CN104318717A - Rainstorm debris flow early warning method under shortage conditions of historical data - Google Patents

Abstract

The invention provides a rainstorm debris flow early warning method under shortage conditions of historical data. The rainstorm debris flow early warning method under the shortage conditions of the historical data includes steps: obtaining enriched sample information by using a sample point as a representative of a surrounding information gap and diffusing information of the sample point around; building a relationship between rainfall conditions and debris flow eruption by using the sample information after being enriched in combination with a BP neural network, and thereby achieving effective debris flow early warning finally. The rainstorm debris flow early warning method under the shortage conditions of the historical data is used for the debris flow early warning in areas short of the historical data, can significantly improve early warning accuracy and reliability, conveniently collects related rainfall data, and is low in cost.

Description

Debris Flow method for early warning under a kind of historical data shortage condition

Technical field

The invention belongs to disaster alarm technical field, particularly relate to the Debris Flow method for early warning under a kind of historical data shortage condition.

Background technology

China is one of country suffering mud-stone flow disaster the most serious.The rubble flow wherein excited by heavy rain, being referred to as Rain-induced Debris Flow, is the most common and break out the highest a kind of rubble flow of frequency, especially in China western mountainous areas, all can have hundreds and thousands of raceway groove generation rubble flow annual rainy season, the direct economic loss caused thus is up to more than 2,000,000,000 yuan.Therefore, the prediction technology of further investigation Rain-induced Debris Flow, for the prevention ability conscientiously improving mud-stone flow disaster, minimizing casualties and economic loss are significant to greatest extent.

The generation of Debris Flow need possess three conditions: the rainfall of sufficient solid matter, the precipitous ditch bed gradient and a large amount of high strength, and rainfall is wherein most active factor.Research both domestic and external shows, the activity of the factors such as quantity of precipitation, precipitation intensity and rainy persistent time and Rain-induced Debris Flow has extremely close relationship.Therefore, relation between research condition of raining and debris flow occurrence, the debris flow early-warning method that people usually use just is become, as Critical Rainfall Weigh Direct Determination, Regression model forecasting method, Artificial Neural Network, clustering methodology etc. thus by rainfall monitoring data prediction forecast rubble flow.These methods in actual use, usually the rainfall data that a large amount of mud-stone flow disasters in the past occurs all is needed, namely sample size is enough large, otherwise the deficiency due to statistics is occurred larger deviation, even predicts the outcome far from each other with actual conditions.But, be but difficult to obtain these abundant history rainfall datas in reality.On the one hand, a lot of cheuch inherently lacks rainfall monitoring and the tracking of long-term sequence; On the other hand, different cheuches, due to landform, that landforms and rock form structure is different, and thus the rainfall data in a region can not be used for the early warning in another region as statistical data; Moreover, Southwestern China area is after suffering the Wenchuan violent earthquake of 2008, the thing source of many cheuches and topographic condition have great change, excite the rainfall condition of rubble flow also to have great change thus, and thus the predictive value of Rainfall data to current rubble flow that accumulated in the past of violent earthquake is little.These factors determine under many circumstances, usually can only obtain very limited historical data, thus adopt above-mentioned technology to be difficult to realize effective early warning of rubble flow.

Summary of the invention

The object of the present invention is to provide the Debris Flow method for early warning under a kind of historical data shortage condition, be intended to solve the problem that prior art is difficult to realize effective early warning of rubble flow.

The present invention is achieved in that the Debris Flow method for early warning under a kind of historical data shortage condition, comprises the following steps:

S1, the rainfall observation data that will obtain, after diffusion of information, obtain normalization diffuseness values;

The historical data of S2, employing locality is trained BP neural network, to obtain suitable weights after the diffusion of information of step S1;

S3, according to the BP neural network after training, using the rainfall data that collects after the diffusion of information of step S1 as network input, after the computing of network each layer, obtain output layer neuronic output z1, i.e. rubble flow possibility occurrence size;

The threshold value that S4, basis preset, judges that whether export z1 exceedes threshold value, in this way, then starts early warning.

Preferably, in step sl, described diffusion of information is on respective domain, adopt normal function to spread respectively each rainfall data.

Preferably, in step s 2, describedly training is carried out to BP neural network adopt BP algorithm to train.

