CN102546492B - Information channel environmental detection method based on support vector machine - Google Patents

Abstract

An information channel environmental detection method based on a support vector machine (SVM) belongs to the technical field of wireless information channels and is characterized in that signal estimated value variable quantity and step amplitude between data blocks are adopted as two attribute values of an information channel, and a label value indicating whether the information channel environment changes serve as an initial assignment of the support vector machine. A penalty parameter c and a function parameter g are determined in a K-CV method, and a radial basis kernel function serves as the parameter of the support vector machine to build a support vector machine model and detect whether the information channel environment changes. The information channel environmental detection method can ensure correctness of information channel environmental detection in conditions that the information channel environment is totally unknown, and is convenient to popularize and use.

Description

A kind of information channel environmental detection method based on SVMs

Technical field

The present invention relates to a kind of information channel environmental detection method using SVMs (Support vector machines).

Background technology

1, wireless channel environment detection background technology

In wireless communication system, the Doppler frequency shift that multipath transmisstion and transmitting-receiving two-end motion due to radio magnetic wave are formed, wireless channel is provided with frequency selectivity and selects characteristic with the two of time selectivity, makes signal experienced by comparatively complicated decline.And in the correct transmission information that demodulates of receiving terminal, then to must carry out channel estimating accurately.

The channel estimation methods that oneself has can be broadly dassified into: by means of the estimation of reference signal (training sequence, pilot tone), blind Channel Estimation, half-blind channel estimating method.According to the characteristic of channel, the basis expansion model by means of reference signal is widely used in the modeling that doubly selective channel is estimated.Different according to the basic function used, its performance then there will be larger difference in different environments, such as, discrete K-L basis expansion model (Karhuen-Loeve BEM, K-L-BEM) although be mean square error optimum, the not match condition of its performance to channel statistic is very responsive; Complex exponential basis expansion model (complex exponential BEM, CE-BEM) independent of channel statistic, but has certain modeling error; Polynomial basis extended model (polynomial BEM, P-BEM) is then comparatively responsive to Doppler frequency shift, usually in low Doppler frequency shift situation, has better performance.

And for actual wireless communication environment complicated and changeable, use above-mentioned single model to carry out channel estimating, then cannot obtain long-time, high-quality estimated performance; And channel estimating is carried out for multi-model, although take into account various wireless communication environment, but also need to judge whether present channel environment there occurs change in time, accurately, and reasonably carry out channel model switching, high-quality estimated performance can be obtained like this.

Existing channel circumstance change is detected and is mainly judged by signal envelope amplitude, if Received signal strength envelope range value is undergone mutation, detects that channel circumstance there occurs change.The main method adopted has: before and after envelope range value change moment and saltus step sampled point under range value known case, the hypothesis testing of available NP (Neyman-Pearson) method, for change moment known but range value is unknown or the change moment is unknown but range value is known and both all unknown situations, then adopt GLRT (Generalized Likelihood Ratio Test) method hypothesis testing.

2, SVMs background technology

SVMs (SVM, Support Vector Machine) is first proposed by Vapnik, can be used for pattern classification and nonlinear regression.The main thought of SVMs sets up an Optimal Separating Hyperplane as decision-making curved surface, and the isolation edge between positive example and counter-example is maximized.The theoretical foundation of SVMs is Statistical Learning Theory, is the approximate realization of structural risk minimization.This principle depend on training error rate and one based on: the error rate of Learning machine in test data (i.e. extensive error rate) item of VC dimension (Vapnik-Chervonenkis dimension) and be boundary, can in merotype situation, SVMs is zero for the value of last item, and Section 2 is minimized.Therefore, although the field internal problem of its not Utilizing question, the Generalization Capability that support vector function provides on pattern classification problem, this attribute is that SVMs is distinctive.

Summary of the invention

The present invention is directed to the complexity of channel circumstance test problems in wireless fading channel, when channel circumstance the unknown, a kind of new training method is proposed, utilize the distribution of the pilot frequency sequence in data block and the feature of channel model algorithm, and the description to channel circumstance, choose wherein effective feature interpretation value, as the property value of SVMs, set up corresponding supporting vector machine model, due to the predictability of supporting vector machine model self, complete the estimation to channel.Through continuing to optimize and improving algorithm, the accuracy detecting channel circumstance change can be ensured, be convenient to actual use.

Due to the operating characteristic of SVMs, be need first to train known sample data set, and set up out supporting vector machine model, real data collection according to set up model, could complete the classification of real data, judgement work.

When carrying out channel estimating, need clear and definite used training set, penalty parameter c sum functions parameter g is reasonably configured, finally complete model training work, when using real data to carry out channel circumstance detection, only by related data by supporting vector machine model, need can obtain channel circumstance testing result.By newly-designed training set, the present invention can judge whether channel circumstance there occurs change accurately, the switching of multi-model provides effective foundation.

