Summary of the invention:
For the defect of prior art, the invention provides a kind of flexible measurement method based on multi-model prediction gas density and system, overcome the defect that single model Forecasting Methodology learning time is long, learn precision and extrapolability difference, improve the precision of prediction of forecast model.
On the one hand, the invention provides a kind of flexible measurement method based on multi-model prediction gas density, comprising:
To obtain under many group mines not gas density data in the same time, by under described many group mines not gas density data be in the same time stored to gas density historical data base;
Using described gas density historical data base as chaos time sequence, C-C method is adopted to calculate time delay and the Embedded dimensions of described chaos time sequence;
According to described time delay and Embedded dimensions, carry out phase space reconfiguration to described chaos time sequence, obtain the learning sample collection of gas density multi-model prediction soft-sensing model and desirable output, described learning sample collection comprises training data and test data;
Using described training data as input quantity, associated ideal exports, and builds gas density multi-model prediction soft-sensing model;
Using described test data as input quantity, according to described gas density multi-model prediction soft-sensing model, export gas density predicted vector.
Alternatively, gas density data are organized in described acquisition more, described many group gas density data are stored to gas density historical data base, comprise:
Firedamp sensor is adopted to obtain under many group mines not gas density data in the same time;
Mobile base station to receive under described many group mines not gas density data in the same time, by underground communication network by under described many group mines not gas density data in the same time transfer to ground monitoring system;
Described ground monitoring system by under described many group mines not gas density data be in the same time stored to gas density historical data base.
Alternatively, empirical mode decomposition method is adopted to carry out self-adaptive solution to the gas density data in described gas density historical data base, using the gas density data after denoising as chaos time sequence, C-C method is adopted to calculate time delay and the Embedded dimensions of described chaos time sequence.
Alternatively, adopt C-C method to calculate time delay and the Embedded dimensions of described chaos time sequence, be specially,
The delay time T of described chaos time sequence is
the t value that first minimal value is corresponding when meeting 0≤t≤200, described in
for the mean value between test statistics, calculated by following formula,
Wherein, S (m, r
k, t) be test statistics, calculated by following formula,
Wherein, r
kfor radius, C
s(m, r
k, be t) correlation integral of gas density sequence, calculated by following formula,
Wherein, δ (x) is impulse function,
X
i,x
jbe respectively input vector;
The embedding dimension m of described chaos time sequence is m=1+ τ
w/ τ;
Wherein, τ
wfor time window, be S
cor(t) t value that global minimum is corresponding when meeting 0≤t≤200, described S
cort () is the index amount of definition, calculated by following formula,
Wherein,
for the mean value of test statistics residual quantity, calculated by following formula,
Wherein, Δ S (m, t) selects the residual quantity between maximum test statistics and the test statistics of minimum two radiuses, is calculated by following formula,
ΔS(m,t)=max{S(m,r
j,t)}-min{S(m,r
j,t)}。
Alternatively, described gas density multi-model prediction soft-sensing model comprises P submodel, and each described submodel comprises C
iindividual sub-submodel.
Alternatively, described structure gas density multi-model prediction soft-sensing model, comprising:
According to subtractive clustering method, described training data is divided into P cluster centre, obtains P submodel, according to hazy condition clustering method, each described submodel is divided into C
iindividual sub-submodel;
According to the fuzzy membership relation of described training data and described submodel, to P submodel after dividing and C
iindividual sub-submodel is selected;
The submodel weighting chosen is exported, obtains described gas density multi-model prediction soft-sensing model.
Alternatively, described to P submodel after division and C
iindividual sub-submodel is selected, and is specially:
According to the fuzzy membership relation of described training data and described submodel, a sub-submodel is chosen from the submodel after each division, according to relative distance measure measuring method, from all sub-submodel chosen, degree of membership is selected to be greater than the sub-submodel of predetermined threshold value.
