multivariate time series;

(b) above-mentioned data sample is divided into two types at random, a part is a learning sample, and another part is a test sample book;

(c) difference evolution algorithm of support vector machine is carried out the initialization setting, comprise difference evolution initial parameter is set, variable number to be optimized is 2, population quantity, zoom factor F, hybridization probability constant C R; Provide initial population at random, different individual corresponding different SVMs parameters, promptly corresponding different nuclear parameter σ and penalty factor C;

The algorithm of (d) evolving according to difference carries out mutation operation and interlace operation; Carry out the training of learning sample with each individual corresponding parameters; The supporting vector machine model that training obtains is predicted test sample book; According to predicting the outcome and according to formula: S (x)=ERMSE+EPA; ERMSE in the formula and EPA are respectively two performance index that the sample in the statistics is estimated: root-mean-square error and precision of prediction, calculate this individual adaptive value S (x);

(e) judge to select individual adaptive value S (x) whether to meet the demands, as not meeting the demands, to carry out the calculating of a new round again, to return step (d), if adaptive value meets the demands, the learning training process finishes;

(f), carry out the prediction of certain sensor information in the future according to importing present polynary Monitoring Data.For following monitoring information prediction of a plurality of sensors, each sensor repeats the process of (c)-(f) respectively;

Step 3, the information of forecasting result who obtains according to field monitoring information and step 2 judge warning; Concrete steps are following:

Judge a certain monitor value information of field monitoring or predict whether value of reaching capacity of a certain following information, report to the police when reaching;

Whether judge a certain monitor value information of field monitoring simultaneously or predict has continuous a days speed to surpass the speed warning value in a certain following information, if any reporting to the police, and a >=1 wherein;

Whether judge field monitoring information simultaneously or predict has the continuous b of a plurality of information days speed to surpass the speed warning value in the following information, if report to the police, and b >=1 wherein.

In the said step 1, following based on the concrete steps of the polynary parameter recognition of country rock of difference evolution algorithm and SVMs:

Its principle steps is following: choose one group of calculation of parameter and obtain surrouding rock deformation and flow physical quantity; If the computational physics amount differs bigger with monitoring physical quantity data, then adjust input parameter and recomputate, differ very little up to calculating with Monitoring Data; At this moment corresponding parameter is the parameter of identification; Physical meaning according to the tunnel surrounding parameter is set bound, if z observed reading arranged in the zone, adopts following formula to carry out constrained optimization to it:

$\left\{\begin{array}{c}\min E(x_1,x_2,...,x_n)=\frac{1}{z}\underset{i=1}{\overset{z}{\mathrm{\Σ}}}{[Y_i^0-Y_i]}^2\\ x_i^a\≤x_i\≤x_i^b(i=\mathrm{1,2},...,n)\end{array}\right.$

In the formula,

Be the measured value of country rock multiple information, Y _iBe the calculated value of country rock multiple information, z is the observed reading number, x _iBe i parameter, n is a number of parameters,

With Be x _iUpper and lower limit;

Through country rock mechanics and hydrologic parameter combination orthogonal design scheme in the certain limit is carried out dimensional Finite Element; Obtain the data sample of the corresponding relation of mechanics, hydrologic parameter and measuring point head, flow, displacement response physical quantity monitoring information; Utilize these sample training SVMs; Thereby set up the SVMs response surface function of the three-dimensional hydrology, the supporting vector machine model between mechanics, hydrologic parameter and the response physical quantity is:

$\left.\begin{array}{c}H=\mathrm{SVM}\left(P\right)\\ P=(p_1,p_2,...,p_n)\\ H=(H_1,H_2,...,H_z)\end{array}\right\}$

In the formula: P is the parameter of country rock zoning, and n is the number of parameter; H is the multiple information of observation, and z is used for responding finding the solution of physical quantity for the number of observation multiple information, the SVM model that learning training is obtained.

In the said step 2, following based on the predicted data sample construction step of the input-output in the country rock multiple information time series forecasting algorithm of difference evolution algorithm and SVMs:

Suppose x ₁, x ₂, x ₃..., x _nSensor has obtained the n metamessage, and each sensor is gathered since t-m constantly, obtains m-1 Monitoring Data; Then following t data can be come out by a prostatitis m-1 data-speculative; Along with the acquisition of 1 sequential in back, the oldest sequential sample that substitutes the front with this new sequential sample carries out next step prediction, obtains predicted value next time; And the like, just formed some groups of I/O mapped sample datas;

Multiple information time series forecasting thinking is: with first sensor is example, the Monitoring Data that it is follow-up, promptly following data x constantly ₁(t), the data x that obtains through n sensor ₁(t-m), x ₁(t-m+1) ..., x ₁(t-1); x ₂(t-m), x ₂(t-m+1) ..., x ₂(t-1); ; x _n(t-m), x _n(t-m+1) ..., x _n(t-1) infer,, replenish into new Monitoring Data, replace old Monitoring Data, realize the data corresponding relation of rolling like this, just can obtain the sample of the relation of historical data and following data along with the acquisition of new Monitoring Data;

The problems referred to above formulation is following: to each sensor monitors amount, selecting the future anticipation step number is 1, i=1, and 2 ..., n is the sensor number, thus historical polynary vector is expressed as:

$\stackrel{\→}{X}(t-1)=[x_1(t-m),x_1(t-m+1),...,x_1(t-1),...,x_i(t-m),x_i(t-m+1),...,x_i(t-1),$

$...,x_n(t-m),x_n(t-m+1),...,x_n(t-1){]}^T$

Wherein: what represented is historical polynary vector; Represent that each sensor constantly begins to gather m data from t-m; The information that comprises the needed time in the past sequence of the following information of prediction in the formula; Input as model; Following arbitrary monitoring information constantly can be expressed as the function of historical polynary vector, is expressed as:

$x_i\left(t\right)=f\left(\stackrel{\→}{X}(t-1)\right)$

X in the formula _i(t) be following prediction physical quantity data, f is the nonlinear function of historical polynary vector sum future anticipation information;

Above-mentioned model input vector has redundant characteristics, and input vector adopts principal component analytical method to simplify, and concrete steps are following:

(1) standardization sample matrix X _{M * n}For

X _{M * n}For

Expression;

(2) svd

is:

$\tilde{X}=\mathrm{U\Σ}{V}^T$

Wherein: ∑=diag [s ₁s ₂s _p0 ... 0]

In the formula: p is the eigenwert number, s ₁>=s ₂>=...,>=s _pBe the singular value of correspondence, U is m * m rank matrix, and ∑ is m * n rank matrix, and V is n * n rank matrix;

(3) formula of definite singular value that need keep is following:

${\mathrm{\η}}_k=\frac{\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{S_j}^2}{\underset{i=1}{\overset{p}{\mathrm{\Σ}}}{S_i}^2}$ j＝1，2，…，k；i＝1，2，…，p.