Preferably, in step s 4 which, being set to of described threshold value: be set in advance as one close to 1 number, or the statistical value of false-alarm and false dismissal in using according to reality, be combined false-alarm and false dismissal loss coefficient that unit determines voluntarily, according to least disadvantage principle dynamically definite threshold.

Preferably, the described number close to 1 comprises 0.9.

The present invention overcomes the deficiencies in the prior art, provides the Debris Flow method for early warning under a kind of historical data shortage condition, in the methods of the invention, for the debris flow early-warning of historical data shortage area, can obtain good effect.The basic thought of the method is, when the sample size obtained is little (when namely history rainfall data is few, wherein the rainfall data of mud-stone flow disaster record is called a sample point each time, assuming that have recorded the rainfall data of N mud-stone flow disaster, then sample size is claimed to be N, namely N number of sample point is had), sample provides the information of understanding rainfall and rubble flow relation to be incomplete, information space is left between sample point and sample point, sample size is less, then this information space is larger.In this case, the observed reading of each sample point does not represent definite information, but has certain ambiguity and uncertainty.Now, each sample point should be regarded as the representative in its peripheral information space, nearer apart from it, then this to represent degree higher.When a sample point is regarded as the representative in peripheral information space, be in fact exactly in order to fuzzy set the conversion of definite value.By this conversion, the information of a sample point is spread towards periphery, thus filled up the information space between sample point and sample point to a great extent, enriched sample information.Sample information after foundation is abundant, in conjunction with BP neural network, can set up the relation between more accurate and reliable condition of raining and debris flow occurrence, thus finally realize effective debris flow early-warning.

Accompanying drawing explanation

Fig. 1 is the flow chart of steps of the Debris Flow method for early warning under historical data of the present invention shortage condition;

Fig. 2 is the structural representation of neural network in the present invention.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

A Debris Flow method for early warning under historical data shortage condition, as shown in Figure 1, comprises the following steps:

S1, the rainfall observation data that will obtain, after diffusion of information, obtain normalization diffuseness values.

Now there are some researches show, Rain-induced Debris Flow be precipitation and the coefficient result of antecedent precipitation then and there, here choose most representativeness, 4 rainfall factors having the greatest impact as rainfall condition, namely rainfall, when daily rainfall, 1 hour maximum raininess and 10 minutes maximum raininess.Prophase programming makes the soil body have higher water cut, the shearing strength of the soil body can be reduced, early-stage preparations are played a part to rubble flow, get the rainfall that first 20 day every day occurred rubble flow, be weighted the cumulative index as prophase programming by attenuation coefficient, computing formula is as follows:

$P=\underset{i=1}{\overset{20}{\mathrm{\Σ}}}{K}^i{P}_i---\left(1\right)$

In formula i represent rubble flow occur before number of days, P _ifor the rainfall amount of first i-th day occurs rubble flow, K is attenuation coefficient, and according to previous karyotype studies, desirable K is 0.84.The effect that debris flow occurrence works as daily rain amount makes the further unstability of the soil body, rubble flow is excited under the common promotion of follow-up short duration raininess, follow-up peak value precipitation then plays direct excitation to rubble flow, because peak value precipitation general persistence is shorter, usually only have several minutes to several tens minutes, therefore select 1 hour maximum raininess on debris flow occurrence same day and 10 minutes maximum raininess as the two indices of peak value precipitation.So, for a sample point, contain 4 observed readings, namely rainfall, when daily rainfall, 1 hour maximum raininess and 10 minutes maximum raininess, reflect relevant rainfall data when certain rubble flow in history occurs.

When the sample point had is few, the information that sample provides is incomplete.For obtaining abundant information from limited historical data, adopt the mode of diffusion of information here, 4 of each sample point observed readings are spread respectively on respective domain.Assuming that have N number of sample point, each sample point is designated as S _i, i=1,2 ... N, S _ibe expressed as:

S _i＝(s _i1,s _i2,s _i3,s _i4) (2)

S in formula _i1, s _i2, s _i3, s _i4represent the rainfall of i-th sample point respectively, when daily rainfall, 1 hour maximum raininess and 10 minutes maximum raininess, four components.To a jth component of sample, its domain is designated as U _j, j=1,2,3,4, U _jrepresent the possible span of all sample point jth components.U _jpolicy setting by such: in the current N number of sample point had, first obtains the maximal value of their a jth component, is designated as MAX _j=MAX{s _1j, s _2j..., s _nj, then get MAX _jtwice about as higher limit, then get its higher limit 1/10th as lower limit, finally decile nine equal portions between lower limit and the upper limit, namely get ten values as U _jdiscrete point, therefore U _jcan be expressed as:

U _j＝[u _j1＝u _j10/10,u _j2,...u _j10≈2MAX _j],j＝1,2,3,4 (3)

After setting 4 components domain separately, 4 of each sample point observed readings are spread respectively on respective domain.For realizing sane effect, adopt the diffusion way of normal function.To pending sample point S _i, i=1,2 ... N, by its jth component s _ij, j=1,2,3,4 are diffused into corresponding domain U _j, j=1,2,3, discrete point u in 4 _jk, k=1,2 ..., the value on 10 is defined as:

$q_{i,j,k}=\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_j}\exp [-\frac{{(u_{\mathrm{jk}}-s_{\mathrm{ij}})}^2}{2{\mathrm{\σ}}_j^2}]---\left(4\right)$

As can be seen from the above equation, the some distance observed reading in domain is nearer, and the diffuseness values of acquisition is larger, otherwise the diffuseness values then obtained is less.σ in formula _jfor the standard deviation of a jth component when spreading, control the diffusion of respective component, σ _jvalue should consider two factors: on the one hand, if sample size is less, to illustrate between sample point and sample point that information space is comparatively large, at this moment for filling up these larger spaces, each observed reading should spread far away; On the other hand, for a certain component, if the value difference of each sample point is comparatively large, namely sample standard deviation is comparatively large, and illustrate that the information space of this component is comparatively large, thus the observed reading of this component should spread far away.According to such principle, definition exponential function below determines σ _jvalue:

${\mathrm{\σ}}_j=A{\tilde{L}}_j\exp (-\sqrt{N}),j=\mathrm{1,2,3,4}---\left(5\right)$

In formula, A is positive constant coefficient, can value be about 40.N is sample size, i.e. sample points. for the sample standard deviation of a jth component, defined by following formula:

${\tilde{L}}_j={\left[\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(s_{\mathrm{ij}}-{\stackrel{\‾}{L}}_j)}^2\right]}^{1/2},j=\mathrm{1,2,3,4}---\left(6\right)$

In formula for the sample average of a jth component, be calculated as follows:

${\stackrel{\‾}{L}}_j=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{s}_{\mathrm{ij}},j=\mathrm{1,2,3,4}---\left(7\right)$

Now, to each sample point, by normal fashion, its 4 observed readings are diffused on 10 discrete points of respective domain respectively.For eliminating the impact brought due to the difference of each component values scope, and meeting the feature of fuzzy set degree of membership, further the diffuseness values of each component being normalized.Sample point S _ijth, j=1,2,3,4 component s _ijat corresponding domain U _jon diffuseness values add up be:

$C_{\mathrm{ij}}=\underset{k=1}{\overset{10}{\mathrm{\Σ}}}{q}_{i,j,k}---\left(8\right)$

So diffuseness values q _{i, j, k}, k=1,2 ... 10 are converted to μ after normalization _{i, j, k}:

${\mathrm{\μ}}_{i,j,k}=\frac{q_{i,j,k}}{C_{\mathrm{ij}}}---\left(9\right)$

Normalized μ _{i, j, k}have characteristic, in fact this reflect observed reading s _ija kind of probability distribution on corresponding domain.

Through above-mentioned process, 4 of each sample point observed readings changed respectively in order to the normalization diffuseness values of 10 on respective domain, this conversion also can be understood as and 4 observed readings has been separately converted to 4 fuzzy sets.In this way, obtain 4 × 10 diffuseness values by 4 original values of sample point, thus greatly enriched sample information.

The historical data of S2, employing locality is trained BP neural network, to obtain suitable weights after the diffusion of information of step S1.