General principle flow chart of the present invention is shown in Fig. 1.

Use an information channel environmental detection method for support vector machines, realize according to the following steps successively,

Step (1), processes training set data according to the following steps, determines supporting vector machine model and suitable penalty parameter c sum functions parameter g:

Transmission convert information is that data block step is as follows by step (1.1):

Transmission information u, be converted into data block u _i, i represents training set data block number, i=1,2 ..., I, each data block comprises two parts, information symbol s _gwith frequency pilot sign p _g, in a data block, evenly insert G frequency pilot sign p _g, g=1,2 ..., G, G are Finite Number, data block u _irepresent with following formula:

$u_i={[p_0^T,s_0^T,p_1^T,s_1^T,...,p_{G-1}^T,s_{G-1}^T]}^T,$

Step (1.2), the channel model step used under utilizing described complex exponential base expansion algorithm, polynomial basis function expansion algorithm and switching channels algorithm to determine present channel environment is as follows:

Step (1.2.1), utilizes described complex exponential base expansion module to calculate channel estimation value following formula and represents:

$h_i^{\mathrm{CE}-\mathrm{BEM}}\left(n\right)=\underset{q=0}{\overset{Q}{\mathrm{\Σ}}}{h}_{i,q}{e}^{{\mathrm{i\ω}}_{q}n},$

Wherein, ω _q=2 π (q-Q/2)/MN, M=2 is the expansion of over-sampling complex exponential base, the item number of Q representative polynomial basic function, q=1,2 ..., the item number sequence number of Q, q representative polynomial basic function, f _maxfor maximum doppler frequency, T _sfor sending the sampling interval of sequence and receiving sequence, i=1,2 ..., I, n=1,2 ..., N represents the position of signal in data block;

Step (1.2.2), utilizes described polynomial basis function expansion module to calculate Signal estimation value following formula and represents:

$h_i^{P-\mathrm{BEM}}\left(n\right)=\underset{q=0}{\overset{Q}{\mathrm{\Σ}}}{h}_{i,q}{(n-n_0)}^q,$

Wherein, i=1,2 ..., I, n=1,2 ..., N represents the position of signal in data block, n ₀represent the point midway of data block, for basis expansion model parameter, d represents differentiate;

Step (1.2.3), utilizes described switching channels algorithm, calculates all channel model algorithm Ω={ Μ _y| y=1,2 ..., hopping amplitude sum E between multiple continuous print data blocks of Y}, and determine with this model algorithm y that engages in this profession ^*for the preferred channels model algorithm under present channel environment, represent with following formula:

$y^{*}=\arg \underset{y}{\min }\left(E_y\right),$

Wherein, y=1,2 ..., Y, Y represent channel model number, i is the number of got multiple consecutive data block, I ∈ [3,5], I<I, h _i(1) the Signal estimation value of first symbol of i-th data block is represented, h _i-1(N) the Signal estimation value of last symbol of the i-th-1 data block is represented;

Step (1.3), determine the label value in training set, property value according to the following steps: hopping amplitude between Signal estimation value variable quantity and data block:

Whether step (1.3.1), there occurs during by judging that channel circumstance and last time are detected and change as training set label value, and described training set label value is divided into channel circumstance to change and channel circumstance does not change;

Step (1.3.2), by Signal estimation value variable quantity algorithm, calculates the Signal estimation value variable quantity r in data block when channel circumstance does not change _irepresent:

r _i＝[h _i(2)-h _i(1),…h _i(n+1)-h _i(n),…h _i(N)-h _i(N-1)] ^T，

Wherein, n=1,2 ..., N-1;

Step (1.3.3), to the Signal estimation value variable quantity r that step (1.3.2) exports _i, ask for all Signal estimation value variable quantity r in described i-th data block _imean value r _icp, and utilize following formula to calculate Signal estimation value variable quantity in the data block of channel circumstance when changing

${\tilde{r}}_i=r_{\mathrm{i\; cp}}\&PlusMinus;\frac{\mathrm{Max}(h_i(n+1)-h_i\left(n\right))-\mathrm{Min}(h_i(n+1)-h_i{\left(n\right)}_i)}{{\mathrm{\&beta;}}_1},$

Wherein, $\left|\frac{\mathrm{Max}\left(h_i(n+1)h_i\left(n\right)\right)-\mathrm{Min}(h_i(n+1)-h_i{\left(n\right)}_i)}{{\mathrm{\&beta;}}_1}\right|>0$ Time, represent that channel circumstance there occurs change, $\left|\frac{\mathrm{Max}(h_i(n+1)-h_i\left(n\right))-\mathrm{Min}(h_i(n+1)-h_i{\left(n\right)}_i)}{{\mathrm{\&beta;}}_1}\right|=0$ Time, represent that channel circumstance does not change, β ₁∈ [1,20] is empirical coefficient, expresses support for the perception sensitivity level that vector machine changes channel circumstance; Signal estimation value variable quantity r when the channel circumstance exported from step (1.3.2) does not change in data block _isignal estimation value variable quantity in data block when changing with the channel circumstance exported from step (1.3.3) be referred to as Signal estimation value variable quantity;