Confirm threshold value 0＜η ₀＜1, if η _k＞η ₀, k singular value before then keeping;

(4) establish k singular value and be retained, then corresponding pivot number is k, reduces transformation matrix V _{N * n}Dimension do System's pivot is:

$\stackrel{\→}{Z}(t-1)\≈\stackrel{\→}{X}(t-1)\tilde{V}$

Input vector

through after the principal component analysis (PCA) effectively reduces dimension, and forecast model becomes following formula like this:

$x_i\left(t\right)=f\left(\stackrel{\→}{Z}(t-1)\right)$

In the formula;

is the input vector after the principal component analysis (PCA); Polynary input vector after just simplifying, what

represented is the polynary input vector before simplifying.Above each sensor of formulate constantly begins to gather m data from t-m; Through above-mentioned principal component analysis (PCA);

that

of n dimension reduced to the k dimension realized the simplification of input information, for SVM prediction provides the foundation.

The present invention compares with prior art; Its advantage is conspicuous; System according to the invention has reduced on-the-spot communication cable and to the requirement of site disposal ability, thereby makes the collection-transmitter of situ part be integrated in a circuit board, and is installed in the sealing metal box; Advantages of small volume, and have waterproof action.The field by using wireless data transmission technology; Can all sidedly the field monitoring data successfully be passed to the data handling system of overlength distance; The result who analyzes reports to the police through note, email and qq, has improved specific aim, dirigibility and the reliability of reporting to the police, and has broken through the restriction of space length.It is on-the-spot to make the related personnel needn't directly contact complicated dangerous country rock, obtains the status information of country rock in the distance safely.

Description of drawings

Fig. 1 is an evolution-algorithm of support vector machine process flow diagram;

Fig. 2 is the process flow diagram of system according to the invention;

The solid coupling parameter identifying process flow diagram of Fig. 3 country rock stream;

The original inputoutput data synoptic diagram in a corresponding moment of Fig. 4;

Fig. 5 is based on the country rock multiple information time series forecasting algorithm of DE-SVM;

Fig. 6 is the structured flowchart of warning device;

Fig. 7 collection in worksite-transmitter hardware composition diagram;

The TCP signal procedure process flow diagram of Fig. 8 GPRS module;

Fig. 9 note transmission flow;

The arrangenent diagram of sensor among Figure 10 embodiment;

The convergence process synoptic diagram is optimized in Figure 11 difference evolution (DE);

Figure 12 SVM prediction model is to the comparison diagram that predicts the outcome of test samples.

Embodiment

In order to solve the automatic monitoring that to carry out multiple information in real time, and can to carry out multiple information fusion treatment and analysis in real time, have the problem of the tunnel warning of overlength distance and advanced prediction function; The present invention proposes two level monitoring systems of host computer-slave computer: slave computer is responsible for on-the-spot multiple information monitoring, just carries out multichannel data acquisition.Host computer carries out the analysis and the processing of data.On-slave computer carries out the information transmission through the data wireless lift-off technology.Host computer requires to have powerful multivariate data analysis processing power, comprises the DE-SVM Intelligent Recognition of the solid coupling parameter of stream; Can carry out the time of fusion prediction SVM prediction and the multistage warning of multiple information; And has a function that warning message mobile phone and Internet overlength distance transmit.

This system can synthetically obtain polynary monitoring information in real time and make identification and prediction; Rationally obtain the state of tunnel stability in time; Thereby the advanced prediction that can be used for the tunnel safety property under the condition of complicated rich pool especially is directed to place Tunnel Engineering dangerous, that face prominent mud gushing water or karst landslide.

Its technical scheme is:

One, makes up the multiple information automatic monitoring system in the tunnel, rich pool of host computer-slave computer two levels; Slave computer mainly is mounted in the static collector of engineering site, is responsible for being captured in the on-the-spot Monitoring Data that multielement bar comprises that displacement meter, taseometer, ventage piezometer, reinforcing bar meter, pressure cell, flowmeter etc. transmit of laying.Host computer is the PC that is placed in rear processing enter machine room, and PC is installed the fusion intellectual analysis warning system of multiple information, carries out the online warning of overlength distance.Host computer-slave computer is through wireless tranmission techniques Data transmission.

Two, the fusion intellectual analysis warning system of host computer multiple information has comprised the DE-SVM intelligent identification Method of country rock parameter; The DE-SVM method of the time series forecasting of polynary Monitoring Data and Realtime Alerts algorithm.Time series self-adaptation rolling DE-SVM prediction through real-time criterion of the speed of multiple information and multiple information, the variation of advanced prediction surrounding rock displacement, and with the warning value contrast, if reach alert if, send warning message.

Three, data processed result sends to scene and other specific places through wireless launcher, realizes forecast fast.The directed alarm module that adopts; At first trigger the acousto-optic warning unit at the nearest place, workmen place of distance detection point; And carry out the locating alarming of SMS through GSM, and will in time warning message be sent on designated person's the QQ and email through GPRS network.

(the SVM principle: (Difference Evolution is a kind of emerging evolutionary computation technique DE) to difference evolutionary optimization algorithm to difference evolution algorithm (DE), has stronger global convergence ability and robustness with SVMs.Difference evolution algorithm fast convergence rate need not differentiated to objective function, therefore is applicable to complicated non-linear implicit expression function.The present invention flows the optimization and the supporting vector machine model Parameter Optimization of solid coupling parameter with it.

Difference evolution algorithm principle is following: supposing needs to optimize n parameter, and then at first producing dimension is Np the vector of n, and Np is called population scale, and each vector is one group of potential separating just, is called individuality.Each individual vector is calculated according to objective function, and result of calculation is as evaluation of estimate.According to DE rule iteration, move closer to optimum solution according to the evaluation of estimate size.The computation optimization process of DE is following:

1) produces initial population.It is n that selection needs parameter (independent variable) number of back analysis, produce at first at random satisfy the constraint of independent variable bound /a NP n-dimensional vector, formula is following:

$x_{\mathrm{ij}}\left(0\right)=\mathrm{rand}*({x_{\mathrm{ij}}}^U-{x_{\mathrm{ij}}}^L)+{x_{\mathrm{ij}}}^L---\left(1\right)$

i＝1，2，…，Np；j＝1，2，…，n.

is respectively i to the flow control j component upper bound and lower bound in the formula; Rand is the random number between [0,1].The number of the corresponding parameter to be identified of vector dimension n.