In step s 2, as shown in Figure 2, whole network is divided into three layers to neural network structure: input layer, hidden layer and output layer, realize totally interconnected connection between the neuron of adjacent two layers.Input layer is containing 41 neurons, wherein x ₁to x ₁₀10 diffuseness values of corresponding sample point the 1st component, x ₁₁to x ₂₀10 diffuseness values of corresponding 2nd component, x ₂₁to x ₃₀10 diffuseness values of corresponding 3rd component, x ₃₁to x ₄₀10 diffuseness values of corresponding 4th component, can unified presentation be:

$x_l=\left\{\begin{array}{cc}{\mathrm{\μ}}_{i,1,l}& ,1\≤l\≤10\\ {\mathrm{\μ}}_{i,2,l-10}& ,11\≤l\≤20\\ {\mathrm{\μ}}_{i,3,l-20}& ,21\≤l\≤30\\ {\mathrm{\μ}}_{i,4,l-30}& ,31\≤l\≤40\end{array}\right.---\left(10\right)$

Neuron x ₀fixing value is 1, and the power of corresponding edge plays threshold value.Hidden layer contains 9 neurons, y _j, j=0,1,2 ... 8 represent each neuronic output valve, wherein y ₀be fixed as 1, and x ₀similar, play threshold value and arrange.Output layer is only containing 1 neuron, and its output valve is z ₁, represent the result that whole network exports, i.e. the possibility size of rubble flow generation, value is in interval (0 1), and output valve, more close to 1, represents that the possibility that rubble flow occurs is larger.Connection weight between input layer and hidden layer neuron is designated as w _ij, i=0,1 ... 40, j=1,2 ... 8, and the connection weight between hidden layer and output layer neuron is designated as w _j1, j=0,1 ... 8.To the output y of hidden layer neuron _j, j=1,2 ... 8, be calculated as follows:

$y_j=f\left(\underset{i=0}{\overset{40}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_i\right)=\frac{1}{1+\exp (-\underset{i=0}{\overset{40}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_i)}---\left(11\right)$

That is, the output of hidden layer neuron is by all neuronic weighted sums of input layer, then through a non-linear S type transfer function conversion gained.Similarly, output layer neuronic output z ₁, calculated by following formula:

$z_1=f\left(\underset{j=0}{\overset{8}{\mathrm{\Σ}}}{w}_{j1}{y}_j\right)=\frac{1}{1+\exp (-\underset{j=0}{\overset{8}{\mathrm{\Σ}}}{w}_{j1}{y}_j)}---\left(12\right)$

The neural network of Fig. 2 is adopted to carry out debris flow early-warning, exactly according to the 4 kinds of rainfall observation datas collected, after diffusion of information, the normalization diffuseness values of gained is inputted as network, through the computing of formula (11) and (12), obtain output layer neuronic output z ₁.Export z ₁namely the rubble flow possibility occurrence size of network calculations gained is represented.

But network is wanted correctly to work, and it respectively connects weights and need arrange rationally, correctly can reflect the actual conditions of local debris flow.Therefore, before real work, need to adopt local historical data to train network, to obtain suitable weights.

To the training of network, adopt BP algorithm.Its principle is exactly, and rainfall data when occurring for each in history rubble flow collected, when inputting as network after being spread, the network expected exports and (uses t ₁represent) be 1, and when 4 kinds of rainfall datas are 0, desired output should be 0.When network weight arranges unreasonable, truly export z ₁with desired output t ₁must have error, for all N number of sample points, this sum of the deviations can be expressed as:

$\mathrm{\ϵ}=\frac{1}{2}\underset{r=1}{\overset{N}{\mathrm{\Σ}}}{(z_{r1}-t_{r1})}^2---\left(13\right)$

Z in formula _r1represent the true output of r sample point, and t _r1represent the desired output of this sample point.To the training of network, be exactly that each connects weights by iteration adjustment, make total error ε level off to 0.Particularly, following six steps are comprised to the training process of network:

1, four kinds of rainfall datas during local each rubble flow generation are in history collected, namely rainfall, when daily rainfall, 1 hour maximum raininess and 10 minutes maximum raininess, four kinds of rainfall datas when each rubble flow occurs form a sample point, its desired output is 1, add the sample point that four kinds of rainfall datas are 0 in addition, its desired output is 0.Total sample points is designated as N, and each sample point is designated as S _i=(s _i1, s _i2, s _i3, s _i4), i=1,2 ... N.

2, according to the sample data obtained, the domain U of each rainfall data is set by formula (3) _j, j=1,2,3,4.