Step (1.3.4), by hopping amplitude algorithm between described data block, hopping amplitude e between data block when calculating channel circumstance does not change _irepresent:

e _i＝h _i(1)-h _i-1(N)，

Wherein, i=2,3 ..., I;

Step (1.3.5), hopping amplitude e between the data block utilizing step (1.3.4) to export _i, ask for hopping amplitude e between whole described data block _imean value e _icp, and utilize following formula to calculate Signal estimation value variable quantity in the data block of channel circumstance when changing

${\tilde{e}}_i=e_{\mathrm{i\; cp}}\&PlusMinus;\frac{\mathrm{Max}(h_i\left(1\right)-h_{i-1}\left(N\right))-\mathrm{Min}(h_i\left(1\right)-h_{i-1}\left(N\right))}{{\mathrm{\&beta;}}_2},$

Wherein, $\left|\frac{\mathrm{Max}(h_i\left(1\right)-h_{i-1}\left(N\right))-\mathrm{Min}(h_i\left(1\right)-h_{i-1}\left(N\right))}{{\mathrm{\&beta;}}_2}\right|>0$ Time, represent that channel circumstance there occurs change, $\left|\frac{\mathrm{Max}(h_i\left(1\right)-h_{i-1}\left(N\right))-\mathrm{Min}(h_i\left(1\right)-h_{i-1}\left(N\right))}{{\mathrm{\&beta;}}_2}\right|=0$ Time, represent that channel circumstance does not change, β ₂∈ [1,20] is empirical coefficient, expresses support for the perception sensitivity level that vector machine changes channel circumstance; Hopping amplitude e between data block when the channel circumstance exported from step (1.3.4) does not change _iand hopping amplitude between data block when changing from the channel circumstance that step (1.3.5) exports be referred to as hopping amplitude between data block;

Step (1.3.6), judges channel circumstance according to the following steps:

Step (1.3.2) and step (1.3.4) is utilized to export the Signal estimation value variable quantity r obtained _iand hopping amplitude e between data block _i, as training set property value when judging that channel circumstance does not change; Step (1.3.3) and step (1.3.5) is utilized to export the Signal estimation value variable quantity obtained and hopping amplitude between data block as training set property value when judging that channel circumstance changes;

Step (1.4), carry out serioparallel exchange:

Hopping amplitude and training set label value between the Signal estimation value variable quantity described in serial input, data block, carry out serioparallel exchange, parallel output is to buffer memory;

Step (1.5), set up buffer memory according to the following steps:

Using the Signal estimation value variable quantity that obtains after described serioparallel exchange first property value as training set, between data block, hopping amplitude is as second property value of training set, with training set label value, is stored into buffer memory;

Step (1.6), determine supporting vector machine model and suitable penalty parameter c sum functions parameter g step as follows:

Step (1.6.1), sets up described supporting vector machine model from described buffer memory, through the Signal estimation value variable quantity of serioparallel exchange input as the property value of training set, and hopping amplitude and training set label value between data block; What export from configuration parameter is less than most ceiling value, and be greater than the penalty parameter c sum functions parameter g of minimum limit value, and described and be SVMs kernel function with RBF, then, set up the described supporting vector machine model under described two property values and a label value;

Step (2), uses supporting vector machine model to detect channel circumstance:

Step (2.1), utilizes cross validation K-CV method to determine and exports best penalty parameter c sum functions parameter g;

Step (2.2), utilizes step (1.3.2) to calculate the Signal estimation value variable quantity of current data block; Step (1.3.2) is utilized to calculate the hopping amplitude of current data block and a upper data interblock;

Step (2.3), the Signal estimation value variable quantity of input current data block, the hopping amplitude of current data block and a upper data interblock, best penalty parameter c sum functions parameter g, utilize supporting vector machine model, export the channel circumstance testing result obtaining current data block;

Step (2.4), judges channel circumstance monitoring result according to the following steps:

Step (2.4.1), judges whether current data block is last data block:

If not last data block, then proceed to step (2.4.2);

If last data block, then terminate;

Step (2.4.2), judges the judgement whether present channel environment changes:

If channel circumstance changes, proceed to step (1.2);

If channel circumstance does not change, proceed to step (2.3).

The present invention is by choosing relevant data value, obtains the effective property value setting up supporting vector machine model; And by regulating detection method for the sensitivity of environmental change, judge whether channel circumstance there occurs change accurately.The present invention when channel circumstance is completely unknown, can ensure the accuracy that channel circumstance detects.

Accompanying drawing explanation

Fig. 1, adjacent two moment SVMs channel estimating general principle flow processs.