2) mutation operation.On the 3rd the vectorial individuality in the difference in the convergent-divergent population between any two vectorial individualities and the population that is added to, form new variation variable.J component of i variation vector is for G+1 generation:

V _i，j(G+1)＝x _r1j(G)+F(x _r2j(G)-x _r3j(G)) (2)

Subscript r1 in the formula, r2, r3 are random integers in [1, NP] and different, and F is a zoom factor, is used for regulating the step-length amplitude of vectorial difference, value in 0～2.Formula (3) is the basic variation mode of difference evolution algorithm, is known as the DE/rand/1 pattern.Difference is evolved and is also had other patterns, like DE/best/1, DE/best/2, DE/rand/2 etc.

3) interlace operation.With object vector x _i(G) with the variation vector v _i(G+1), generate new sample vector u according to following rule hybridization _i(G+1), its j representation in components is:

R in the formula _j∈ [0,1] be with to flow control j component random number corresponding; CR ∈ [0,1] is hybridization probability constant; Rn _iFor 1,2 ..., integer of random choose among the D is to guarantee the vectorial V that makes a variation _i(G+1) in, has one-component at least by sample vector u _i(G+1) adopt.

4) select.Adopt greedy search method to carry out selection operation.With sample vector u _i(G+1) and object vector x _i(G) relatively, call finite element program and calculate, the objective function that adopts formula (1) is as evaluation function, if u _i(G+1) the less target function value of correspondence is then selected u _i(G+1); If instead, x _i(G) the less target function value of correspondence then keeps x _i(G).

5) loop iteration.Repeat 2)-4) calculating, up to i from 1 to Np, j accomplishes from 1 to n circulation, promptly accomplishes the iteration of population.Circulation is then ended iteration up to reaching greatest iteration step number or adaptive value less than setting value, and the vector of exporting current correspondence is the parameter of being discerned.The more detailed content list of references of relevant DE algorithm ([1] Rainer Storm; Kenneth Price.Differential Evolution-A Simple and Efficient Adaptive Scheme for Global Optimization over Continuous Spaces [J] .Technicai Report TR-95-012.International Computer Science Institute; 1995. [2] Xie Xiaofeng; Zhang Wenjun; Zhang Guorui etc. the experimental study of differential evolution [J]. control and decision-making, 2004,19 (1): 49-52.).

Algorithm of support vector machine is based on the machine learning algorithm of Statistical Learning Theory, also is the outstanding non-linear self study instrument after neural network.It is based on the theories of learning of strict structural risk minimization, overcomes neural network and crosses problem concerning study, is considered to very outstanding machine learning model.

The thinking of SVMs is mapped to linear problem with the high dimensional nonlinear problem, replaces inner product to calculate with kernel function then.Through to the data sample to { x _i, y _i, i=1 ..., n, x _i∈ R ^d, y _iThe study of ∈ R can be set up regression function and done

$f\left(x\right)=(w\·x)+b=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i)(x_i\·x)+b^{*}---\left(4\right)$

Here α _i,

Corresponding is the support vector of sample.

Inner product operation in the following formula can replace with the kernel function that meets certain condition, and the inner product operation that generally can adopt radially basic gaussian kernel function (9) to substitute in (8) is realized the nonlinear function match.

$k(x,x_i)=\exp \{-\frac{{|x-x_i|}^2}{{\mathrm{\σ}}^2}\}---\left(5\right)$

Similar with the method for neural network model, SVMs also comprises training and predicts two stages.SVMs training algorithm commonly used has Chunking algorithm, decomposition algorithm, sequence minimum optimization SMO method, closest approach iteratively faster (NPA) algorithm, least square method etc.Wherein the SMO algorithm has been avoided the unstable and problem consuming time of numerical solution under the multisample situation, does not also need big matrix stores space, has effectively improved and has found the solution efficient, adopts the SMO algorithm here.But relevant theoretical more in detail and principle list of references ([3] VapnikV.The Nature of Statistical Learning Theory [M], Springer-Verlag, New York.1995.[4] Zhang Xuegong. about Statistical Learning Theory and SVMs [J]. robotization journal, 2000).

The present invention mainly utilizes the nonlinear fitting ability and the self-learning capability of SVMs, and be used for expressing rich pool and flow the nonlinear relationship that is coupled admittedly between parameters for tunnel and the monitoring information, and the nonlinear relationship of historical Monitoring Data and following Monitoring Data.Guarantee the fitting precision of the solid coupling nonlinear mechanical behavior of rich pool stream like this, and, greatly improved computing velocity through replacing the numerical evaluation model with supporting vector machine model.

Theory and practice proves that SVM prediction model accuracy and its parameter (nuclear parameter σ and penalty factor C) have confidential relation, and general SVMs parameter selection needs artificial tentative calculation, bears the character of much blindness.Here the difference evolution algorithm of introducing with preceding text (DE) is optimized the parameter of SVMs, obtains the best model of precision, forms evolution-supporting vector machine model.Evolution-algorithm of support vector machine is seen Fig. 1.

Multiple information intelligent wireless warning system is monitored in a kind of tunnel as shown in Figure 2 automatically, comprises the steps:

Step 1, based on the country rock parameter recognition of difference evolution algorithm and SVMs:

Underground works has complicacy and uncertain characteristics, and rich pool tunnel surrounding is all the more so, and prior country rock mechanics and hydrologic parameter are difficult to confirm.Country rock mechanics and hydrologic parameter are again basis and the foundations of confirming information alarm thresholds such as displacement, flow, stress.Back analysis is meant mechanics parameter and the hydrologic parameter that utilizes field observation physical quantity inverse to go out country rock.Back analysis is divided into analytical method and numerical method, and numerical method is divided positive back analysis, converse analysis again and schemed three kinds of spectrometries.Wherein positive back analysis method can utilize to have now is just drilling calculation procedure, has very strong adaptability.Its principle steps is following: choose one group of calculation of parameter and obtain surrouding rock deformation and flow physical quantity; If the computational physics amount differs bigger with monitoring physical quantity data; Then adjust input parameter and recomputate, differ very little up to calculating with Monitoring Data, at this moment corresponding parameter is the parameter of identification; This shows that parameter back-analysis is optimization problem in essence.Physical meaning according to the tunnel surrounding parameter is set bound, if z observed reading arranged in the zone, adopts following formula to carry out constrained optimization to it:

$\left\{\begin{array}{c}\min E(x_1,x_2,...,x_n)=\frac{1}{z}\underset{i=1}{\overset{z}{\mathrm{\Σ}}}{[Y_i^0-Y_i]}^2\\ x_i^a\≤x_i\≤x_i^b(i=\mathrm{1,2},...,n)\end{array}\right.---\left(6\right)$

In the formula,

Be the measured value of country rock multiple information, Y _iBe the calculated value of country rock multiple information, z is the observed reading number, x _iBe i parameter, n is a number of parameters,

With

Be x _iUpper and lower limit;

Y in the following formula _iCalculating need be undertaken by the dimensional Finite Element program of the solid coupling of stream; The dimensional Finite Element of the solid coupling of stream need consume the plenty of time; If in the objective function optimization process, carry out this dimensional Finite Element process in large quantities, this optimizing process is to be difficult to realize so.The present invention adopts the SVMs response surface model of preceding text introduction to replace the three-dimensional flow FEM calculation that is coupled admittedly.

Through certain limit interior mechanics and hydrologic parameter combination orthogonal design scheme are carried out dimensional Finite Element; Obtain the data sample of the corresponding relation of mechanics, hydrologic parameter and measuring point head, flow, displacement physical quantity monitoring information; Utilize these sample training SVMs; Thereby set up the SVMs response surface function of the three-dimensional hydrology, the supporting vector machine model between mechanics, hydrologic parameter and the response physical quantity is:

$\left.\begin{array}{c}H=\mathrm{SVM}\left(P\right)\\ P=(p_1,p_2,...,p_n)\\ H=(H_1,H_2,...,H_z)\end{array}\right\}$

Wherein P is the parameter of country rock zoning, and n is the number of parameter; H greatly improves computing velocity for the multiple information of observation, z for the number of observation multiple information, the SVM model that learning training is obtained is used for calculating finding the solution of monitoring information.Research shows, different corresponding the different support vectors with choosing of nuclear parameter σ of penalty factor c have considerable influence to the SVM fitting precision, and these selection of parameter also can be regarded optimization problem as.

The difference evolution algorithm is used for choosing of the solid coupling parameter identification of country rock stream and supporting vector machine model penalty factor c and nuclear parameter σ; Thereby set up the algorithm (as shown in Figure 3) based on the parameter three-dimensional numerical value feedback identification of difference evolution-SVMs (DE-SVM), step is following:

(1) according to engineering problem, confirm the span of country rock mechanics and hydrologic parameter, and according to the numerical procedure of Orthogonal Experiment and Design principles of construction parameter combinations;

(2) adopt the three-dimensional flow Finite Element Method that is coupled admittedly that each parameter combinations scheme of structure is calculated, obtain the physical quantity information of the corresponding observation station of each scheme, and each numerical procedure is constituted one group of learning sample with corresponding information calculations value;

(3), adopt best SVMs penalty factor c and the nuclear parameter σ of difference evolution algorithm search based on above-mentioned sample;

(4) use the best SVMs parameter that obtains,, set up and treat the country rock parameter of inverting and the non-linear supporting vector machine model between the polynary monitoring information, replace flowing the finite element solving that is coupled admittedly top sample learning training;

(5) according to treating the mechanics of inverting and the supporting vector machine model between hydrologic parameter and the polynary observation information; Error with typical position observed reading and calculated value is an objective function; Adopt the search of difference evolution algorithm to treat the country rock parameter of inverting; Utilize the parameter of inverting further to confirm the mechanical property of country rock, thereby adjust the warning value of country rock monitoring information.

Step 2, based on the country rock multiple information time series forecasting of difference evolution algorithm and SVMs:

The essence of advanced prediction is to utilize the historical data of each monitoring to predict as shown in Figure 5 to following data.With Fig. 4 is how the example explanation makes up sample through the historical Monitoring Data of monitoring, and carries out following data prediction.Suppose x ₁, x ₂, x ₃..., x _nIndividual sensor has obtained the n metamessage, and each sensor is gathered since t-m constantly, obtains m-1 Monitoring Data, and then following t data can be come out by a prostatitis m-1 data-speculative, the principle of this time series forecasting just.With the monobasic time series is example; Along with the acquisition of 1 sequential in back, the oldest sequential sample that substitutes the front with this new sequential sample carries out next step prediction, obtains predicted value next time; And the like, just formed some groups of I/O mapped sample datas;

Because the coupling of multiple information, the future value of a certain monitoring information not only depends on self-sensor device historical data, also depends on the Monitoring Data of other sensors, so the multiple information historical data has more rationality than simple monobasic information data.Multiple information time series forecasting thinking is: the Monitoring Data that certain sensor is follow-up, the following data x constantly of the 1st sensor ₁(t), the data x that obtains through n sensor ₁(t-m), x ₁(t-m+1) ..., x ₁(t-1); x ₂(t-m), x ₂(t-m+1) ..., x ₂(t-1); ; x _n(t-m), x _n(t-m+1) ..., x _n(t-1) infer,, replenish into new Monitoring Data, replace old Monitoring Data, realize the data corresponding relation of rolling like this, just can obtain the sample of the relation of historical data and following data along with the acquisition of new Monitoring Data;

The problems referred to above formulation is following: to each monitoring variable, select the future anticipation step number to be 1 (also can other constants), and i=1,2 ..., n is the sensor number, thus historical polynary vector is expressed as:

$\stackrel{\→}{X}(t-1)=[x_1(t-m),x_1(t-m+1),...,x_1(t-1),...,x_i(t-m),x_i(t-m+1),...,x_i(t-1),---\left(7\right)$

$...,x_n(t-m),x_n(t-m+1),...,x_n(t-1){]}^T$

Wherein: what formula (7)

was represented is historical polynary vector; Represent that each sensor constantly begins to gather m data from t-m; The information that comprises the needed time in the past sequence of the following information of prediction in the formula; Input as model; Following arbitrary monitoring information constantly can be expressed as the function of historical polynary vector, is expressed as:

$x_i\left(t\right)=f\left(\stackrel{\→}{X}(t-1)\right)---\left(8\right)$

The x of (8) in the formula _i(t) be following prediction physical quantity data, f () is the nonlinear function of historical polynary vector sum future anticipation information; This inputoutput data corresponding relation with shown in Figure 4 also be consistent.