3, by formula (5), (6), (7), determine that the irradiated standard of each rainfall data is poor, and by formula (4), four of each sample point component datas are spread on 10 discrete points of respective domain, according to formula (8), (9), diffuseness values is normalized, obtain 40 normalized diffuseness values by four raw data of sample point thus, be designated as μ _{i, j, k}, i=1,2 ... N, j=1,2,3,4, k=1,2 ... 10.

4, netinit: to all connection weights of network, is set to the random number that (0 0.5) are interval.

5, a less iteration ends threshold value is set, is designated as T, usually desirable 1 × 10 ^-8left and right.Note t is the step number of iteration, and initial value is set to 0, and note ε (t) is the total error that t walks between the true output of network and desired output, and initial value ε (0) is set to 0, note w _ij(0), i=0,1 ... 40, j=1,2 ... 8, w _j1(0), j=0,1 ... 8 networks when being initial connect weights.T=t+1, enters circulation:

A, to all N number of sample points, by formula (10), its 40 normalization diffuseness values are inputted as network one by one, obtained the output of hidden layer neuron by formula (11), be designated as y _rj, r=1,2 ... N, j=1,2 ... 8, r representative sample point sequence number in formula, j represents the sequence number of hidden layer neuron.Again by formula (12), obtain the neuronic output of output layer, be designated as z _r1, r representative sample point sequence number in formula, the desired output of respective sample point is designated as t _r1.

B, according to network Output rusults, drawn the total error ε (t) of this step by formula (13).

C, calculate distance, delta between this step error ε (t) and previous step error ε (t-1) _ε=| ε (t)-ε (t-1) |.If Δ _ε< T, jumps out circulation, forwards the 6th step to, otherwise down performs.

D, upgrade connection weights between hidden layer and output layer:

$w_{j1}\left(t\right)=w_{j1}(t-1)+\mathrm{\&eta;}\underset{r=1}{\overset{N}{\mathrm{\&Sigma;}}}(t_{r1}-z_{r1})z_{r1}(1-z_{r1})y_{\mathrm{rj}},j=\mathrm{0,1},...8,$ In formula, η is a constant, controls the speed of iteration change, can be taken as 0.2.

E, upgrade connection weights between input layer and hidden layer:

$w_{\mathrm{ij}}\left(t\right)=w_{\mathrm{ij}}(t-1)+\mathrm{\&eta;}\underset{r=1}{\overset{N}{\mathrm{\&Sigma;}}}{y}_{\mathrm{rj}}(1-y_{\mathrm{rj}})(t_{r1}-z_{r1})(1-z_{r1})w_{j1}{x}_{\mathrm{ri}},i=\mathrm{0,1},...40,j=\mathrm{1,2},...8,$ In formula, η is taken as 0.2 equally.

F, t=t+1, forward A step to.

6, training terminates.

Namely network after training can be used for the debris flow early-warning in real work.

S3, according to the BP neural network after training, using the rainfall data that collects after the diffusion of information of step S1 as network input, after the computing of network each layer, obtain output layer neuronic output z ₁, i.e. rubble flow possibility occurrence size.

In step s3, by the four kinds of rainfall datas collected, diffusion is carried out and normalization according to the determined domain when network training and irradiated standard difference, by formula (10), obtain 40 normalization diffuseness values are inputted as network, calculated the output of hidden layer neuron by formula (11), then calculate the neuronic output of output layer by formula (12).

The threshold value that S4, basis preset, judges to export z ₁whether exceed threshold value, in this way, then perform step S5.

In step s 4 which, threshold value is set to 0.9 usually, also can in practice, according to the statistical value of false-alarm and false dismissal, is combined false-alarm and false dismissal loss coefficient that unit determines voluntarily, according to least disadvantage principle dynamically definite threshold.

S5, startup early warning.

As output valve z ₁when exceeding the threshold value of setting in advance, can debris flow early-warning be started, otherwise refuse early warning.

For showing performance of the present invention, apply the present invention to the debris flow early-warning of For It In Beibei, Chongqing.For It In Beibei, Chongqing is located in East Sichuan Basin equal ridge-valley region, belongs to southwestern col pleater tape, and landform, based on low mountain, hills, has Pingba and along river Valley.The geologic hazard that rubble flow Shi Gai district is important, excites primarily of heavy rain, and be a kind of typical Rain-induced Debris Flow, spot is mainly distributed in the low mountain of entrance, kwan-yin gorge and exit, Wen Tang gorge.Larger in history and mud-stone flow disaster that is that have relevant Rainfall data to record is as shown in table 1, have 8 times.