Fig. 2, channel model algorithm schematic diagram when mobile station speed is 50km/h, in figure---actual channel value,------channel estimation value.

Fig. 3, channel model algorithm schematic diagram when mobile station speed is 100km/h, in figure---actual channel value,------channel estimation value.

Fig. 4, channel model algorithm schematic diagram when mobile station speed is 200km/h, in figure---actual channel value,------channel estimation value.

Fig. 5, diagram data interblock hopping amplitude schematic diagram when mobile station speed is 50km/h, in figure---and the channel of non-plus noise,-----channel after plus noise ... polynomial basis function extended model algorithm ,-.-.-.-complex exponential basis expansion model algorithm.

Fig. 6, diagram data interblock hopping amplitude schematic diagram when mobile station speed is 1800km/h, in figure---and the channel of non-plus noise,-----channel after plus noise ... polynomial basis function extended model algorithm ,-.-.-.-complex exponential basis expansion model algorithm.

Fig. 7, each channel model algorithm performance figure when travelling carriage does accelerated motion, in figure polynomial basis function extended model algorithm, complex exponential basis expansion model algorithm, based on the Multiple Models Algorithm that SVMs channel circumstance detects.

Fig. 8, travelling carriage is each channel model algorithm performance figure when plus-minus moves, in figure polynomial basis function extended model algorithm, complex exponential basis expansion model algorithm, based on the Multiple Models Algorithm that SVMs channel circumstance detects.

Embodiment

Information channel environmental detection method based on SVMs of the present invention mainly comprises the following step:

Based on SVMs channel circumstance monitoring for travelling carriage do respectively constant speed, acceleration, retarded motion judged result in table 1.

Table 1 based on the channel circumstance monitoring of SVMs for the judged result of travelling carriage different motion state

Ruili channel that the initial velocity that meets travelling carriage and final speed, signal to noise ratio are 20dB is generated respectively, by the transmission data block of convert information and the Received signal strength through channel according to table 1,100 data blocks are generated respectively at initial velocity and final speed, be inserted with 10 frequency pilot signs in each data block respectively, data block length is 200.

For often kind of different initial velocity and final speed situation, following steps circulate 16 times respectively.

Step one: determine the channel model that present channel environment uses

Complex exponential basis expansion model algorithm and polynomial basis function extended model algorithm is utilized to try to achieve the Signal estimation value of all channel models of front 4 data blocks respectively.Obtain Signal estimation value by complex exponential basis expansion model algorithm can be represented by the formula:

$h_i^{\mathrm{CE}-\mathrm{BEM}}\left(n\right)=\underset{q=0}{\overset{Q}{\mathrm{\&Sigma;}}}{h}_{i,q}{e}^{{\mathrm{i\&omega;}}_{q}n},$

In formula, ω _q=2 π (q-Q/2)/MN, M=2 is the expansion of over-sampling complex exponential base, the item number of Q representative polynomial basic function, q=1,2 ..., the item number sequence number of Q, q representative polynomial basic function, f _maxfor maximum doppler frequency, T _sfor sending the sampling interval of sequence and receiving sequence, i=1,2 ..., I, n=1,2 ..., N represents the position of signal in data block.

Obtain Signal estimation value by polynomial basis function extended model algorithm can be represented by the formula:

$h_i^{P-\mathrm{BEM}}\left(n\right)=\underset{q=0}{\overset{Q}{\mathrm{\&Sigma;}}}{h}_{i,q}{(n-n_0)}^q,$

In formula, i=1,2 ..., I, n=1,2 ..., N represents the position of signal in data block, n ₀represent the point midway of data block, for basis expansion model parameter, d represents differentiate.

Do not considering under the impact that noise causes, according to the distribution of the pilot frequency sequence in data block and the characteristic discover of channel model algorithm, the Signal estimation value at last symbol place of each data block and information symbol place the most easily and real signal value there is maximum error amount; And the Signal estimation value at first of each data block symbol place and frequency pilot sign place and actual signal are worth error minimum.For continuous print two data blocks, think that the variable quantity between last symbol and first symbol of second data block of first data block is negligible, because these two symbol places have employed twice channel model algorithm respectively, according to the description to channel model algorithm characteristics, find that between data block, channel model can exist saltus step, the Signal estimation value at these two symbol places then has larger difference, i.e. hopping amplitude between data block.Between data block, hopping amplitude is less, then illustrate that this channel model algorithm more can be followed the tracks of channel accurately.Calculate all channel model algorithm Ω={ Μ _y| y=1,2 ..., hopping amplitude sum between multiple continuous print data blocks of Y}, and determine channel model algorithm y with this ^*for the preferred channels model algorithm under present channel environment, represent with following formula:

$y^{*}=\arg \underset{y}{\min }\left(e_y\right),$

In formula, y=1,2 ..., Y, Y represent channel model number, h _i(1) the Signal estimation value of first symbol of i-th data block is represented, h _i-1(N) the Signal estimation value of last symbol of the i-th-1 data block is represented.