Can see from Fig. 4 and formula (8); Import historical polynary vector

and comprised n dimension multiple information; Amount to m * n data; Between this n dimension data certain correlativity is arranged; If all data directly cause data redundancy to a certain degree to repeat as input, reduced the learning efficiency of model.

In order to improve counting yield, should reduce the relevant input vector of repetition as far as possible, carry out the processing of data redundancy, reduce the dimension of input data.The effective ways that reduce the input dimension of multiple information are the principal component analytical methods (PCA) in the statistics, and the present invention adopts the PCA method to import the simplification of historical polynary vector.

PCA is known a kind of multivariate data dimension-reduction treatment mathematical method.Data to input are carried out linear combination, make the output parameter of acquisition have bigger variance, just possess the ability of better pattern being distinguished.Input vector adopts principal component analytical method to simplify, and concrete steps are following:

(1) standardization sample matrix X _{M * n}For

X _{M * n}For Expression; The former is the matrix that m Monitoring Data of n sensor acquisition constitutes, and the latter is the canonical matrix after simplifying.

(2) svd

is:

$\tilde{X}=\mathrm{U\Σ}{V}^T---\left(9\right)$

Wherein: ∑=diag [s ₁s ₂s _p0 ... 0] (10)

In the formula: p is the eigenwert number, s ₁>=s ₂>=..., s _pBe the singular value of correspondence, U is m * m rank matrix, and ∑ is m * n rank matrix, and V is n * n rank matrix;

(3) owing to The noise, can not have strictness is zero singular value.It has been generally acknowledged that less singular value is to be caused by noise, calculating only needs to keep bigger singular value, the pivot composition in the corresponding input sample matrix.Adopt the formula of the definite singular value that need keep of a kind of simple but effective method following in the literary composition:

${\mathrm{\η}}_k=\frac{\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{S_j}^2}{\underset{i=1}{\overset{p}{\mathrm{\Σ}}}{S_i}^2}$ j＝i，2，…，k；i＝1，2，…，p. (11)

Confirm threshold value O＜η ₀＜1, if η _k＞η ₀, k singular value before then keeping;

(4) establish k singular value and be retained, then corresponding pivot number is k, reduces transformation matrix V _{N * n}Dimension do

System's pivot is:

$\stackrel{\→}{Z}(t-1)\≈\stackrel{\→}{X}(t-1)\tilde{V}---\left(12\right)$

Input vector

through after the principal component analysis (PCA) effectively reduces dimension, and forecast model becomes following formula like this:

$x_i\left(t\right)=f\left(\stackrel{\→}{Z}(t-1)\right)---\left(13\right)$

In the formula;

is the input vector after the principal component analysis (PCA); Input vector after just simplifying; What

represented is to simplify input vector before; Represent that each sensor constantly begins to gather m data from t-m; Through above-mentioned principal component analysis (PCA);

that

of n dimension reduced to the k dimension realized the simplification of input information, for SVM prediction provides the foundation.

It is the nonlinear relationship of a complicacy that the forecast model of formula (13) is expressed, and conventional polynomial regression model effect is undesirable, and neural network model existed problem concerning study.Follow the principle of structural risk minimization in SVMs (SVM) theory, therefore can overcome neural network crosses problem concerning study.SVMs can be advantageously applied to the nonlinear function fitting problems, and the present invention adopts SVMs to carry out the time series forecasting of multiple information.SVM prediction is divided into 2 steps, at first carries out the study and the training of sample, sets up input and the mapping relations of exporting.Then, the sample of input is predicted.

The time series forecasting process of SVMs is following:

(a) through multichannel sensor country rock is carried out the time series collection of polynary monitoring information, structure input and output sample data is utilized the principal component analytical method dimensionality reduction, has formed the predicted data sample of the input-output of multivariate time series;

(b) above-mentioned data sample is divided into two types at random, a part is a learning sample, and another part is a test sample book, (the former is used for training forecast model, the latter be used for testing model precision of prediction);

(c) difference evolution algorithm of support vector machine is carried out the initialization setting, comprise difference evolution initial parameter is set, variable number to be optimized is 2, population quantity NP, zoom factor F, hybridization probability constant C R; Provide initial population at random, different individual corresponding different SVMs parameters promptly provides different nuclear parameter σ and penalty factor C;

The algorithm of (d) evolving according to difference carries out mutation operation and interlace operation; SVMs is carried out the training of learning sample with each individual corresponding parameters; The supporting vector machine model that training obtains is predicted test sample book; According to predicting the outcome and according to formula: S (x)=ERMSE+EPA; ERMSE in the formula and EPA are respectively two performance index that the sample in the statistics is estimated: root-mean-square error and precision of prediction, calculate this individual adaptive value S (x);

(e) judge to select individual adaptive value S (x) whether to meet the demands, as not meeting the demands, to carry out the calculating of a new round again, to return step (d), if adaptive value meets the demands, the learning training process finishes;

(f) according to present input multivariate data, carry out the prediction of information in the future, (for example, carry out learning training through to existing surrouding rock deformation and flow monitoring information structuring sample.Utilize to obtain model information such as the surrouding rock deformation in future or flow are carried out rolling forecast, algorithm is of the back; ) for a plurality of monitoring informations, each detection information repeats the process of (c)-(f) respectively;

Step 3, the information of forecasting result who obtains according to field monitoring information and step 2 judge warning; Concrete steps are following:

Judge a certain monitor value information of field monitoring or predict whether value of reaching capacity of a certain following information, report to the police when reaching;

Whether judge a certain monitor value information of field monitoring simultaneously or predict has continuous a days speed to surpass the speed warning value in a certain following information, if any reporting to the police, and a >=1 wherein;

Whether judge field monitoring information simultaneously or predict has the continuous b of a plurality of information days speed to surpass the speed warning value in the following information, if report to the police, and b >=1 wherein.

Warning value is comprehensively to confirm according to standard, engineering experience and theoretical analysis.For example surrounding rock displacement and rate of displacement can be confirmed according to country rock parameter character according to " Design of Railway Tunnel standard TB10003-2005 ", " vcehicular tunnel design specifications (JTGD70-2004) " etc.According to accumulation monitoring variable warning value and rate of deformation peaked 100%, 80% and 50% three grades of alarming values are set respectively, to carry out the expression of hazard level different conditions.The information warning value of regulation and stipulation need be according to the parameter character and the classification of country rock, but because complicacy and the country rock parameter character uncertainty of geologic body, a lot of parameters are also indeterminate in advance.And some information can't provide through standard, and the polynary parameter recognition algorithm that need be coupled admittedly based on the country rock stream of DE-SVM obtains parameter, and carries out the part Theoretical Calculation, thereby confirms the alarm threshold value of rich pool tunnel surrounding.