Table 1 For It In Beibei, Chongqing is debris flow occurrence Rainfall data (unit: mm) in history

The data of these 8 mud-stone flow disasters is divided into two parts, before several times rainfall data as training sample, network is trained, then the rainfall observed reading of several times below to be predicted according to the network after training, judges whether consistent with actual conditions.In addition, when testing, also will several times rainfall observed reading respectively divided by 10, predict for generating new data below, to check this method whether effectively real.Due to very little divided by the rainfall data after 10, all substantially can not excite rubble flow in theory with in reality, thus now desirable network output valve should be tending towards 0.

Test is divided into 6 kinds of situations, be respectively using in table above the data of 2 to 7 rubble flow as training sample, corresponding using the rainfall data of 6 to 1 time below as test sample book.The measured result of test result and BP networks method is as shown in table 2 below:

Table 2 test result

As can be seen from data in table, in all cases, to the original rainfall data of test sample book, by the output valve of this method substantially all more than 0.95, be in close proximity to 1, the threshold value according to 0.9 is arranged, and can both realize correctly early warning at every turn.To test sample book raw data divided by 10 situation, then output valve is substantially all about 0.02, close to 0, thus can not early warning, consistent with actual conditions.This illustrates, adopts this method, although under the very limited condition of historical data, still can realize correct early warning, substantially can stop the possibility of false dismissal and false-alarm.Investigate BP networks method, outwardly, to the original rainfall data of test sample book, its output valve is all more than 0.99, and seeming can early warning effectively.But analyze a little, be not difficult to find out the two problems of existence, one is in each case, and the output valve of each test sample book is all identical, this illustrates that the rainfall data of BP networks method to input is very insensitive, can not make the judgement of debris flow occurrence possibility size according to different rainfall datas.The more important thing is in addition, test sample book raw data divided by 10 situation, its output valve is still basic more than 0.9, with the output valve under raw data condition closely.This means, no matter which kind of rainfall condition, no matter be can excite or can not excite rubble flow, the method all can start early warning substantially, is in fact exactly can not normally work at all.Cause the reason of these problems, be just that historical data data is very few, thus effectively can not also realize effective classification by training network.Illustrate thus, under the condition of historical data shortage, the fuzzy neural network combined by our diffusion of information technology and BP network, still reliably can be realized the early warning of rubble flow, effectively overcome the intrinsic defect of prior art and deficiency.

Compared to the shortcoming and defect of prior art, the present invention has following beneficial effect: the present invention is used for the debris flow early-warning of historical data shortage area, significantly can improve accuracy and the reliability of early warning, and relevant rainfall data collection is convenient, and cost is low.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, all any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within protection scope of the present invention.

Claims (5)

1. the Debris Flow method for early warning under historical data shortage condition, is characterized in that, comprise the following steps:

S1, the rainfall observation data that will obtain, after diffusion of information, obtain normalization diffuseness values;

The historical data of S2, employing locality is trained BP neural network, to obtain suitable weights after the diffusion of information of step S1;

S3, according to the BP neural network after training, using the rainfall data that collects after the diffusion of information of step S1 as network input, after the computing of network each layer, obtain output layer neuronic output z ₁, i.e. rubble flow possibility occurrence size;

The threshold value that S4, basis preset, judges to export z ₁whether exceed threshold value, in this way, then start early warning.

2. the Debris Flow method for early warning under historical data shortage condition as claimed in claim 1, it is characterized in that, in step sl, described diffusion of information is on respective domain, adopt normal function to spread respectively each rainfall data.

3. the Debris Flow method for early warning under historical data shortage condition as claimed in claim 1, is characterized in that, in step s 2, describedly carries out training employing BP algorithm to BP neural network and trains.

4. the Debris Flow method for early warning under historical data shortage condition as claimed in claim 1, it is characterized in that, in step s 4 which, being set to of described threshold value: be set in advance as one close to 1 number, or the statistical value of false-alarm and false dismissal in using according to reality, be combined false-alarm and false dismissal loss coefficient that unit determines voluntarily, according to least disadvantage principle dynamically definite threshold.

5. the Debris Flow method for early warning under historical data shortage condition as claimed in claim 4, it is characterized in that, the described number close to 1 comprises 0.9.