On the other hand, the invention provides a kind of hard measurement system based on multi-model prediction gas density, comprising:
Acquisition module, for obtaining under many group mines not gas density data in the same time, by under described many group mines not gas density data be in the same time stored to gas density historical data base;
Time delay and Embedded dimensions computing module, for using described gas density historical data base as chaos time sequence, adopt C-C method to calculate time delay and the Embedded dimensions of described chaos time sequence;
Learning sample collection acquisition module, for according to described time delay and Embedded dimensions, carry out phase space reconfiguration to described chaos time sequence, obtain the learning sample collection of gas density multi-model prediction soft-sensing model and desirable output, described learning sample collection comprises training data and test data;
Gas density multi-model prediction soft-sensing model builds module, for using described training data as input quantity, associated ideal exports, and builds gas density multi-model prediction soft-sensing model;
Prediction module, for using described test data as input quantity, according to described gas density multi-model prediction soft-sensing model, export gas density predicted vector.
Alternatively, described system, also comprises:
Pretreatment module, carries out self-adaptive solution for adopting empirical mode decomposition method to the gas density data in described gas density historical data base.
Alternatively, described gas density multi-model prediction soft-sensing model builds module, specifically for:
According to subtractive clustering method, described training data is divided into P cluster centre, obtains P submodel, according to hazy condition clustering method, each described submodel is divided into C
iindividual sub-submodel;
According to the fuzzy membership relation of described training data and described submodel, to P submodel after dividing and C
iindividual sub-submodel is selected;
The submodel weighting chosen is exported, obtains described gas density multi-model prediction soft-sensing model.
As shown from the above technical solution, flexible measurement method based on multi-model prediction gas density of the present invention and system, described method is by obtaining many group gas density data, described many group gas density data are stored to gas density historical data base, using described gas density historical data base as chaos time sequence, C-C method is adopted to calculate time delay and the Embedded dimensions of described chaos time sequence, and then adopt phase space reconfiguration to obtain the learning sample collection of gas density multi-model prediction soft-sensing model, using the input quantity of described learning sample collection as gas density multi-model prediction soft-sensing model, build gas density multi-model prediction soft-sensing model, export gas density predicted vector, thus, overcome single model Forecasting Methodology learning time long, the defect of study precision and extrapolability difference, improve the precision of prediction of forecast model.
Embodiment:
Below in conjunction with drawings and Examples, the specific embodiment of the present invention is described in further detail.Following examples for illustration of the present invention, but are not used for limiting the scope of the invention.
Fig. 1 shows the flexible measurement method schematic flow sheet based on multi-model prediction gas density that first embodiment of the invention provides, and as shown in Figure 1, the method for the present embodiment is as described below.
101, to obtain under many group mines not gas density data in the same time, by under described many group mines not gas density data be in the same time stored to gas density historical data base.
In this step, gather gas density data by gas concentration sensor, and the gas density data collected are stored to gas density historical data base X
lib=x (k) k=1,2 ..., l}, concrete steps comprise,
Firedamp sensor is adopted to obtain under many group mines not gas density data in the same time;
Mobile base station to receive under described many group mines not gas density data in the same time, by underground communication network by under described many group mines not gas density data in the same time transfer to ground monitoring system;
Described ground monitoring system by under described many group mines not gas density data be in the same time stored to gas density historical data base.
102, using described gas density historical data base as chaos time sequence, adopt C-C method to calculate time delay and the Embedded dimensions of described chaos time sequence.
In this step, described employing C-C method calculates time delay and the Embedded dimensions of described chaos time sequence, is specially,
The delay time T of described chaos time sequence is
the t value that first minimal value is corresponding when meeting 0≤t≤200, described in
for the mean value between test statistics, calculated by following formula,
Wherein, S (m, r
k, t) be test statistics, calculated by following formula,
Wherein, r
kfor radius, C
s(m, r
k, be t) correlation integral of gas density sequence, calculated by following formula,
Wherein, δ (x) is impulse function,
X
i,x
jbe respectively input vector;
The embedding dimension m of described chaos time sequence is m=1+ τ
w/ τ;
Wherein, τ
wfor time window, be S
cor(t) t value that global minimum is corresponding when meeting 0≤t≤200, described S
cort () is the index amount of definition, calculated by following formula,
Wherein, Δ S (m, t) selects the residual quantity between maximum test statistics and the test statistics of minimum two radiuses, is calculated by following formula,
Wherein, Δ S (m, t) selects the residual quantity between minimum and maximum two radius r test statistics, is calculated by following formula,
ΔS(m,t)=max{S(m,r
j,t)}-min{S(m,r
j,t)}。
103, according to described time delay and Embedded dimensions, carry out phase space reconfiguration to described chaos time sequence, obtain the learning sample collection of gas density multi-model prediction soft-sensing model and desirable output, described learning sample collection comprises training data and test data.