Utilize above-mentioned country rock multiple information time series forecasting algorithmic procedure to realize the follow-up advanced prediction of each monitoring information based on DE-SVM.The advanced prediction value of each information and the warning value of each information are compared, if surpass warning value then report to the police.

Said system is applied in the real-time monitoring system of tunnel, rich pool two level multiple informations; Its advantage is more outstanding: to the characteristics in tunnel, rich pool; Consider to relate to multiple information such as surrouding rock deformation, surrouding rock stress, structural internal force, seepage flow flow, pore water pressure, flow in the constructing tunnel; Lay multichannel sensors such as displacement transducer, pressure cell, reinforcing bar meter, water turbidity degree sensor, ventage piezometer, flowmeter at the scene, these sensors pass to the slave computer of advantages of small volume-also be on the static data collector to data through serial cable.The static data collector utilizes the powerful data processing function of PC computer to carry out real-time analysis, the advanced prediction of multivariate data through the long-range host computer-also be on the PC computer of data processing centre (DPC) of sending to of transmitter module.Computer possesses man-machine interaction's function, and this is quite necessary and useful for handling similar complicated tunnel stability problem.Send to the acoustic-optic alarm in on-the-spot specific place according to computer data result and alarm mode through directed wireless transmitter module, alerting signal inserts Ihternet and communication network, in time is dealt on related personnel's mobile phone, QQ and the email.Through on two levels-layout of slave computer monitoring system, solved site space nervous with the contradiction of analysis processor volume between greatly, and increased the ability of adaptation complicacy that makes the monitoring analysis function of host computer greatly.Can make full use of the powerful data processing function of the computer at rear.Radio Transmission Technology, the cable that data transmit is saved in the restriction of breakthrough space length, and can make the transmission target of alerting signal flexible more and extensive.The hardware block diagram of native system device is seen Fig. 6.On-the-spot collection emitting appearance has the function of automatic collection multiple information and based on the function of GPRS wireless transmit transmission data, each The data serial communication transmission of collection is according to the ID numbering identification of each sensor.The special-purpose GPRS modulation module of emission The data, its hardware composition diagram is seen Fig. 7.

The SIM300 chip that GPRS modulation module adopts Siemens Company to produce; It is embedded ICP/IP protocol, complete compatible AT instruction provides the standard serial interface and the MCU of 9 pins to get in touch; Realize the data communication of duplex, GSM short message, voice and GPRS data service can be provided.The hardware circuit of communication module comprises power interface, TTL-RS232 level shifter interface circuit, RF radio-frequency antenna circuit, speech IO interface.The TCP signal procedure flow process of the GPRS module of collection-transmitter is seen Fig. 8.

Along with the development of digital mobile communication network (GSM), can in time not understand the firsthand information in forefront of the production yet, in time the accident in the processing production process is the modern control system Development Trend.Gsm system is present the most ripe perfect, coverage rate is the widest, function is the strongest, the user is maximum GSM in the mobile communication system based on tdma.GSM host will provide multiple business such as voice, short message, data.Advantages such as the GSM short message service is simple with its connection, cost, wide coverage, realization convenience have obtained using widely.The module that GSM is commonly used have TC35 series, the Ericsson of Siemens DM10/DM20, in emerging ZXGM18 series, and SIM300 mentioned above etc., and these functions of modules usage difference are little.The SMS alarm information source is sent by two positions, and the one, collection-transmitter calls polynary monitoring information speed Realtime Alerts flow process, directly carries out Realtime Alerts, sends GSM note and sound and light alarm.The 2nd, through data processing centre (DPC), call polynary monitoring information SVMs advanced prediction algorithm, carry out advanced early warning.The email and the qq that send GSM note, GPRS report to the police and on-the-spot sound and light alarm.Text commonly used and PDU pattern at present send SMS message.The Text code is simple, but does not support Chinese, and PDU then can support Chinese to send.GMS note transmission flow is seen Fig. 9.

Embodiment: with certain mountain tunnel is that example describes.Tunnel trunk passes through high terrace, the Yellow River and V level country rock, loess hills district, and topographic relief is big, and relative relief reaches more than the 300m.Face of land cheuch is grown, and cuts darker relatively.Except that importing and exporting and the face of land, cheuch location is distributed with the round gravel soil, all the other faces of land, location mostly have loess to cover, and massif is domatic, and to go up vegetation more sparse.The tunnel mileage is DK7+284～DK10+382, and total length 3098m is a double track tunnel.The minimum buried depth 35m in tunnel, the maximum buried depth in tunnel can reach 300m.The tunnel adopts three steps, seven steps excavation method to carry out.Shotcrete bolt construction method technical requirement design is pressed in the tunnel, adopts composite lining, and combined bolting and shotcrete is adopted in preliminary bracing.Pneumatically placed concrete adopts wet spraying process.The sensor that the tunnel is laid after a lining comprises: pressure cell, hydrostatic level, pore pressure gauge, multipoint displacement meter and reinforcing bar meter.The transducer arrangements in tunnel is seen Figure 10 (A is a pressure cell, and B is a hydrostatic level, and C is a pore pressure gauge, and D is the reinforcing bar meter, and E is a multipoint displacement meter).Monitoring Data is through the collection of data acquisition Multielement Information System and analyze.The side wall displacement that multipoint displacement meter obtains finally is 4.32cm, and 4.73cm, hydrostatic level record crown and sink to being 3.83cm.It is 0.26MPa that pore pressure gauge records pore water pressure, the flow 30m of face place ³/ day.The mechanics parameter that at first obtains through the polynary parameter recognition algorithm inverting that is coupled admittedly based on the country rock of DE-SVM stream is: E=4.104GPa, c=0.589MPa, φ=25 °.The parameter substitution numerical model of inverting is calculated and the measured displacements contrast.The convergence curve that DE optimizes in the parameter recognition process is seen Figure 11.Parameter according to identification is carried out numerical evaluation, and combines " Design of Railway Tunnel standard TB10003-2005 ", " vcehicular tunnel design specifications (JTGD70-2004) " to wait three grades of alarming values of this umbrella arch sedimentation 17mm, 27.2mm and 34mm respectively.