Step 2: determine training set

Training set is the necessary associated data set set up to set up supporting vector machine model, and comprising property value and label value, property value represents the characteristic value of things itself, and label value represents the classification between different things.

Due to the randomness of the channel circumstance such as speed, noise change, inconvenience sets up label value according to the factor such as speed, signal to noise ratio, therefore can according to current detected channel circumstance result, whether there occurs when judging that channel circumstance and last time are detected and change as label value, be namely divided into channel circumstance changed and channel circumstance do not change.Label value needs corresponding with corresponding property value, sets up supporting vector machine model so that follow-up.

Owing to changing the speed of travelling carriage, can have influence on Doppler frequency shift, channel variation also can be compressed or be stretched, and the speed of travelling carriage is larger, and channel variation is faster; Converted quantity between the Signal estimation value in each moment also can increase thereupon.Use suitable Signal estimation algorithm in friction speed section, the converted quantity between the Signal estimation value obtained as shown in Figure 2, Figure 3, Figure 4.The difference of the Signal estimation value in each moment can as the property value weighed channel circumstance and whether change, because channel circumstance can not change in the short time, therefore can think that channel circumstance is consistent in front 10 data blocks, Signal estimation value variable quantity r when channel circumstance does not change in data block _irepresent:

r _i＝[h _i(2)-h _i(1),…h _i(n+1)-h _i(n),…h _i(200)-h _i(199)] ^T，

Wherein, n=1,2 ..., 199, i=2,3 ..., 10.

In order to meet set up training set label value, the change amount signal in data block when not only needing channel circumstance constant, training set property value when also needing channel circumstance to change.Can constant according to channel circumstance time data block in all Signal estimation value variable quantity r _imean value r _icp, and peak signal estimated value variable quantity Max (h _i(n+1)-h _i(n)) and minimum signal estimated value variable quantity Min (h _i(n+1)-h _i(n) _i) the data block of difference when determining that channel circumstance changes in Signal estimation value variable quantity

${\tilde{r}}_i=r_{\mathrm{i\; cp}}\&PlusMinus;\frac{\mathrm{Max}(h_i(n+1)-h_i\left(n\right))-\mathrm{Min}(h_i(n+1)-h_i{\left(n\right)}_i)}{{\mathrm{\&beta;}}_1},$

Wherein, $\left|\frac{\mathrm{Max}\left(h_i(n+1)h_i\left(n\right)\right)-\mathrm{Min}(h_i(n+1)-h_i{\left(n\right)}_i)}{{\mathrm{\&beta;}}_1}\right|>0$ Time, represent that channel circumstance there occurs change, $\left|\frac{\mathrm{Max}(h_i(n+1)-h_i\left(n\right))-\mathrm{Min}(h_i(n+1)-h_i{\left(n\right)}_i)}{{\mathrm{\&beta;}}_1}\right|=0$ Time, represent that channel circumstance does not change, β ₁=8 is empirical coefficient, and numerical value is larger, expresses support for the perception that vector machine changes channel circumstance sensitiveer, too high β ₁learning state can be caused to occur.

When channel circumstance keeps relative stability and employs the channel model algorithm being comparatively suitable for present channel environment, the hopping amplitude between each data block can be stabilized in certain difference range substantially.When channel circumstance there occurs change, the channel model algorithm before used certainly will be affected at aspect of performance, no matter make the more accurate of its channel tracking, or larger error can be caused to Signal estimation value, channel model hopping amplitude between each data block all can change, and between data block at various speeds, hopping amplitude as shown in Figure 5, Figure 6.Second the property value whether hopping amplitude thus between each data block can change as measurement channel circumstance.Hopping amplitude e between data block when channel circumstance does not change _irepresent:

e _i＝h _i(1)-h _i-1(200)，

Wherein, i=2,3 ..., 10.

In order to meet set up training set label value, hopping amplitude between data block when not only needing channel circumstance constant, training set property value when also needing channel circumstance to change.

Can constant according to channel circumstance time ask for the hopping amplitude e of all described 10 data interblocks _imean value e _icp, and utilize maximum hopping amplitude Max (h between data block _i(1)-h _i-1(200) minimum transition amplitude Min (h) and between data block _i(1)-h _i-1(200) hopping amplitude between data block when difference) determines that channel circumstance changes

${\tilde{e}}_i=e_{\mathrm{i\; cp}}\&PlusMinus;\frac{\mathrm{Max}(h_i\left(1\right)-h_{i-1}\left(200\right))-\mathrm{Min}(h_i\left(1\right)-h_{i-1}\left(200\right))}{{\mathrm{\&beta;}}_2},$

Wherein, $\left|\frac{\mathrm{Max}(h_i\left(1\right)-h_{i-1}\left(N\right))-\mathrm{Min}(h_i\left(1\right)-h_{i-1}\left(N\right))}{{\mathrm{\&beta;}}_2}\right|>0$ Time, represent that channel circumstance there occurs change, $\left|\frac{\mathrm{Max}(h_i\left(1\right)-h_{i-1}\left(N\right))-\mathrm{Min}(h_i\left(1\right)-h_{i-1}\left(N\right))}{{\mathrm{\&beta;}}_2}\right|=0$ Time, represent that channel circumstance does not change, β ₂=8 is empirical coefficient, and be empirical coefficient, numerical value is larger, expresses support for the perception that vector machine changes channel circumstance sensitiveer, too high β ₁learning state can be caused to occur.