In this step, adopt phase space reconfiguration to obtain the learning sample collection of gas density multi-model prediction soft-sensing model, be specially: set up phase space reconfiguration X (k) respectively and walk the learning sample collection { X (k) of predicted vector Y (k) as gas density multi-model prediction soft-sensing model with p; Y (k) k=1+ (m-1) τ ..., n+1}, wherein, described gas density multi-model prediction soft-sensing model be input as X (k)=[x
1(k), x
2(k-τ) ..., x
m(k-(m-1) τ)], the output of described gas density multi-model prediction soft-sensing model is Y (k)=[x (k+p)]
.
104, using described training data as input quantity, associated ideal exports, and builds gas density multi-model prediction soft-sensing model.
In this step, it should be noted that described gas density multi-model prediction soft-sensing model comprises P submodel, each described submodel comprises C
iindividual sub-submodel.Described gas density multi-model is predicted the integration that the final output of soft-sensing model is exported by submodel and obtains, this model is when learning, for a certain learning sample, by selecting mechanism to select suitable sub-submodel to learn this sample from different submodels, thus each learning sample is made to have one or more sub-submodel to process this learning sample.
Particularly, described structure gas density multi-model prediction soft-sensing model, comprising:
According to subtractive clustering method, described training data is divided into P cluster centre, obtains P submodel, according to hazy condition clustering method, each described submodel is divided into C
iindividual sub-submodel;
According to the fuzzy membership relation of described training data and described submodel, to P submodel after dividing and C
iindividual sub-submodel is selected;
The submodel weighting chosen is exported, obtains described gas density multi-model prediction soft-sensing model.
Further, described to P submodel after division and C
iindividual sub-submodel is selected, and is specially:
According to the fuzzy membership relation of described training data and described submodel, a sub-submodel is chosen from the submodel after each division, according to relative distance measure measuring method, from all sub-submodel chosen, degree of membership is selected to be greater than the sub-submodel of predetermined threshold value.
The flexible measurement method based on multi-model prediction gas density of the present embodiment, by obtaining under many group mines not gas density data in the same time, be stored to gas density historical data base, using described gas density historical data base as chaos time sequence, C-C method is adopted to calculate time delay and the Embedded dimensions of described chaos time sequence, and then the learning sample collection adopting phase space reconfiguration to obtain gas density multi-model prediction soft-sensing model exports with desirable, using described training data as input quantity, associated ideal exports, build gas density multi-model prediction soft-sensing model, using described test data as input quantity, according to described gas density multi-model prediction soft-sensing model, export gas density predicted vector, thus, overcome single model Forecasting Methodology learning time long, the defect of study precision and extrapolability difference, improve the precision of prediction of forecast model.
Fig. 2 shows the flexible measurement method schematic flow sheet based on multi-model prediction gas density that second embodiment of the invention provides, and as shown in Figure 2, the method for the present embodiment is as described below.
201, gas density data are obtained, stored in gas density historical data base by firedamp sensor.