Based on the country rock multiple information time series forecasting algorithm of DE-SVM to the data time series analysis of October 1 to October 25; The crown of choosing 1# hydrostatic level (Figure 10 is right to be surveyed) position sinks as prediction output; Choose 1# hydrostatic level and adjoining ventage piezometer, pressure cell, multipoint displacement meter multiple information measured value as the prediction cuit, the input data constitute 4 dimension information, and choosing historical step number is 5; Sample input raw data is 5 * 4=20, is output as 1.The Monitoring Data of 20d before utilizing is carried out rolling forecast to the distortion of back 5d.Adopt 15 learning samples of preceding method structure of the present invention, utilize the PCA principal component analytical method, these 15 sample inputs are reduced to dimension 14 by dimension 20.

Utilize above-mentioned sample, employing DE-SVM evolution SVMs is trained and is predicted that the initial parameter F=0.8 of difference evolution algorithm is set, and CR=1, initial population number are 30, and maximum iteration time is 600.The operation calculation procedure, when iteration finished, the selected penalty factor C that searches for supported vector machine was 1021.3, and gaussian kernel function parameter σ is 213.6, and the SVM prediction model prediction result of correspondence sees Figure 10.

Visible by Figure 12, the prediction maximum absolute error on October 20 is 1.2mm, and relative error is 3.75%, and effect is gratifying.The alarm condition of system is following: October 20, data center makes 5 days red advanced early warning in advance according to above-mentioned SVM prediction result.Workmen, equipment and material are prepared in advance, find October 25 that change in displacement was certain and early warning information is approaching, carry out surrounding rock consolidation immediately, have controlled the development of follow-up surrouding rock deformation well.

Apparatus of the present invention (Fig. 6) warning course of work is following: elder generation is according to buried depth, cavern's size and the rock mass classification of country rock, based on the warning values at different levels of experience initial setting multiple information.Lay a plurality of sensors and gather multivariate data automatically, through the polynary monitoring information of collection in worksite, obtain the country rock parameter back-analysis again, warning values at different levels are adjusted based on the DE-SVM recognizer at the scene, tunnel.Country rock multiple information time series forecasting algorithm based on DE-SVM predicts that to specific physical quantity information information of forecasting and warning value at different levels compare.Predicted value is compared with setting warning value,, explain that then country rock is safe if less than warning value.If predicted value is greater than warning value then carry out dangerous classification and judge, and carry out sound and light warning, and adopt that multipath is directed reports to the police, the email from GPRS internet to the managerial personnel in a distant place and the qq that carry out GSM mobile handset note, report to the police.

Two levels of this device design host computer-slave computer; Adopt wireless transmit to transmit data; Reduced to the local communication cable with to the requirement of field instrumentation data-handling capacity, thereby made the collection-transmitter of situ part be integrated in a circuit board, and be installed in the sealing metal box; Advantages of small volume, and have waterproof action.The field by using wireless data transmission technology; Can all sidedly the field monitoring data successfully be sent to the data handling system of the rear computer of overlength distance; The result who analyzes reports to the police through note, email and qq; Improve specific aim, dirigibility and the reliability of reporting to the police, broken through the restriction of space length.It is on-the-spot to make the related personnel needn't directly contact complicated dangerous country rock, obtains the status information of country rock in the distance safely.

The above; Be merely the preferable embodiment of the present invention; But protection scope of the present invention is not limited thereto; Any technician who is familiar with the present technique field is equal to replacement or change according to technical scheme of the present invention and inventive concept thereof in the technical scope that the present invention discloses, all should be encompassed within protection scope of the present invention.

Claims (3)

1. multiple information intelligent wireless warning system is monitored in a tunnel automatically, it is characterized in that comprising the steps:

Step 1, based on the country rock parameter recognition of difference evolution algorithm and SVMs, concrete steps are following:

(1) according to engineering problem, confirm the span of country rock mechanics and hydrologic parameter, and according to the numerical procedure of Orthogonal Experiment and Design principles of construction parameter combinations;

(2) adopt the three-dimensional flow Finite Element Method that is coupled admittedly that each parameter combinations scheme of structure is calculated, obtain the physical quantity information of the corresponding observation station of each scheme, and each numerical procedure is constituted one group of learning sample with corresponding observation station information calculations value;

(3), adopt best SVMs penalty factor c and the nuclear parameter σ of difference evolution algorithm search based on above-mentioned sample;

(4) use the best SVMs parameter that obtains,, set up and treat the country rock parameter of inverting and the non-linear supporting vector machine model between the polynary monitoring information, replace flowing the finite element solving that is coupled admittedly top sample learning training;

(5) according to treating the mechanics of inverting and the supporting vector machine model between hydrologic parameter and the polynary observation information; Error with typical position observed reading and calculated value is an objective function; Adopt the search of difference evolution algorithm to treat the country rock parameter of inverting, utilize the parameter of inverting further to confirm the mechanical property of country rock;

Step 2, based on the country rock multiple information time series forecasting of difference evolution algorithm and SVMs, concrete steps are following:

(a) through the time series collection of multichannel sensor to polynary monitoring information, structure input and output sample data is utilized the principal component analytical method dimensionality reduction, has formed the predicted data sample of the input-output of multivariate time series;

(b) above-mentioned data sample is divided into two types at random, a part is a learning sample, and another part is a test sample book;

(c) difference evolution algorithm of support vector machine is carried out the initialization setting, comprise difference evolution initial parameter is set, variable number to be optimized is 2, population quantity, zoom factor F, hybridization probability constant C R; Provide initial population at random, different individual corresponding different SVMs parameters, promptly corresponding different nuclear parameter σ and penalty factor C;

The algorithm of (d) evolving according to difference carries out mutation operation and interlace operation; Carry out the training of learning sample with each individual corresponding parameters; The supporting vector machine model that training obtains is predicted test sample book; According to predicting the outcome and according to formula: S (x)=ERMSE+EPA; ERMSE in the formula and EPA are respectively two performance index that the sample in the statistics is estimated: root-mean-square error and precision of prediction, calculate this individual adaptive value S (x);

(e) judge to select individual adaptive value S (x) whether to meet the demands, as not meeting the demands, to carry out the calculating of a new round again, to return step (d), if adaptive value meets the demands, the learning training process finishes;

(f), carry out the prediction of certain sensor information in the future according to importing present polynary Monitoring Data; For following monitoring information prediction of a plurality of sensors, each sensor repeats the process of (c)-(f) respectively;

Step 3, the information of forecasting result who obtains according to field monitoring information and step 2 judge warning; Concrete steps are following:

Judge a certain monitor value information of field monitoring or predict whether value of reaching capacity of a certain following information, report to the police when reaching;

Whether judge a certain monitor value information of field monitoring simultaneously or predict has continuous a days speed to surpass the speed warning value in a certain following information, if any reporting to the police, and a >=1 wherein;

Whether judge field monitoring information simultaneously or predict has the continuous b of a plurality of information days speed to surpass the speed warning value in the following information, if report to the police, and b >=1 wherein.