By the Signal estimation value variable quantity r obtained _iand hopping amplitude e between data block _i, be judged as training set property value when channel circumstance does not change, and carry out corresponding with corresponding label value; The Signal estimation value variable quantity obtained and hopping amplitude between data block be judged as training set property value when channel circumstance changes, and carry out corresponding with corresponding label value.

Step 3: determine supporting vector machine model

The present invention utilizes the forecast function of SVMs to complete detection to channel circumstance, and forecast function is that under being based upon prerequisite that supporting vector machine model determined, therefore according to training set, the foundation completing supporting vector machine model is very important.

For the foundation of model, not only need the property value of training set and the label value of training set to set up corresponding relation, also need a large amount of training set data as the foundation of training pattern, training set data amount is larger, model will be all the more accurate, but training burden also can rise thereupon.

Take RBF as SVMs kernel function, the penalty parameter c sum functions parameter g of reasonable disposition, too high c can cause learning state to occur, the accuracy rate of the classification namely during the very high and practical application of training set classification accuracy is very low (generalization ability of grader reduces), so choose suitable c and corresponding g, and according to the channel circumstance of front 10 data blocks, by the different conditions of training data block channel circumstance, complete the foundation of supporting vector machine model.

Step 4: use the monitoring of supporting vector machine model channel circumstance

The method of cross validation (CV, Cross Validation) can be used, find parameter c best under definite meaning and g, effectively can avoid the generation of study and deficient learning state.CV is used to a kind of statistical analysis technique verifying classifier performance, and basic thought training set data is classified under certain meaning, and a part is as training set, and another part is as checking collection.Its method is first trained grader with training set, and recycling checking collection tests the model of training and obtaining, with the performance index of the classification accuracy obtained as classification of assessment device.

The present invention is the penalty parameter c sum functions parameter g using K-fold Cross Validation (K-CV) method to determine.Training set data is divided into 3 groups, each subset data is made one-time authentication collection respectively, remaining 2 groups of subset data is as training set simultaneously, 3 models can be obtained like this, then c and g grid division is calculated respectively, by the average of the classification accuracy of the final checking collection of these 3 models performance index as grader under this K – CV.

For the use of supporting vector machine model, also need to input the current Signal estimation value variable quantity r of data block and the hopping amplitude e between current data block and previous data block, wherein Signal estimation value variable quantity r following formula represents:

r＝[h ₁₁(2)-h ₁₁(1),…h ₁₁(n+1)-h ₁₁(n),…h ₁₁(200)-h ₁₁(199)] ^T，

Between data block, hopping amplitude e following formula represents:

e＝h ₁₁(1)-h ₁₀(200)，

For the channel circumstance drawn through SVMs, also need to judge whether current data block is last data block, if not last data block, also need the channel circumstance of the current data block judging to obtain, if channel circumstance does not change, then the channel circumstance of supporting vector machine model to next data block set up is utilized to judge; If channel circumstance changes, then redefine training set, penalty parameter c, function parameter g, and new model need be set up; If be last data block, then terminate.

After the algorithm proposed the present invention is tested, acquired results can be monitored channel value environment more accurately.Fig. 7 is travelling carriage each channel model algorithm performance when doing accelerated motion, and Fig. 8 is travelling carriage each channel model algorithm performance when doing retarded motion.

The present invention realizes being use Matlab language to programme on PC, and it is a kind of senior matrix language, comprises control statement, function, data structure, input and output and object based programming feature.MATLAB is a set comprising a large amount of computational algorithm.It has the mathematical operation function will used in more than 600 engineering, can realize the various computing functions needed for user easily.The algorithm used in function is all the newest research results in scientific research and engineering calculation, and before have passed through various optimization and fault-tolerant processing.Under normal conditions, programming on bottom layer language can be replaced, as C and C++ with it.When calculation requirement is identical, use the programing work amount of MATLAB to greatly reduce, and allow user to write can to carry out mutual C or C Plus Plus program with MATLAB.Libsvm tool box used in the present invention uses C language programming, therefore, before the use, need to carry out dependent compilation in the Command Window window of Matlab, compiling work only need complete before first time uses libsvm tool box, and reuse libsvm tool box can without the need to compiling later again.Use the natural language of the program of support vector machines information channel environmental detection method as follows:

Start

Generation signal to noise ratio is 20dB, initial velocity is 50km/h, 70km/h, 100km/h, 130km/h, 160km/h, 180km/h respectively, final speed distinguishes corresponding 50km/h, 70km/h, 100km/h, 130km/h, 160km/h, 180km/h separately, data block in often kind of situation, Rayleigh channel, Received signal strength through channel, often group comprises 100 data blocks, be inserted with 10 frequency pilot signs in each data block respectively, data block length is 200;

Below % program cycle 16 times, the initial velocity of setting and final speed situation are as table 1

Utilize in step one formula and formula, calculates the Signal estimation value of all channel models;

Utilize y in step one ^*formula, calculates the preferred channels model algorithm under present channel environment;

Generate the label value that channel circumstance does not change and changed;

Utilize r in step 2 _iformula, calculates Signal estimation value variable quantity when channel circumstance does not change in data block;

Utilize in step 2 formula, calculates Signal estimation value variable quantity when channel circumstance changes in data block;

Utilize e in step 2 _iformula, to calculate when channel circumstance does not change hopping amplitude between data block;

Utilize in step 2 formula, to calculate when channel circumstance changes hopping amplitude between data block;

In data block when not changed by channel circumstance, between Signal estimation value variable quantity, data block, hopping amplitude carries out corresponding with corresponding label value;

In data block when being changed by channel circumstance, between Signal estimation value variable quantity, data block, hopping amplitude carries out corresponding with corresponding label value;

SVMs kernel function uses RBF;

Penalty parameter c sum functions parameter g is set in range of choice;

Obtain supporting vector machine model;

Traversal combination penalty parameter c sum functions parameter g;

SVMs training set is equally divided into 3 parts, allow every part predict as test set respectively, remaining part is trained grader as training set, gets the average of all classification accuracies finally obtained;

Get penalty parameter c sum functions parameter g when classification accuracy average is the highest in traversal combination;

Utilize r formula in step 4, calculate the Signal estimation value variable quantity in current data block;

Utilize e formula in step 4, calculate the hopping amplitude between current data block and previous data block;

Test set (test set comprises: the Signal estimation value variable quantity in current data block, the hopping amplitude between current data block and previous data block) is input to SVMs, obtains the channel circumstance monitoring result under present case;

Judge whether current data block is last data block;

If not;

Judge whether the channel circumstance of the current data block obtained there occurs change;

If not;

Then use supporting vector machine model, continue the channel circumstance judging next data block;

If;

Then need to re-establish the supporting vector machine model meeting situation at that time;

If so, then terminate.

Claims (1)

1. use an information channel environmental detection method for support vector machines, realize according to the following steps successively,

Step (1), processes training set data according to the following steps, determines supporting vector machine model and suitable penalty parameter c sum functions parameter g:

Transmission convert information is that data block step is as follows by step (1.1):

Transmission information u, be converted into data block u _i, i represents training set data block number, i=1,2 ..., I, evenly inserts G frequency pilot sign p within the data block _g, g=1,2 ..., G, G are Finite Number, make each data block comprise two parts, information symbol s _gwith frequency pilot sign p _g, data block u _irepresent with following formula:

Step (1.2), the channel model step used under utilizing complex exponential basis expansion model algorithm, polynomial basis function extended model algorithm and switching channels algorithm to determine present channel environment is as follows:

Step (1.2.1), utilizes complex exponential basis expansion model algorithm to calculate Signal estimation value following formula and represents:

Wherein, ω _q=2 π (q-Q/2)/MN, M=2 is the expansion of over-sampling complex exponential base, the item number of Q representative polynomial basic function, q=0,1,2 ..., the item number sequence number of Q, q representative polynomial basic function, f _maxfor maximum doppler frequency, T _sfor sending the sampling interval of sequence and receiving sequence, i=1,2 ..., I, n=1,2 ..., N, n represent the position of signal in data block;

Step (1.2.2), utilizes described polynomial basis function extended model algorithm to calculate Signal estimation value following formula and represents:

Wherein, i=1,2 ..., I, n=1,2 ..., N, n represent the position of signal in data block, n ₀represent the point midway of data block, for basis expansion model parameter, d represents differentiate;

Step (1.2.3), utilizes described switching channels algorithm, calculates all channel model algorithm Ω={ Μ _y| y=1,2 ..., hopping amplitude sum E between multiple continuous print data blocks of Y}, and determine channel model algorithm y with this ^*for the preferred channels model algorithm under present channel environment, represent with following formula:

Wherein, E _ybe y channel model algorithm multiple continuous print data blocks between hopping amplitude sum, y=1,2 ..., Y, Y represent channel model number, j is the number of got multiple consecutive data block, J ∈ [3,5], j<J, h _j(1) the Signal estimation value of first symbol corresponding to a jth data block is represented, h _j-1(N) the Signal estimation value of last symbol corresponding to jth-1 data block is represented;