In this step, gas radio detection sensor is placed on it rib front, extractive equipment and operating personnel, truly to reflect the most truth with the forward position gas emission that works in the continuous motion process of excavator, and mobile base station reception gas density information is set at 50 ~ 100 meters, and by underground communication network, gas density is sent to the historical data base X of ground monitoring system outward
libin, there is X
lib=x (k) k=1,2 ..., l}.The gas density adopted in the present embodiment is subordinate to the raw data in database, as shown in table 1 below, totally 500 groups of data, and containing a large amount of noise signals in these data, as shown in Figure 3,
Table 1 gas density raw data set
0.4010 |
0.3725 |
0.3013 |
0.3717 |
0.2523 |
0.1966 |
0.1542 |
0.1412 |
0.1007 |
0.1002 |
0.1474 |
0.1478 |
0.1996 |
0.1679 |
0.1108 |
0.1703 |
0.2513 |
0.2207 |
0.2712 |
0.2510 |
0.2019 |
0.1663 |
0.2091 |
0.2998 |
0.2087 |
0.1865 |
0.1432 |
0.1373 |
0.1209 |
0.1112 |
0.2099 |
0.2467 |
0.2011 |
0.1632 |
0.2412 |
0.1533 |
0.1002 |
0.1542 |
0.1988 |
0.2699 |
0.1103 |
0.0832 |
0.1321 |
0.3743 |
0.4323 |
0.3565 |
0.3612 |
0.3003 |
0.3988 |
0.2421 |
… |
… |
… |
… |
… |
… |
… |
… |
… |
… |
0.1474 |
0.1478 |
0.1996 |
0.1679 |
0.1108 |
0.1703 |
0.2153 |
0.2207 |
0.2712 |
0.2510 |
0.1103 |
0.0832 |
0.1321 |
0.3743 |
0.4323 |
0.3565 |
0.3612 |
0.3003 |
0.3988 |
0.2421 |
0.1903 |
0.2533 |
0.1993 |
0.2005 |
0.2002 |
0.1995 |
0.2009 |
0.1662 |
0.1668 |
0.2199 |
0.1112 |
0.0833 |
0.1320 |
0.3741 |
0.4323 |
0.3565 |
0.3613 |
0.3003 |
0.3986 |
0.2420 |
0.4012 |
0.3728 |
0.3110 |
0.3712 |
0.2527 |
0.1963 |
0.1547 |
0.1416 |
0.1017 |
0.1012 |
0.2493 |
0.2366 |
0.2223 |
0.2497 |
0.2213 |
0.2318 |
0.2105 |
0.1997 |
0.202 |
0.2103 |
0.1903 |
0.2543 |
0.1993 |
0.2005 |
0.2032 |
0.1999 |
0.2109 |
0.1665 |
0.1678 |
0.2139 |
202, empirical mode decomposition method is adopted to carry out self-adaptive solution to the gas density data in described gas density historical data base.
In this step, carry out self-adaptive solution to the gas density raw data collected in step 201, concrete steps are as follows,
1) all local minizing points of signal x (k) are found out, coenvelope line and the lower envelope line of signal is formed by method of interpolation, as shown in Figure 4, the upper and lower envelope of original gas density data that Fig. 4 provides for second embodiment of the invention, calculates the mean value m of two envelopes
1, note h
1for x (k) and m
1residual quantity, i.e. h
1=x (k)-m
1, repeat said process, until h
1meet intrinsic mode functions (Intrinsic Mode Function is called for short IMF) definition.If the result of kth step is h
1k=h
1(k-1)-m
1k, note h
1k=c
1, then c
1first IMF decomposing out exactly from original gas density data;
2) by r
1=x (k)-c
1as data to be decomposed, repeat above-mentioned decomposable process, be constant when surplus or stop for during monotonic quantity, finally signal decomposition can be become a n IMF and residual components r
n, namely
as shown in Figure 5, Fig. 5 shows the original time of gas density data empirical mode decomposition and the corresponding relation figure of amplitude that second embodiment of the invention provides, and be divided into as can be seen from Figure 5 and solve 7 small scale intrinsic mode functions, last is residual components.
3) choose front k IMF component denoising, adopt adaptive threshold T
i=(T
i=(i-1)
2σ
i)/n
2, 1≤i≤k, wherein σ
iit is the standard deviation of i-th intrinsic mode function.
4) standard deviation of i-th signal adjacent signals is calculated
if
then adopt low-pass filter denoising.As shown in Figure 6, what Fig. 6 showed that second embodiment of the invention provides adopts the gas density historical data after low-pass filtering denoising to front 3 the small scale intrinsic mode functions after decomposing, and table 2 shows the gas density data set after denoising.