2. multiple information intelligent wireless warning system is monitored in a kind of tunnel according to claim 1 automatically, it is characterized in that: in the said step 1, following based on the concrete steps of the polynary parameter recognition of country rock of difference evolution algorithm and SVMs:

Its principle steps is following: choose one group of calculation of parameter and obtain surrouding rock deformation and flow physical quantity; If the computational physics amount differs bigger with monitoring physical quantity data, then adjust input parameter and recomputate, differ very little up to calculating with Monitoring Data; At this moment corresponding parameter is the parameter of identification; Physical meaning according to the tunnel surrounding parameter is set bound, if z observed reading arranged in the zone, adopts following formula to carry out constrained optimization to it:

$\left\{\begin{array}{c}\min E(x_1,x_2,...,x_n)=\frac{1}{z}\underset{i=1}{\overset{z}{\mathrm{\Σ}}}{[Y_i^0-Y_i]}^2\\ x_i^a\≤x_i\≤x_i^b(i=\mathrm{1,2},...,n)\end{array}\right.$

In the formula,

Be the measured value of country rock multiple information, Y _iBe the calculated value of country rock multiple information, z is the observed reading number, x _iBe i parameter, n is a number of parameters, With

Be x _iUpper and lower limit;

Through country rock mechanics and hydrologic parameter combination orthogonal design scheme in the certain limit is carried out dimensional Finite Element; Obtain the data sample of the corresponding relation of mechanics, hydrologic parameter and measuring point head, flow, displacement response physical quantity monitoring information; Utilize these sample training SVMs; Thereby set up the SVMs response surface function of the three-dimensional hydrology, the supporting vector machine model between mechanics, hydrologic parameter and the response physical quantity is:

$\left.\begin{array}{c}H=\mathrm{SVM}\left(P\right)\\ P=(p_1,p_2,...,p_n)\\ H=(H_1,H_2,...,H_z)\end{array}\right\}$

In the formula: P is the parameter of country rock zoning, and n is the number of parameter; H is the multiple information of observation, and z is used for responding finding the solution of physical quantity for the number of observation multiple information, the SVM model that learning training is obtained.

3. multiple information intelligent wireless warning system is monitored in a kind of tunnel according to claim 1 automatically; It is characterized in that: in the said step 2, following based on the predicted data sample construction step of the input-output in the country rock multiple information time series forecasting algorithm of difference evolution algorithm and SVMs:

Suppose x ₁, x ₂, x ₃..., x _nSensor has obtained the n metamessage, and each sensor is gathered since t-m constantly, obtains m-1 Monitoring Data; Then following t data can be come out by a prostatitis m-1 data-speculative; Along with the acquisition of 1 sequential in back, the oldest sequential sample that substitutes the front with this new sequential sample carries out next step prediction, obtains predicted value next time; And the like, just formed some groups of I/O mapped sample datas;

Multiple information time series forecasting thinking is: with the 1st sensor is example, the Monitoring Data that it is follow-up, promptly following data x constantly ₁(t), the data x that obtains through n sensor ₁(t-m), x ₁(t-m+1) ..., x ₁(t-1); x ₂(t-m), x ₂(t-m+1) ..., x ₂(t-1); ; x _n(t-m), x _n(t-m+1) ..., x _n(t-1) infer,, replenish into new Monitoring Data, replace old Monitoring Data, realize the data corresponding relation of rolling like this, just can obtain the sample of the relation of historical data and following data along with the acquisition of new Monitoring Data;

The problems referred to above formulation is following: to each sensor monitors amount, selecting the future anticipation step number is 1, i=1, and 2 ..., n is the sensor number, thus historical polynary vector is expressed as:

$\stackrel{\→}{X}(t-1)=[x_1(t-m),x_1(t-m+1),...,x_1(t-1),...,x_i(t-m),x_i(t-m+1),...,x_i(t-1),$

$...,x_n(t-m),x_n(t-m+1),...,x_n(t-1){]}^T$

Wherein: what represented is historical polynary vector; Represent that each sensor constantly begins to gather m data from t-m; The information that comprises the needed time in the past sequence of the following information of prediction in the formula; Input as model; Following arbitrary monitoring information constantly can be expressed as the function of historical polynary vector, is expressed as:

$x_i\left(t\right)=f\left(\stackrel{\→}{X}(t-1)\right)$

X in the formula _i(t) be following prediction physical quantity data, f is the nonlinear function of historical polynary vector sum future anticipation information;

Above-mentioned model input vector has redundant characteristics, and input vector adopts principal component analytical method to simplify, and concrete steps are following:

(1) standardization sample matrix X _{M * n}For

X _{M * n}For

Expression;

(2) svd is:

$\tilde{X}=\mathrm{U\Σ}{V}^T$

Wherein: ∑=diag [s ₁s ₂s _p0 ... 0]

In the formula: p is the eigenwert number, s ₁>=s ₂>=...,>=s _pBe the singular value of correspondence, U is m * m rank matrix, and ∑ is m * n rank matrix, and V is n * n rank matrix;

(3) formula of definite singular value that need keep is following:

${\mathrm{\η}}_k=\frac{\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{S_j}^2}{\underset{i=1}{\overset{p}{\mathrm{\Σ}}}{S_i}^2}$ j＝1，2，…，k；i＝1，2，…，p.

Confirm threshold value 0＜η ₀＜1, if η _k＞η ₀, k singular value before then keeping;

(4) establish k singular value and be retained, then corresponding pivot number is k, reduces transformation matrix V _{N * n}Dimension do

System's pivot is:

$\stackrel{\→}{Z}(t-1)\≈\stackrel{\→}{X}(t-1)\tilde{V}$

Input vector

through after the principal component analysis (PCA) effectively reduces dimension, and forecast model becomes following formula like this:

$x_i\left(t\right)=f\left(\stackrel{\→}{Z}(t-1)\right)$

In the formula;

is the input vector after the principal component analysis (PCA); Polynary input vector after just simplifying; What

represented is the polynary input vector before simplifying; Above each sensor of formulate constantly begins to gather m data from t-m; Through above-mentioned principal component analysis (PCA);

that of n dimension reduced to the k dimension realized the simplification of input information, for SVM prediction provides the foundation.

Images