Step (1.3), determine the label value in training set, property value according to the following steps: hopping amplitude between Signal estimation value variable quantity and data block:

Whether step (1.3.1), there occurs during by judging that channel circumstance and last time are detected and change as training set label value, and described training set label value is divided into channel circumstance to change and channel circumstance does not change;

Step (1.3.2), by Signal estimation value variable quantity algorithm, calculates the Signal estimation value variable quantity r in data block when channel circumstance does not change _in () represents:

r _i(n)＝h _i(n+1)-h _i(n)；

Wherein, n=1,2 ..., N-1; I=2,3 ..., 10;

Step (1.3.3), to the Signal estimation value variable quantity r that step (1.3.2) exports _in (), asks for all Signal estimation value variable quantity r in i-th data block _ithe mean value r of (n) _icp, and utilize following formula to calculate Signal estimation value variable quantity in the data block of channel circumstance when changing

Wherein, time, represent that channel circumstance there occurs change, time, represent that channel circumstance does not change, β ₁∈ [1,20] is empirical coefficient, expresses support for the perception sensitivity level 1 that vector machine changes channel circumstance; Signal estimation value variable quantity r when the channel circumstance exported from step (1.3.2) does not change in data block _isignal estimation value variable quantity in (n) and data block when changing from the channel circumstance that step (1.3.3) exports be referred to as Signal estimation value variable quantity;

Step (1.3.4), by hopping amplitude algorithm between data block, hopping amplitude e between data block when calculating channel circumstance does not change _irepresent:

e _i＝h _i(1)-h _i-1(N)，

Wherein, i=2,3 ..., 10;

Step (1.3.5), hopping amplitude e between the data block utilizing step (1.3.4) to export _i, ask for hopping amplitude e between whole described data block _imean value e _icp, and utilize following formula to calculate hopping amplitude between the data block of channel circumstance when changing

Wherein, time, represent that channel circumstance there occurs change, time, represent that channel circumstance does not change, β ₂∈ [1,20] is empirical coefficient, expresses support for the perception sensitivity level 2 that vector machine changes channel circumstance; Hopping amplitude e between data block when the channel circumstance exported from step (1.3.4) does not change _iand hopping amplitude between data block when changing from the channel circumstance that step (1.3.5) exports be referred to as hopping amplitude between data block;

Step (1.3.6), judges channel circumstance according to the following steps:

Step (1.3.2) and step (1.3.4) is utilized to export the Signal estimation value variable quantity r obtained _ihopping amplitude e between (n) and data block _i, as training set property value when judging that channel circumstance does not change; Step (1.3.3) and step (1.3.5) is utilized to export the Signal estimation value variable quantity obtained and hopping amplitude between data block as training set property value when judging that channel circumstance changes;

Step (1.4), carry out serioparallel exchange:

Hopping amplitude and training set label value between the Signal estimation value variable quantity described in serial input, data block, carry out serioparallel exchange, parallel output is to buffer memory;

Step (1.5), set up buffer memory according to the following steps:

Using the Signal estimation value variable quantity that obtains after described serioparallel exchange first property value as training set, between data block, hopping amplitude is as second property value of training set, with training set label value, is stored into buffer memory;

Step (1.6), determine supporting vector machine model and suitable penalty parameter c sum functions parameter g step as follows:

Step (1.6.1), sets up described supporting vector machine model, exports the Signal estimation value variable quantity as the property value of training set, hopping amplitude and training set label value between data block from described buffer memory; Export from configuration parameter and be less than most ceiling value, and be greater than the penalty parameter c sum functions parameter g of minimum limit value, and be SVMs kernel function with RBF, then, set up the described supporting vector machine model under described two property values and a label value;

Step (2), uses supporting vector machine model to detect channel circumstance:

Step (2.1), utilizes cross validation K-CV method to determine and exports best penalty parameter c sum functions parameter g; Step (2.2), utilizes step (1.3.2) to calculate the Signal estimation value variable quantity of current data block; Step (1.3.4) is utilized to calculate the hopping amplitude of current data block and a upper data interblock;

Step (2.3), the Signal estimation value variable quantity of input current data block, the hopping amplitude of current data block and a upper data interblock, best penalty parameter c sum functions parameter g, utilize supporting vector machine model, export the channel circumstance testing result obtaining current data block;

Step (2.4), judges channel circumstance monitoring result according to the following steps:

Step (2.4.1), judges whether current data block is last data block:

If not last data block, then proceed to step (2.4.2);

If last data block, then terminate;

Step (2.4.2), judges whether present channel environment changes:

If channel circumstance changes, proceed to step (1.2);

If channel circumstance does not change, then use the supporting vector machine model of foundation to judge the channel circumstance of next data block, proceed to step (2.3).