Gas density data set after table 2 denoising
0.1473 |
0.2498 |
0.3138 |
0.3452 |
0.3502 |
0.3348 |
0.3052 |
0.2672 |
0.2272 |
0.1911 |
0.1640 |
0.1472 |
0.1408 |
0.1446 |
0.1555 |
0.1702 |
0.1866 |
0.2026 |
0.2163 |
0.2261 |
0.2316 |
0.2327 |
0.2296 |
0.2232 |
0.2147 |
0.2054 |
0.1964 |
0.1891 |
0.1847 |
0.1844 |
0.1889 |
0.1958 |
0.2021 |
0.2036 |
0.1968 |
0.1828 |
0.1656 |
0.1503 |
0.1432 |
0.1507 |
… |
… |
… |
... |
… |
… |
… |
… |
… |
… |
0.1987 |
0.2443 |
0.2982 |
0.3472 |
0.3757 |
0.373 |
0.3523 |
0.3398 |
0.3258 |
0.2923 |
0.2515 |
0.2213 |
0.2083 |
0.2038 |
0.2009 |
0.1969 |
0.1912 |
0.1893 |
0.1955 |
0.2076 |
0.221 |
0.2321 |
0.238 |
0.2374 |
0.232 |
0.2232 |
0.2123 |
0.2022 |
0.197 |
0.2003 |
0.213 |
0.2242 |
0.2225 |
0.2127 |
0.2037 |
0.2014 |
0.1997 |
0.1887 |
0.1584 |
0.0989 |
203, the gas density data after denoising are processed as chaos time sequence, utilize C-C method to calculate time delay and the Embedded dimensions of described chaos time sequence.
In this step, according to Takens theorem, with suitable Embedded dimensions m and time delay τ, be dynamics equivalence with original system under " track " differomorphism meaning of reconstruction attractor in embedded space.Regard the gas density data after denoising as chaos time sequence, adopt C-C method to calculate the parameter of phase space reconfiguration: Embedded dimensions m and time delay τ.
204, the gas density time series data after phase space reconfiguration is utilized to obtain the learning sample collection of gas density multi-model prediction soft-sensing model and desirable output.
In this step, according to the Embedded dimensions m and the time delay τ that calculate gained in step 203, set up phase space reconfiguration X (k) respectively based on the gas density data sequence after denoising and walk the training sample set { X (k) of predicted vector Y (k) as gas density multi-model prediction soft-sensing model with p; Y (k) k=1+ (m-1) τ ..., n+1}.
It should be noted that, the Embedded dimensions calculated in the present embodiment is m=5, corresponding time delay is τ=2, namely described gas density multi-model predicts being input as [x (t) x (t-2) x (t-4) x (t-6) x (t-8)] of soft-sensing model, the gas concentration in described gas density multi-model prediction soft-sensing model prediction lower 2 moment, namely the output of described gas density multi-model prediction soft-sensing model is Y (k)=[x (t+2)]
For example, the learning sample collection of the prediction of the gas density multi-model after phase space reconfiguration soft-sensing model and ideal is adopted to export as shown in table 3 below, totally 489 groups of data.It should be noted that the present embodiment selects front 300 groups of data as the training data of gas density multi-model prediction soft-sensing model, rear 189 groups of data are as the test data of gas density multi-model prediction soft-sensing model.
Table 3 gas density multi-model prediction soft-sensing model learning sample data set and desirable output
205, according to training data and the desirable output of described gas density multi-model prediction soft-sensing model, gas density multi-model prediction soft-sensing model is built.
In this step, the training data of the gas density multi-model prediction soft-sensing model according to table 3 in step 204, build gas density multi-model prediction soft-sensing model, as shown in Figure 7, Fig. 7 shows the gas density multi-model prediction soft-sensing model structural drawing that second embodiment of the invention provides, concrete steps are as follows
1) submodel and sub-submodel is divided, specific as follows,
In option table 3 give front 300 groups of training datas of data to build gas density multi-model prediction soft-sensing model, first by the idea output y shown in table 3
kcarry out K-means cluster, if its cluster centre is V respectively
y, 1=y
min, V
y,p=y
max, make { V
y, 2 ...,v
y, p-1, then total P cluster centre.Goal set Y is built P fuzzy set.Can be divided into P submodel according to this result, expression formula is as follows,
Wherein, f
ikfor sample y
kto i-th (i=1 ..., P) degree of membership of individual fuzzy set;
For each objective fuzzy collection, the X of input is carried out fuzzy clustering, and expression formula is as follows,
Wherein, V
ijbe i-th (i=1 ..., P) individual objective fuzzy set pair answer jth (j=1 ..., C
i) cluster centre of individual input fuzzy set, C
ibe the input fuzzy clustering number that the i objective fuzzy set pair is answered, a, b are constant, and a, b value is generally 2,
be the subdivision matrix of the input fuzzy clustering that i-th objective fuzzy set pair is answered, meet:
Described training data can be divided into C by said method
tindividual sample set, expression formula is as follows,
Each sample set builds a sub-submodel, and (a corresponding P submodel, uses N to be divided into P group
eTirepresent, i=1 ..., P), each sub-submodel (uses N
eTijrepresent, j=1 ..., C
i) process the input of corresponding training data respectively.
2) selection of submodel and sub-submodel, specific as follows,
According to above-mentioned 1) described in step, directly there is certain membership in training data and submodel, produces close output principle according to close input, if input X
kdistance sample center V
ijclosely, then X
kbe under the jurisdiction of N
eTijpossibility large, adopt relative distance measure measuring method to X
kbe under the jurisdiction of N
eTijpossibility calculate, expression formula is as follows,
Wherein, J
ifor choosing the performance index function of sub-submodel from submodel, w
ijfor X
kto N
eTijdegree of membership,
d
ijfor inputting right relative distance measure, d
ij=|| X
k-V
ij||/da
ij, da
ijfor N
eTijthe evaluation distance of training data,
n
ijfor N
eTijsample number.
Make above-mentioned performance index function J
iminimize, can w be obtained by Lagrangian multiplier method
ij, expression formula is as follows,
Wherein, i=1 ..., P, j=1 ..., C
i, obviously can find out, d
ijlarger, w
ijless, X
kbe under the jurisdiction of N
eTijpossibility less; Otherwise, d
ijless, w
ijlarger, X
kbe under the jurisdiction of N
eTijpossibility larger.Choose sub-submodel according to maximum membership degree rule, selected sub-submodel is designated as N
eTis, make w
is=1, w
ij, j ≠ s=0, like this, a sub-submodel can be selected in each submodel for the treatment of X
k, the output of each submodel is N
eTisoutput.But this selection is a kind of primary election, not all N
eTis(i=1 ..., P) be all applicable to process X
k,but selected sub-submodel screens, and screening technique is as follows,
Screen sub-submodel performance index function to from the sub-submodel selected, expression formula is as follows,
Wherein, w
ifor X
kto N
eTidegree of membership,
d
ifor inputting right relative distance measure, d
i=|| X
k-V
i||/da
i, da
ifor N
eTithe evaluation distance of training data,
n
ifor N
eTisample number.
Make above-mentioned performance index function J minimize, can w be obtained by Lagrangian multiplier method
i, expression formula is as follows,
It should be noted that owing to only having a N in each submodel
eTisprocess X
k, w
iin fact be also X
kto N
eTisdegree of membership.Consider the overlapping features of sample, a degree of membership threshold k is set, meets w
ithe submodel of>=K is to X
kprocess, by above-mentioned system of selection, for given training sample X
k, with X
kthe difference of distributing position and the difference of K value, participate in the sub-network having quantity not wait X
kprocess.
3) integration of sub-submodel output, specific as follows,
For training data X
k,make w={w
1..., w
p, if w
i< K, then make w
i=0 (i=1 ..., P), be then normalized w, if total output of gas density multi-model prediction soft-sensing model is Y, expression formula is as follows,
Wherein, y
ifor N
eTito X
koutput, w
ifor N
eTiweights (i-th component after w normalized), for there is no selected submodel, its weight w
i=0, namely not selected submodel exports without contribution to total, total weighted sum exported as selected submodel.
206, using test data as input quantity, according to described gas density multi-model prediction soft-sensing model, carry out prediction gas density.
In this step, for example, the gas density multi-model constructed by step 205 is predicted that the optimum configurations of soft-sensing model is P=4, C={3,4,4,3}, namely whole gas density multi-model prediction soft-sensing model has 4 groups of submodels, and often organizing submodel has 3,4 respectively, 4,3 sub-submodels, in each sub-submodel, hidden nodes is 8, in step 205 2) described in sub-submodel Selection parameter K=0.1.Fig. 8 shows predicted value and the actual comparison design sketch of the gas density multi-model prediction soft-sensing model that second embodiment of the invention provides, and table 4 is the gas density multi-model prediction predicted value of soft-sensing model and the contrast of actual value.
The table 4 gas density multi-model prediction predicted value of soft-sensing model and the contrast of actual value
Pass through said method, the present invention can realize Mine Methane accurately predicting, meet the technical requirement of current mine gas administration system, compared with the gas density Forecasting Methodology existed at present, on the one hand, the present invention is directed to the problem usually containing much noise impact prediction model prediction accuracy in current mine gas density data, by empirical mode decomposition, history gas density sequence data is resolved into multiple small scale intrinsic mode function, again by decompose after small scale intrinsic mode function through low-pass filter Adaptive Wavelet Thrinkage, and then reduce learning sample to the impact of precision of forecasting model.On the other hand, the present invention is directed to the defect of single model forecast model learning time length, precision of prediction and extrapolability difference, build the multi-model forecast model that has hierarchical structure, this mode input information is by the integrated process of multiple different submodels, the precision of forecast model can be improved, improve the robustness of forecast model.
Fig. 9 shows the hard measurement system architecture schematic diagram based on multi-model prediction gas density that third embodiment of the invention provides, as shown in Figure 9, the hard measurement system based on multi-model prediction gas density of the present embodiment comprises: acquisition module 91, time delay and Embedded dimensions computing module 92, learning sample collection acquisition module 93, gas density multi-model prediction soft-sensing model build module 94, prediction module 95;
Described acquisition module 91 for obtaining under many group mines not gas density data in the same time, by under described many group mines not gas density data be in the same time stored to gas density historical data base;
Described time delay and Embedded dimensions computing module 92 for using described gas density historical data base as chaos time sequence, adopt C-C method to calculate time delay and the Embedded dimensions of described chaos time sequence;
Described learning sample collection acquisition module 93 is for according to described time delay and Embedded dimensions, phase space reconfiguration is carried out to described chaos time sequence, obtain the learning sample collection of gas density multi-model prediction soft-sensing model and desirable output, described learning sample collection comprises training data and test data;
Gas density multi-model prediction soft-sensing model build module 94 for using described training data as input quantity, associated ideal exports, and builds gas density multi-model prediction soft-sensing model;
Described prediction module 95 for using described test data as input quantity, according to described gas density multi-model prediction soft-sensing model, export gas density predicted vector.
Further, the described hard measurement system based on multi-model prediction gas density, also comprises:
Pretreatment module, carries out self-adaptive solution for adopting empirical mode decomposition method to the gas density data in described gas density historical data base.
In concrete implementation procedure, aforesaid gas density multi-model prediction soft-sensing model builds module 94 in the process realizing building gas density multi-model prediction soft-sensing model, specifically comprises:
According to subtractive clustering method, described training data is divided into P cluster centre, obtains P submodel, according to hazy condition clustering method, each described submodel is divided into C
iindividual sub-submodel;
According to the fuzzy membership relation of described training data and described submodel, to P submodel after dividing and C
iindividual sub-submodel is selected;
The submodel weighting chosen is exported, obtains described gas density multi-model prediction soft-sensing model.
The hard measurement system based on multi-model prediction gas density that the present embodiment provides, by the gas density gathered in mine historical data is considered as chaos time sequence, first utilize empirical mode decomposition that analyzed history gas density time series data is resolved into multiple small scale intrinsic mode function, again by decompose after small scale intrinsic mode function through low-pass filter Adaptive Wavelet Thrinkage, and phase space reconfiguration is carried out to the sequence data after denoising, thus obtain the learning sample collection of multi-model soft-sensing model, secondly for the gas density data be collected in the mine of zones of different, the invention provides a kind of multi-model forecast model with hierarchical structure, this soft-sensing model can not only process by the submodel different to the gas density data selection of zones of different, and each gas density data are by model multiple sub-submodel associated treatment, improve the precision of prediction of model, decision support is provided for realizing Safety of Coal Mine Production,
Last it is noted that above each embodiment is only in order to illustrate technical scheme of the present invention, be not intended to limit; Although with reference to foregoing embodiments to invention has been detailed description, those of ordinary skill in the art is to be understood that: it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein some or all of technical characteristic; And these amendments or replacement, do not make the essence of appropriate technical solution depart from the scope of the claims in the present invention.